A dismissed review now returns a successful exit_plan_mode result, so
the plan policy must not treat any success as leaving plan mode. Also
update the user-facing description of a dismissed review.
Group built-in settings before optional and installed plugins, and open the full experimental catalog on a separate page.
Preserve detail navigation and unsaved configuration when installation finishes, and use the visible More label as the accessible name.
Choosing "Request changes" on a plan review rejected the question, and
exit_plan_mode turned that into a failed tool result asking the model to
stop. The loop then spent one more model request only to have the model
wait. A dismissed review now returns { approved: false } and concludes the
turn, so the user's next message starts the following request.
Silero cannot back-date the first segment of a cold detector, so it
reports a start index several frames after speech actually began. Every
recording therefore lost its opening audio: measured across 50
utterances the median loss was 102 ms and the worst 230 ms — a full
syllable, which surfaced as dropped leading words ("麻烦把发票…" →
"把发票…", "用 Python…" → "写一个 Python…").
Only the first segment is affected; later segments aligned within 3 ms
because the detector retains history by then. Tuning cannot fix it:
lowering minSpeechSeconds from 0.25 s to 0.05 s left the median loss at
110 ms, and lowering the threshold to 0.2 still lost up to 174 ms while
risking false triggers.
Restore the dropped onset from the recording itself, bounded by the new
vadOnsetPaddingSeconds (default 0.3 s, zero disables). Measured on the
same corpus with identical DSH settings, this recovers the full no-VAD
accuracy and in the technical domain exceeds it, since segmentation also
trims silence:
语料 条件 无VAD VAD丢起音 VAD+回填 改善
audio clean 12.50% 16.33% 12.50% 23.5%
audio noisy10 17.74% 19.15% 17.94% 6.3%
audio_tech clean 10.34% 15.65% 9.81% 37.3%
audio_tech noisy10 25.73% 30.50% 26.26% 13.9%
Inference cost is unchanged (2.6 s vs 2.6 s on both clean corpora).
Use the package's own `SubprocessTerminalActivity['state']` instead of a derived
alias, label the samples as `state:revision` so a reader cannot mistake the
suffix for a count, name the summariser's bound as a group count, and state only
the two observables the prompt count supports.
Refs #5755
An activity poll that times out reports only "unknown" or "Matcher did not
succeed in time". The Linux standby has failed that way twice, at different
polls, and the log shows nothing about what `inspect()` read, which returns
`unknown` for several distinct conditions.
Every activity poll now records the state and revision it samples and counts the
prompts printed while it waits, so a timeout names both. Repeated samples
collapse and the printed list is bounded. Timeouts, intervals and expectations
are unchanged, and where the poll budget equals the test budget the test timer
can still preempt this message.
Refs #5755
An unexpected fork exit can close the worker's channel under a send, which
Vitest reports as a runner-level worker error instead of the clean exit. It
then tears the worker down before the ChildProcess exit event, and
DiagnosticForkWorker treated that teardown as an orderly shutdown: it cleared
the observer and marked the exit expected, so the coverage log lost the pid,
file, code and signal it exists to report.
A worker error observed before a stop now keeps the observer, and a stop
requested after that error no longer marks the exit as expected.
Vitest reports one such exit either as the clean exit path or as a
runner-level worker error, so the spec accepts both forms and compares the
record against whichever appeared.
Refs #5813
pi-ai sends Anthropic tool changes natively only when the initial tool list
has an active declaration. When every projected declaration is deferred,
prepend a non-deferred DeferredToolPlaceholder instead of failing the request.
Document the placeholder, the catalog-owned update capabilities, and the
image role limit as separate README entries.
Master removed `backgroundMode` and the `run_in_background` argument, so
every delegation starts a managed activation. Mount the working-directory
owner SubagentRuntime injects and the JSONL persistence root the spawn
provider requires, and synchronize on the child's cancellation, which the
parent teardown or the manager drain raises instead of a parent-only cause.
The Typert workspace pass scans the analyzer's symbol table, and every
symbol asked `registrationForFile`/`packageNameForFile` for its owning
package. Both looked up the real path and then scanned every registration
with `isWithin`, and `isWithin` re-ran `path.resolve` through `realPath`
for both arguments on each call. A CPU profile of the host pass put
17.0 s of the 30.9 s sampled in that scan, almost all of it in
`realPath`/`resolve`/`existsSync`.
The registration set and the analyzer's face are fixed for the analyzer's
lifetime, so each source path's answer cannot change: memoize both
lookups per analyzer. `resolve` depends only on the requested path and the
process working directory, so memoize the normalization step too; the
existing canonical-path cache keeps its "existing paths only" rule.
tsdown host pass, interleaved A/B: 39.89/37.18 s -> 25.70/25.06 s.
Isolated generator phase: discover 8.69/8.59 s -> 7.33/7.44 s,
generate 17.19/16.96 s -> 4.21/4.21 s. Every emitted `typert.*` file
hashes identically before and after each run.
waitForTitleAfterTurnEnd and waitForEventAfterTurnEnd now name the session,
expected record, and deadline when the deadline passes before the first log
harvest settles, and keep the vi.waitFor failure as cause. Previously they
surfaced Vitest's generic 'Timed out in waitFor!', which failed the title
diagnostic spec on the macOS Sandbox lane at master 090348e4.
The title diagnostic spec uses isolateDiagnosticTimeout like its siblings,
replacing the Windows-only 5s timeout.
Refs #5631
The upload warns instead of silently ignoring missing files, and the spec
pins the upload after the benchmark step. benchmarks/AGENTS.md and the
report Agent Note now name the artifact CI publishes.
Refs #5840
The node-24-bench job uploads benchmarks/.dsh-report after the benchmark
step, including when a budget assertion fails, so the report the Vitest
globalSetup renders survives the runner. #5838 deferred this step because
the authoring token had no workflow scope.
Refs #5838
Extension failures, unreadable package identities, a blocked session-log
upload, and degraded Messages replay recur on every request. Log the first
occurrence per field, route, module, or blocking event instead.
Co-authored-by: turtle1999 <[email protected]>
Request-extension fields are auxiliary. A contributor whose preparation
throws is now logged and omitted while the other fields are still sent,
and a post-2xx acceptance failure is logged without failing the request.
Package inventory is best effort: a named manifest without a version, or
a bare package whose manifest is missing or unreadable, reports version
"unknown" instead of failing preparation.
A failing request-extension contributor surfaced only
'DeepSeek request extension preparation failed'; the cause stayed on
LlmError.cause, which the turn failure record and retry notice drop.
Append the cause chain to preparation and acceptance messages.
Every recorded case states what it measures and which user scenario it
affects, shown on hover in the report. Endpoints whose workers report
process CPU time default to the measured non-CPU share and busy
parallelism; the Session corpus worker now reports CPU time per
endpoint. Machine profiles carry logical CPU counts, and the page
shows each default share with its source and can reset overrides.
The General font-size row gains a More font settings disclosure that expands
a font field for each of text, code, and the sidebar terminal, plus code
(10-16 px) and terminal (10-20 px) size steppers. Font lists are normalized,
stored in the ui-theme namespace, embedded by the boot script, and placed
ahead of each role's built-in stack; the code size moves the code typography
tokens; the terminal applies font and size live and refits without replacing
the shell. ThemeRuntime adopts Host sections only after its own writes settle.
Refs #5217
Two fixture helpers in surface.spec.ts returned ': SessionEvent', a
declaration-merge-extensible mapped union. Spreading such a value (the
rejection cases build '{ ...event, data: ... }') materialises an anonymous
union per fixture, and the following 'as SessionEvent' casts then compare
union against union, which dominates the root type-check program.
Return the single variant each helper actually builds, spelled as the
canonical indexed access SessionEvent<'system/message'> /
SessionEvent<'user/message'>, drop the casts that only existed because data
was a union, and trim the comment of the member count that drifts with every
declaration merge.
Measured on the host build: check 110.9 s -> 84.6 s (median, -23.7 %),
wall -17 % to -24 %; fixture types alone (clean buildinfo, root program):
check 53.4/45.8 s (Extract alias) vs 45.3/49.1 s (indexed access).
Two declaration forms relied on a checker binding a name the permissive way:
- cordis-host-runner declared 'export function CordisDynamicPluginId'
alongside 'import type { CordisDynamicPluginId }' of the branded type, so
the call sites resolved through the type-only import. Export the four
constructors under separate local names; the public names, JSDoc and the
type names in signatures are unchanged, and the typert signature metadata
keeps the original type names.
- views-type-chain.client.spec.tsx spread one slots.register call over three
lines, which lets a checker anchor the rejected overload on the argument
line instead of the call line, outside the @ts-expect-error directive.
Keeping the call on one line keeps the negative assertion effective.
Let a title provider replace one that is still draining, so switching the
titles bundle during an in-flight request no longer leaves no provider.
Omit group container rows from bundle inventories, since they have no
plugin entry to switch. Drop the obsolete preset-target product-use case,
fix the Claude Code row id example, and refresh stale Agent Note text.
Session restore and projection run after the file closes, so they use
no storage share; the page keeps one share per endpoint. Record the
repeat projection-list budget, drop the post-GC page heap from the
memory view, remove the unused profile group, and add the owning
Agent Note.
Remove the `{ preset, insert }` / `{ preset, id, ...override }` profile patch
operation added in #5773, together with the app-boot compiler and reload drain,
config-editor filtering, plugin-manager preset row inventory and retained-module
removal checks, the read-only preset rows on the Plugins page, and the
agent-preset-registry row identities that served them.
Session search, Ralph, persistent terminals, Codex, and Claude Code bundles now
insert ordinary Host rows whose tools register in the global tool layer, like the
earlier optional bundles. Every preset, including minimal, sees them, and toggling
a bundle updates live Agents without new composition generations.
Record budgeted benchmark endpoints and peak memory, and render a
self-contained interactive report that rescales measured times to
reviewed laptop and cloud profiles by single-thread CPU rating and
default-disk latency, and compares peaks with 4 GB and 8 GB of
available memory.
The packed rehearsal now names each stage on the child's stdout and bounds the
manager's removal, so an operation that never settles reports the profile lock
and the still-live pnpm process instead of only a test timeout. That evidence
located pnpm 11.7.0 leaving its worker pool alive after a short-circuit return
(pnpm issue 12297, fixed in 11.23.0 and carried by this stack's base).
Three attempts to keep pnpm off that path — a private pnpm store, installing
both bundles before removing either, and pre-declaring the profile's age
exemptions — did not change the outcome, so none of them remain.
- The summary lists only the catalog roots as installable specs; the other 29
archives stay in the run manifest, where the installer resolves them itself.
- Only a pnpm entrypoint receives pnpm's arguments, a catalog root counts as a
required edge for platform admission, and a signal that stops packing reports
the interrupted exit instead of an exception.
- The packed rehearsal names the Agent SDK's musl artifact on musl hosts again,
and its comments match what the fixture does.
- The cache test registers its pnpm children for termination, and a new case
keeps one profile and its real lockfile: pnpm refuses a new run's URL with
ERR_PNPM_TARBALL_URL_MISMATCH until the stale resolution is cleared with
pnpm's own `pnpm clean --lockfile` remedy, which the desktop README now
documents for same-version iteration.
The teardown asserted the Web process's SIGTERM exit inside `finally`, so a
body failure was reported as a shutdown timeout instead of itself. The
assertions now run only when the body passed, the earlier failure is rethrown
after teardown, and the listener readiness became a named step that no longer
clones the release pack coordinator's listen idiom.
Both are failure paths the registry owns: a signal that stops packing must
report as a cancellation rather than as a pack failure, and a client that
disconnects mid-tarball must not stop later requests. execa reports the first
either as a resolved canceled result or as a thrown error, so the packing step
now names both.
A packaged application installs an Official on-demand bundle as
`name@<running DSH version>`, and the plugin manager asks a custom registry
alone, so the unsigned test DMG cannot install a locally built provider before
its release reaches npm: each packed manifest still names the rest of the
repository by version, and ordinary bundle tarballs therefore resolve against
a registry that does not have them yet.
`pnpm run dev:bundle-registry` validates the Official catalog and the workspace
inventory, builds the official artifacts, packs the catalog's complete local
dependency closure with the release families' own payload and manifest checks,
records the run in `dist/test-bundles/<run-id>/bundle-registry.json` as its
completion marker, and only then serves it from 127.0.0.1 on an allocated port
under a fresh namespace. Repository-owned names are answered from this run or
refused locally; third-party packages keep npm's real metadata and bytes
through a fixed redirect. The run publishes nothing, changes no global
npm/pnpm configuration, and leaves DMG packaging untouched.
The packed installer rehearsal now installs that closure into an isolated
profile with no `file:` override, checkout link, ignore list, or scoped
registry standing in for a package the registry should answer, and it asserts a
transitive repository package served only by this run.
The fixture workspaces declare packageManager, so corepack resolves that
version when pack-release runs pnpm pack inside them; leaving 11.7.0 there
would exercise the version this change replaces.
pnpm 11.28.5 runs verifyDepsBeforeRun before a script, so `pnpm run install`
inside the installed node-pty directory first installed node-pty's own
dependencies into the package and ran its `prepare` build. Both reached the
packaged payload: the SEA blob then carried a node-gyp tree and tripled pkg's
file records, and postject aborted while injecting it. The step now runs the
package's install script alone, as it did with pnpm 11.7.0.
pnpm writes allowBuilds placeholders for blocked dependency scripts only when
it can prompt, so a managed run that always captures output left
pendingBuilds empty and the approval flow could not offer anything. The failed
run's own report is the surviving record: the profile policy now records those
names, which keeps approval, persistence, and the wildcard and decision guards
unchanged.
pnpm's legacy deploy links the deployed project into the target's node_modules
when the target sits inside that project. Materializing that link copies the
payload into itself and fails with ERR_FS_CP_EINVAL, which broke the
release-shaped Python runtime matrix on Linux and Windows. The staged payload
never carries the project source, so the staging tree's owner drops the entry
and leaves the deploy root's files alone.
pnpm 11.7.0 can finish its work, print `Done in Xs`, and then never exit
(pnpm/pnpm#12297, pnpm/pnpm#13617): finishWorkers() clears the worker pool
reference, and a later worker API call recreates a pool that is never
finished, so idle worker threads keep the event loop alive. The upstream fix
(pnpm/pnpm#13226) shipped in 11.23.0, and 11.25.0 fixed the hoisted-linker
rename race as well.
The hang surfaces in this repository as a packed-profile rehearsal that runs
`pnpm remove` through the plugin manager and then waits for the test deadline,
holding the profile write lock for the whole time.
Bump packageManager and devDependencies.pnpm in the root, the pinned pnpm in
apps/desktop, and the native/system packageManager; relock; exempt the fresh
package manager from the release-age window; follow the development guide
pair.
Review found the same Host profile write plus inventory reload chain still
uncovered in worktree-opt-in.e2e.ts, and still written as 10_000 literals in
plugin-manager.e2e.ts. Both now use PLUGIN_TOGGLE_SETTLE_MS, the README keeps
only the convention, and the constant's JSDoc names the store's load() as the
reader that republishes the switch.
A plugin-card switch republishes `aria-checked` only after the Host writes
the profile and the panel re-reads its inventory. That chain measured
0.5-0.8 s on an idle host and 2.0-4.8 s across 32 samples in a 16-worker
lane, so Vitest's 1-second `expect.poll` default truncated it and failed
the switch assertions of the CoT translation scenarios in CI.
Every poll that follows a plugin-card switch click now passes the shared
PLUGIN_TOGGLE_SETTLE_MS budget; assertions, retries, and product code are
unchanged.
A reasoning route always declares at least one level, so the adapter's
"no floor" path is defensive; the resolver rejects an empty list before any
consumer sees it. Mark that branch with the repo's v8-ignore convention and
state why.
`purpose` is a label for replay and transport metadata, so no adapter should
branch on one caller's value. Exact-model metadata now reports
`reasoning.floorEffort`, the least effort a route accepts, and the title
service selects it for its request. `llm-pi-ai` and `llm-deepseek` dispatch
the effort they receive and no longer recognize `session-title`; the
`minimizeReasoning` field is gone. A route with no registered adapter, such
as the recorded replay lanes serve, reports no floor, so the title request
keeps the adapter default instead of failing before dispatch.
Replaces the `minimizeReasoning` field from the previous commit.
`purpose` is a provider-neutral label for replay and transport metadata, so
an adapter must not branch on one caller's value. `GenerateOptions` gains
`minimizeReasoning`, which any auxiliary caller sets when reasoning would
consume the output budget it reserved for a short visible result; both
shipped adapters dispatch such a request at the least reasoning the model
supports, and `session-title-llm` sets it instead of relying on the
adapter recognizing `'session-title'`. The pasted type in the subsystem
page and the generated Cordis API catalog follow.
Prompted by review: llm-pi-ai must not know about session-title.
A request body also carries the session log, whose recorded title event
embeds the framed prompt, so a body-wide marker search matches the main
request too; both mocks now read the request's own messages. The headless
scenarios also assert the dispatched cap, which is the base-bundle ceiling
on the 256000-token route and the route override's 1024 in the pi-ai
compatibility scenario. The extraction cases cover asterisks that are not
emphasis.
The lowest-level dispatch keeps reasoning off the title's output ceiling,
not a small cap's visible text, and it deliberately skips the
requested-effort check. The README pair records that the dispatched cap
is the smaller of the configured ceiling and the latest folded
request/header, including the tighter-main-cap and override consequences,
and the accepted-title comment states the deliberate whole-line emphasis
rule.
Headless stream-json and session expected outputs record the title
request's maxTokens, and every one of them still recorded the previous
base-bundle cap.
The README pair states that an accepted result is the first non-empty
line of the model text, with one wrapping emphasis pair removed, and that
the dispatched cap is the smaller of maxOutputTokens and the cap the
routed adapter declares. The pair is re-recorded with both sides.
The headless and Web mock providers recognize a title request by its
max_tokens, which still assumed the previous 256-token base-bundle cap.
The recorded expected outputs carry that cap too and follow with it.
The prompt asks for a short title, so the cap only stops runaway output:
reasoning a model cannot switch off also counts as output, and 256 tokens
truncated titles on the HFAI gateway. The measured worst title response
was 153 output tokens, so 4096 is about 25 times the observed maximum.
Claude wraps some titles in asterisk emphasis and writes the title, a
blank line, then an explanation when the messages say little. Joining the
text blocks made that explanation part of the saved title. Take the first
non-empty line, drop one emphasis pair around it, and ask the model for a
short best-effort title instead of an explanation.
The dispatched cap is now the smaller of the configured cap and the cap
the session's current main request recorded, so a route the deployment
capped below the shipped ceiling is never asked for more than that cap.
A session-title request dispatches off exactly as a selected off does,
which on protocols that omit the reasoning field leaves the gateway
default in effect. The README now says so instead of implying reasoning
is always switched off. requestReasoningLevel follows the
resolveReasoningLevel it calls and names when it returns no level, and
the Loader composition helper takes provider config and further entries
as separate parameters.
Headless stream-json and session expected outputs record the title
request's maxTokens, and the headless and Web mock providers recognize a
title request by its max_tokens. Both still assumed the previous 64-token
base-bundle cap.
The base bundle capped title generation at 64 output tokens, sized for a
route that disables thinking. Models whose reasoning cannot be switched
off spend output on it even at their lowest level: on the HFAI gateway,
GPT-6 Sol and GPT-6 Astra used up to 87 and 121 output tokens for
five-word titles at reasoning low, so Sol still finished with max-tokens
on most first prompts. Raise the shipped cap to 256; routes that disable
thinking stop after the title and are unaffected.
Session-title calls carry a 64-token output cap and rely on the adapter to
keep reasoning out of it. llm-deepseek maps purpose 'session-title' to
thinking disabled, but llm-pi-ai ignored the purpose, so pi-ai routes sent
title calls at the route's default effort. On HFAI's adaptive-thinking
Claude models (profile reasoning: max) the thinking consumed the cap, the
call finished with max-tokens, and the session kept its fallback title.
The pi-ai adapter now dispatches a session-title request at the model's
lowest supported level: off when the model offers it, otherwise the
lowest declared level. GenerateOptions.purpose documents that obligation
for every adapter.
A successful verifier message was treated as terminal even though the
Worker Thread was still alive. The parent immediately terminated the
worker and detached cancellation, so an error, nonzero exit, or caller
cancellation after that message could be ignored and verification could
resolve successfully.
Keep the first successful response pending until an observed natural exit
with code 0. A missing response or nonzero exit rejects after the worker
has already exited. Error and abort callbacks retain the existing
terminate-and-join path, including its first-error and AggregateError
behavior. The scheduler stays unchanged: its permit remains held until
the verification promise settles, now after successful worker quiescence.
The worker already closes its parent port in finally, so success needs
neither a forced stop nor another protocol message.
Replace success-termination expectations with deterministic message/exit
barriers and cover post-response errors, nonzero exit, cancellation, and
permit handoff. Remove the obsolete hand-written termination thenable and
its retired unknown-assertion baseline entry. Add real source workers for
plain and Zstandard input, and require the built plain-Node migration smoke
to complete with zero termination calls. Document the settlement condition.
Native Windows CI recorded process access violation 0xC0000005 while a
migration fixture was active. Removing unnecessary successful-worker
termination is consistent with that evidence, but this commit does not
claim that the native faulting module or mechanism has been established.
The deterministic regressions prove the lifecycle defect independently;
the next native CI run must assess whether the observed crash recurs.
Validation:
- Original source failed all new message-to-exit regression scenarios;
the updated focused lifecycle file passes 19 tests with 100% coverage.
- Node 24 owning package: 19 files, 741 passed, 1 skipped; every included
package source file reaches 100% statements/branches/functions/lines.
- Real source workers passed for none/zstd; built plain-Node smoke passed
with zero terminate calls and no timeout or process signal.
- Exact Node 22.19 source-mode none/zstd workers naturally exited 0 with
zero terminate calls, exercising the async Module.register fallback.
- Host TypeScript build, focused lint, no-new-unknown-assertion gate,
README translation pairing, and all 42 doc-sync gates passed.
Related to #4387 and #4394.
The master cwd integration updated the shipped headless profile, but three
current snapshot scenarios merged without textual conflicts and retained
expectations from before the directory service was composed. Full CI
replay found the stale product-subagent-both persona and missing current
working-directory context in the no-message-tool and structured-parent
parent logs; those two scenarios also owned stale child/header sidecars.
Refresh only those three scenarios through the real built headless profile.
Record the required runtime cwd section in their four parent/child V4
logs, remove static cwd text from three prompt sidecars, and include the
new cwd start parameters and working_directory tool in three schema
sidecars. Preserve every model response and event count. No production
code, scenario input, workspace oracle, or historical V0-V3 fixture changes.
Validation: selective refresh passed 3/3. Full DSH_EXAMPLE_MODE=lib pnpm
run test:snapshot passed all 8 files: 197 passed and 2 PowerShell scenarios
skipped. The first full local run selected macOS CPython 3.9.6 and failed
the two Python PTC scenarios; the final full run prepended the existing
CPython 3.12.13 bin directory to PATH, with no runtime or fixture workaround.
SHA256 guards confirmed all 363 retained historical inputs unchanged;
JSON comparison confirmed the four changed logs differ only by cwd context.
TextMate scanning cost grows superlinearly with line length, so a
generated line of tens of thousands of characters blocked the page for
seconds when the sidebar previewed the file. Lines of 1,000 or more
UTF-16 units now render as one plain run in every highlighted surface.
Integrate the observed master checkpoint 7ab215e23e5d977aae7eb5d4fb91782d39969dbb without rewriting the earlier published checkpoints. Master adds current-directory events, directory-aware consumers, and retained Git worktrees; its old subagent startup paths overlapped the unified activation implementation.
Resolve directory selection within managed startup after acquiring parent ownership. Preserve the original Session origin while setting each new local child directory, retain persisted directory state on cold resume, and pass a required absolute cwd to external providers. Keep unified local/external ownership, full completion notices, parked-input host completion, SDK wait/error/shutdown behavior, and ACP cancellation semantics. Retain removal of the old in-process driver and continuation managers.
Combine SDK wait with the new directory requests, preserve the new master directory/worktree fixtures, regenerate derived catalogs from the combined sources, and refresh only the four stale prompt/schema sidecars. Preserve every protected historical schema, archived note, catalog-v2 acknowledgement, and recorded predecessor input.
Validation: full build; 59 CI static gates; 855 subagent tests with 100% coverage on changed core/external sources; affected consumer suites including 119 repaired-fixture tests; SDK 11 and headless 16 recorded scenarios; two local headless expected cases; six built nested SDK/ACP cases; lint, strict cast inventory, workspace constraints, and whitespace checks. Native Windows worker diagnostics remain additive and require the next CI run to classify the original crash.
Repeated native Windows coverage runs stopped a Vitest fork while a JSONL
migration suite was unfinished. Vitest's generic unexpected-exit error
discarded the process exit code and signal, preventing classification of
the failure despite identifying the missing test file from the inventory.
Reuse the public ForksPoolWorker diagnostic implementation and protocol
fixtures from independently reviewed PR #4388, pinned at
77e215ff76711ec75ed8aa08aa5b414780e539ae. The runtime helper is unchanged.
It reports PID, project, worker identity, outstanding file request, code,
and signal before stock Vitest reports the original failure. It clears
ownership only on the typed testfileFinished response, suppresses expected
stop, and leaves cancellation and failure handling with the original pool.
Wire that existing pool into generated coverage partition configs. Keep
the original worker counts, partition scheduling, test timeouts, coverage
thresholds, and failure verdicts. No environment, raw argv, or test object
is serialized. This commit adds diagnostic evidence; it does not claim to
repair the still-unidentified Windows crash. PR #4388 remains the separate
reviewable owner; no duplicate diagnostic framework or PR is introduced.
Adapt the copied fixtures to the current no-new-unknown-assertions rule:
declare the private preload exit function, validate consumed logger fields
with the public logger type, and assign matcher/rejection values to typed
unknown locals without assertions. Add a concise bilingual testing note.
Validation:
- Diagnostic suite: 10 passed, including a real configured Vitest fork
exiting 23, stock-pool negative control, successful silent shutdown,
unexpected exit 0, early exit, independent forks, and signal metadata.
- Coverage coordinator suite: 51 passed.
- Targeted lint and focused strict TypeScript check passed.
- pnpm exec tsc -b tsconfig.host.json passed.
- verify-no-unknown-casts passed; 1430 existing assertions unchanged.
- doc-sync passed 40 gates; the two initially failing gates were repaired
and rerun successfully: doc-typecheck and verify-doc-budgets.
- Translation pairing and git diff --check passed.
Refs #4387, #4388, #4394.
Co-authored-by: Chinesezjc <[email protected]>
The cwd regression used a fake Context cast through unknown. Updating its service lookup for descendant shutdown changed the exact legacy assertion fingerprint, which the no-new-unknown-casts CI gate correctly rejected.
Exercise initialize, prompt, and wait with the real Context, Agent registry, Session store, and a scripted LLM adapter. Spy on the real service methods to keep the cwd, reasoning effort, and token-limit assertions. Remove both obsolete unknown assertions and prune only their recorded allowances; do not add or broaden an exception.
Validation: the cwd regression passes; all 54 SDK server tests pass with 100% server coverage. The full check:ci:static aggregate passes 59/59, including the strict unknown-cast inventory. No production behavior changes.
The gated final-flush fixture awaited the real flush but dropped its boolean return value. The regression ran correctly under Vitest, but the full Host build rejected the mock against Sessions.flush Promise<boolean>.
Return the real flush result after the gate. This retains the observation order and matches the service signature without a cast or altered production behavior.
Validation: targeted terminal-capture-failure regression passes; the complete build is rerun after this typing correction.
SDK shutdown disposed its root Agent handles without stopping managed
Activations below those exact roots. The manager owns descendant handles
in a separate context, so an interrupted child with parked input remained
registered after shutdown resolved. Its settlement watcher could then
outlive the SDK root and race later projection teardown.
Close scoped descendant admission and await child-first drain before
releasing root handles or the server adapter. Keep draining limited to the
server's exact root Agents so unrelated parent trees remain live. Collect
drain failures without skipping the existing root/adapter teardown, and
preserve the final single-error or aggregate-error result.
Add real AgentLoop/SubagentRuntime regressions for parked children at
one and two levels, child-first disposal, preservation of an unrelated
parked tree, and descendant/root teardown failure aggregation. Correct an
existing Context.get test stub to return its LLM only for the LLM service.
Document the lifecycle order in the server JSDoc and bilingual README.
Validation: 52 SDK server tests with 100% server.ts statements, branches,
functions, and lines; four final shutdown regressions; host TypeScript
build; full lint; 21 quick documentation checks; staged pairing/diff checks.
Doc-sync passed 41/42 checks; the pre-existing config catalog source link
for subagent-acp/index.ts is stale at base and left to integration refresh.
Interrupting a local child can preserve unclaimed steering while leaving the Agent idle. Activation residency intentionally continues in that state, but completion barriers joined its result forever. The real SDK server reproduced the hang for both direct children and a nested parked leaf.
Join progressing work and cleanup while retaining parked inboxes and their activation results. Observe local activity and maintenance, external execution, unpublished ownership holds, and materialization; rescan after changes. Recheck resumed Agents after concurrent idle observations. Inbox insertion and ownership publication wake pending observations, so late non-waking injection cannot strand a result wait. Existing descendants remain responsible for their own work; no second state store or polling timer is introduced.
Document that persistent SDK/ACP hosts keep these children available for later wake-up, while headless shutdown discards unclaimed input. Simplify the final settlement recheck to retry changed readiness uniformly. Update generated API and event references from their sources.
Validation: real SDK negative controls fail at both depths on the prior manager with clean teardown; 545 tests pass across 27 files. Manager, local activation, and headless runner each reach 100% statements/branches/functions/lines. Six deterministic races cover revival, preparation publication/rollback, flush-time injection, and late children. doc-sync 42/42, targeted build, lint, and whitespace checks pass.
Removing the duplicate ACP directory helpers moved its Config declaration by one line. The generated English and Chinese catalogs still referenced the old line, so the exact-source freshness gate rejected the otherwise unchanged configuration.
Regenerate only the affected source links. No configuration fields, defaults, examples, or translated prose change.
Validation: pnpm run gen-config-catalog produced exactly these two link updates; the combined doc-sync run verifies freshness.
A child completion can wake the parent after its first completed turn. If
that summary cannot append turn/start or turn/end, AgentLoop reports an
agent/error and contains the rejected driver. Whole-agent idleness then
resolves, but the headless runner previously reused the older completed
terminal and exited successfully, sometimes printing the older answer.
Retain failures from the exact owned Agent with the current Session log
offset. Only a later terminal committed by the same Session clears that
failure. After work and persistence drain, fail the run before emitting a
final answer when no such terminal exists. Dispose both observers with
the run. Ordinary durable error outcomes and later successful recovery
continue to use the log's final terminal.
Add real AgentLoop regressions for missing summary start/end commits in
plain and JSON output, ordinary hook failures with and without recovery,
recovery after persistence failures, and unrelated Agent errors. Update
the bilingual exit contract and generated event consumer matrix.
Validation: 106 headless package tests; 100% runner statements, branches,
functions, and lines; 9 focused terminal-failure tests after the typed
append wrapper correction; package TypeScript build; full lint including
host build; 21 quick documentation checks; 42 doc-sync checks; diff check.
Local activation results were derived only from committed Session events.
When the first turn/end append failed, no ending existed and the result
defaulted to completed. When the first turn/start append failed, the initial
message remained unclaimed and kept the idle activation resident forever.
Both paths emitted agent/error but had no durable terminal to project.
Retain the log offset of the latest exact-Agent live failure on the local
activation. Clear it only when the same Session commits a turn/end at or
after that offset, so a later recovery can supply its own outcome and a
reentrant older terminal notification cannot erase a newer failure. The
child scope owns both observers and removes them with its Agent handle.
Capture error while that live failure remains. Let the failed local
activation settle despite its own unclaimed inbox, while preserving the
existing requirement to wait for owned descendants. Wake its settlement
observer when a live failure changes readiness. No public API, raw Agent
idle semantics, or descendant teardown policy changes.
Add regressions through real local activation creation for first start/end
append failures, recovery by later durable endings, exact-Agent isolation,
and an error emitted reentrantly during an older terminal notification.
Update the bilingual package contract and generated event consumer graph.
Validation:
- Both first-append regressions reproduced before the fix: turn/start
remained pending; turn/end returned completed instead of error.
- Six focused behavior cases passed after the fix.
- 423 subagent package tests passed; activation.ts and manager.ts reached
100% statements, branches, functions, and lines under scoped coverage.
- pnpm exec tsc -b packages/subagent/subagent/tsconfig.json passed.
- Focused source/test oxlint and git diff --check passed.
- doc-sync passed 41 gates; its sole stale event-graph failure was corrected
with gen-doc-graphs, then verify-doc-graphs passed for all seven graphs.
- README translation pairing regenerated and passed the doc-sync check.
Caller delivery changes result ownership, but does not make local workflow children private or remove their continuable catalog entries. The workflow README stated durable discovery without making the message and cold-resume consequences explicit.
Document that resident workflow children accept continuation and that cold continuation requires persistence and Session query. An ephemeral child therefore returns NOT_RESUMABLE after disposal. This describes current supported behavior and does not add fallback storage or a new child mode.
Validation: checked against startActivation, sendMessage, and cold-resume admission; English and Chinese pairing updated. Full doc-sync follows the combined review fixes.
ACP retained a private copy of the external-provider directory helpers after the other providers adopted the shared implementation. The copies differed on absolute-path validation and imposed separate filesystem branches and maintenance.
Use resolveChildCwd and validateConfiguredCwd from the existing subagent package. Preserve configuration-time relative-path resolution, parent workspace inheritance, missing-directory failures, and ACP diagnostic normalization. This deletes the duplicate directory implementation without introducing an abstraction or dependency.
Validation: 85 passing tests across the ACP provider and shared out-of-process helper suites, including real child process cwd, configured overrides, and invalid-directory rejection.
Parent completion notices now always carry the child closing answer, even when the child can send progress messages. Two assembled headless expectations still asserted the removed omission and made the snapshots-and-artifacts CI lane fail.
Require CHILD_RESULT in the settlement assertion and refresh only the owning parent and stream output expectations. Preserve released replay inputs, event ordering, sandbox denial evidence, and the recorded model responses.
Validation: DSH_EXAMPLE_MODE=lib pnpm run test:expected apps/cli/tests/profiles/headless/tests/headless.expected.e2e.ts apps/cli/tests/profiles/headless/tests/subagent-inheritance.expected.e2e.ts (18 passed).
The subsystem result paragraph still said that a parent log records the
final cleanup-failure status and that the result had already been returned.
The manager instead captures the execution result, performs cleanup, sends
that preserved result in the parent notice, and then resolves the public
result. Cleanup failure separately rejects dispose and marks the live
subagent/end event as error.
Correct the bilingual paragraph to name those actual recipients and timing.
Do not imply that the successful execution notice becomes a failed notice,
or change the runtime's independent execution/cleanup outcomes.
Validation: the updated bilingual pair passes. The statement is checked
against finishDisposal/notifySettlement and the existing cleanup-failure
regressions; final documentation gates run before publication.
The initial-cancellation regression supplied a synchronous content array to
defineContentToolFixture, whose executor contract requires a Promise of
content blocks. Vitest executes the fixture, but the strict repository
Host build correctly rejects its signature with TS2739.
Return the same blocks through Promise.resolve. Keep the deterministic
stream/output gates and assertions unchanged; do not widen the shared
fixture API merely to accept this one caller.
Validation: the focused initial-cancellation regression passes. The full
Host build is rerun before publication to validate the fixture signature.
Cancelling during the prompt's initial whenIdle() wait could retain its
slot indefinitely. If a child delivered a result while the aborted root
was unwinding, Agent.whenIdle() followed the replacement driver and waited
for the newly autonomous summary. The earlier child-wait recheck could
not help because settlement had not yet reached that child wait.
This is an inherited latent issue: the pre-PR ACP bridge also awaited the
same replacement-following whenIdle() after cancellation. Existing
continuable spawn children could expose it; the unified asynchronous fork
default makes the ordering possible for fork delegations as well.
Give the existing InflightPrompt one cancellation-completion promise,
replacing the local cancellation wake-up and its listener. The private
admission AbortSignal is the cancellation flag instead of duplicated
boolean state. The exact root's emitted status identifies the interrupted
activity's retirement and synchronously adopts that output queue. Later
root work and later output cannot replace the first captured completion.
Record whether cancellation began at public idle. Maintenance has that
public status too, so its next running transition establishes release of
work behind maintenance; the ordinary whenIdle promise provides a
fallback when no transition occurs or plugin disposal detaches observers.
Both observations settle the same promise. Settlement races that promise
against initial idle and child waits, then waits only for the interrupted
activity/output on the cancellation path. No future work is re-cancelled,
child admission remains open, and session close keeps its separate full
quiescence and descendant teardown.
Add two deterministic real-Agent regressions. One gates aborted model
cleanup, previously queued image delivery, and the later autonomous
summary independently. The other gates cancelled maintenance cleanup and
the later summary. Both failed before this fix; both now wait for old work
and output, return cancelled while later work remains gated, and permit a
subsequent ACP prompt to execute another real local delegation. Existing
plugin-disposal tests also verify the idle fallback after observer cutoff.
Update the bilingual ACP semantics and generated event-consumer matrix.
Validation:
- Both focused regressions reproduced pending cancellation before the fix.
- pnpm exec vitest run packages/acp/acp/tests --coverage --coverage.include=packages/acp/acp/src/session.ts --coverage.include=packages/acp/acp/src/index.ts --coverage.reportsDirectory=/tmp/pr4394-acp-owned-cancel-89ec33e5788-consumer-v2
(150 passed; both changed source files have 100% statements, branches,
functions, and lines)
- pnpm exec tsc -b packages/acp/acp/tsconfig.json
- pnpm exec tsdown --config-loader native --env.DSH_BUILD_FACE host --filter '@deepseek-ai/dsh-acp'
- DSH_EXAMPLE_MODE=lib pnpm run test:e2e packages/subagent/subagent-dsh-sdk/tests/nested-delegation.e2e.ts -t '^external tasks.*acp'
(3 passed, 3 unselected)
- Focused lint, named translation pairing, generated graph verification,
and git diff --check passed.
Independent ACP rereview found no remaining issue, passed seven focused
turn/disposal cases, and verified a prepended idle listener can start a
new hanging root before ACP handles the old idle event without retaining
the cancelled prompt.
Preserve the published 78ae4fc2 integration checkpoint and merge the later master tip separately. This master delta changes only the browser-GIF skill to use the managed DSH credential store through normal application configuration; it changes no product source, schema, or test behavior. The existing credential-provider recording overlay already follows that guidance.
Seed a fixed POSIX effective directory before the first request and disable ambient instruction discovery in the two pressure fixtures. Preserve the original project and existing output, headroom, retention, and summary limits while keeping shadowed token counts independent of temporary workspace path length.
Supply the required directory collaborator to current runtime fixtures, verify absolute result navigation and deployment fallback, and retain the valid V4 working-directory recording instead of the unreleased V3 input that contains a V4-only event.
The completion coverage audit left two uncovered paths: replaying an ACP
abort that occurred before listener registration, and ignoring an Agent
error from outside the SDK server's owned root sessions.
Remove the redundant ACP abort replay. The private cancelPrompt method
is the sole owner of admissionController.abort(), sets cancelRequested
first, and never resets it. Settlement checks that sticky cancellation
state after its initial idle observation and before awaiting the local
cancelled promise. A pre-listener abort therefore never needs to resolve
that promise; the listener still handles cancellation during child waits.
Add a real local subagent failure regression for SDK ownership. The child
emits agent/error through the normal AgentLoop path and returns an error
result, while the SDK root's wait and a subsequent prompt both complete.
The test also rejects observer warnings, so dropping the ownership guard
fails on the swallowed TypeError rather than merely executing the branch.
No fake Agent identities or private-state mutation are used.
Validation:
- Focused descendant test passed; removing the ownership guard as a
temporary negative control failed on the unexpected observer error.
- pnpm exec vitest run packages/acp/acp/tests packages/sdk/server/tests packages/subagent/subagent-dsh-sdk/tests --coverage --coverage.include=packages/acp/acp/src/session.ts --coverage.include=packages/sdk/server/src/server.ts --coverage.reportsDirectory=/tmp/pr4394-completion-coverage-00243b2f89e-consumer
(251 tests passed; measured statements, branches, functions and lines 100%)
- Focused lint and git diff --check passed.
The early-cancellation audit also reproduced a separate initial-idle wait
liveness issue. It is intentionally left for its own authorized follow-up
commit rather than being combined with this coverage change.
The SDK subagent provider now uses session/wait to keep a delegated child
runtime alive through nested work. A deployment that explicitly selects an
older dshBin can still initialize and accept a prompt but cannot satisfy
that completion request; it must report an execution failure rather than
silently treating the first idle as success.
State the required operation in the provider's owning failure/recovery
contract and update the current activation upgrade guide to tell users to
upgrade a separately configured runtime with the provider. This adds no
capability negotiation, compatibility fallback, or new runtime behavior.
Validation: upgrade-guide structure/word budget and both bilingual pairs
pass. The requirement matches the provider's direct session/wait request
and the SDK server's explicit unknown-method rejection.
A summary can fail while appending turn/start or turn/end and still leave
the root idle with a stable Session sequence. The SDK wait operation then
returned {}, letting the external provider treat the preceding completed
terminal as the result of the entire task. A listener installed only for
the wait would also miss failures between prompt admission and the wait
request.
Retain one latest live failure on each existing SDK Session record, scoped
to the exact owned Agent. Record the Session's next log offset when the
failure occurs. A turn/end from the exact owned Session clears it only
when that terminal was committed at or after that offset. This preserves
failures raised during terminal publication, keeps ordinary represented
turn errors in the durable outcome, and permits later successful activity
to recover. At stable root/descendant completion, session/wait rejects any
remaining failure. Raw session.status and HarnessSession.run behavior are
unchanged; no generic failure history or new notification is introduced.
Add seven owner-local cases for later start/end failures before and during
wait, subsequent successful prompts, represented errors, and publication
ordering including an undefined thrown value. Six failure cases returned
success before the fix. Strengthen listener teardown coverage to assert
all actual subscriptions are disposed after a disposer fails.
Extend the existing recorded subagent-continuable SDK scenario to exercise
session/wait. It still asserts that run() returns DONE at parent turn 1,
then requires child completion and parent turn 2/idle plus its final
summary to be delivered before the wait response. Bypassing the wait made
that ordering assertion fail. No recorded Session or golden output was
invented or refreshed. Update the protocol/server bilingual READMEs and
regenerate the event-consumer matrix for the new error listener.
Validation:
- pnpm exec vitest run packages/sdk/server/tests/wait.spec.ts before the fix:
6 failed, 1 passed (the represented-error case)
- pnpm exec vitest run packages/sdk/server/tests/server.spec.ts packages/sdk/server/tests/wait.spec.ts (40 passed)
- pnpm run test:e2e packages/subagent/subagent-dsh-sdk/tests/nested-delegation.e2e.ts packages/subagent/subagent-dsh-sdk/tests/loader-composition.e2e.ts (8 passed)
- DSH_EXAMPLE_MODE=lib pnpm run test:e2e packages/subagent/subagent-dsh-sdk/tests/nested-delegation.e2e.ts packages/subagent/subagent-dsh-sdk/tests/loader-composition.e2e.ts (8 passed)
- pnpm run test:snapshot snapshots/sdk/sdk.snapshot.ts -t 'replays subagent-continuable through dsh --profile sdk' (1 passed, 24 unselected)
- pnpm run doc-sync: Host compile/bundle, doc-typecheck and 40 other gates
passed; the generated event graph was stale, then regenerated and passed
pnpm run verify-doc-graphs (all 7 graph docs current)
- pnpm run lint:contracts-ready and focused lint of all changed SDK code
- named translation pairing checks and git diff --check
Independent SDK rereview and 7/7 focused tests found no remaining issue.
The parent task owns unchanged Python advanced/recovery validation on the
combined branch; this commit changes no Python or high-level SDK results.
Master adds the claude-code-mods recorded scenario, whose full assembled
prompt and tool-schema pins still describe foreground fork calls and the
removed run_in_background switch. Git merged this new scenario cleanly,
but its actual tool composition uses this PR's managed activation tools,
so CI rejects the stale request header and shared delegation guidance.
Refresh only the scenario's system-prompt and tool-schema pins. Preserve
its Session recording, model replies, Mods behavior, and runtime code.
The expectation remains a strict comparison with the actual shipped tool
assembly; no comparison rule is relaxed.
Validation: the owning lib-mode scenario passes refresh and then strict
replay. The original mismatch is recorded in the current-head snapshots
and artifacts CI job.
The two-stage real-stack workflow test still expected an empty parent
catalog after the caller-delivery discovery fix. That obsolete assertion
fails deterministically in both Linux and Windows coverage jobs: the two
local children now correctly leave continuable membership records.
Compare catalog identities and modes with independently observed workflow
agent-start identities. Keep the assertions that neither parent followup
nor steer was called, and that both child Agents reached disposal. This
checks the intended separation between durable discovery and result
notification rather than restoring the old missing-navigation behavior.
Validation: all three workflow integration cases pass. The original empty
catalog assertion is the exact failure in both current-head coverage logs.
A later summary can fail while appending turn/start, before AgentLoop
advances phase.turn. Its agent/error event therefore carries the previous
completed turn number. ACP ignored that matching number and reported the
old completed outcome. A failed turn/end append was also ignored and
returned cancelled despite no cancellation.
Retain every live Agent failure observed while the prompt owns queued
work. A subsequently committed turn/end supersedes that live failure and
provides the authoritative terminal reason. If the start/end append fails,
no terminal event clears the error, so ACP rejects the prompt. Remove the
unused turn-number argument from the bridge callback; numeric equality is
not evidence that a failure has a durable outcome.
Add deterministic regressions for later turn/start and turn/end append
failures after an earlier completed turn. Both failed before this change:
start failure returned end_turn, while end failure returned cancelled.
Normal represented model/turn errors retain their existing error mapping,
and the existing cancellation/recovery cases remain covered. Update the
bilingual ACP behavior documentation.
Validation:
- pnpm exec vitest run packages/acp/acp/tests/turns.spec.ts -t 'rejects a later'
rejected both pre-fix outcomes (2 regressions reproduced)
- pnpm exec vitest run packages/acp/acp/tests (148 passed)
- pnpm exec tsc -b packages/acp/acp/tsconfig.json
- pnpm run verify-translation-pairing --write packages/acp/acp/README.md
- git diff --check
The ACP completion loop previously observed root idle and no children,
then awaited outputTail outside that loop. An autonomous root turn could
start while an earlier image update was blocked in conversion. Turn
correlation adopted the new turn and cleared endReason, but settlement
did not wait for that turn. Releasing the old image therefore returned
cancelled without any cancellation while the root was still running.
Keep committed output delivery inside the normal completion observation.
After output drains, re-read root status and Session sequence before
settling; any new root activity returns to the idle/descendant checks.
Read mutable cancellation and idle state through local observation
functions so each post-await decision uses the current values. Preserve
the separate cancellation path and its prompt output drain.
Add a deterministic regression that blocks the initial image update,
allows the first idle/no-child check, starts a gated autonomous root turn,
and releases the original image. The prompt remains pending until the
later turn completes, then reports end_turn. A subsequent ACP prompt
runs another real local subagent through the shared delegation fixture.
The bilingual README defines work accepted during output draining as
part of the still-open ACP activity interval.
Validation:
- The new focused regression failed before the fix: it observed cancelled
while the later root turn remained gated.
- pnpm exec vitest run packages/acp/acp/tests (146 passed)
- pnpm exec vitest run packages/acp/acp/tests/turns.spec.ts (39 passed)
- pnpm run test:e2e packages/subagent/subagent-dsh-sdk/tests/nested-delegation.e2e.ts (6 passed)
- pnpm exec tsc -b packages/acp/acp/tsconfig.json
- pnpm run lint:contracts-ready
- pnpm run doc-sync (41 gates passed; doc-typecheck identified an optional
Agent in the extracted test fixture, corrected with an explicit guard)
- pnpm run doc-typecheck (Host build and all 79 documentation blocks pass)
- pnpm run verify-translation-pairing --write packages/acp/acp/README.md
- git diff --check
Independent rereview confirmed both completion races are fixed and that
an original image-delivery failure remains attributed to the prompt even
when a later root turn has already started.
An ACP prompt can be waiting for descendants while its root is idle.
Cancellation aborts that wait and calls agent.cancel(), but a child's
settlement message can wake the root in the next microtask. The wait race
then returns with the root running, causing settlement to enter another
whenIdle() before checking cancelRequested. An unrelated long-running
summary consequently retains the cancelled prompt slot indefinitely.
Recheck cancellation immediately after the descendant wait, before any
new root idle wait. Cancellation stops work present at the cancellation
call and releases that prompt; child messages accepted afterward remain
autonomous work. Do not re-cancel those later turns or drain descendants,
which would either race further wakeups or close future child admission.
Document this cancellation boundary in the bilingual ACP README.
Add a deterministic regression that delivers a child message in the
microtask after agent.cancel(), gates the new root turn, and verifies the
original prompt returns cancelled while that turn is still gated. After
releasing the autonomous turn, a second ACP prompt invokes a real local
subagent activation through a tool and completes without another cancel.
Declare the test's subagent dependency and retain only its workspace
importer change in the lockfile.
Validation:
- The new focused regression failed before the fix: the prompt result was
undefined while the late root turn remained gated.
- pnpm exec vitest run packages/acp/acp/tests/turns.spec.ts (38 passed)
- pnpm run verify-translation-pairing --write packages/acp/acp/README.md
- git diff --check
A registered provider can be used in a composition without the agents
service. startActivation checked the absent registry against the supplied
parent first, so it reported UNAUTHORIZED instead of identifying the
missing runtime dependency with CONTINUATION_UNAVAILABLE. This misdirected
callers toward parent identity even though no activation manager existed.
Require the manager before checking the live parent. Keep provider lookup,
capability validation, and parent identity enforcement in place; provider
startup still cannot run in an incomplete composition. Document the two
failure conditions in the owning bilingual README.
Validation: the new missing-runtime regression failed with UNAUTHORIZED
before the change and passes afterward. All 51 service tests pass, including
unsupported providers/capabilities and invalid parent identity. The changed
README pair verifies successfully.
The unified delegation tool returns as soon as an activation accepts its
prompt. SDK and ACP child runtimes could therefore reach root idle while
an admitted fork was still running. Both external providers interpreted
that idle edge as successful task completion and disposed the runtime,
returning a waiting message while cancelling the nested work.
Add the opt-in SDK session/wait request for existing SDK-owned sessions.
It waits for root idle, managed descendants, and any resulting root turns,
then requires stable idle state and Session sequence across the final
check. Preserve raw session.status notifications and HarnessSession.run's
receipt-to-next-idle API; only the task-owning SDK provider uses the new
completion operation. Subscribe before prompt admission, validate and
fold queued root events through the FIFO wait response, and retain
accepted partial output on cancellation or transport failure.
Keep ACP prompt ownership through descendant completion and later root
summary turns, including their stop reasons and output-delivery failures.
Race child waiting against the existing prompt cancellation controller;
cancellation releases the prompt without permanently closing descendant
admission or repinning the original model selection onto later turns.
Session close retains child-first descendant teardown.
Add real Loader/profile regressions with a gated nested fork for SDK and
ACP completion, cancellation, and final-summary failure. Add focused
coverage for unknown/malformed wait requests, SDK root summary ordering,
ACP cancellation followed by another prompt, and later-summary image
output draining/failure. Update the five owning bilingual README pairs.
Validation:
- pnpm run build
- pnpm run typecheck
- pnpm exec vitest run packages/subagent/subagent-dsh-sdk/tests/subagent-dsh-sdk.spec.ts packages/sdk/server/tests/server.spec.ts packages/acp/acp/tests/turns.spec.ts (123 tests at initial fix)
- pnpm exec vitest run packages/acp/acp/tests/turns.spec.ts packages/sdk/client/tests/sdk-client.spec.ts (80 tests)
- pnpm exec vitest run packages/acp/acp/tests/turns.spec.ts packages/sdk/server/tests/server.spec.ts (70 tests)
- pnpm exec vitest run packages/acp/acp/tests (144 tests)
- pnpm run test:e2e packages/subagent/subagent-dsh-sdk/tests/nested-delegation.e2e.ts (6 tests)
- DSH_EXAMPLE_MODE=lib pnpm run test:e2e packages/subagent/subagent-dsh-sdk/tests/nested-delegation.e2e.ts (6 tests)
- pnpm run test:docs (21 gates)
- pnpm run doc-sync (42 gates)
- pnpm run lint:contracts-ready
- git diff --check
The unchanged SDK next-idle API remains covered by its existing client
tests; no raw notification or persisted Session event format changes.
A structured local activation now accepts continuation messages, but its
terminal predicate checked only delegated children. A correction accepted
during the model request could therefore remain in either Inbox queue while
the earlier response committed structured_output. The loop later consumed
the correction, but the terminal tool guard prevented revising the captured
value, so callers could receive the obsolete answer as completed work.
Require both Inbox queues to be empty as well as all owned children to be
released before staging a structured result. Keep input admission closed
from that synchronous staging point through authoritative result policy,
including the enclosing PTC result. Derive this admission state from the
existing staged, pending, and captured records instead of introducing a
second lifecycle controller. Failed results delete their stage and reopen
input; successful capture keeps the existing terminal behavior. Caller-owned
delivery, complete final text, descendant ownership, and cold resume remain
unchanged.
Add deterministic barriers for accepted steering and queued input, and for
Native/PTC post-execute acceptance and rejection. All six new cases failed
against the unchanged runtime for stale results or incorrectly accepted
late input, then passed with the fix. Extend the authored headless structured
scenario to reject answer 41, consume a parent correction, and commit 42.
Its two own schema sidecars also need the current master bash/edit/write
definitions: retaining their previous definitions fails verifyHeaders on
this pinned base. Update the owning bilingual result-admission documentation.
Validation:
- pnpm run build:native-system
- pnpm exec vitest run packages/subagent/subagent/tests/structured.spec.ts
packages/subagent/subagent/tests/continuation.spec.ts
packages/subagent/subagent/tests/external-activation.spec.ts (214 passed)
- pnpm run test:snapshot:refresh snapshots/session/headless.snapshot.ts
-t subagent-structured-parent
- pnpm run test:snapshot snapshots/session/headless.snapshot.ts
-t subagent-structured-parent (1 selected scenario passed)
- pnpm run test:docs (21 gates passed)
- pnpm run doc-sync (42 gates passed)
- pnpm run lint completed its host build; after correcting test deferred
generic types, pnpm run lint:contracts-ready passed
- git diff --check
SubagentRun.id previously promised uniqueness only within one parent,
but the unified manager indexes every external activation by that id in
one runtime-wide resident map. Publication also checks the live Agent and
Session registries. Two compliant-looking provider results could therefore
collide even when they belonged to different parents.
Document the existing shared Session id namespace across all parents,
providers, and local Sessions. Update the package reference, bilingual
type-equivalent declaration, and the current upgrade guide so custom
providers know to replace parent-local counters with fresh execution ids.
The shipped ACP, DSH SDK, Codex, and Claude Code providers already mint
UUIDs; no runtime mapping or additional retained identity state is needed.
Add a cross-parent collision test through startActivation. It proves that
publication rejects the duplicate, disposes only the new external handle,
leaves the second parent catalog empty, and preserves the first catalog
entry and execution result through normal cleanup.
Validation: all seven external-lifecycle tests pass, as do scoped bilingual
pairing, verify-upgrade-guides, doc-sync, and full lint. Cordis catalog and
inspection-catalog regeneration produces no tracked artifact changes.
The activation refactor removes tool-subagent configuration keys, changes
the model-visible result from foreground or Job-backed output to an
activation receipt, and removes a published in-process driver package.
Updated package references describe the new behavior but do not tell an
existing deployment or package consumer how to migrate those surfaces.
Add the bilingual v0.2.1-alpha.1 upgrade guide and pairing record. Explain
which cordis.yml keys and tool arguments to remove, how to consume parent
completion notices, when host code should await caller-delivery results,
and how publication changes cancellation ownership. Name the existing
Spawn/Fork providers and prepareContinuable API as the driver replacement
without adding a compatibility adapter or promising a foreground mode.
Validation: verify-upgrade-guides, scoped verify-translation-pairing,
verify-md-links, and doc-sync pass. A temporary runtime probe executed the
guide's exact YAML and TypeScript fences and confirmed the caller result,
catalog membership, and absence of a parent notice. The four existing
parent-result tests pass. The probe was removed after execution.
The retained continuation decision still claims a public followup method,
a bare MessageId start result, no activation result/cancellation ownership,
and no human Queue/Steer operations. Those statements contradict the current
startActivation receipt and adjacent-Agent/host message operations already
documented in the same record. The subsystem introduction also says the
activation result can settle before cleanup, despite the manager resolving
that result after teardown and notification.
Correct the current realization in the English/Chinese active documents.
Keep the original single-inbox and residency rationale, distinguish accepted
message identity from the activation receipt, and state the actual result
and disposal timing. Do not change implementation, archived decisions,
provider-level result timing, or runtime APIs.
Validation: doc-sync passed 41 gates on the first run. Its remaining pairing
gate found one Chinese README target using the English locale; correct that
target, re-record the pair, and verify both changed pairs successfully.
The two subagent scenarios introduced by this PR retain tool-schema pins
from before master's Bash deletion/path guidance and filesystem argument
ordering changes. Their actual runtime behavior and recorded model replies
still replay, but strict assembled request-header comparison now rejects
those stale schema descriptions and required-argument order.
Refresh only the three owning schema-pin files against the merged shipped
profile. Preserve every Session recording, model stream, event reference,
timestamp, prompt, and runtime implementation. This keeps the real header
assertion instead of weakening the normalizer or rolling back master tools.
Validation: both affected headless scenarios pass in refresh mode and then
in strict lib-mode replay; the other 33 selected replays passed unchanged.
The recording workflow currently points only to a repository .env file.
A valid key can instead live in ~/.dsh/.credentials.yaml, so missing
DEEPSEEK_API_KEY and .env do not establish that real-model recording is
blocked.
Name the managed credential source and route it through the existing
credentials-local path override when the recording uses a scratch
DSH_HOME. Preserve application-state isolation and keep secret resolution
inside the normal provider rather than printing or manually extracting
key values. This changes only recording instructions.
The collapsed Windows caption toggle and New Session kept a 50% circle after
the unified radius standard moved compact icon buttons to --dsw-radius-sm, so
the toggle's hover fill changed shape when the sidebar collapsed.
Integrate master 78ae4fc2d4d8ecaebb4b674028f781a358b92cc9 after the four
independent review fixes. Keep this base update separate from those fixes
and from subsequent review changes so the integration is independently
reviewable and the observed-master checkpoint remains preserved.
Retain activation-owned execution, caller-independent local discovery,
and unconditional parent final-answer delivery. Preserve master's preset
scoped scheduling tools and all delegated-child schedule deny rules. Move
the new scheduling regression from the removed driver package into the
surviving subagent package and run it through a real Spawn activation.
Preserve master's removal of runtime invariant plugins, dependencies,
exports, and test mounts. Keep context disposal before fixture directory
cleanup. Combine Python workflow ordering with activation/catalog events
without inventing unsupported source-plane import paths.
Keep all frozen Agent Note archives byte-for-byte identical to master;
do not accept Git's automatic edits to renamed archived notes. Retain
current lifecycle and delivery facts in active documentation, repair its
links, and regenerate catalogs/persistence schemas from merged source.
All 62 master persistence roots remain; only subagent/catalog gains the
PR's external membership extension. Keep master's third-party lock graph
and update only workspace dependency changes.
Validation before this checkpoint: 683 focused tests passed, seven
Windows-only cases skipped; 2,772 archived artifacts verified. Generated
catalog and active-pair checks passed. Fresh artifact build and remaining
build-dependent documentation checks continue before publication.
A local child can have send_message available and still finish with a plain
text answer or use the tool only for progress. Selecting completion content
from tool availability drops that captured answer and leaves the parent
with a status-only notice. The recorded spawn case already demonstrates
this: its child returns CHILD_OK without calling the available tool.
Remove the includeClosingText policy and pass every captured final text
through the existing text-only notice renderer. Preserve structured
results, safe diagnostics, external continuation wording, and caller-only
result delivery. No delivery ledger or additional lifecycle state is
needed. A child that sends its final answer explicitly may repeat it in the
completion notice; that cost is preferable to making result delivery
depend on model compliance.
Extend the real-loop parent-result regression with an available-but-unused
messaging tool, restore the unconditional lifecycle expectation, and pin
text alongside structured results and diagnostics. The new regression
fails on the prior implementation because CHILD_RESULT_42 is missing.
Update the owning bilingual descriptions and the exact parent notice
content in headless and TypeScript SDK recordings. Keep replay streams,
released generations, and unrelated recorded clocks unchanged.
Validation: 317 focused lifecycle/workflow tests; full build; 26 focused
headless/SDK replays; Python SDK advanced/recovery smoke (unchanged
fixtures); test:docs (20 gates); doc-sync (42 gates); full lint;
git diff --check.
Caller delivery previously suppressed both settlement messages and local
catalog membership. Workflows therefore stopped publishing the durable
parent-child edge that the parent header and recursive discovery require.
The workflow panel only opens running members, and the ordinary sidebar
hides subagent Sessions, so completed workflow transcripts had no normal
navigation path even though their logs still existed.
Record every successfully admitted local child in its parent's catalog,
independently of the result recipient. Membership answers which local
Sessions belong to the parent; delivery answers who consumes this activation's
result. Keeping those decisions separate restores completed-child and
descendant discovery without duplicating workflow completion messages or
adding return guidance to caller-owned work. Publication failures retain
the existing rollback transaction for both delivery modes.
These local children intentionally retain the unified continuable mode:
list_agents includes them, the parent catalog opens their transcripts after
reload, and the existing direct-parent controls can cold-resume them. No
one-shot execution path or additional lifecycle state is introduced.
Caller-owned external executions still use their consumer's membership
records because they have no local Session to open.
Update the existing bilingual service/workflow documentation and decision
records, refresh the owning workflow/Ralph/advanced Session recordings and
Python SDK event expectations, and preserve unrelated recorded stream times.
Validation:
- Full repository build passed.
- 443 subagent, workflow-host, and Ralph tests; 14 list_agents tests passed.
- Four focused headless workflow/advanced snapshot refreshes and replays passed.
- Three shipped-Web workflow tests passed, including completed-child catalog
navigation after reload and the continuable composer.
- Python SDK advanced and scheduler-recovery snapshots refreshed and replayed
through the built dsh CLI with Python 3.12.
- Three discovery regressions fail with the original catalog exclusion and
pass after restoring this fix.
- test:docs (20 gates), doc-sync (42 gates), lint:contracts-ready, Host and
Client TypeScript builds, and git diff --check passed.
Portal the session-header job popover to the document body and place it
with useAnchoredPosition, so the clipping center column and the right
sidebar no longer cut off its right edge.
Restore six expected-output files to the PR merge-base versions so the
subagent activation refactor does not carry unrelated recording churn.
The persistent-tools SDK notifications contain the same parsed records;
only JSON object key order changed. The multi-turn, empty-response-retry,
error-finish, packed-chunks, and text-turn writer oracles were refreshed
from V3 to V4 representations together with volatile stream clocks and
hook timing. These cases do not exercise the changed subagent behavior.
The snapshot harness explicitly accepts older supported writer oracles
and restores them through the strict Session catalog before comparison.
Keeping the existing outputs retains the same assertions and historical
coverage without rewriting canonical released Session generations.
Validation:
- pnpm install --frozen-lockfile completed.
- pnpm run build passed, including the host addon and runtime artifacts.
- Six focused lib-mode keyless replays passed: SDK multi-turn and
persistent-tools; headless empty-response-retry, error-finish,
packed-chunks, and text-turn (2 files; 6 passed, 155 filtered out).
- Parsed persistent-tools protocol records are equal before and after.
- All six files exactly match merge base 8af775c480d73c4692813387622fe16bf8eadf37;
no canonical Session recordings or production files changed.
- git diff --check passed.
Remove the second child_process mock added during subagent integration.
The existing mock already isolates Git from concurrent workspace writes
and reads the fixture's sourceCommit and dirtyFiles controls. The later
factory replaced those controls with a constant commit and clean status,
so the Windows tests could no longer detect changes made while awaiting
packaging confirmation.
Keep the original mock and its fail-loud command handling. This restores
the existing git-head and git-status rejection cases without changing
packaging behavior or adding a second fixture abstraction.
Validation:
- Owning Vitest file: 3 passed, 7 Windows-x64-only cases skipped on macOS.
- Prior isolated duplicate-mock probe: 2 failures with the duplicate,
2 passes with only the original state-controlled factory.
- Original exact-head Windows CI fails the same two Git-change cases.
JavaScript object enumeration moves canonical array-index property keys ahead
of other string keys. Describing every numeric-looking Loader ID as reordered
is too broad: leading-zero and other noncanonical numeric strings remain
ordinary string properties.
Name the exact key category in the inventory fixture's local comment. The
inventory-* IDs exclude every canonical array-index key, and the real Loader
activation, fixture order, output assertions, and production code are unchanged.
This addresses the automatic review's nonblocking terminology clarification.
Validation:
- Type-aware Oxlint passed for the owning inventory.spec.ts.
- git diff --check passed.
- No behavior or test inputs changed, so the already observed 46-test and
74/74-branch evidence remains applicable.
`collectCliHelp` stops propagating an alias when it reaches a page
already on the current path, so a `help [command]` row that prints its
parent's help no longer recurses until the stack overflows. `serve|s`
command rows are split into separately collected names.
The CLI reference and upgrade guide describe the forwarded-argument
boundary that Commander applies, including a leading pnpm option, and
args.spec.ts pins that `--profile` placed after a leading pnpm option is
forwarded. Generated pages use the repository's two-line generated
header, and docs/AGENTS.md lists the CLI help among the generated
references.
Refs #4114
`scripts/gen-cli-help.ts` runs the source `dsh` launcher with temporary
Harness and Agents homes and captures the launcher, `dsh plugin`, and every
shipped profile's text help (profile names come from `PROFILE_TEMPLATES`).
It recursively discovers subcommands from each page's `Commands:` section,
shows identical alias help once with every invocation, and writes
`docs/cli-help.md`, its Chinese page, and their pairing record. The record
resolves links between the two pages as written, so a first run already
matches `verify-translation-pairing`. `verify-cli-help` checks all three
files and runs as the `cli-help` doc-sync leaf, which the static CI lane
also runs on an unbuilt tree.
`dsh plugin` now owns `-h`/`--help` and reads its own options only before
the forwarded pnpm arguments: `dsh plugin --help` prints its help instead of
failing on the missing `--profile`, while everything after the pnpm command
(including `-h`, `--help`, and `--profile`) reaches pnpm verbatim. A
`--profile` placed after the pnpm command is no longer accepted; the
v0.2.1-alpha.1 upgrade guide covers the migration.
Refs #4114
The parameterized versioned/name-only inventory test reversed plugin arrival
order but left Loader entry IDs random. EntryTree generates hexadecimal IDs
and enumerates its object-backed store with Object.values(). Numeric-only
IDs become JavaScript integer property keys and enumerate before other keys,
so the fixture's declared arrival order did not determine the comparator's
operand order. Both parameter cases could exercise the same orientation
while retaining the correct sorted output, intermittently leaving the right
version's nullish fallback uncovered by the per-file 100% CI gate.
Use the typed root-group update API with fixed non-integer inventory-* IDs
for this fixture. Both requested arrival orders now reach the inventory
collector without integer-key reordering. Keep real Loader imports and
activation, duplicate entries, and the existing output assertions. Explain
the non-obvious identifier requirement locally in the test. Production
sorting, package identity rules, and request fields are unchanged.
A controlled negative uses numeric IDs for versioned entries: all 46 tests
pass, but package source branch coverage is 73/74 (98.64%), with index.ts:186
uncovered exactly as in the observed node 24 coverage failure. The final
non-integer IDs pass the same owning file with 74/74 branches (100%). No
global Math.random interception, retry, timeout, or coverage exception is
introduced.
Validation:
- pnpm install --frozen-lockfile --ignore-scripts installed the existing lockfile.
- pnpm exec vitest run packages/llm/plugin-package-inventory-deepseek/tests/inventory.spec.ts --coverage --coverage.include='packages/llm/plugin-package-inventory-deepseek/src/**/*.ts' --coverage.reporter=text --coverage.reportsDirectory=/tmp/dsh-merge-5719/inventory-negative
reproduced the controlled coverage failure with 46 passing tests.
- The same command with reportsDirectory=/tmp/dsh-merge-5719/inventory-fixed
passed 46 tests and 100% statements (80/80), branches (74/74), functions
(15/15), and lines (68/68) for the package's executable source.
- pnpm --silent run change-scope --base 70996a53b84212db9efafe21a4db3dddd46b47ec
confirms the sole changed path is the owning inventory.spec.ts.
- git diff --check passed.
The capture clone shares one source with the caller's body stream, so the
caller key alone keeps the Request reachable on every path, including when
capture has already stopped at its byte limit. Cover that path in the
regression test and document the process-wide --expose-gc flag.
Refs #5614
When a visible shared file became a content duplicate, reconciliation emitted
a removal notice and discarded its metadata. Its durable state then remained
remove, while the metadata fast path accepted only candidates without any
previous record. Every subsequent unchanged pass therefore reread the file,
even after a later read had populated the cache. This also affected the path
where an edited earlier winner made an unchanged later file a duplicate.
Retain the still-present duplicate's metadata with the removal's deferred
version update. Pass the already-cached state when removal follows an unchanged
metadata probe, and the newly read state when the duplicate's content changed.
Apply this metadata only when the associated removal survives rendering, using
the existing exact-transition filter; confirmed absence still clears metadata.
Allow both never-rendered candidates and previous removals to skip content reads
when unchanged provider metadata and an earlier kept digest identify the file as
a duplicate. Continue probing metadata. Once the winner changes or disappears,
read and promote the shared file normally, then suppress unchanged follow-ups.
Document the metadata guarantee in the local cache types and bilingual README.
Model-visible messages, source records, ordering, and byte budgets stay the same.
Add four production-plugin regressions covering duplication after either file
changes, then promotion after the winner is edited or deleted. They inspect the
provider's actual content reads across two unchanged passes and require exactly
one shared read for promotion, rather than trusting the context's self-report.
Validation:
- All four new cases failed before the fix because each unchanged pass reread
the shared file despite unchanged metadata.
- pnpm exec vitest run packages/context/agent-instructions/tests/agent-instructions.spec.ts -t 'caches a removed duplicate': 4 passed after the fix.
- README pairing was rerecorded; git diff --check passed.
- The removal regression narrows its source kind and compares the change as a
plain typed value; no matcher carrying any is assigned into an expected record.
- Type-aware Oxlint passed for the three affected packages and the new
metadata-failure fixture; focused cache regressions passed again after the
assertion correction.
A content-deduplicated shared file has no visible change record. If its
metadata became unavailable while the harness-home winner changed or vanished,
the reconciler skipped the failed candidate and still published the winner's
replacement or removal. Guidance that remained in the shared file disappeared
from the model's effective instructions. Content-read failures had the same
partial-group behavior even when metadata probing succeeded.
Rollback the probed candidate group whenever any metadata probe is unavailable
or its bounded content read cannot complete. Include candidates without a
visible predecessor and candidates previously removed as duplicates. Reuse
one rollback operation for proposed messages, deferred version updates,
metadata mutations, and retained digest bookkeeping, so a later successful pass
can reconstruct the complete group independent of cache warmth.
A filesystem path can also appear as a project instruction when the Harness
home is the project root. Releasing observed paths on rollback let that later
alias publish the very content withheld by the failed global group. A transient
failure of the first member also left failed and unvisited paths unreserved.
Factor the existing scope-to-file path calculation and reserve every eligible
member's absolute path for the rest of that pass when its group rolls back.
Project aliases therefore cannot bypass group preservation, without additional
filesystem operations. The shared resolver keeps probing and rollback path
calculations identical. Add metadata, content-read, and one-shot first-member
failure regressions for this supported same-file layout and its recovery.
Keep cancellation propagating to the caller. Keep explicit baseline-budget
exclusions outside observation rollback, because their removals are deliberate
policy decisions. No new session fields, baseline identities, polling, retries,
or process-global state are needed. Clarify the metadata/content and hidden
candidate guarantee in the bilingual README.
Add four real-plugin outage/recovery tests for metadata or content failure while
the retained home file is edited or deleted, plus a cold-cache reconciliation
regression. Add an authored headless session scenario with instance-local
provider failure injection: request two keeps baseline A while shared metadata
is unavailable; request three receives the home B replacement and shared A
addition together after recovery. The fixture restores the provider method on
disposal and asserts the real home/shared files before restoring observation.
Declare its test-only filesystem dependency without changing runtime packages.
Regenerate the persistence catalog's source anchors after the shared path
helper import shifts AgentInstructionSource by one line. Its persisted fields
and SHA-256 type fingerprint remain unchanged; no type acknowledgement or
Session migration is required for a source-location-only update.
Validation:
- All five new regressions failed with the intended partial-group emissions
before the source fix.
- pnpm exec vitest run packages/context/agent-instructions/tests/agent-instructions.spec.ts -t 'hidden duplicate has|cold global group|unavailable active candidate|unavailable scope': 8 passed.
- pnpm exec vitest run packages/context/agent-instructions: 174 passed.
- The focused authored headless snapshot refresh and replay passed.
- Restoring the early skip for an unavailable hidden candidate makes the
headless snapshot fail with the premature harness-home replacement.
- README translations were updated together and their pairing rerecorded.
- Three project-alias regressions failed on the initial group rollback fix,
then passed after reserving failed and unvisited candidate paths.
- pnpm exec vitest run packages/context/agent-instructions/tests/agent-instructions.spec.ts -t 'frozen across a project alias|hidden duplicate has|cold global group|caches a removed duplicate': 12 passed.
A baseline can deduplicate identical harness/shared instructions before the
byte renderer omits their retained winner. Removing every deduplicated scope
from excludedScopes then made the shared copy eligible on the next unchanged
step. It injected the broader instructions the baseline had deliberately
omitted, contradicting the retained-baseline budget policy. Project sibling
copies had the same problem.
Allow a duplicate to leave the exclusion set only when the included baseline
contains its candidate group and trimmed digest. Match the digest as well as
the group: a different sibling may fit the budget while the duplicate's winner
does not. Use the included files' original content so truncated and empty
represented winners retain their ordinary duplicate-promotion behavior.
Copies of omitted content inherit that omission for the visible baseline.
Duplicates of represented content remain probed and can be promoted when their
winner changes or disappears. Keep both README languages, the existing shared
root decision, and the local result-field documentation consistent with this
rule. No durable event fields or baseline identities change.
Add plugin regressions for identical global copies across repeated steps and
resumes, and an omitted project digest with a different retained sibling. Add
an authored recorded-session scenario that runs the shipped headless profile
with identical 1900-byte global files and a 700-byte budget: its two model
requests retain one project-only baseline and never inject the shared copy.
Validation:
- Both new plugin regressions failed on the reviewed implementation with two
contexts instead of one, then passed with the fix.
- pnpm exec vitest run packages/context/agent-instructions/tests/agent-instructions.spec.ts -t 'duplicate budget-omitted|different digest fits|content-deduplicated shared|changed shared file': 4 passed.
- pnpm run test:snapshot:refresh snapshots/session/headless.snapshot.ts -t agent-instructions-budget: passed.
- pnpm run test:snapshot snapshots/session/headless.snapshot.ts -t agent-instructions-budget: passed.
- Reverting only src/index.ts to the reviewed head makes that snapshot fail
with an extra agents-global set and 304 bytes of omitted global content.
- pnpm run test:snapshot scripts/session-snapshot-corpus.corpus.ts: 3 passed.
- Scoped bilingual pairing was rerecorded and verified; git diff --check passed.
The shared-root Agent Note still describes agentsHome as an accepted row field
and gives explicit row configuration precedence. The decision retained those
statements from the configurable-root proposal, although the instruction loader
now exposes neither dshHome nor agentsHome. Maintainers could therefore add an
unsupported row override or mistake the skill provider's root options for the
instruction loader's API.
State that instruction loading resolves the harness home from DSH_HOME or
~/.dsh and the shared agents root from DSH_AGENTS_HOME or ~/.agents. Keep the
shared resolver and its tilde/blank-environment behavior, while distinguishing
dsh-skill-filesystem's retained explicit home fields for its own skill roots.
Update the English and Chinese Decision/Consequences passages together and
record their confirmed pairing hashes. Root order, durable scope keys, and the
other decision rationale are unchanged.
Validation:
- pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-10-05-shared-agents-root-instructions.md
- pnpm run verify-translation-pairing .agents/notes/implemented/architecture/2026-10-05-shared-agents-root-instructions.md: one named pair consistent
- pnpm run verify-agent-note-format: all 276 Agent Notes conform
- git diff --check, scoped to the shared-agents-root-instructions EN/ZH/sidecar triplet: passed
The user-global instruction scope now reads two roots: $DSH_HOME/AGENTS.md
first, then the shared agents root that other agent tools use
($DSH_AGENTS_HOME, else ~/.agents). Both resolve from the process environment
through @deepseek-ai/dsh-home-paths; no config field is added, because the
stack base (PR #5718) removed the per-row home override.
The two files are candidates of one candidate group: identical trimmed content
renders once from the harness home, an unobservable candidate preserves the
group's last-good state, and reconciliation probes both roots so external
edits, removals, and new files become visible change notices. A candidate
dropped as a duplicate keeps being reconciled, so deleting or changing the
retained candidate promotes it. The harness-home file keeps its recorded
`user-global` scope key; the shared file carries `agents-global`.
A user-global delta states `These user-global instructions apply to all work.`
instead of naming its reconciliation scope, and a project path whose first
component would collide with a user-global scope directory renders with a
leading `.` so it keeps its own scope.
The removed dshHome row key still redirected instruction loading away from
DSH_HOME. Schemastery's non-strict object validator preserves unknown keys,
and resolveConfig forwarded the validated row wholesale to the library
discovery normalizer, which intentionally accepts an already-resolved home.
Removing the public field and schema entry therefore left the old runtime
override active and contradicted the upgrade guide.
Copy only the declared discovery controls when normalizing a plugin row.
Home resolution now follows process policy even when the row retains a stale
dshHome or another undeclared home field such as agentsHome. This allowlist
also applies when the shared-agents-root stack layer adds its internal home
parameter, without widening the deployment-facing configuration.
Keep the exported discovery and baseline-loading helpers' explicit home
parameter: it carries a library owner's already-resolved path, rather than
granting a plugin row its own home. Extend that regression through baseline
loading to preserve this distinction. Document the ignored stale row key in
both package READMEs and refresh their pairing record. No Session format,
candidate ordering, budget, or instruction rendering changes are required.
Add a production Cordis/schema regression that mounts a row with both stale
home keys and checks model-visible baseline and refresh content. Before the
fix, its baseline contained the stale row instruction instead of the process
home instruction. Both stages now use the process home, and the same test
will reject an agentsHome leak after propagation into the upper stack layer.
Validation:
- pnpm exec vitest run packages/context/agent-instructions/tests/agent-instructions.spec.ts -t 'ignores a stale row home'
failed before the source fix with the intended content mismatch.
- pnpm exec vitest run packages/context/agent-instructions/tests/agent-instructions.spec.ts -t 'ignores a stale row home|uses an already-resolved home'
passed both focused tests after the fix.
- pnpm exec vitest run packages/context/agent-instructions/tests/agent-instructions.spec.ts
passed all 157 tests with both stale row keys present.
- pnpm exec tsc -b packages/context/agent-instructions/tsconfig.json passed.
- pnpm run doc-sync passed all 42 gates.
- pnpm run test:snapshot snapshots/session/headless.snapshot.ts -t 'agent-instructions'
passed the selected recorded-session replay.
- pnpm run verify-translation-pairing --write packages/context/agent-instructions/README.md
refreshed the bilingual record; doc-sync verified it.
- git diff --check passed.
`resolveDiscoveryConfig` accepted `DiscoverOptions.dshHome` but resolved the
home from the environment alone, so a caller passing an already-resolved home
silently received the process home while the option comment claimed otherwise.
It now normalizes the passed value, which keeps the plugin's mount-time
resolution authoritative for discovery and reconciliation. A test pins that an
owner-resolved home wins over `$DSH_HOME`.
Review follow-ups:
- The guide claimed the removed field was the last per-row home override;
other providers (skill-filesystem, shell-env, credentials-local,
subagent-dsh-sdk, attachment-local) keep their own fields, so it now states
that the removal covers the instruction loader, and step 1 also names the
home-level `$DSH_HOME/cordis.patch.yml` layer.
- `DiscoverOptions` regains an already-resolved `dshHome` that `apply()` passes
from its mount-time `resolveConfig()`, so baseline discovery and
reconciliation share one resolution point instead of reading the environment
separately.
- The package README says the loader reads the process home rather than "the
same home as sessions, skills, and storage", and the resolver Agent Note
states what the skill provider's fields actually select.
- `discoverBaselineInstructionFiles` JSDoc no longer lists a removed option.
The agent-instructions row accepted `dshHome`, letting one consumer read a
different harness home than sessions, skills, and storage. Home is process
policy: the plugin now resolves `$DSH_HOME`, else `~/.dsh`, through
`resolveDshHome()` from `@deepseek-ai/dsh-home-paths`, and the config field and
the library's explicit-home option are gone.
Tests and both snapshot fixtures pin the home with `vi.stubEnv('DSH_HOME', ...)`
and the harness's existing `DSH_HOME` export instead of row configuration.
`dsh-skill-filesystem` keeps its own `dshHome`/`agentsHome` fields.
The removed cordis.yml key gets an upgrade guide.
Node's fetch forwards a caller's AbortSignal only while the Request it was
given is reachable. The fetch observer creates that Request locally, so after
garbage collection an abort after response headers no longer cancelled the
body: capture never finished, Network.loadingFinished was never sent, and the
connection stayed open.
Fixes#5614
Review follow-up: `const closed` beside the boolean `this.closed` reads as the
flag, so name the promise `closing` at every site; the test-side comment
repeated the production mechanism verbatim, so each test keeps only the
ordering assumption its fake encodes.
The EPERM path in the Windows coverage failure also matches the HMR
exact-config watcher: it watches the temp harness home with chokidar's
default 2 s write-settle window, and that spec constructs it on the very
directory its afterEach deletes. Apply the same containment to that
watcher and to the skill-filesystem root watcher, the remaining owners
that enable awaitWriteFinish.
A sloppy-mode .cjs or .js file can spell the name as '\141ppendPluginRecord',
which TypeScript decodes to appendPluginRecord, but the prefilter only
removed the backslash and skipped the file. The prefilter now decodes a
legacy octal escape as TypeScript does, taking three digits only after
0 to 3, so the value stays within 0o377. The spec rejects two spellings
that use three-digit octal escapes, and tests the decoder itself for
the digit limit, which no spelling of the name, whose letters are all
0o1xx, can exercise.
chokidar's close() drops every listener while a scheduled write-settle poll
can still run. That straggler stats a document the home teardown is
deleting, which Windows reports as EPERM, and an 'error' emission with no
listener is rethrown as an uncaught exception: the whole coverage
partition fails while every test file passes.
Re-attach the provider's error listener across the close so the straggler
stays inside the provider, and make it a no-op after disposal.
The caller gate parsed only files whose text held appendPluginRecord or a
\u escape, so a release package could reach the function through
S['\x61ppendPluginRecord'], a \u{61} code point escape, a needless
character escape, or a line continuation, which TypeScript reads as the
same string. The gate now also parses a file when its text with every
escape unescaped holds the name, and its spec rejects each spelling.
- Make the record commit an ES private method, so appendPluginRecord()
is its only runtime entry.
- Document PluginRecordType and PluginRecord on the session subsystem
page, and state in the README and JSDoc that the V3-to-V4 edge renames
unknown ignorable V3 events into the plugin: namespace and that plugins
outside this repository must not call the writer.
- The caller gate parses only files that contain the name or a \u
escape, documents that wildcard re-exports are not reported, and its
spec reads a temporary repository instead of this checkout.
- Cover Session.fromRestore in the record test and name SessionEventMap
in the catalog generator's diagnostic.
classShape dropped only #private properties; methods and accessors with a private name reached the declaration text the cordis_inspect catalog serves. It now drops every member whose name is a private identifier, as the public-api projection in verify-type-equiv does.
Add appendPluginRecord(), which appends one ignorable event whose type is
`plugin:` followed by lowercase slash-separated segments, and
pluginRecordOf(), which reads such an event back. Records are not
SessionEventMap members, so the persistence catalog is unchanged; a build
that does not recognize one retains and skips it, and records never enter
the model-visible surface. Resume and fork carry them with the log, and a
format migration keeps them on a best-effort basis.
Only production source under packages/experimental/ may call the writer.
The verify-plugin-record-callers hygiene gate checks every reference in
repository sources, and the persistence catalog generator rejects a
SessionEventMap member in the plugin: namespace.
Session.append() and the record writer share one commit path, so a record
follows the same reentrancy, detach, and observer-containment rules. The
ignorable-events decision, the session README, the format-version
cookbook, and the plugin-author practices describe the records.
A class's static initialization block has no type-level meaning, and the
public-api projection that verify-type-equiv checks already drops it.
The generated declaration text, which the model-visible cordis_inspect
catalog serves, printed it with the class's public members.
pi-ai 1.0.1 and 1.0.2 bring model catalog updates and adapter fixes:
Anthropic native tool changes define later tools inline, Bedrock binds
replayed thinking blocks, OpenAI-compatible requests can pick sampling
parameters by thinking level, and "model is at capacity" is retried.
The runtime dependencies change only in @anthropic-ai/sdk (0.124.0 to
0.129.0) and @earendil-works/pi-telemetry (^1.0.2).
1.0.2 still re-parses the whole accumulated tool-argument JSON on every
stream delta (earendil-works/pi#9265), so the patch is regenerated
against the 1.0.2 sources: it removes the same six per-delta parses as
the 1.0.0 patch, and only the bedrock-converse-stream.js hunk moves, by a
line.
State the Pi-to-DSH direction of the built-in tool translation, how setActiveTools treats Agent-scoped extension tools, that module-scope extension state is shared by concurrent Sessions, which format edges retain plugin records, and that SessionStartSource already declares clear and compact. Date the npm package count and narrow the plugin note's opening sentence.
Record two proposals from the Pi 1.0.0 compatibility investigation.
The Pi extensions bridge note states the design of an experimental bridge
that loads unmodified Pi extension sources on existing DSH extension
points, and ranks the DSH core mechanisms that are narrower than Pi's
extension API.
The Pi plugins note defers a compatibility layer for Pi's Chord-facet
plugins, whose API upstream declares unstable, and defines six levels of
support with the upstream trigger and the DSH work for each, plus the
runtime differences the plugin model exposes.
pi-ai 1.0.0 adds Sign in with ChatGPT to the `openai` provider, and that
login refuses to start unless the app supplies a stable installation UUID
through LoginOptions.getDeviceId. The harness keeps no installation ID
meant for a third-party account sign-in, so the route offers its API-key
login only; a Loader-composed test pins the offered methods.
catalogProvider(), catalogModels(), and catalogProviderIds() now read one
index that leaves out providers without a chat model, so a hand-declared
route under such a key inherits nothing from pi-ai. The browser-worker
stub gains the builtinModels refusal llm-pi-ai imports and drops the
getBuiltinProviders reader nothing imports. The tool-argument streaming
spec also covers the patched Anthropic Messages path.
The chat-model filter left a provider out whenever pi-ai listed no chat
model for it. The browser-worker preview replaces pi-ai with a stub that
lists no models at all, so every provider was dropped and the empty
directory registration failed the llm-pi-ai mount. Only a provider whose
catalog lists models but no chat model is left out now; the stub gains
the `getAllBuiltinModels` reader the filter calls.
pi-ai 1.0.0 adds classifier and image models to the built-in catalog and
ships `typesafe` with classifier models only. The adapter dispatches chat
requests only, so a catalog provider without a chat model gets no entry in
the configurable-provider directory and no sign-in flow.
Move the pi-ai provider adapter from 0.87.1 to 1.0.0 and re-create the
local tool-argument patch for the new dist files. pi-ai 1.0.0 still
re-parses the whole accumulated argument JSON on every stream delta in
the Anthropic, Bedrock, Mistral, OpenAI Completions, OpenAI Responses,
and pi-messages adapters (upstream earendil-works/pi#9265), so the patch
removes the same six per-delta parses; arguments are still parsed once at
toolcall_end.
The release-age exemption moves to the exact 1.0.0 pi-ai and
pi-telemetry versions, and the generated third-party notices and README
links name the new patch file.
Replace the persisted background marker with a client-only rule. A
foreground run writes run-end before its Tool Step ends and a background
run outlives that Step, so only a Turn closed after a crash or at a fork
boundary shows a run without run-end as interrupted.
A background workflow run outlives the Tool Step that started it, so the
Web workflow node inferred interruption as soon as that Step closed and
showed live members as interrupted. Mark background runs on
tool-workflow/run-start and infer interruption only for foreground runs.
Refs #5638
Review follow-up on the watchdog rewrite:
- Register the deadline reaction inside `next()` and detach it when the
demand settles. Racing against one watchdog-lifetime promise kept every
settled race reachable through that promise's reactions, retaining each
resolved `StreamChunk` for the length of the stream: a 10,000-value probe
kept 100/100 sampled values before this change and 0/100 after.
- Take the neighboring stall case's 1,000ms idle budget in the new
transport-recovery case: the retry performs a real local HTTP/SSE round
trip, and a 50ms budget would let coverage-lane scheduling decide it.
- Document the contract where it is described: `packages/util/timeout/README`
(both languages) and the bounded-recovery and timeout-library notes no
longer claim that only the transport observes the watchdog signal.
An aborted read can stay pending forever: undici drops the error when the
abort lands on an in-flight response-body read (observed on the coverage
lane's Node v24.13.0), and the stalled body of
`llm-retry/tests/transport-recovery.spec.ts` then never settles. The idle
watchdog delegated termination to the iterator observing its signal, so the
request outlived its own deadline and the agent stayed `running` until
Vitest's 90s budget; awaiting the abandoned request generator's `return()`
queued behind that demand and would have re-hung the settled outcome.
`idleWatchdog.next` now races its demand against its own deadline and
rejects with the watchdog's TimeoutReason, so the deadline — not the
transport — decides when an outstanding demand ends. Both adapters that
consume the watchdog stop awaiting the abandoned generator's drain; the
abort they already issue owns socket and SDK teardown.
Evidence: instrumented stress harness at ~1s idle on Node v24.13.0 hung 4/160
iterations before this change (each with the abort delivered at ~1002ms and
the underlying read never settling) and 0/160 after; Node v24.19.0 never
hung. The new case in transport-recovery.spec.ts pins the behavior
deterministically: it fails (times out) with the watchdog change alone and
with no product change at all.
Review follow-up: the developer message history is the net tool-list change
between two consecutive requests, not one entry per successful call, so a
remove/add pair batched into one step logs no message. Derive the expected
history from the recorded request headers, search for the later call after the
earlier one, and accept an already-removed tool in the removal prompt so the
REMOVED reply stays required.
The dynamic tool cache e2e asserted the exact recorded call sequence, so a
model that also removed the temporary tool inside the add turn failed an
otherwise valid run. Assert the calls the cache contract depends on and their
relative order instead, and derive the expected transcript registry history
from those recorded calls so one developer message per successful registry
change is still required.
Typing at the end of the highlighted command token lands in the styled
node. The decoration split the overflow back out, but splitText copied
the token style onto it, so Lexical's normalization merged the halves
again and the transform looped until Lexical aborted the update. The
DOM kept the native insertion while the editor state rolled back, which
left the composer desynchronized for every later keystroke.
Clear the overflow style in the same transform pass.
Refs #4490
Review follow-ups: the Electron payload smoke allocates the way the product
does on Windows, so the packaged tree resolves the console host assets it now
depends on; the README states that dependency and its loud failure as a known
limitation; the terminal handle takes the platform this method already
resolved; and the note links the earlier tail-grace note, records why that
grace cannot cover a malformed tail, and drops the claim that the smoke
exercises the operating system host.
The per-platform case reads the third spawn argument to assert the console-host
option is absent on POSIX; an untyped vi.fn() declares no parameters, so the
indexed call argument fails typecheck.
The README, the Ctrl-C comment, and the bare-kill comment still described
Windows' own console host as the one backing the PTY. Name the host the code
now selects, record the exit-event behaviour of both teardown paths measured on
the pool machine, state that a missing conpty asset fails loud, and move the
note to the bug-fix class next to the other Windows console fix. The terminal
case now asserts the option key is absent on POSIX rather than not true.
The note records which categories carry this change: the per-platform unit
case, the master push's Windows standby run as the end-to-end signal, and the
desktop payload smoke for the packaged runtime tree.
The console host Windows provides, on the images the master standby pool runs
(Server 2022, Windows 10 22H2), replays a prompt line that scrolled out of the
viewport as `\ndsh>\x1b[1C`: the trailing space becomes a cursor move, so the
persistent-pwsh readiness path never sees its `dsh> ` tail and every send waits
out the silence tier. The same case passes on the Blacksmith Windows image,
whose console host emits the tail verbatim.
Allocate Windows PTYs with `useConptyDll: true`, so the PTY is backed by the
OpenConsole the node-pty package ships. The option is set only for `win32`;
the terminal spec pins it per platform.
Probing the pool machine with the bundled console host rendered the `dsh> `
tail for all seven prompts of the failing case's command sequence, and
`pty.pid` still resolves to `pwsh.exe` and equals the shell's own `$PID`.
Historical team/message/queued and team/message/delivered records remain
decodable and projected, but Team no longer drains or acknowledges them
after restart. Remove the compatibility dispatcher, its disposal wait,
and its tests, and document that historical pending mail stays undelivered.
The installed-wheel smoke checks real model tool use and externally verifies
the created file, a fresh host-only challenge, the copied receipt, and the
unchanged source. DeepSeek can return the requested acknowledgement before
a file-details paragraph; requiring it to be the last line makes correct
SDK execution fail according to the placement of explanatory prose.
Accept the exact acknowledgement on any standalone line, retaining completed
turns, model-requested tool calls, exact file bytes, and session-log checks.
Replay the observed response through both turns and keep negative coverage
for absent or embedded acknowledgements and incorrect files.
@@ -78,7 +78,7 @@ Each adapter exposes a validated `streamIdleTimeoutMs` configuration field with
`@deepseek-ai/dsh-timeout` exposes a rearmable idle-watchdog primitive. One stable local `AbortController` is fused with the caller signal and passed to the transport for the whole adapter call; each outstanding `next()` arms the watchdog, resolution disarms it, and the next demand rearms it. Out-of-band transport activity calls `pulse()` to rearm an outstanding demand without yielding a value. Timeout aborts that stable controller with a capability-owned `TimeoutReason`, and `finally` clears the timer. The adapter classifies its watchdog as `TIMEOUT` and an earlier upstream abort as `ABORTED`. The existing one-shot `deadline()` is not presented as a sliding timer.
Boundary tests prove termination at both actual transports. The hand-written adapter aborts its fetch/reader, and the pi-ai adapter maps the stable signal through the SDK and proves the SDK closes the response. A timer that merely rejects a consumer promise while leaving the request running does not satisfy the contract.
Boundary tests prove termination at both actual transports. The hand-written adapter aborts its fetch/reader, and the pi-ai adapter maps the stable signal through the SDK and proves the SDK closes the response. A timer that merely rejects a consumer promise while leaving the request running does not satisfy the contract, so the watchdog does both: the signal aborts the request, and the deadline settles the outstanding demand even when the transport never observes that abort.
### Keep attempts separate in the existing log
@@ -117,7 +117,7 @@ If recovery is exhausted, the final failure is stored once on `turn/end.reason`
- Real agent-loop tests cover failure before chunks, partial chunks then failure, thrown and in-band failures, retry to success inside the same turn, exhaustion to structured `turn/end.reason`, and composition with `dsh-compaction-basic` context-overflow recovery.
- The partial-chunk integration test proves failed chunks remain attributed to the failed step, no assistant message or tool side effect is committed for that step, and the successful retry records its own chunk seqs and provider/model route.
- The plugin-owned `llm/retry` event is non-surface, survives a JSONL round trip, is ignored by message derivation, and drives TUI and Web retraction plus scheduled-retry rendering. Client tests cover complete wire validation, clock-independent countdown, cancellation versus completed retry labels, and trajectory attribution; keyless UI snapshots cover Web scheduling and success, real Web composition tests cover partial transport failure through recovery and exhausted recovery's terminal error row beside the settled retry chain, and ACP automation snapshots confirm that a discarded attempt stays off the wire while the recovered reply is emitted.
- Idle-watchdog tests prove the stable signal is rearmed only while `next()` is outstanding, disarmed during consumer think time and in `finally`, and classified separately from a total-call deadline and an earlier caller abort; adapter tests prove the signal stops the underlying request rather than merely detaching it.
- Idle-watchdog tests prove the stable signal is rearmed only while `next()` is outstanding, disarmed during consumer think time and in `finally`, and classified separately from a total-call deadline and an earlier caller abort; adapter tests prove the signal stops the underlying request rather than merely detaching it, and that the idle deadline settles a demand the transport abandoned.
- Direct `ctx.llm.stream()` callers remain single-attempt and receive the same structured failure facts.
@@ -18,7 +18,7 @@ The assembled system prompt had four defects, all of one family: facts the harne
## Decision
**One principle: every fact in the prompt has exactly one owner.** The model name and workspace are config/session facts → the harness exposes them as variables and the persona references them. Per-tool semantics and when-to-use → the tool's `description`. Cross-call habits a description cannot carry → the tool package's prompt section. The product name and SDK identity line → the static `harness:identity` section. Deployment role and behavior → the deployment's persona.
**One principle: every prompt fact has exactly one owner.** The loop exposes route facts as variables referenced by the persona. The working-directory service supplies the current directory as user context. Per-tool semantics and when-to-use belong to the tooldescription; cross-call habits belong to its package prompt section. The product name and SDK identity line belong to the static `harness:identity` section. Deployment role and behavior belong to the deployment persona.
### Assemble context
@@ -28,7 +28,7 @@ The assembled system prompt had four defects, all of one family: facts the harne
Plugins register `{{name}}` values through `ctx.systemPrompt.variable(name, provider)`. Assembly resolves them into the waterfall-visible variable map. Rendering rejects unknown own-property references, registered providers that return `undefined`, malformed complete references, and unbalanced references that still contain a closing `}}`; a lone unmatched `{{` remains prose, and substituted values are not rescanned. Registration rejects invalid or duplicate variable names, and section names are unique. Sections may set `interpolate: false` to preserve generated documents literally; `tools:sdk` does so because tool descriptions and schemas may document their own `{{…}}` syntax.
`dsh-agent-loop` registers the two built-ins, both pure projections of the context agent: `model` (= `options.model`) and `cwd` (= `session.header.cwd`). The example personas write `powered by the {{model}} model`— the model name is stated once, in the `model:` config key. `{{cwd}}` is demonstrated in the ACP example only: every ACP session carries the client's cwd, while config-pre-created stdio agents have none (a persona claiming `{{cwd}}` there fails the turn — by design). The variables stay on the loop plugin (unlike the sections below): they are runtime facts of the agents THIS loop drives, and a replacement loop supplies its own.
`dsh-agent-loop` registers `provider` (= `options.provider`) and `model` (= `options.model`) as pure projections of the context agent. Example personas reference `{{model}}`so the route owns the model name. These variables remain on the loop plugin because they describe the agents it drives; a replacement loop supplies its own. The [Session working-directory decision](2026-09-13-session-working-directory.md) owns directory state and its literal user-context placement; the loop does not register `cwd`.
### Persona as the order-0 section
@@ -54,7 +54,6 @@ Per-tool semantics and selection guidance live in tool descriptions. Prompt sect
## Out of scope
- Further variables (`date`, platform, git state) — the registry makes each a one-line contribution by whichever plugin owns the fact; none is claimed here.
- A config `cwd` for pre-created stdio agents (would let the stdio persona use `{{cwd}}` and partition persistence by real path) — deferred until the session-cwd story is revisited.
## Shipped invariants
@@ -68,5 +67,5 @@ Per-tool semantics and selection guidance live in tool descriptions. Prompt sect
- Every fact in the assembled prompt now has exactly one owner, and the hand-maintained tool prose in leaf YAML is gone: loading or dropping a tool plugin no longer means editing any deployment's persona.
-`{{model}}` reflects `AgentOptions.model` at assembly time. A plugin that switches models in the `agent/request` waterfall makes the prompt's claim stale for that step, and one that SUPPLIES the model there (options.model unset — the loop's documented fallback) leaves the variable valueless at render, failing a `{{model}}` persona before the waterfall runs. Both have the same remedy, and it is the ownership rule itself: the plugin that owns the late-bound model fact states it early on the `system-prompt/assemble` waterfall (`assembly.variables['model'] = …`) — one owner, both statements; a loop test pins the supply path end-to-end. Accepted.
- While a bound provider is absent (not yet activated, unloaded, mid-HMR-reload), the subagent tool does not exist and a model request in that window simply lacks it. That is the honest state — the alternative was a registered tool whose description or execution could not be trusted.
- Strictness means a persona can fail a turn at render (e.g. `{{cwd}}` on a cwd-less session). The failure is contained — the turn ends `error`, the loop survives — and it is an authoring error we WANT loud.
- Strict interpolation rejects missing or valueless variables before a model request. The turn ends with `error` while the loop remains available for later turns; a persona referencing `{{model}}` requires its composing runtime to supply that value.
- Inline escapes remain unsupported in interpolated text; literal sections need no escaping. PTC unit tests cover both modes and runtime languages, and the recorded `ptc-turn` scenario preserves tool-template examples in the model-visible prompt.
@@ -101,7 +101,7 @@ The signal only *notifies*; termination is always the listener's job, and the li
- `SpawnSpec.timeoutMs` and `SpawnOutcome.timedOut`/`aborted` were removed rather than kept as always-zero/always-false vestiges: with `runBash` owning no timer and the executor owning classification, they were read nowhere. An always-0 field read by nothing is dead weight under the per-file coverage gate.
- web_fetch shed its bespoke controller/timer/listener/reason-recovery; the classifier now keys off the deadline signal (`timeoutOf` + `aborted`) rather than the thrown error's shape, which is robust across both the request-phase reject-with-reason and the read-phase bare-`AbortError`.
- `AbortSignal.any` and `using`/`Symbol.dispose` enter the repo for the first time here (Node ≥ 24 baseline, already met).
- Model streams now share one rearmable timer contract without turning a sliding idle interval into a total-call deadline or charging consumer think time. Adapters that can observe out-of-band transport activity may pulse an outstanding demand; suppressed activity remains invisible to the watchdog. The primitive still only notifies; adapter tests prove their transports observe its stable signal and terminate.
- Model streams now share one rearmable timer contract without turning a sliding idle interval into a total-call deadline or charging consumer think time. Adapters that can observe out-of-band transport activity may pulse an outstanding demand; suppressed activity remains invisible to the watchdog. The one-shot `deadline()` still only notifies; adapter tests prove their transports observe the rearmable watchdog's stable signal and terminate. The watchdog additionally settles the one demand it drives when its idle interval elapses, because a transport that drops the error of an aborted in-flight read would otherwise hold the caller past that interval.
Out of scope, named to mark the boundary: `web_search` can gain an optional model-facing `timeout_ms` once its tool-schema/snapshot coverage is planned; the ripgrep-backed fs discovery tools ([packaged ripgrep search](../../archived/architecture/2026-08-01-packaged-ripgrep-search.md)) consume the same provider-owned deadline shape through `dsh-tool-call-timeout-policy` and `exec.signal`; a `tools/execute` waterfall middleware could arm a default deadline for every tool call by driving `exec.signal` — that would be a plugin that *consumes* this library and still only notifies, the hard kill remaining each capability's job.
@@ -111,6 +111,6 @@ Out of scope, named to mark the boundary: `web_search` can gain an optional mode
**Per-tool ad-hoc timeout, no shared code (the prior status quo, and Claude Code's choice).** Rejected because it was already producing divergence and duplicated correctness burden: web_fetch hand-rolled the exact controller/reason logic that future network/process-backed tools would each have to re-derive, and the fusion + `signal.reason` recovery are the error-prone parts. Claude Code tolerates full duplication; this repo has a single shared abort channel (`exec.signal` on every `execute`) that makes a small shared primitive strictly cleaner, so the cost/benefit differs.
**A `withTimeout(promise, ms)` wrapper instead of a signal factory.** Rejected because racing a promise against a timer resolves the *tool-call* promise on deadline without stopping the underlying work — the child process or fetch socket leaks on. Handing out a signal and requiring the capability to listen is what forces a real termination path to exist. This mirrors the "dispose must reach quiescence, not just request it" defensive rule.
**A `withTimeout(promise, ms)` wrapper instead of a signal factory.** Rejected because racing a promise against a timer resolves the *tool-call* promise on deadline without stopping the underlying work — the child process or fetch socket leaks on. Handing out a signal and requiring the capability to listen is what forces a real termination path to exist. This mirrors the "dispose must reach quiescence, not just request it" defensive rule. `idleWatchdog` is the deliberate exception, and only for the single iterator demand it drives: the signal it hands the transport still owns releasing that read (an abort destroys the socket), while the deadline ends the demand when the transport never settles it.
**Keep separate bash timeout and cancellation triggers.** Rejected because one deadline signal removes the bespoke timer and standardizes classification. Racing causes report whichever abort arrived first, while the provider-owned termination path is independent of which cause won.
@@ -44,19 +44,19 @@ The deploy root is [`python/sdk-runtime/package.json`](../../../../python/sdk-ru
The deploy root includes `@deepseek-ai/dsh-mcp-client` as an explicitly supported custom-configuration plugin even though no shipped preset mounts it. An external config can therefore connect to user-supplied stdio and Streamable HTTP MCP servers and register their tools; the distribution does not carry those servers. The optional resource service provides MCP Resources; MCP Prompts remain unsupported. The executable and installed-wheel smokes start a temporary stdio server, discover its tool, and complete one model-requested call.
The [Office kit](2026-09-14-independent-libreoffice-kit.md) and its installed dependency tree live in `<executable-stem>-office/node_modules`, preserving complete package contents and helper permissions. Top-level `pkg.ignore` excludes kit packages from the SEA archive, including assets selected by dependency `files` fields; small dependencies shared with other consumers remain available in the VFS. The private Python bootstrap uses Node `registerHooks` to resolve the kit entry and its package manifest from the real directory. Native executables and URL Workers therefore retain their ordinary paths without modifying the kit or the public CLI. Missing required packages and missinginstalled package manifests fail the copy; absent optionals preserve the kit's engine selection.
The [Office kit](2026-09-14-independent-libreoffice-kit.md) and its installed dependency tree live in the explicitly downloaded sidecar’s `node_modules`, preserving complete package contents and helper permissions. Top-level `pkg.ignore` excludes kit packages from the SEA archive, including assets selected by dependency `files` fields; small dependencies shared with other consumers remain available in the VFS. The private Python bootstrap uses Node `registerHooks` to resolve the kit entry and its package manifest from the real directory. Native executables and URL Workers therefore retain their ordinary paths without modifying the kit or the public CLI. The build locks the selected engine and dependency locations to exact npm tarballs; downloads verify their integrity before publication.
### Build pipeline and artifacts
[`scripts/build-exe-for-python-sdk.ts`](../../../../scripts/build-exe-for-python-sdk.ts): runtime closure verification → `pnpm run build` → (after clearing) `pnpm --filter dsh-python-runtime-closure deploy --legacy --prod --config.allow-unused-patches=true --config.node-linker=hoisted --config.auto-install-peers=false --config.link-workspace-packages=true`**directly into**`python/sdk-runtime/src/deepseek_harness_runtime/runtime/node/`, including [`python/sdk-runtime/runtime-bootstrap.mjs`](../../../../python/sdk-runtime/runtime-bootstrap.mjs) as the carrier-root `runtime-bootstrap.mjs` → restore direct workspace packages omitted by legacy deploy and reject any remaining manifest gap → replace staged dependency symlinks with their target bytes, remove package-manager `.bin` links, and fail if any symlink remains → verify the deployed bootstrap and inject pkg configuration with that bin plus assets covering dynamic profile, bundle, frontend, preset, native-library, and configuration reads → stage the target `node-pty` addon → invoke `pkg --sea` once per target → write `deepseek-harness-sdk-runtime-<platform>-<arch>` under `dist-exe/` and copy it with its target-specific Office directory into the runtime directory. The Python runtime owns that bootstrap. It calls the public CLI export for ordinary launches and dispatches a provider-private selection to the same `@deepseek-ai/dsh-subprocess-local/runner` core without changing CLI grammar or adding another executable; the [native-containment reference](../../../../packages/subprocess/subprocess-local/README.md) owns that private path. Linux CI rebuilds `pty.node` inside the matching manylinux 2.28 container because legacy deploy omits that install side effect. Every target copies its native `@vscode/ripgrep` binary beside the executable as the required `-rg` sidecar; pkg runtimes select that sidecar through `process.pkg`, while ordinary Node execution uses `@vscode/ripgrep` directly. macOS uses its target prebuild and also emits the required `-spawn-helper`. The deploy flags preserve the measured runtime layout: `--legacy` is the mandatory path with inject-workspace-packages off; hoisted gives pkg a stable single-instance layout that the explicit materialization pass makes symlink-free; disabling automatic peer installation prevents undeclared peers from expanding the closure; link-workspace-packages selects direct workspace dependencies. [`pnpm-workspace.yaml`](../../../../pnpm-workspace.yaml) overrides the transitive `@deepseek-ai/cosmokit` and `@deepseek-ai/schemastery` semver requests to the pinned vendor sources so legacy deploy never resolves those unpublished names from a registry. Production deployment permits unused workspace patches because development-only tools can be absent from its closure; included-package patch failures still fail deployment. Workspace installs retain strict unused-patch validation.
[`scripts/build-exe-for-python-sdk.ts`](../../../../scripts/build-exe-for-python-sdk.ts): runtime closure verification → `pnpm run build` → (after clearing) `pnpm --filter dsh-python-runtime-closure deploy --legacy --prod --config.allow-unused-patches=true --config.node-linker=hoisted --config.auto-install-peers=false --config.link-workspace-packages=true --config.hoist-workspace-packages=false`**directly into**`python/sdk-runtime/src/deepseek_harness_runtime/runtime/node/`, including [`python/sdk-runtime/runtime-bootstrap.mjs`](../../../../python/sdk-runtime/runtime-bootstrap.mjs) as the carrier-root `runtime-bootstrap.mjs` → restore direct workspace packages omitted by legacy deploy and reject any remaining manifest gap → replace staged dependency symlinks with their target bytes, remove package-manager `.bin` links, and fail if any symlink remains → verify the deployed bootstrap and inject pkg configuration with that bin plus assets covering dynamic profile, bundle, frontend, preset, native-library, and configuration reads → stage the target `node-pty` addon → invoke `pkg --sea` once per target → write `deepseek-harness-sdk-runtime-<platform>-<arch>` under `dist-exe/` and copy it with its lightweight resource manifest and Office skill assets into the runtime directory. The Python runtime owns that bootstrap. It calls the public CLI export for ordinary launches and dispatches a provider-private selection to the same `@deepseek-ai/dsh-subprocess-local/runner` core without changing CLI grammar or adding another executable; the [native-containment reference](../../../../packages/subprocess/subprocess-local/README.md) owns that private path. Linux CI rebuilds `pty.node` inside the matching manylinux 2.28 container because legacy deploy omits that install side effect. Every target copies its native `@vscode/ripgrep` binary beside the executable as the required `-rg` sidecar; pkg runtimes select that sidecar through `process.pkg`, while ordinary Node execution uses `@vscode/ripgrep` directly. macOS uses its target prebuild and also emits the required `-spawn-helper`. The deploy flags preserve the measured runtime layout: `--legacy` is the mandatory path with inject-workspace-packages off; hoisted supplies package directories that staging makes symlink-free; staging removes nested workspace copies only when pnpm records the same dependency identity at the payload root, preserving singleton services; disabling workspace hoisting excludes unrelated workspace source packages; disabling automatic peer installation prevents undeclared peers from expanding the closure; link-workspace-packages selects direct workspace dependencies. [`pnpm-workspace.yaml`](../../../../pnpm-workspace.yaml) overrides the transitive `@deepseek-ai/cosmokit` and `@deepseek-ai/schemastery` semver requests to the pinned vendor sources so legacy deploy never resolves those unpublished names from a registry. Production deployment permits unused workspace patches because development-only tools can be absent from its closure; included-package patch failures still fail deployment. Workspace installs retain strict unused-patch validation.
CI: [`.github/workflows/build-exe-for-python-sdk.yml`](../../../../.github/workflows/build-exe-for-python-sdk.yml) runs [installed-wheel validation](../testing/2026-08-23-installed-python-wheel-black-box-ci.md) on Linux/Windows x64 for pull requests and Linux ARM64 plus both macOS architectures for master pushes. The [public publication workflow](../../archived/process/2026-08-11-python-publication-workflow.md) calls it for all five targets; `workflow_dispatch` can still select a subset. Native builds run on linux-x64 / linux-arm64 (`ubuntu-24.04-arm`) / macos-arm64 / macos-x64 (`macos-15-intel`) / win-x64 (`windows-2025`), with `~/.pkg-cache` cached where applicable, and pkg handles macOS ad-hoc signing. Each leg installs the release-shaped SDK and runtime wheels into a clean venv outside the checkout, verifies their package and executable origin, then drives the complete keyless scenario set through the public SDK and direct NDJSON JSON-RPC. Trusted pull requests and master pushes additionally run a real DeepSeek two-turn tool smoke on their selected targets; fork and Dependabot heads receive no key. Linux inspects the executable and native addon's GLIBC requirements and runs an additional manylinux 2.28 smoke, while macOS checks the runtime, ripgrep, and PTY helper architectures and verifies that all three deployment targets fit the wheel tag. A full five-target run retains six artifacts, each containing one release file: the platform-independent SDK wheel and five native runtime wheels; a subset dispatch retains the SDK wheel and selected runtime wheels. Bare executables and source bundles remain intermediate test inputs. [`.gitlab-ci.yml`](../../../../.gitlab-ci.yml) accepts `python-v<repository-version>` tag pipelines whose version matches the root `package.json`, builds one SDK wheel and five native runtime wheels, then a single serialized job checks and publishes all six to the project PyPI registry. The [`python/sdk-runtime` README](../../../../python/sdk-runtime/README.md) owns the Windows target and the explicit exclusion of Windows arm64.
### Python SDK distribution: two carriers, exe for production, node for development
The Python SDK lives at [`python/`](../../../../python/README.md): `python/sdk` is the client and `python/sdk-runtime` is the runtime carrier package. The runtime package's data directory holds the build-injected platform executable with its required `-rg` sidecar, `-office/` directory, and optional macOS helper, plus the build-injected `runtime/node/` closure tree for repository development. `resolve_bundled_launch_args()` selects the executable by default; explicit `DSH_RUNTIME_MODE=node` runs `runtime/node/node_modules/@deepseek-ai/dsh/lib/bin.js` on system Node 22.19 or newer. The node carrier never enters wheel distributions, and neither carrier uses a checked-in complete `cordis.yml`.
The Python SDK lives at [`python/`](../../../../python/README.md): `python/sdk` is the client and `python/sdk-runtime` is the runtime carrier package. The runtime package's data directory holds the build-injected platform executable with its required `-rg` sidecar, lightweight download metadata and Office skills, and optional macOS helper, plus the build-injected `runtime/node/` closure tree for repository development. `resolve_bundled_launch_args()` selects the executable by default; explicit `DSH_RUNTIME_MODE=node` runs `runtime/node/runtime-bootstrap.mjs` on system Node 22.19 or newer. The node carrier never enters wheel distributions, and neither carrier uses a checked-in complete `cordis.yml`.
[`scripts/build-python-release.py`](../../../../scripts/build-python-release.py) reads the authoritative `X.Y.Z` or prerelease version from the repository root `package.json`, converts prereleases to their PEP 440 spelling, and stages both packages at that wheel version, with `deepseek-harness-sdk` depending exactly on the matching `deepseek-harness-runtime-bin`. An optional `python-v<repository-version>` release tag is a consistency assertion and is rejected when it differs from the repository version; the source `pyproject.toml` development sentinel never determines a release version. Staging also carries the repository license into both wheels and the third-party notices into the bundled runtime wheel. The SDK is a `py3-none-any` wheel; each wheel-only runtime package contains one exe, its Office directory, and its architecture-matched ripgrep sidecar, and the macOS wheel also contains its architecture-matched spawn helper. Runtime wheels use `py3-none-manylinux_2_28_x86_64`, `py3-none-manylinux_2_28_aarch64`, `py3-none-macosx_14_0_arm64`, `py3-none-macosx_14_0_x86_64`, or `py3-none-win_amd64`; the Hatch hook rejects sdists, universal tags, mixed-platform payloads, missing or extra sidecars, and unsupported platforms. Both macOS tags deliberately declare a conservative 14.0 installation floor: the packaged Node 24 executables declare macOS 13.5, and the x64 PTY helper declares 10.7, but release validation proves the complete payload only against the 14.0 wheel claim rather than promising each observed component minimum as a supported host. The two macOS wheels remain architecture-specific; no universal2 wheel is published.
[`scripts/build-python-release.py`](../../../../scripts/build-python-release.py) reads the authoritative `X.Y.Z` or prerelease version from the repository root `package.json`, converts prereleases to their PEP 440 spelling, and stages both packages at that wheel version, with `deepseek-harness-sdk` depending exactly on the matching `deepseek-harness-runtime-bin`. An optional `python-v<repository-version>` release tag is a consistency assertion and is rejected when it differs from the repository version; the source `pyproject.toml` development sentinel never determines a release version. Staging also carries the repository license into both wheels and the third-party notices into the bundled runtime wheel. The SDK is a `py3-none-any` wheel; each wheel-only runtime package contains one exe, lightweight resource metadata and skills, and its architecture-matched ripgrep sidecar, and the macOS wheel also contains its architecture-matched spawn helper. Runtime wheels use `py3-none-manylinux_2_28_x86_64`, `py3-none-manylinux_2_28_aarch64`, `py3-none-macosx_14_0_arm64`, `py3-none-macosx_14_0_x86_64`, or `py3-none-win_amd64`; the Hatch hook rejects sdists, universal tags, mixed-platform payloads, missing or extra sidecars, and unsupported platforms. Both macOS tags deliberately declare a conservative 14.0 installation floor: the packaged Node 24 executables declare macOS 13.5, and the x64 PTY helper declares 10.7, but release validation proves the complete payload only against the 14.0 wheel claim rather than promising each observed component minimum as a supported host. The two macOS wheels remain architecture-specific; no universal2 wheel is published.
The Python client launches the packaged `dsh` command with the selected profile (`sdk` by default), ordered patch files, and an explicit Harness home. The profile owns JSON-RPC serving and application composition; missing homes, profiles, bundles, patches, and server rows fail without an external complete-config fallback.
@@ -74,7 +74,7 @@ The verification surface has three tiers. Mechanism tier: the measured conclusio
Manual-driving caveat: the bin treats stdin EOF as "the client is gone" and disposes immediately, so a short-lived pipe aborts an in-flight turn — pipe-driven runs must keep stdin open until the turn ends.
The installed-wheel `sdk-office` scenario copies the target's complete payload to a private temporary directory, requires exactly one target backend (the declared native target, or WASM when no native target is declared), and converts DOCX once through the shipped SDK profile with that engine. It checks that the kit module resolves inside the relocated directory, that conversion reports the expected backend, and that the output contains PDF bytes. Wheel validation checks the declared engine assets and native helper execute bit; CI uploads the completed wheel, so nested helper permissions stay inside the ZIP archive.
The installed-wheel `sdk-office` scenario relocates the lightweight wheel payload and explicitly downloads the complete npm Office sidecar with its independent Node. It requires exactly one target backend, verifies that the SDK loads the kit from the downloaded sidecar, and converts DOCX through the skill CLI with an empty PATH. Python Office operations use the same CLI. Explicit downloads reuse the existing build preparation implementation; SDK startup performs no downloads. The [runtime reference](../../../../python/sdk-runtime/README.md) owns cache and activation behavior.
@@ -36,6 +36,8 @@ The adapter calls pi-ai's `streamSimple()` so each catalog model chooses its reg
pi-ai's common stream options do not expose stop sequences. `dsh-llm-pi-ai` rejects a defined Harness `stop` option with `UNSUPPORTED_OPTION` rather than silently ignoring it or growing a second provider-specific payload implementation. `dsh-llm-deepseek` continues to support `stop` through its native request serializer.
pi-ai owns protocol-specific system-message placement and the concatenation fallback for models without in-history system support. Harness preserves later system messages in transcript order and reports catalog capabilities to prompt admission and shared tool-history projection. Duplicating pi-ai's fallback or applying Anthropic's placement rules to every protocol would create a second serialization policy that can drift when pi-ai changes. Capability reporting remains a Harness responsibility because the loop needs it before serialization.
### Recorded assistant route and replay state
Assistant messages carry the request's `provider` and `model`, plus an optional JSON-serializable adapter replay state. A successful `assistant/message` session event records those fields and `deriveMessages()` returns them with the assistant message. User, system, context, and tool-result messages carry no assistant route fields. The provider/model fields are authoritative loop data; an adapter owns only its opaque replay-state payload.
@@ -21,7 +21,7 @@ One resolver owns the harness home, in `@deepseek-ai/dsh-home-paths`, single-roo
explicit configured path > $DSH_HOME > ~/.dsh
```
An empty or whitespace-only `$DSH_HOME` is treated as unset; otherwise `resolve('')` would silently place the home at the current working directory. The harness keeps all user data under one root; there is no XDG config/data/cache split. `dshHomePath(...segments)` joins deployment-owned children onto that root, and `dsh-app-boot` exposes it to Loader `!!js` config expressions before mounting entries, so shipped compositions derive `sessions` and `storages` without copying the resolver. `dshHomeDisplay()` names a resolved root symbolically for user-facing paths — `~/.dsh` for the default home, `$DSH_HOME` for any configured home — so the user-global `AGENTS.md` label never leaks an absolute machine path. It replaces agent-instructions's bespoke default-vs-`$DSH_HOME` check.
An empty or whitespace-only `$DSH_HOME` is treated as unset; otherwise `resolve('')` would silently place the home at the current working directory. The harness keeps all user data under one root; there is no XDG config/data/cache split. `dshHomePath(...segments)` joins deployment-owned children onto that root, and `dsh-app-boot` exposes it to Loader `!!js` config expressions before mounting entries, so shipped compositions derive `sessions` and `storages` without copying the resolver. `dshHomeDisplay()` names a resolved root symbolically for user-facing paths — `~/.dsh` for the default home, `$DSH_HOME` for any configured home — so the user-global `AGENTS.md` label never leaks an absolute machine path. It replaces agent-instructions's bespoke default-vs-`$DSH_HOME` check. Home is process policy rather than row configuration: `dsh-agent-instructions` accepts no home override, so a composition cannot point instruction loading at a root that sessions, skills, and storage would not read. `resolveAgentsHome()` and `agentsHomeDisplay()` apply the same precedence, tilde expansion, and symbolic labelling to the shared agent configuration root (`$DSH_AGENTS_HOME` or `~/.agents`), which stays outside the harness home because other agent tools share it ([shared agents root](2026-10-05-shared-agents-root-instructions.md)); `dsh-skill-filesystem` still accepts its own `dshHome` and `agentsHome` fields, which select its own skill roots and resolve through the shared resolver in the instruction-loading change.
`dshCachePath(...segments)` derives paths below the resolved home's `cache` directory. An initial `{ dshHome }` option preserves a provider's explicit home override. It resolves paths without creating directories; callers own directory creation. `attachment-local` uses this helper for regenerable request-image variants while retaining durable attachment objects in their versioned storage tree, so clearing the cache cannot remove Session attachments. Existing request-image cache entries are left in place and are not read or copied; a cache miss regenerates the variant from its durable attachment.
@@ -13,9 +13,11 @@ The web GUI host serves `/api` over plain loopback HTTP (default `127.0.0.1:3080
Enforce browser trust once, at the carrier, for the entire `/api` prefix — two halves:
- **Media-type fence (dsh-client-connection)**: every `/api` POST must declare `application/json`, else 415 before parsing. Cross-site "simple" requests thereby stop existing: any cross-site attempt is forced into a CORS preflight this server never answers.
- **Authority fence (dsh-client-connection, `src/api-request-trust.ts`)**: every request must present a `Host` that is loopback or matches a `trustedHosts` entry (exact on `host:port`, any port on port-less entries, WHATWG-normalized; rebinding defense). Deliberately no shortcut for unmarked requests: over plain HTTP a browser attaches neither `Origin` nor Fetch-Metadata to reads (EventSource, images, navigations — those headers go only to trustworthy destinations), so an unmarked request may be a rebound browser read whose response the page can read, and Host is the one header rebinding cannot forge; non-browser clients pass via loopback, the derived LAN IP literals, or a declared authority. An attached `Origin` must equal the Host authority; `sec-fetch-site: cross-site` is refused outright. A `trustedHosts` entry that is not a bare, canonical authority fails the plugin load — WHATWG parsing would otherwise quietly authorize the hostname inside a typo or broaden an exact-port grant. `host.pickDirectory` loses its bespoke guard and rides the same fence.
- **Authority fence (dsh-client-connection, `src/api-request-trust.ts`)**: every request must present a `Host` that is loopback or matches a `trustedHosts` entry (exact on `host:port`, any port on port-less entries, WHATWG-normalized; rebinding defense). Deliberately no shortcut for unmarked requests: over plain HTTP a browser attaches neither `Origin` nor Fetch-Metadata to reads (EventSource, images, navigations — those headers go only to trustworthy destinations), so an unmarked request may be a rebound browser read whose response the page can read, and Host is the one header rebinding cannot forge; non-browser clients pass via loopback, a declared authority, or the listener's own bind literal that [Bind one concrete Web address](2026-09-19-concrete-web-bind-address.md) admits. An attached `Origin` must equal the Host authority; `sec-fetch-site: cross-site` is refused outright. A `trustedHosts` entry that is not a bare, canonical authority fails the plugin load — WHATWG parsing would otherwise quietly authorize the hostname inside a typo or broaden an exact-port grant. `host.pickDirectory` loses its bespoke guard and rides the same fence.
Reachability is the webserver binding's policy (`host: 127.0.0.1 | 0.0.0.0`), and this fence is a confused-deputy defense rather than identity. Connection applies the separate [browser token authentication](2026-08-24-browser-token-authentication.md) after the fence. The fence does not inspect peer socket addresses: binding expresses reachability, `trustedHosts` names accepted authorities, and the socket address adds nothing the Host/Origin checks need.
Reachability is the webserver binding's policy (`host`; [Bind one concrete Web address](2026-09-19-concrete-web-bind-address.md) owns which literals it accepts), and this fence is a confused-deputy defense rather than identity. Connection applies the separate [browser token authentication](2026-08-24-browser-token-authentication.md) after the fence. The fence does not inspect peer socket addresses: binding expresses reachability, `trustedHosts` names accepted authorities, and the socket address adds nothing the Host/Origin checks need.
`HostConnectionHandle.allowsRemoteAuthorities` projects whether any validated `trustedHosts` entry names a non-loopback hostname: a loopback listener behind a proxy can still serve remote browsers, so the directory picker requires a policy without remote authorities before it picks the native chooser.
## Alternatives considered
@@ -26,6 +28,6 @@ Reachability is the webserver binding's policy (`host: 127.0.0.1 | 0.0.0.0`), an
## Consequences
- Any future `/api` method is covered by construction; there is no per-route trust decision left to forget.
- A custom non-loopback composition must trust its serving authorities or requests are refused, then satisfy browser authentication like every loopback request. The shipped CLI rejects `--host 0.0.0.0`;`--trusted-host` only extends the Host/Origin fence and grants no identity.
- A non-loopback listener accepts its own bind IP directly ([Bind one concrete Web address](2026-09-19-concrete-web-bind-address.md)); other remote authorities require configured trust. Every accepted authority still requires browser authentication.`--trusted-host` extends only the Host/Origin fence and grants no identity.
- Clients must label POST bodies `application/json` (ours always did; raw-fetch tests gained the header).
- Host and Origin remain request-routing evidence only. The process token and signed cookie establish the browser identity used by every Host method.
@@ -26,7 +26,7 @@ The package ships the runtime alongside the protocol; it remains independently b
## Wire contract
Frames are JSON-lines on fd 3, one object per line, leaving stdout/stderr free for the program's own output. Child → host: `boot-ack`, `call`, `log`, `done`. Host → child: `boot` (first frame), `run` (after `boot-ack`), and one `reply` per `call`. The `log` frame's `truncated` flag marks the frame that IS the child ledger's own truncation marker, so the host stops capturing at the same point the child did instead of inferring it from its own budget. The `log` frame's `open` flag marks an unterminated line committed by an explicit flush: the host holds it and appends the next frame to the same entry, so an explicit flush followed by more text reads back as one line rather than a fake newline. The one exception is truncation: when a later over-budget frame trips the ledger, the already-billed prefix is committed as its own entry and the truncation marker follows it (marker last, no re-charge). The merged entry's wire cost is billed exactly once, split incrementally across its fragments on both sides (O(k) for k fragments, never a re-walk of the whole hold): the FIRST fragment pays the full JSON-string cost plus the separator, each continuation and the closing frame pay only their content; the host's exact-cost caps are `logBudget - 1` for a first fragment (the ledger's reserved byte, matching `admit`) and `logBudget + 2` for a continuation or closing frame (billed without the two quotes), and `jsonStringCostUpTo` returns `undefined` below a 2-byte cap; the child keys its split billing off `_open_started` alone, so a closing frame bills as the merged tail. `done.error.kind` is one of `exception`, `invalid-output`, `output-limit`; wall/CPU budgets, aborts, and substrate death are observed host-side, not carried as frames.
Frames are JSON-lines on fd 3, one object per line, leaving stdout/stderr free for the program's own output. Child → host: `boot-ack`, `call`, `log`, `done`. Host → child: `boot` (first frame), `run` (after `boot-ack`), and one `reply` per `call`. The `log` frame's `truncated` flag marks the frame that IS the child ledger's own truncation marker, so the host stops capturing at the same point the child did instead of inferring it from its own budget. The `log` frame's `open` flag marks an unterminated line committed by an explicit flush: the host holds it and appends the next frame to the same entry, so an explicit flush followed by more text reads back as one line rather than a fake newline. The one exception is truncation: when a later over-budget frame trips the ledger, the already-billed prefix is committed as its own entry and the truncation marker follows it (marker last, no re-charge). The merged entry's wire cost is billed exactly once, split incrementally across its fragments on both sides (O(k) for k fragments, never a re-walk of the whole hold): the FIRST fragment pays the full JSON-string cost plus the separator, each continuation and the closing frame pay only their content; the host's exact-cost caps are `logBudget - 1` for a first fragment (the ledger's reserved byte, matching `admit`) and `logBudget + 2` for a continuation or closing frame (billed without the two quotes), and `jsonStringCostUpTo` returns `undefined` below a 2-byte cap; the child keys its split billing off `_open_started` alone, so a closing frame bills as the merged tail. `done.error.kind` is one of `exception`, `invalid-output`, `output-limit`, or `timeout` from the CPU check after model code returns; wall deadlines, aborts, and substrate death are observed host-side. The post-return CPU check attempts a diagnostic on fd 3 before mandatory signal termination because a sandbox launcher can convert that signal into an ambiguous exit code. Ordinary soft-limit expiry retains default `SIGXCPU` termination without a Python handler, so native calls cannot postpone enforcement until Python resumes. These frames remain untrusted outcome reports; kernel resource limits enforce the CPU ceiling independently.
@@ -520,7 +520,7 @@ A consumer may import a Remote contract that is not currently mounted on the Hos
Connection's general channel API must suit both the current HTTP carrier and a future WebSocket carrier. If the Client Remote or Gateway exposes `fetch`, an HTTP request, or a route handle, WebSocket migration will pierce the Remote layer again. Those physical objects must therefore remain internal to Connection.
Remote endpoints use Connection's `trusted-host` authority. Loopback is accepted by default and LAN callers require an explicit trusted-host configuration, but this layer adds no per-method caller authorization; every trusted host can invoke a mounted Remote endpoint.
Remote endpoints use Connection's `trusted-host` authority. Loopback and the listener's own bind address are accepted by default; any other LAN authority requires an explicit trusted-host configuration. This layer adds no per-method caller authorization; every accepted authority can invoke a mounted Remote endpoint.
`hasSeen()` favors strict-definition safety over SRC availability. While a strict descriptor is withdrawn, such as during HMR, the Gateway continues to claim the endpoint and reports it unavailable instead of falling back to a weak SRC descriptor. Re-registration restores it; only a Typert registry restart forgets the historical strict definition.
@@ -12,9 +12,11 @@ The earlier shipped composition avoided this mismatch by keeping forked children
## Decision
The shared activation entry point, external execution ownership, and caller-versus-parent result delivery are governed by [Unified subagent activations](../simplification/2026-09-17-unified-subagent-activations.md). This record retains the independent rationale described below.
The model-facing `send_message` tool is registered globally for every Agent in a composition. A continuable forked child therefore receives the same tool name, description, schema, and ordering as its parent. Its initial task is appended after the inherited Session seed, and the task includes the direct parent id plus guidance to return results with `send_message({ agent_id, message })` when that tool is visible to the child.
The base and headless compositions retain one-shot fork as their conservative lifecycle policy. The `cordis`, `standard`, and `ptc` CLI presets may bind fork to the continuable lifecycle because that binding no longer inserts child-only request-head fields. `ForkInProcessProvider.prepareContinuable()` and `ctx.subagents.startContinuable()` remain the implementation seam for those presets.
The base bundle and the `cordis`, `standard`, and `ptc` CLI presets bind fork to the continuable lifecycle. Base-backed profiles, including headless, inherit that default. `ForkInProcessProvider.prepareContinuable()` and `ctx.subagents.startActivation()`preserve the same request-prefix behavior across these compositions.
Byte-identical prefix reuse is qualified by explicit deployment choices. A fork delegation that applies a child persona or `toolFilter` may still change the request head. In particular, filtering out `send_message` removes both the schema and the return guidance from the child; the runtime does not bypass an explicit allow-list.
@@ -32,8 +34,8 @@ Byte-identical prefix reuse is qualified by explicit deployment choices. A fork
- Parent and continuable-fork child expose byte-identical ordered tool schemas when the delegation does not request a persona or tool filter.
- The inherited Session seed precedes the child's initial task and return guidance.
- The base and headless profiles keep one-shot fork, while selected CLI presets exercise continuable fork without a child-only request-head addition.
-A child sends zero or more messages to its direct parent explicitly; its final answer is not implicitly copied. The manager-owned settlement notice remains unconditional and separate.
- The base bundle and full CLI presets use local continuable fork for both parent and caller delivery, without a child-only request-head addition.
-The child can send messages to its direct parent through `send_message`. Parent delivery adds a separate runtime settlement notice; caller delivery returns the result through the activation receipt.
- Keyless snapshots and package tests pin schema equality, inherited-history ordering, parent-id guidance, and child-to-parent delivery through the same `send_message` operation used in the other direction.
@@ -10,15 +10,17 @@ Session logs must be upgradable after release, and the runtime that ships first
## Decision
The [reader-compatibility review decision](../process/2026-10-08-session-reader-compatibility-review.md) supersedes the structural bump criterion. This note owns the integer, unknown-event marker, direction-aware refusal, and generation-preservation decisions.
**One monotonic integer, no major/minor split.** Whether a version step is auto-upgradable is a property of that step — expressed by whether its upgrader exists — not something a two-level numbering scheme should promise in advance; design time rarely reveals whether the next change will turn out "major".
**The writer decides bumps, not the reader.** A bump is required exactly when an old runtime could no longer handle a new log with full semantic correctness. "Parses without error" is not the bar: silently skipping content that shapes reconstruction is a wrong read. Only structural changes qualify — header shape, event envelope, core event semantics, the surface mechanism (`SurfaceEventType` set, `SurfaceOp` variants). When unsure, bump: a near-identity upgrader is almost free, a missed bump silently corrupts old readers.
**Bump when existing discriminators cannot prevent unsafe interpretation.** Successful parsing alone does not establish compatibility: silently skipping content required for reconstruction is a wrong read. Same-version changes require safe old-reader handling and preservation of historical meaning by new readers. The linked reader-compatibility decision defines the review obligations.
**Read rules by direction.** Equal version: read normally. Newer than the reader: refuse, name the direction ("written by a newer harness — upgrade"), and point at the raw log artifact so the user can still see the text (`SessionFormatUnsupportedError`, distinct from `SessionPersistenceCorruptionError` because nothing is damaged). Older than the reader: every event-body operation first runs the complete adjacent chain in memory and leaves the source path, bytes, and inode unchanged. Read handles may consume that current logical result directly; a write open exclusively publishes the final current generation under its canonical versioned filename before append. Header-only listing remains non-mutating and reports the numerically highest canonical generation. Catalog generation and module initialization reject a missing adjacent step, so a published first-party build never exposes a partial historical chain. Retained lower generations are not automatic fallback or a downgrade compatibility promise.
**A per-event `ignorable` marker covers vocabulary growth, so ordinary event additions never bump the version.** The event vocabulary is decided by which plugins are mounted, which a single version integer cannot describe. A reader meeting an unrecognized event type refuses to interpret the log unless the event carries `ignorable: true` in its envelope. The default is *required*: forgetting the marker over-refuses a resumable session (an inconvenience), while a default of ignorable would make the same mistake silently resume a gutted one (a safety failure). The architecture makes this sound: model-visible content flows only through the four `surfaceOp`-marked surface event types plus the `request/header`/`request/context` folds, so the dangerous unknowns are exactly the non-surface events that change how the rest of the log is read (`session/end-seed` is the existing example).
The [accepted compatibility checkpoint](../../../../docs/session-format-status.md#finalization-record) prevents an accepted transition from reusing its header increase for later breaking changes. Backward-compatible evolution uses new acknowledgement records at the same Session version; breaking changes require a higher writer. The checkpoint does not change the unknown-event protocol.
**Accepted machine records remain immutable.**The [accepted compatibility checkpoint](../../../../docs/session-format-status.md#finalization-record) preserves accepted declarations and schemas. Later changes receive new acknowledgements, and a version-bump decision needs its own increasing header transition. The checkpoint does not change the unknown-event protocol.
@@ -14,7 +14,7 @@ Both values belong only on the official DeepSeek adapter path. Adding them to `G
## Decision
`@deepseek-ai/dsh-deepseek-llm-api-extensions` registers `ctx.deepseekLlmApiExtensions`, an additive registry of top-level fields for `deepseek-official` request bodies. A contributor claims one declaration-merged field with `register()`. The adapter invokes `prepare()` after serializing the exact wire messages, passes the request cancellation signal, rejects preparation or base-field collision before HTTP, merges the detached fields, and calls the captured `accept()` transaction after HTTP 2xx. A merged body that fails to serialize is sent without extension fields or acceptance under the [bounded-upload decision](2026-09-24-bounded-session-log-upload.md). The registry stops awaiting preparation after cancellation even if a contributor ignores the signal. Acceptance failures remain request failures under `REQUEST_EXTENSION`; transport and non-2xx failures never accept a contribution. A composition without the registry retains the reusable base adapter. Shipped compositions mount the registry and both contributors: both package metadata and Session-log upload are enabled by default; `session-log-deepseek.enabled: false` disables log upload under the [default-upload decision](2026-09-14-session-log-upload-default.md). Keyless `deepseek-official` replay invokes preparation with a synthetic empty base body and the same acceptance transaction before its first recorded chunk, preserving post-2xx extension side effects rather than field bytes.
`@deepseek-ai/dsh-deepseek-llm-api-extensions` registers `ctx.deepseekLlmApiExtensions`, an additive registry of top-level fields for `deepseek-official` request bodies. A contributor claims one declaration-merged field with `register()`. The adapter invokes `prepare()` after serializing the exact wire messages, passes the request cancellation signal, rejects base-field collision before HTTP, merges the detached fields, and calls the captured `accept()` transaction after HTTP 2xx. A merged body that fails to serialize is sent without extension fields or acceptance under the [bounded-upload decision](2026-09-24-bounded-session-log-upload.md). The registry stops awaiting preparation after cancellation even if a contributor ignores the signal. Extension fields are auxiliary and best effort: the registry logs and omits a contributor whose preparation throws, and the adapter logs an acceptance failure without failing the model request; transport and non-2xx failures never accept a contribution. A composition without the registry retains the reusable base adapter. Shipped compositions mount the registry and both contributors: both package metadata and Session-log upload are enabled by default; `session-log-deepseek.enabled: false` disables log upload under the [default-upload decision](2026-09-14-session-log-upload-default.md). Keyless `deepseek-official` replay invokes preparation with a synthetic empty base body and the same acceptance transaction before its first recorded chunk, preserving post-2xx extension side effects rather than field bytes.
The provider-neutral `llm` package and `llm-pi-ai` contain no extension type, service lookup, field merge, or acceptance call.
@@ -50,7 +50,7 @@ The process-lifetime manifest-identity cache remains separate because in-process
## Verification
Registry tests pin duplicate ownership, effect-scoped disposal, detached field values, concurrent and abortable preparation, receiver-preserving acceptance, one acceptance settlement, and failure aggregation. Session tests pin default-on and explicit-off policies, full-first/suffix-later delivery, direct complete event envelopes independent of base-body messages, incremental watermark folding, persisted restart recovery, fork identity fencing, and out-of-order acceptance. Package-inventory tests pin default-on and explicit-off policies, host and standing-preset discovery, conflicting Loader resolution bases, manifest resolution, lifecycle filtering, and exact name/version ordering. The direct adapter mock proves pre-HTTP preparation failure, cancellation, non-2xx non-acceptance, 2xx acceptance before a later stream failure, and field collision. Keyless replay pins post-2xx extension acceptance, and the TypeScript JSON-RPC plus Python packaged-runtime snapshots project the acceptance event through both SDKs. Real Loader composition pins default package metadata and Session upload plus explicit upload disablement, one real-API request mounts both shipped extensions and proves the official endpoint accepts them, and pi-ai tests retain their unchanged wire requests.
Registry tests pin duplicate ownership, effect-scoped disposal, detached field values, concurrent and abortable preparation, receiver-preserving acceptance, one acceptance settlement, and failure aggregation. Session tests pin default-on and explicit-off policies, full-first/suffix-later delivery, direct complete event envelopes independent of base-body messages, incremental watermark folding, persisted restart recovery, fork identity fencing, and out-of-order acceptance. Package-inventory tests pin default-on and explicit-off policies, host and standing-preset discovery, conflicting Loader resolution bases, manifest resolution, lifecycle filtering, and exact name/version ordering. The direct adapter mock proves pre-HTTP preparation rejection, cancellation, non-2xx non-acceptance, 2xx acceptance before a later stream failure, and field collision. Keyless replay pins post-2xx extension acceptance, and the TypeScript JSON-RPC plus Python packaged-runtime snapshots project the acceptance event through both SDKs. Real Loader composition pins default package metadata and Session upload plus explicit upload disablement, one real-API request mounts both shipped extensions and proves the official endpoint accepts them, and pi-ai tests retain their unchanged wire requests.
## Alternatives considered
@@ -85,7 +85,7 @@ About 98% of the measured v1 real-session events were `assistant/chunk`. Omittin
## Consequences
Official DeepSeek requests carry active package versions to their resolved `baseURL`, including configured gateways. Unless Session-log upload is disabled, each eligible request also carries newly unaccepted Session events, up to `maxBytes` per request. The fields are model-hidden and add no prompt tokens or KV-cache changes, but can substantially increase HTTP body size. Manifest resolution, field collision, acceptance logging, or provider schema rejection fails the model request rather than silently dropping metadata. A merged body that fails to serialize is the exception: the request proceeds without extension fields and logs a warning.
Official DeepSeek requests carry active package versions to their resolved `baseURL`, including configured gateways. Unless Session-log upload is disabled, each eligible request also carries newly unaccepted Session events, up to `maxBytes` per request. The fields are model-hidden and add no prompt tokens or KV-cache changes, but can substantially increase HTTP body size. Field collision or provider schema rejection fails the model request. A contributor preparation failure, an acceptance failure, or a merged body that fails to serialize logs a warning and lets the request proceed; a package without a usable version reports only its name.
The `delivery-accepted` event becomes part of the canonical log and is itself delivered on a later request. Crash recovery can duplicate a suffix but does not infer acceptance from assistant output or create a second local cursor store. Direct calls without a live Session omit the session field; host package inventory remains available.
@@ -46,7 +46,7 @@ Direct SDK use follows normal Harness-home resolution: explicit `dshHome`, inher
### Python runtime
The Python runtime wheel stages [`python/sdk-runtime/runtime-bootstrap.mjs`](../../../../python/sdk-runtime/runtime-bootstrap.mjs) as the `dsh-python-runtime-closure` entry. Its ordinary branch calls the public CLI export; a provider-private selector dispatches to the internal subprocess runner before CLI parsing and is not an application entry point. The [native-containment reference](../../../../packages/subprocess/subprocess-local/README.md) owns that private dispatch. The Python client selects `dsh --profile sdk` by default, ordered patch files, and an explicit Harness home; the runnable example under `python/sdk/examples` selects `sdk-minimal`. The installed `dsh` console command exposes the same profile grammar, including the `web` profile.
The Python runtime wheel stages [`python/sdk-runtime/runtime-bootstrap.mjs`](../../../../python/sdk-runtime/runtime-bootstrap.mjs) as the `dsh-python-runtime-closure` entry. Its ordinary branch calls the public CLI export; a provider-private selector dispatches to the internal subprocess runner before CLI parsing and is not an application entry point. A private `DSH_RUNTIME_DOWNLOAD` selector prepares optional resources before application launch and is removed before spawning preparation subprocesses; ordinary CLI arguments do not select downloads. The [native-containment reference](../../../../packages/subprocess/subprocess-local/README.md) owns that private dispatch. The Python client selects `dsh --profile sdk` by default, ordered patch files, and an explicit Harness home; the runnable example under `python/sdk/examples` selects `sdk-minimal`. The installed `dsh` console command exposes the same profile grammar, including the `web` profile.
The executable family is `deepseek-harness-sdk-runtime-<platform>-<arch>`. The SDK wire, wheel and import distribution names, sidecar names, and wire identity `deepseek-harness-sdk-runtime` remain stable. The SDK package family is `@deepseek-ai/dsh-sdk-client`, `@deepseek-ai/dsh-sdk-protocol`, and `@deepseek-ai/dsh-sdk-jsonrpc-server`; `@deepseek-ai/dsh-acp` remains the ACP protocol plugin. There is no Python-specific Node application, checked-in complete config, compatibility package, forwarding executable, fallback parser, or SDK/ACP launcher alias. [docs/architecture.md](../../../../docs/architecture.md) owns this launch, and the [`python/sdk-runtime` README](../../../../python/sdk-runtime/README.md) owns the Windows carrier.
@@ -14,13 +14,13 @@ The Web Host runs tool-capable Sessions with the current operating-system user's
Each Host process generates a random launch token, retained by the root application context across Connection hot reloads. `dsh-web-app` prints and opens the normal root URL with that token in the query once per process. `frontend-static` asks Connection to authorize index responses: only `GET /?token=...` exchanges the process token for a cookie, then redirects to clean `./`; the token is not accepted on API paths or in an Authorization header. An obsolete token paired with a valid cookie redirects to clean `./`. Missing and invalid credentials receive one minimal 401 response. Static non-index assets remain public.
The cookie is a signed, authority-bound bearer. Its deterministic name and signed payload both include the normalized hostname plus port, so one Harness home can run independent Web ports without cookie collisions. The payload carries safe-integer issue and expiry times under an absolute lifetime; `cookieMaxAgeDays` defaults to 30. The cookie is host-only, `Path=/`, `HttpOnly`, and `SameSite=Strict`. It omits `Secure` because the shipped server uses loopback HTTP. There is no logout operation.
The cookie is a signed, authority-bound bearer. Its deterministic name and signed payload both include the normalized hostname plus port, so one Harness home can run independent Web ports without cookie collisions. The payload carries safe-integer issue and expiry times under an absolute lifetime; `cookieMaxAgeDays` defaults to 30. The cookie is host-only, `Path=/`, `HttpOnly`, and `SameSite=Strict`. The HTTP listener omits `Secure`. There is no logout operation.
The HMAC secret is a versioned `grant` record at `client-connection/browser-session` in `ctx.credentials`; the local provider stores it in `$DSH_HOME/.credentials.yaml`. Connection loads or creates the record during activation and retains the secret for synchronous request verification. An active Connection continues using its loaded secret if the durable record changes; the next activation loads the replacement or creates a missing record, so deleting the record and restarting the process revokes every existing cookie. Invalid owner payloads fail loud instead of being replaced. The launch token itself is never persisted and changes on every process start, while an unexpired cookie remains valid across restarts on the same authority.
The in-page Web Worker preview exposes no network socket. Its page-owned `postMessage` tunnel enters the real route first, then retries a 401 or 403 through the worker-local fetch handler. This keeps Connection interceptors while limiting the authentication bypass to the page that created the Host worker.
The shipped CLI continues to reject`--host 0.0.0.0`. Authentication does not imply supported network deployment, TLS, forwarding-header interpretation, or proxy configuration.
The shipped CLI accepts concrete local bind addresses but rejects wildcard addresses such as`--host 0.0.0.0`. Authentication does not configure TLS, forwarding-header interpretation, or proxies.
## Verification
@@ -42,6 +42,8 @@ Unit coverage pins process-token retention across Connection reloads, one secret
Possession of the browser cookie authorizes the complete tool-capable Host API, matching the authority the Web application exposes after Session creation. `Host` does not grant a higher method tier, and a method migration between API Proxy and Typert Remote cannot change its caller set.
The persistent secret makes cookies survive restarts but gives a stolen cookie up to the configured absolute lifetime. Deleting the record and restarting the process is the global revocation mechanism; the active Connection intentionally avoids credential-provider work on each request. Omitting `Secure` preserves loopback HTTP and permits plaintext transmission if an operator makes the same cookie authority reachable over an unencrypted network. The startup URL contains a process credential and must be treated as sensitive output; runtime diagnostics do not repeat it.
The persistent secret makes cookies survive restarts but gives a stolen cookie up to the configured absolute lifetime. Deleting the record and restarting the process is the global revocation mechanism; the active Connection intentionally avoids credential-provider work on each request. An HTTP listener permits plaintext cookie transmission if its authority is reachable over an unencrypted network. The startup URL contains a process credential and must be treated as sensitive output; runtime diagnostics do not repeat it.
The decision partially supersedes the authentication deferral and unauthenticated non-loopback consequences in the [browser trust note](2026-07-28-api-browser-trust-boundary.md). That note remains active authority for media-type, Host, Origin, Fetch-Metadata, and configured-authority validation. No active Agent Note is archived: the overlap is partial and both security rules retain future decision value.
The [native HTTPS listener decision](2026-09-23-native-https-web-listener.md) supersedes the HTTP-only transport and omitted-`Secure` assumptions here. Listener protocol now determines `Secure`; the token exchange, signed-cookie authority, and credential lifetime remain owned by this note.
@@ -81,7 +81,7 @@ Test releases require `DOWNLOAD_TEST_RELEASE_ID` from the platform dotenv file.
Core dsh and the private Desktop Host come only from the signed application resource tree. Plugin installation forwards package specs to pnpm, including local and remote sources, but never accepts raw pnpm commands. pnpm owns dependency resolution and the profile’s `allowBuilds` policy; the Host loads activated bundles.
Electron release artifacts are signed; macOS artifacts are notarized. Package and upload commands read release settings from the Git-ignored target `.env.windows` or `.env.macos`; subprocesses receive the fields selected by orchestration. The target file exclusively owns release fields so stale shell or system credentials cannot override local selection, and loading does not mutate the parent environment. Packaging validates the mode-specific application ID, update origin, signing identity, and local files before builds, downloads, or release-record cleanup; macOS also requires one complete notarization strategy. The separate `check:package` command runs the same validation without accessing the token or Apple; actual signing and notarization still verify authentication. Configuration loading rejects missing or malformed identifiers and incomplete notarization credentials, while macOS packaging requires signing so certificate discovery cannot silently select another installed identity or emit an unsigned release. Runtime preparation verifies the exact Authority and Team ID plus the timestamp and hardened-runtime flags on every embedded Mach-O file. An after-sign hook performs Apple's deep strict application verification and requires the same leaf Authority and Team ID before artifact creation continues. The fixed-target installer command uses [isolated App copies for parallel notarization](../process/2026-09-09-parallel-macos-notarization.md): the ZIP contains a stapled App, while the signed DMG carries the ticket covering its unstapled inner App. The DMG artifact-completion hook requires the configured identity, a valid ticket, and Gatekeeper acceptance. Both artifact lanes must succeed before the command promotes their outputs and writes the release completion record; directory-only commands still notarize and staple the App. DMG blockmaps are disabled because macOS updates consume the signed ZIP, and stapling would otherwise invalidate an already-generated DMG blockmap.
Electron release artifacts are signed; macOS artifacts are notarized. Release package and upload commands read release settings from the Git-ignored target `.env.windows` or `.env.macos`; subprocesses receive the fields selected by orchestration. The target file exclusively owns release fields so stale shell or system credentials cannot override local selection, and loading does not mutate the parent environment. Packaging validates the mode-specific application ID, update origin, signing identity, and local files before builds, downloads, or release-record cleanup; macOS also requires one complete notarization strategy. The separate `check:package` command runs the same validation without accessing the token or Apple; actual signing and notarization still verify authentication. Configuration loading rejects missing or malformed identifiers and incomplete notarization credentials, while macOS release packaging requires signing so certificate discovery cannot silently select another installed identity or emit an unsigned release. Runtime preparation verifies the exact Authority and Team ID plus the timestamp and hardened-runtime flags on every embedded Mach-O file. An after-sign hook performs Apple's deep strict application verification and requires the same leaf Authority and Team ID before artifact creation continues. The fixed-target installer command uses [isolated App copies for parallel notarization](../process/2026-09-09-parallel-macos-notarization.md): the ZIP contains a stapled App, while the signed DMG carries the ticket covering its unstapled inner App. The DMG artifact-completion hook requires the configured identity, a valid ticket, and Gatekeeper acceptance. Both artifact lanes must succeed before the command promotes their outputs and writes the release completion record; directory-only commands still notarize and staple the App. DMG blockmaps are disabled because macOS updates consume the signed ZIP, and stapling would otherwise invalidate an already-generated DMG blockmap.
The [pinned osx-sign patch](../../../../patches/@[email protected]) uses `lstat` in both published module builds, so Framework file and directory aliases do not trigger duplicate signing. The patch remains necessary until the selected upstream release skips those aliases. PAK files are resources sealed by the enclosing bundle; individual signatures add serial timestamp requests without additional resource integrity. Desktop preserves all locale files and skips only their standalone signatures. Executable code retains Developer ID signatures, secure timestamps, and hardened runtime. The [signer traversal regression](../../../../apps/desktop/tests/macos-signing-walk.spec.ts) exercises the installed dependency with real Framework aliases; release qualification still requires strict application verification, notarization, and startup.
@@ -89,7 +89,7 @@ Windows release packaging supplies the public EV leaf certificate named by `DSH_
Windows package invocations force `ELECTRON_BUILDER_7Z_FILTER=BCJ`. The bundled 7-Zip 24.09 encoder automatically selects ARM64 filters for ARM64 PE files, but the NSIS decoder from `nsis-resources-3.4.1` omits those entries during extraction. A native extraction probe with the actual NSIS plugin loses both `node-pty` ARM64 binaries under automatic filtering and restores both byte-for-byte with BCJ. Keeping a compatible filter preserves dependency contents and runtime integrity instead of removing architecture-specific files or weakening verification.
Local Windows installation testing uses an explicit `--unsigned` package invocation with the same build and runtime preparation. It strips certificate inputs, isolates artifacts in `unsigned-artifacts`, and omits updater configuration and the release completion record. The regular package command explicitly selects signed mode even when its parent environment requests unsigned mode. This separation permits installation diagnosis without an EV token while preventing local test output from qualifying for release upload.
Local Windows and macOS installation testing uses an explicit `--unsigned` package invocation with the same build and runtime preparation. It strips certificate inputs, isolates artifacts in `unsigned-artifacts`, and omits updater configuration and the release completion record. macOS local builds ignore dotenv files and inherited release settings, use the standard application ID, and omit mandatory-update policy so local testing needs no deployment configuration. Their App receives a certificate-free ad-hoc signature; their DMG is unsigned and unnotarized. Windows local builds retain their platform configuration requirements. The regular package command explicitly selects signed mode even when its parent environment requests unsigned mode. This separation permits installation diagnosis without release credentials while preventing local test output from qualifying for release upload.
Windows application replacement follows the [directory-installation reference](../../../../apps/desktop/README.md): extract beside the destination with a command-line tool that returns failure status, then rename complete directories on the same volume. The installer retains the old directory through staging and restores it if promotion fails. Registration, shortcuts, and the signed uninstaller remain owned by electron-builder.
本地 Windows 安装测试使用显式的 `--unsigned` 打包调用，并执行相同的构建和运行时准备。它清除证书输入，将产物隔离到 `unsigned-artifacts`，并省略更新器配置和发布完成记录。即使父进程环境请求未签名模式，常规打包命令也会显式选择签名模式。这样既能在没有 EV Token 时诊断安装问题，也能防止本地测试产物通过发布上传校验。
Windows 应用替换遵循[目录安装参考](../../../../apps/desktop/README.zh.md)：使用能返回失败状态的命令行工具解压到目标旁边，再在同卷内改名替换完整目录。安装器在暂存期间保留旧目录，正式替换失败时恢复旧目录。注册信息、快捷方式和签名卸载器仍由 electron-builder 持有。
@@ -16,6 +16,10 @@ The canonical `SessionEvent` envelope retains `ignorable?: true`, and every repr
The field is removable only after a replacement supports the current third-party plugin across event production, persistence, reload, and transport, with an explicit cutover for sessions already containing the marker. The [session log versioning decision](2026-08-10-session-log-version-mechanism.md) continues to own the default-required safety rule and format-version policy.
`appendPluginRecord()` in `@deepseek-ai/dsh-session` is the one first-party writer of the marker. It appends ignorable records whose type is in the `plugin:` namespace, which the V3-to-V4 edge already assigns to unknown ignorable V3 events, and only production source under `packages/experimental/` may call it; the `verify-plugin-record-callers` gate enforces that, and the persistence catalog generator rejects a `SessionEventMap` member in the namespace. Experimental packages keep state there that their owners can lose: a format migration retains records on a best-effort basis, so a release package declares its events instead.
Experimental packages declare record names and payload types in the separate `PluginRecordMap`. The map types writes and supplies the dedicated [experimental persistence catalog](../../../../docs/experimental-persistence-catalog.md), making plugin-owned persistence discoverable from current source without expanding its payload schemas or recording released type history. It does not add names to `KNOWN_SESSION_EVENT_TYPES` or change stored-event admission. The reader returns unknown payloads for owner validation, including records whose declarations have changed or disappeared.
Historical format migration is deliberately stricter in the alpha implementation. The v0-to-v1 edge refuses every unknown v0 type, including an ignorable one, because an opaque payload may contain references that a format edge cannot validate. The [alpha historical-event decision](2026-08-31-alpha-historical-unknown-event-refusal.md) owns that bounded exception; equal-version append and reload continue to follow this note.
## Alternatives considered
@@ -26,8 +30,16 @@ Historical format migration is deliberately stricter in the alpha implementation
**Treat every repository-external event as ignorable.** Rejected because a reader cannot infer that an unknown durable event is informational. An external event may change later reconstruction or plugin-owned state.
**Declare a `SessionEventMap` member for each experimental package's state.** Rejected because each member is required-on-read and enters the released-schema inventory and type history: a build without the experimental package refuses the Session, and removing the package leaves a released type behind.
**Keep no catalog of plugin records.** Rejected because maintainers need to find in-repository persistence writers. The dedicated experimental catalog provides that inventory from current-source declarations while their payloads retain best-effort compatibility; a full schema history would impose the released-event obligations that these records deliberately avoid.
**Let any package call the record writer.** Rejected because records survive a format migration only on a best-effort basis, and a release package must not keep state that an upgrade can drop.
**Restrict the writer with a runtime capability.** Not adopted because the writer runs in its caller's process, where a runtime check cannot tell experimental code from release code; the repository gate checks every reference instead.
**Register mounted plugin event names as known.** Not adopted as the removal mechanism because event-name registration alone does not classify whether absence is safe, and acceptance would depend on the reader's current composition rather than the stored record.
## Consequences
Third-party informational events can remain reloadable when their stored records carry the explicit marker, while unknown required events still fail loudly. The field remains part of the public event envelope, JSONL representation, transport types, generated references, and their tests until a replacement satisfies the cutover condition.
Third-party informational events can remain reloadable when their stored records carry the explicit marker, while unknown required events still fail loudly. The field remains part of the public event envelope, JSONL representation, transport types, generated references, and their tests until a replacement satisfies the cutover condition, which now also covers plugin records written by experimental packages.
@@ -10,7 +10,7 @@ An Agent's Cordis Context owns registrations and their cleanup. Agent identity i
The reverse association also required compensating mechanisms after type erasure. Host Remote forwarding inspected a routed subject for its Context, creation inferred runtime parentage from the caller Context, and adapters maintained reverse identity scans. These mechanisms duplicated identity already present in typed requests and obscured which caller owned an Agent at runtime.
Without an explicit owner, `SubagentContinuationManager` creates and resumes children through its private plugin Context, so Context-based inference classifies every continuable child as a runtime root even though the manager holds its exact parent. Root-only consumers could then attach scheduling tools, grant direct-human goal authority, or route user questions as if the child were top-level.
Without an explicit owner, `SubagentManager` creates and resumes children through its private plugin Context, so Context-based inference classifies every continuable child as a runtime root even though the manager holds its exact parent. Root-only consumers could then attach scheduling tools, grant direct-human goal authority, or route user questions as if the child were top-level.
## Decision
@@ -18,7 +18,7 @@ Runtime interfaces carry Agent identity at the point that owns it. `AgentSetup`
Scope-aware registries continue to use the opaque scope key only for registration membership. Tool-subagent does not classify that key or resolve an Agent from Context. A direct `AgentSetup` passes the unpublished Session explicitly and installs through the supplied Context before publication. For a settings-backed standing preset, the event payload supplies the Agent, its Session supplies the policy target, and its Context owns the registrations.
`SubagentContinuationManager` puts the exact parent in both fresh-creation and cold-resume options. A live continuable child is therefore excluded from `AgentRegistry.roots()` and satisfies `isOwnedBy(child.id, parent)`. Durable `parentSession` metadata does not substitute for this relation: a fork or resumed Session may be a runtime root when no live Agent owns it.
`SubagentManager` puts the exact parent in both fresh-creation and cold-resume options. A live continuable child is therefore excluded from `AgentRegistry.roots()` and satisfies `isOwnedBy(child.id, parent)`. Durable `parentSession` metadata does not substitute for this relation: a fork or resumed Session may be a runtime root when no live Agent owns it.
The [scope reference](../../../../packages/core/scope/README.md) and [Agent loop reference](../../../../packages/core/agent-loop/README.md) document registration and lifecycle behavior. The [initiator-scope decision](2026-07-15-agent-initiator-scope.md) owns private-chain rationale; explicit runtime ownership does not change causal initiation or registration lifetime.
@@ -12,15 +12,17 @@ The child descriptor remains necessary for recovery and composition, but it cann
## Decision
The parent Session's required `subagent/catalog` events are the persistent authority for direct-child discovery. Each event is one successful creation fact containing `childId`, `childCreatedAt`, mode, and the mode-discriminated label. Remote one-shot runs without a local Session remain outside this catalog. Invalid own facts, including unsupported payload versions, reject projection restoration because silently dropping a required fact would return an incomplete catalog.
The shared activation entry point, external execution ownership, and caller-versus-parent result delivery are governed by [Unified subagent activations](../simplification/2026-09-17-unified-subagent-activations.md). This record retains the independent rationale described below.
Creation publishes only successful facts. A one-shot run appends the catalog event after its provider returns a local child and before the run reaches its caller. A continuable run admits the initial prompt, appends the catalog event, then returns the child id. If admission orcatalog append fails, creation fails and releases the activation; there is no compensating catalog event or rollback protocol.
The parent Session's required `subagent/catalog` events are the persistent authority for direct-child discovery. Every local child enters the catalog independently of its result recipient; external children enter only with parent delivery. Workflow membership records describe execution progress and do not replace local Session discovery. Each event contains `childId`, `childCreatedAt`, mode, and the mode-discriminated label. Local children use `continuable`; external executions use `external` and cannot be opened as local Sessions. Historical `one-shot` and `unknown` entries remain readable. Invalid own facts, including unsupported payload versions, reject projection restoration because silently dropping a required fact would return an incomplete catalog.
Creation publishes only successful facts. A local activation admits the initial prompt before appending its catalog event. An external activation appends its event after provider startup and before returning the receipt. Failed admission or publication releases the unpublished execution; there is no compensating catalog event or rollback protocol.
The child header and `subagent/descriptor` remain authoritative for recovery and composition. An Activation and the exact parent relationship remain authoritative for authorization and delivery. Mode and label are snapshotted once and the same detached values reach the parent catalog fact and child descriptor.
The registered `subagentCatalog` projection materializes the parent facts. It delegates storage, append, iteration, and checkpoint validation to [`dsh-chunked-list`](../../../../packages/util/chunked-list/README.md), which stores facts in a persistent stack of 64-entry chunks, so an append copies at most the head chunk in bounded O(1) work. Materialization visits chunks from oldest to newest and preserves parent catalog event order in O(D) time for D facts. Concurrent creation is ordered by successful catalog append, independent of child timestamps and ids. A projection checkpoint clones the state once in O(D); projection-cache writes remain asynchronous and use the existing mandatory creation, turn-end, and disposal points.
The utility owns chunk layout and its shared capacity constant; the catalog owns event validation, fork filtering, and row conversion. Catalog projection state version 2 stores generic chunk values, so the projection registry rebuilds incompatible caches from Session events. Session event payloads and public catalog rows retain their formats.
The utility owns chunk layout and its shared capacity constant; the catalog owns event validation, fork filtering, and row conversion. Catalog projection state version 6 accepts external membership, so the projection registry rebuilds incompatible caches from Session events. Payload v0 and v1 remain unchanged; v2 records external executions.
Fork isolation uses the exact `Session.inheritedEventCount` supplied to projection initialization. The fold ignores `subagent/catalog` events below that offset. The state stores the inherited offset but not each event seq because acceptance is decided during folding.
@@ -22,12 +22,20 @@ Each stream receives a fresh 256-bit TLS pre-shared key through private administ
Collected stdout and stderr carry bounded tail snapshots through handlers that only update output observations. Capture continues while a snapshot is waiting for transport. Final snapshots preserve raw-byte offsets, and completed spill files use the local provider’s retained-output storage after connection disposal.
Readiness verifies the installed helper digest and, when PTC is configured, the installed Node bootstrap digest. These checks pin expected deployment artifacts; they do not authenticate a malicious remote operating system. The helper disables Node debugger activation through `SIGUSR1`: file-effect confinement can still permit same-user signals, which must not expose the helper's unrestricted filesystem and process services. File-effect confinement delegates to the remote local sandbox provider and retains its full/partial disclosure and platform limitations.
Readiness verifies the installed helper script or executable digest and, when a script deployment supplies a PTC bootstrap, that bootstrap's digest. Executable deployments embed the PTC worker and require no separate bootstrap. These checks pin expected deployment artifacts; they do not authenticate a malicious remote operating system. The helper disables Node debugger activation through `SIGUSR1`: file-effect confinement can still permit same-user signals, which must not expose the helper's unrestricted filesystem and process services. File-effect confinement delegates to the remote local sandbox provider and retains its full/partial disclosure and platform limitations.
Path canonicalization belongs where the files exist. The shared policy resolver preserves absolute execution-world spelling; enforcing providers resolve symlinks and `..` on their own filesystem. Headless records and validates cwd through `ctx.fs`. Host-path projection remains unavailable for SSH, so Node execution requires an explicitly installed remote bootstrap.
Path canonicalization belongs where the files exist. The shared policy resolver preserves absolute execution-world spelling; enforcing providers resolve symlinks and `..` on their own filesystem. Headless records and validates cwd through `ctx.fs`. Host-path projection remains unavailable for SSH; PTC uses an explicitly installed remote bootstrap for script deployments or the helper executable's embedded worker.
A lost connection invalidates pending operations without reconnect or replay. Helper EOF, signals and a heartbeat lease initiate remote native cleanup. The client cannot turn lease expiry into an observed successful termination: an interrupted mutation or launch can have an unknown outcome. Administrative request deadlines, consumer execution deadlines and cleanup remain separate owners.
### Helper deployment
The explicit `launch.kind` selects remote `node-script` or `executable` invocation independently of the local Harness carrier. A local `process.pkg` flag cannot describe the remote installation. Protocol 2 includes the remote runtime kind in its readiness response and rejects protocol or launch-kind mismatches, so a client cannot apply script bootstrap rules to an embedded worker. Client and helper protocol upgrades must be deployed together.
The executable embeds Node, the helper, managed subprocess runners and the PTC worker. It reuses the [SDK executable packaging decision](2026-07-10-single-file-executable-sdk-runtime-distribution.md): `pkg --sea` preserves a package tree and dynamic module resolution instead of maintaining a separate static plugin registry. That record owns the comparison with bare Node SEA. The embedded runtime serves these private entry points; project commands still depend on the remote machine's development tools.
The four release targets cover Linux and macOS on x64 and arm64, matching the POSIX providers and their supported native resources. Packaging does not add Windows or musl support. Landlock and macOS PTY launchers remain real files beside the executable because the operating system cannot execute paths inside pkg's virtual filesystem. The installation unit is therefore the complete archive directory, with native resources and integrity metadata, rather than the executable alone. The [helper runtime reference](../../../../packages/ssh/ssh-helper-runtime/README.md) owns the build matrix, deployment baselines and archive verification commands.
## Alternatives considered
**Frame every stream over SSH exec stdout/stdin.** This avoids Unix-socket forwarding, but requires per-stream credits, queue bounds and independent EOF semantics inside the application protocol. Independent SSH channels use the transport’s existing flow control and keep program output away from administrative framing.
@@ -48,7 +56,7 @@ The portable-consumer decision remains active; this note supplies its SSH realiz
## Verification
Required protocol and lifecycle evidence covers malformed frames, bounds, reservation cancellation, authenticated stream publication and disconnect errors. Native SSH acceptance must exercise remote file guards and symlink identity, Bash confinement, fd 7 binary traffic, control progress under paused output, terminal operations, LSP and Node execution. Live checks require an explicitly configured disposable remote workspace; keyless tests do not provision one.
Required protocol and lifecycle evidence covers malformed frames, bounds, reservation cancellation, authenticated stream publication and disconnect errors. Native SSH acceptance must exercise remote file guards and symlink identity, Bash confinement, fd 7 binary traffic, control progress under paused output, terminal operations, LSP and Node execution. External-host checks require an explicitly configured disposable remote workspace. Packaged Linux acceptance provisions a disposable OpenSSH container without Node; native-host checks own confinement evidence.
Security evidence requires both a same-user connector and a replaced socket listener that cannot claim a reserved stream, learn its key, alter another run or receive its plaintext output. Process-lifecycle evidence distinguishes direct exit from managed-range quiescence and tests cancellation both before launch publication and after the payload starts. Source and built profile checks verify the installed helper/bootstrap arrangement without transferring host credentials to program environments.
@@ -22,7 +22,7 @@ The host preserves `ELECTRON_RUN_AS_NODE` for child startup; the bootstrap remov
File mode, observed denial and enforcement completeness travel in `PtcRunResult.sandbox` separately from the program outcome. Restricted execution fails if the required sandbox backend cannot launch. Full access is an explicit policy mode. Program success does not prove full enforcement, and neither the `process` descriptor nor the extra control pipe claims multi-tenant isolation.
The private Python provider keeps its existing execution implementation and configured wall deadline. Its resolver accepts cwd but rejects an explicit file policy or per-call timeout override; a capability descriptor never silently grants unsupported protection.
The [experimental Python provider](../../../../packages/experimental/ptc-runtime-python/README.md) shares the resolved file policy and sandbox diagnostics while retaining its local Unix execution and configured wall deadline. Per-call timeout overrides remain unsupported.
English | [中文](2026-09-13-session-working-directory.zh.md)
## Problem
A fixed Session creation directory cannot represent a task that changes checkout during its conversation. Independent tools resolving paths from different sources can read and write different projects, while changing the process directory also affects unrelated Sessions. Putting mutable directory information in the system prompt changes the reusable prompt prefix.
## Decision
`dsh-working-directory` owns one logged effective directory per Session. Its projection initializes from the immutable original project and folds `working-directory/change`; consumers read that owner instead of deriving execution paths from the header or sandbox root. Directory changes validate through the filesystem provider, serialize within the Session, and publish only after the new directory is available.
The `dsh-base` composition mounts `working_directory({ cd? })` globally. The tool reads or changes the Session directory, so every Agent preset can select or recover it, including presets without file-editing tools. The Web composition keeps action tools in preset scopes.
Required user-context snapshots carry the current value; each committed change attempts to queue a notice through the ordinary Agent inbox. Cancellation or disposal may discard unadmitted notices while the directory event survives for subsequent requests. The original project is the recovery destination when the selected directory disappears. Missing recovery destinations produce an error. Write permissions remain independent.
Subagents capture the parent's current directory or an explicit override and commit their selected directory after inherited history. Existing processes, terminals, and persistent shells keep their process-local directory. Project instructions, skills, and completion follow the effective directory without changing Session identity, grouping, or runtime composition.
Live Session skill catalogs and file completion validate through the same directory owner. These reads can commit recovery and queue its notice before discovery, keeping discovered files and subsequent operations tied to the committed Session directory. Cold skill catalogs read the recorded projection without activating an Agent.
Web sidebar terminal environment lookup reads the effective directory without filesystem validation or recovery, so a missing directory does not prevent reconnecting to a retained terminal. New terminal creation validates and recovers through the directory owner before spawning; the process keeps that directory independently of later Session changes.
Playwright MCP and Chrome DevTools MCP validate the effective directory when acquiring a Session's MCP connection and use it to launch the server. Later directory changes do not restart the retained server or change its process directory.
File links and command directory labels use result metadata captured when each operation runs, so historical rows remain correct after later directory changes.
The [prompt-variables note](2026-07-05-prompt-variables-and-tool-guidance-ownership.md) retains strict interpolation, route-variable ownership, and tool-guidance ownership. The [environment-suffix note](../../archived/bug-fix/2026-09-06-environment-prompt-suffix.md) retains ordered prefix/suffix placement for deployment, Harness source, and Web guidance. This decision owns directory state and placement in user context.
## Alternatives considered
**Rewrite the Session header.** The header is immutable identity metadata used by storage, grouping, and permission roots. Rewriting it would conflate current execution state with the original project.
**Let worktree tools own directory state.** Ordinary directory changes, filesystem tools, and subagents need the same value. A Git-specific owner would duplicate selection and recovery.
**Change the process directory or synchronize every shell.** Concurrent Sessions and existing subprocesses have independent lifetimes. A Session change affects new directory-based operations, while a shell's `cd` remains local.
**Keep directory text in the system prompt.** Mutable operational context belongs after retained history, using the same user-context mechanism as permission state.
## Consequences
The runtime exposes one directory owner and preserves the original project. Consumers migrate together, including filesystem, shell, instructions, skills, subagents, and both SDKs. Directory existence can change outside the harness, so validation and logged recovery occur at use. Required runtime context remains visible even when optional context is suppressed.
`{{cwd}}` has no loop-provided value; custom personas that depend on that built-in fail strict interpolation. Deployments remove those references using the [system-prompt migration guidance](../../../../packages/core/system-prompt/README.md). When both the current directory and its recovery destination are unavailable, prompt assembly rejects model turns; the [directory service recovery guidance](../../../../packages/session/working-directory/README.md#known-limitations-and-deferred-work) requires restoring the directory or selecting an existing absolute directory through the SDK or host service.
Verification covers independent Sessions, relative changes, replay, cancellation, recovery, tool disposal, user-context rendering, provider inheritance, and real directory-based operations.
[提示词变量记录](2026-07-05-prompt-variables-and-tool-guidance-ownership.zh.md)保留严格插值、路由变量归属与工具指导归属。[环境后缀记录](../../archived/bug-fix/2026-09-06-environment-prompt-suffix.md)保留部署、Harness 源码与 Web 指导的有序前缀／后缀位置。本决策拥有目录状态及其用户上下文位置。
English | [中文](2026-09-19-concrete-web-bind-address.zh.md)
## Problem
The Web listener accepted `127.0.0.1` or `0.0.0.0`. A wildcard bind exposed every interface of the machine, so the bundle sampled the LAN IPv4 addresses at bind time into a `webRuntime` service and appended them to `trustedHosts`, granting authorities for interfaces the operator never named and mixing inferred reachability into configured trust. A container or workstation that should expose exactly one address had no way to say so, IPv6 addresses were not accepted at all, and a browser dialing the exposed address sent it as `Host` only to be refused unless the sampled list happened to contain it.
## Decision
`webserver.config.host` accepts one concrete IPv4 or IPv6 address literal, loopback by default; the unspecified address in every spelling (`0.0.0.0`, `::`, `::ffff:0.0.0.0`, and their expansions) is rejected when the config loads, and the shipped `dsh web` command refuses the same spellings for `--host` before any consumer activates. An IPv6 `%zone` is kept for `listen`; classification through the exported `isWildcardHost` and `isLoopbackHost` reads the address alone.
The browser Host/Origin fence accepts the listener's own bind IP literal as `Host` on any port, independently of `trustedHosts` and ignoring the zone. A browser dialing that literal presents it as `Host`, the literal is no rebinding target, and the Origin and cross-site checks still apply. DNS names resolving to the address, other interfaces, and proxy authorities still require `--trusted-host`. Connection reads the invocation authorities directly from the `webStartup` provider; the `webRuntime` service and LAN sampling no longer exist.
The advertised URL uses IPv4 for mapped loopback binds, brackets other IPv6 binds, drops a loopback zone, and requires `--public-url` for a zoned non-loopback bind because a zone cannot appear in a URL. A non-loopback bind serves plain HTTP and warns at startup even when an HTTPS `--public-url` fronts it. `web-runtime` is a required startup entry, so a bind failure fails startup instead of leaving a silent listener-less process. Desktop pins `--host 127.0.0.1`, the address Electron dials.
## Alternatives considered
**Keep sampling LAN addresses into `trustedHosts`.** Rejected: the listener already owns the concrete bind address. Sampling grants authorities for other interfaces and mixes inferred reachability with configured trust; direct bind-IP acceptance preserves that distinction without another runtime service.
**Accept the bind IP only on the bound port.** Rejected: a port-less `trustedHosts` entry already matches any port, a proxy on the same address may listen elsewhere, and the port adds nothing to the rebinding defense the Host check provides.
**Keep `0.0.0.0` with a warning.** Rejected: binding every interface exposes remote code execution to every network the machine joins, and the operator can name each address that should be exposed.
## Consequences
An operator exposes exactly the interfaces named by `--host` and reaches the listener through that literal without further configuration; every other remote authority is a `--trusted-host` decision. Compositions that bound `0.0.0.0` fail at load with a message naming the rejected spelling. IPv6-only hosts are supported end to end, including the printed and opened URL.
The [native HTTPS listener decision](2026-09-23-native-https-web-listener.md) supersedes this note's HTTP-only transport assumption. The non-loopback warning applies only without TLS; bind-address validation and authority admission remain unchanged.
## Testing
`packages/host/webserver/tests/webserver.spec.ts` classifies wildcard and loopback spellings and rejects wildcard config; `packages/bundle/web-app/tests/startup.spec.ts` refuses wildcard `--host` before consumers activate; `packages/client/connection/tests/api-request-trust.host.spec.ts` and `node-half.host.spec.ts` admit the bind literal on any port without admitting other authorities; `packages/bundle/web-app/tests/web-app.spec.ts` covers the bracketed IPv6 URL and the plain-HTTP warning; `apps/cli/tests/profiles/web/tests/public-url.expected.e2e.ts` boots the built profile on a non-loopback and a zoned loopback address; `web-failure-matrix.expected.e2e.ts` and `packages/boot/app-boot/tests/app-boot.spec.ts` report the required `web-runtime` entry.
| the `projections` block of an `api-session/added` summary (`manager.handleSessionAdded`) | by the block's `kind`; the summary comes from a live session, so in practice `sequenced` |
| the history first page's `projections` (`projections.seed` in `session.ts`) | sequenced |
| the control baseline, live sessions only (`manager.replaceControlBaseline`) | sequenced |
| the `session.projections` result of `refreshProjections`(a body observation) | sequenced |
| the `session.projections` result of `refreshProjections`| sequenced for an exact read; cached only when [migration is required](2026-09-29-explicit-session-history-preparation.md) |
| push frames (the `projection` frame handled by `manager`) | sequenced |
| the `title` after a successful rename (`session.ts`) | sequenced |
@@ -10,9 +10,13 @@ The Web plugin page offered two optional bundles, Agent Teams and voice input. A
## Decision
`OPTIONAL_BUNDLES`lists Agent Teams, voice input, Auto review, and the Inspector profile. Every optional bundle declares `icon` and exports `./locale/*.json` with `meta.title` and `meta.description`, so the Official group renders a localized title, description, and image; `verify-default-product-isolation` rejects an optional bundle without them. The Inspector profile owns the package-named bundle patch. The standalone Inspector's `cordis.patch.yml` mounts its local built entry as an explicit repository overlay; [Session Inspector](../feature/2026-09-24-session-inspector.md) owns the inspection behavior and activation choices.
`OPTIONAL_BUNDLES`owns the installation's default-off selections, including maintained optional capabilities and experimental capabilities. Every bundle declares `icon` and exports `./locale/*.json` with `meta.title` and `meta.description`; the Official group displays that metadata. A bundle's presentation does not determine its [experimental status](../../../../packages/experimental/README.md#status). The isolation check validates admission and still rejects experimental dependencies of nonexperimental bundles.
An optional bundle is a runtime dependency of the installation, so its dependency graph is downloaded by every `dsh` install. The list weighs the added installation cost: Auto review reuses the installation's closure, while the Inspector profile adds inspection packages, `serve-static`/`open`/`ws`, and 535 mirrored DevTools resources (about 11.95 MiB uncompressed), without another browser binary. The Inspector profile appears in the default plugin list but remains disabled until selected. The browser-use and computer-use providers stay explicit compositions: Playwright MCP, Chrome DevTools MCP, and the native Cua Driver runtime binaries add about 85 MB and 21 packages to every install whether or not the bundle is enabled, and a shipped Cua Driver MCP switch would offer a capability whose executable the installation does not carry. Those provider packages keep locale display metadata for their component rows. Two more packages stay out for other reasons: `ptc-runtime-python` replaces the PTC runtime, and `workflow-ptc` rejects a non-TypeScript runtime at load while the Web presets carry `workflow-ptc` rows a bundle patch cannot reach; `browser-use-stagehand-native` requires a native model name and API key at schema validation with no configuration form on the page.
An optional bundle's dependencies are downloaded with every `dsh` installation. [Reasoning translation](2026-10-05-anonymous-reasoning-translation.md) uses three experimental packages without adding third-party runtime dependencies. Admission therefore considers installation cost and whether a switch supplies a usable composition without an additional configuration form. Browser and computer providers remain explicit compositions because they add substantial runtimes or require an external executable. Stagehand requires model credentials; Python PTC replaces a core provider and conflicts with TypeScript workflow consumers. Demonstrative hooks, webhooks, and mods stay outside the list. [Session Inspector](../feature/2026-09-24-session-inspector.md) owns the Inspector's distinct activation behavior.
Bundles insert ordinary Host rows; optional model tools register in the global tool layer, so every preset, including minimal, receives them. Plugin management, profile composition, and reload need no bundle-specific behavior. Search, Ralph, and terminals supply their own isolated services in a Host group. Search opens one in-memory index on demand; the Host search policy stays unchanged. Badge skills and title providers act on the Host. Switching a bundle registers or removes its tools for live Agents, and later user patch layers retain precedence. A replacement title provider may register once the previous provider's disposal starts; the previous provider's late results never commit.
Claude Code and Codex install on demand through the separate `ON_DEMAND_BUNDLES` catalog; the [on-demand native-bundle decision](2026-10-05-official-on-demand-bundles.md) owns the cost, version, and offline-discovery trade-offs.
## Alternatives considered
@@ -20,6 +24,10 @@ An optional bundle is a runtime dependency of the installation, so its dependenc
**Mount the registration-only `computer-use` and `browser-use` services in `dsh-base`.** The shared composition would carry rows that only an optional provider bundle needs; a provider bundle can insert the service row from its own patch and dependencies.
**Register preset extensions at runtime.** A separate registration path would need its own ordering, module resolution, generation updates, and inspection behavior. Profile patches keep one effective composition for boot, dumps, validation, and user overrides.
**Insert rows into named presets with a `preset` patch operation.** Limiting tools to full presets required a second patch language across profile composition, reload, plugin management, and the plugin page, plus retained-generation removal checks. Global tools reach the same Agents through the existing Host composition.
## Consequences
The Official group carries four optional-bundle entries, each tagged experimental by its experimental name. Their runtime dependencies are installed even while the bundles are disabled. A browser-use or computer-use provider still needs its Service Definition and the provider in a profile patch or composition.
Official discovery does not imply experimental status or a different plugin lifecycle. Switching a bundle off keeps its installed files; failed activation retains the existing saved-selection behavior. Bundle rows are ordinary Host rows with their usual switches. Individual and combined composition tests cover the Host rows and every preset's tool catalog; live switching and recorded Sessions cover activation and model-visible behavior.
English | [中文](2026-09-23-native-https-web-listener.zh.md)
## Problem
A directly reached Web listener needs to encrypt launch tokens and session cookies without requiring a separate reverse proxy. Advertising an HTTPS URL does not encrypt the listener or establish certificate trust.
## Decision
WebServer owns optional TLS through `tls: { certFile, keyFile }`, exposed by `dsh web` as paired `--tls-cert` and `--tls-key` flags. Certificate and unencrypted private-key PEM files are loaded before binding; unreadable, empty, invalid, or mismatched material fails activation without an HTTP fallback. HTTP remains the default. Both transports share routing, upgrades, and disposal; disposal also owns incomplete TLS handshakes.
`webServer.protocol` is the source for the bind-address URL, the non-loopback plain-HTTP warning, and Connection's `Secure` cookie attribute. Neither forwarded headers nor `publicUrl` control encryption or that attribute. Explicit Host-port grants normalize under the listener's scheme, while Origin comparison uses the Origin's HTTP(S) scheme so HTTPS proxies can retain HTTP upstreams. TLS grants no additional Host authority and does not replace authentication.
The API authority check does not equate a proxy's upstream transport with the browser page's scheme. An HTTPS listener can accept a matching HTTP Origin, but `sec-fetch-site: cross-site` still rejects the request and authentication remains required. Schemeful `SameSite=Strict` cookies prevent cross-scheme browser credential delivery; this is not support for serving a public browser frontend over HTTP with a TLS upstream.
HTTP cookie audiences retain their existing spelling so enabling TLS support does not invalidate unrelated HTTP sessions. HTTPS cookie names and signed audiences include the `https:` scheme, preventing replay of an HTTP credential on any HTTPS port even if the cookie is renamed. Default ports follow URL normalization without a separate port-443 spelling.
Certificate issuance, trust, renewal, and replacement belong to the operator. Files are read once per activation; replacement requires a listener reload or process restart. The certificate must cover the hostname or IP the TLS client dials. The default port remains 3080; `--port` selects another port, subject to operating-system permission.
## Alternatives considered
**Require a TLS-terminating proxy.** This remains supported, but requires another service for a directly reached listener.
**Derive transport or cookie security from `publicUrl` or forwarding headers.** An advertisement does not describe the receiving socket, and request headers cannot establish trusted proxy identity. Native listener state keeps this decision independent of proxy policy.
**Issue, renew, or generate certificates.** This would add certificate lifecycle and trust-distribution responsibilities. An automatically generated certificate would not establish browser trust.
**Fall back to HTTP on certificate errors.** This would expose credentials despite an explicit request for encryption; activation must fail instead.
## Consequences
Direct HTTPS needs no TLS-terminating proxy. An HTTPS proxy in front of an HTTP listener still owns external cookie rewriting. An HTTPS listener emits `Secure` cookies, so public browser-facing deployments must retain HTTPS rather than depend on local-browser exceptions for HTTP. Deployment procedures belong to the [public deployments guide](../../../../docs/user/guide/public-deployments.md).
This partially supersedes the HTTP-only assumptions in [browser launch-token authentication](2026-08-24-browser-token-authentication.md) and [concrete bind addresses](2026-09-19-concrete-web-bind-address.md). Those notes remain active for cookie payloads and lifetimes, and bind admission respectively.
## Testing
The WebServer suite covers verified HTTPS, invalid material, and disposal. Connection covers listener-owned cookie attributes, cross-protocol credential replay, and exact-port grants; Web startup rejects unpaired flags. The built CLI's `public-url.expected.e2e.ts` authenticates over verified TLS with a `Secure` cookie while retaining HTTP and proxy cases; its failure matrix verifies unreadable TLS material exits without opening a listener. Desktop's startup suite covers matching WS/WSS credentials and refusal on a different transport.
@@ -25,6 +25,6 @@ This supersedes complete-suffix delivery and fail-closed serialization in the [r
## Consequences
A backlog drains at up to 8 MiB per accepted request by default. At the end-to-end rate measured for the 269 MiB upload, about 8 MiB/s, an 8 MiB field adds about 1 s to each request while a backlog drains, and a 0.86 GB backlog drains in about 100 requests. In the same 96,341-event Session, the median per-step append is 11 KiB and the largest event is 414 KiB, so ordinary steps stay far below the limit and only an unusual event or a small configured limit reaches the oversized-event path. While a backlog drains, the endpoint receives a `throughSeq` behind the Session's current tip. Preparation failures, acceptance failures, and provider rejections still fail the model request.
A backlog drains at up to 8 MiB per accepted request by default. At the end-to-end rate measured for the 269 MiB upload, about 8 MiB/s, an 8 MiB field adds about 1 s to each request while a backlog drains, and a 0.86 GB backlog drains in about 100 requests. In the same 96,341-event Session, the median per-step append is 11 KiB and the largest event is 414 KiB, so ordinary steps stay far below the limit and only an unusual event or a small configured limit reaches the oversized-event path. While a backlog drains, the endpoint receives a `throughSeq` behind the Session's current tip. Provider rejections still fail the model request; preparation and acceptance failures are logged and do not.
Plugin tests cover the exact byte boundary, UTF-8 counting, backlog draining across accepted requests, the oversized-event warning with its unchanged watermark, an event past the string limit, and the 8 MiB default through plugin configuration. Adapter tests cover the base-body fallback through the provider plugin's logger and through an adapter without a reporter.
# Agent Note: Open the Session when migration is required
Status: implemented
English | [中文](2026-09-29-explicit-session-history-preparation.zh.md)
## Problem
Expanding a subagent catalog or forking from the sidebar can trigger an old Session's format migration.
## Decision
The WorkspaceBrowser row-menu fork request carries `allowMigration: false`, which the Host handles by checking `stat` format information before calling the original fork flow. Only a source format requiring migration is refused, with an app-wide Toast offering to open the original Session. That action only navigates and does not retry fork. Neither an absent live Session nor unopened Client history justifies refusal.
Other fork callers omit that option and allow automatic migration.
`session.projections` also checks the stored format first; current-format cold history still uses the original observation flow for exact projections. Success returns `kind: sequenced` and its real `asOfSeq`; required migration returns `kind: migration-required` hints. The Client retains that read state, and cached values cannot replace sequenced values published by history or control.
The `formatStatus` returned by `stat` carries the migration library's existing classification; `header` remains the current logical format. It describes the stored format, not migration-task progress or prepared in-memory history.
When migration is required and `subagentCatalog` is absent, Web marks the Session's row “Migration required” and keeps its history-opening actions without descendant expansion or changing its subagent mode. Opening the Session supplies its catalog through history and control projections. Available catalogs remain usable, including empty catalogs. Session-list lineage resolves navigation addresses, not missing catalog membership. Ordinary loading, cancellation, and storage errors do not become this migration branch.
Create/adopt, full-text search, and ordinary history opening allow automatic migration.
This decision adds a migration exception to [catalog projection reads](../../archived/simplification/2026-09-08-web-subagent-catalog-projections.md). Exact reads and leases in [Session observations](2026-08-25-session-observations-and-projection-owned-client-state.md), and identity checks and cached/sequenced ordering in [projection-cache reads](2026-09-19-projection-cache-listing-identity-and-cached-rows.md), still apply.
## Alternatives considered
**Treat every cold Session as needing migration.** Residency and stored format are different; this would reject otherwise valid forks and discard catalog information available on disk.
**Globally forbid forks that automatically migrate on the Host.** The sidebar's interaction choice does not govern other callers, so its request explicitly carries the restriction.
**Automatically open and retry the sidebar fork.** Opening is an explicit user action, and fork does not run automatically after migration.
**Derive catalogs from Session-list parent ids.** This adds a second membership derivation; explicit Session opening obtains the catalog.
**Give cached hints `asOfSeq: -1`.** Hints have no event cut in current history and cannot participate in sequence comparisons.
## Consequences
Sessions requiring migration use the existing opening flow to prepare history; Sessions that need no migration remain cold-readable and forkable. Catalog hints may lag the log and yield to subsequent exact projections.
A separate follow-redirection proposal discusses future single-Session migration progress; this decision does not implement that flow.
# Agent Note: Anonymous translation remains separate from model history
Status: implemented
English | [中文](2026-10-05-anonymous-reasoning-translation.zh.md)
## Problem
Readers need reasoning in their chosen language without another model credential or replacement of the original assistant output. Translation needs an owner for external requests, cancellation, and failures independent of Agent execution.
## Decision
The experimental translator owns explicit Google or Bing routing, provider-language normalization, and bounded network requests. Its Service Definition and both providers share one package because their current consumers use the same operation and lifetime. The optional GUI consumer owns settings, paragraph scheduling, per-disclosure caches, and translated display. Its authenticated Remote calls the translator; the agent loop and Session event types remain independent.
Bing is the configured default. Enabling the bundle permits transmission of expanded reasoning to the chosen service, disclosed in plugin metadata and settings. Failure never selects another provider. Original remains available; translated text neither replaces persisted assistant output nor enters model input. The [optional-bundle admission policy](2026-09-21-experimental-capabilities-as-optional-bundles.md) keeps the feature disabled until selected.
The GUI registers a translation wrapper in the standard `conversation.chat.reasoning.body` Slot and calls the official `conversation.chat.reasoning.content` Factory for original or translated text. The default body is also an ordinary registration; the host supplies no inline Markdown fallback or plugin action state. The translation wrapper owns its toolbar and passes its localized labels to the Factory, which supplies default labels when callers omit them and fixes compact typography. The [reasoning Body Slot and Content Factory design](2026-10-09-reasoning-content-slot-factory.md) supersedes this record's body composition mechanism; external requests remain separate from model history. Consumers do not replace the assistant renderer or alter Session projection. Only expanded bodies start translation. Collapsing, changing selection, or unloading aborts outstanding work and excludes late results.
## Alternatives considered
**Translate through the conversation's LLM provider.** This adds model requests and depends on model credentials, pricing, and routing. Anonymous translation has separate availability and requires none of those credentials.
**Rewrite reasoning in the Session or provider stream.** This changes the original trajectory and can affect model reasoning passback. Display translation preserves the original model history.
**Silently try another endpoint after failure.** This sends potentially private text to another operator without an explicit selection. Retry uses the same provider; switching providers is a user action.
## Consequences
Anonymous translation adds no third-party translation library or model-token usage. Anonymous endpoints provide no developer availability guarantee. Both providers send source text in POST request bodies. Translations may alter Markdown or code formatting and do not translate collapsed summaries.
The translator retains Session-bound results under the [shared machine-translation persistence decision](2026-10-05-persistent-machine-translation.md); its stateless calls do not record results.
# Agent Note: Discover Official bundles without installing their runtimes
Status: implemented
English | [中文](2026-10-05-official-on-demand-bundles.zh.md)
## Problem
Native subagent integrations bring their own runtimes. Shipping those dependencies with every DSH installation charges that cost to people who never delegate to them, while requiring a package name hides the integrations from the Plugins page.
## Decision
`ON_DEMAND_BUNDLES` admits the Claude Code and Codex bundles to an installation-owned catalog. The catalog embeds their package-owned localized metadata and icons, so discovery works offline without resolving or importing either package. Official identifies project maintenance; optional delivery and experimental maturity remain independent properties.
Selection uses the ordinary third-party bundle installer, profile dependency, Host row patches, compatibility evaluation, progress, cancellation, registry fallback, and build-script approval. Native authentication and configuration remain provider-owned.
The running installation owns the target. Published applications offer the exact DSH release version and save an exact registry dependency pin. Source Web and development Desktop offer ordinary `link:` dependencies into the same checkout as the running CLI, using its prepared workspace dependencies and peers. The source target follows the actual installation, never the current working directory or a prerelease suffix; an incomplete checkout does not fall back to the registry. A development checkout can share a version string with different published code, so replacing an installed registry copy or another checkout’s link compares the source as well as the version. Replacement is explicit, preserves selection, and reports required restarts; startup does not download providers. [Plugin Manager](../../../../packages/boot/plugin-manager/README.md#use-this-package) owns the operational details.
Every Official package remains in the DSH release family. Packed DSH dependencies, optional dependencies, and peers use the exact family version. Publication prerequisites place catalog packages before the DSH package advertising them without adding their runtimes to DSH production dependencies. Catalog verification rejects missing metadata, unsupported membership, and stale generated content.
The [shipped optional-bundle policy](../process/2026-09-15-shipped-optional-bundles.md) still owns lightweight installation-provided bundles. The [optional-composition decision](2026-09-21-experimental-capabilities-as-optional-bundles.md) owns shared patch behavior and global optional tools. Native bundles insert ordinary Host rows, so any profile can select them; profiles that keep the package without its bundle mount the same rows explicitly. The [identity compatibility decision](2026-10-06-web-desktop-plugin-identity-compatibility.md) owns retained names, selectors, and deferred naming cleanup.
## Alternatives considered
**Ship the native runtimes with every installation.** This makes first activation offline, but increases the default download for unused integrations.
**Query the registry to build the discovery list.** This avoids embedded metadata, but makes discovery depend on network availability and does not bind the offered package version to the running application.
**Give Official packages a separate installer or activation registry.** A second path would duplicate package-operation and profile-composition behavior. Installation-owned discovery data can reuse the existing third-party mechanisms.
**Always install the registry package with the running version.** Version equality identifies a published release, but cannot identify unreleased source changes. It can deliver an older provider API and a provider-only patch to a development Host that expects current tool contributions.
**Build a separate development registry or artifact cache.** Ordinary local links already keep provider code and its dependencies in the running checkout. A second delivery system would add lifecycle and invalidation work without improving that development promise.
## Consequences
Registry delivery requires a reachable registry; development delivery requires a prepared, built checkout and normal rebuild/restart discipline. Equal source version strings do not certify identical build outputs. Off only deselects; Remove remains separate and refuses while the bundle's rows are in use. Activation failure can retain the saved dependency and selection; there is no rollback. Offline metadata, release pins, source-target selection, same-version source mismatches, clean installed artifacts, and release ordering have dedicated checks. Native-provider authentication is not part of catalog discovery.
English | [中文](2026-10-05-paid-native-flash-reasoning-translation.zh.md)
## Problem
Readers may choose model-based translation in exchange for additional charges. Reusing their conversation request would inherit its history and generation controls, while routing through an arbitrary configured model would conceal the translation's billing and receiver choices.
## Decision
Paid translation is an explicit alternative to the [default anonymous path](2026-10-05-anonymous-reasoning-translation.md). Only active built-in `deepseek-account` and `deepseek-official` owners with configured credentials and the exact `deepseek-flash` catalog ID qualify. Display names and versions do not participate, and pi-ai routes are excluded. Each fragment receives one fresh text-only query, a simple translation instruction and disabled thinking; no previous fragment or main history is supplied.
The original Session keeps `plugin:translator/request`, the [shared translation request record](2026-10-05-persistent-machine-translation.md), with exact model-visible input in optional `metadata.modelRequest` before dispatch. The existing `appendPluginRecord` writer snapshots its JSON payload and marks it ignorable without replacing main history. Read-only cache lookup precedes native admission and needs no live Session. A cache miss uses the active Session writer and flushes before dispatch; the GUI joins normal Session activation when needed, with its ordinary startup and resume effects. Format migration retains these informational records on a best-effort basis. Translation does not supply the main Session identity to the native request extensions, so a fragment cannot authorize uploading its full conversation log. The configured paid choice is enforced by the Host before billing.
## Alternatives considered
**Reuse the conversation's model call.** Its history, model and thinking settings are unrelated to a translation fragment, and session-log delivery could transmit more than the selected text. A fresh native Flash call makes these choices explicit.
**Offer every model or version label.** This requires additional receiver and billing decisions. The stable `deepseek-flash` ID supports future versions while preserving the requested built-in route restriction.
**Replace the original assistant output.** Translated text is retained separately as a shared result record; original reasoning remains the source of main model history.
## Consequences
Paid queries consume additional input and output tokens. The GUI discloses billing for uncached fragments. Reopening reuses persisted successful results, including when native credentials or model configuration are unavailable. Failures never choose another route. Local audit retention is independent of provider log delivery, and anonymous translation remains the default.
# Agent Note: Machine translation results use shared Session records
Status: implemented
English | [中文](2026-10-05-persistent-machine-translation.zh.md)
## Problem
Readers need a saved translation when they reopen reasoning or restart the application. A disclosure-only cache repeats external requests, can change an already displayed result and can incur repeated model charges. Storing only model inputs leaves successful translations unavailable after restart.
## Decision
The translator owns `plugin:translator/request` and `plugin:translator/result` for every provider. A request records its selected provider, exact source text, source and target languages, and recipe identity. Optional metadata carries provider-owned request details. A result references the request's Session sequence and retains the completed translated text. Requests are durable before dispatch, and results are durable before they return to the GUI.
The browser binds displayed translations to the disclosure's Session, and the Host validates the configured provider and explicit target language. The authenticated Client supplies the original text and Session identity; the Host does not verify that the text belongs to that Session. Successful records are reusable across disclosure lifetimes and restarts. Identity includes the provider and recipe, so changing the receiver, language or translation settings cannot substitute another result. Identical concurrent translations wait for one result; distinct fragments can run in parallel.
Read-only lookup precedes provider work. New records use the active Session's existing writer through `appendPluginRecord` and targeted flush. An uncached inactive Session rejects before provider work; the GUI consumer joins normal Session controller activation and retries once.
These records accompany the [original reasoning](2026-10-05-anonymous-reasoning-translation.md) and never replace its events or model input. They use the existing experimental record writer and best-effort format-migration retention. Provider availability gates new requests; a saved result remains readable without another inference. Stateless translator calls remain available without a Session.
## Alternatives considered
**Keep only disclosure-local results.** Reopening loses the displayed value and sends the same text again, with possible repeat charges.
**Give anonymous and model translations different result formats.** The reader needs the same source-to-translation relationship from every provider. Provider-owned audit details belong in optional metadata.
**Replace original assistant reasoning.** A translation is derived data. Keeping it alongside the original preserves model history and access to the source.
**Open a separate writer for historical translations.** This competes with normal Session activation and requires coordination between two writer owners. The GUI already activates opened conversations; using its existing controller keeps that ownership in one place.
## Consequences
The Session retains translated text and request identity in addition to original reasoning. Only completed persisted results supply cached text; failed or incomplete attempts do not. A completion that has already entered the append-only log may survive a late caller cancellation. Records do not introduce released Session schema history, and future format migrations may drop experimental state.
Saved results remain readable without activation. A new translation may wait for normal Session activation, including preset startup hooks, resume markers, interrupted-turn repair, and any required generation migration. These ordinary lifecycle effects can append context; translation does not bypass them. Cache reads leave older generations unpublished. The translator owns no second writer and requires no core lifecycle changes.
English | [中文](2026-10-05-shared-agents-root-instructions.zh.md)
## Problem
DSH loads one user-global instruction file, `$DSH_HOME/AGENTS.md`. People who run several agent tools keep the instructions they share across those tools in the shared agent configuration root (`~/.agents/AGENTS.md`, or `$DSH_AGENTS_HOME`) that `dsh-skill-filesystem` already scans for skills. Those instructions reached DSH only when the person copied or symlinked them into the harness home, which duplicates one source of truth into DSH-owned data and leaves no way to tell which copy a session loaded.
## Decision
The user-global instruction scope has two ordered roots: `$DSH_HOME/AGENTS.md` first, then `<agentsHome>/AGENTS.md`. The instruction loader resolves the harness home from `$DSH_HOME` or `~/.dsh`, and `agentsHome` from `$DSH_AGENTS_HOME` or `~/.agents`. `resolveAgentsHome` and `agentsHomeDisplay` in `dsh-home-paths` own that resolution for both `dsh-agent-instructions` and `dsh-skill-filesystem`, so the shared root has one resolver whose tilde expansion and blank-environment handling match the harness home.
The two files are candidates of one candidate group. They deduplicate by trimmed content, so identical copies render once from the harness home, and an unobservable candidate preserves the group's last-good state. A duplicate of budget-retained content keeps being reconciled, so deleting or changing the retained candidate promotes it; a byte-budget omission leaves that candidate and its content duplicates unprobed. The harness-home file keeps the `user-global` scope key that earlier sessions recorded; the shared file carries `agents-global`, and reconciliation probes both roots on resume, so an external edit, a removal, or a newly created file appears as a change notice instead of a new baseline.
A user-global scope directory is an internal reconciliation key, so it stays out of model-visible prose: a global delta states `These user-global instructions apply to all work.` instead of naming a scope directory, and a project path whose first component would collide with one renders with a leading `.` so it keeps its own scope.
`USER_GLOBAL_DIRECTORIES` fixes the root order, and discovery, baseline reconciliation, and deduplication all iterate it. The harness home therefore decides content duplicates and keeps its position in the baseline, while distinct content from the shared root follows it.
## Alternatives considered
**Replace the harness-home file with the shared root.** The harness home owns DSH-specific configuration, and its `AGENTS.md` is the user-global file the package documents, tests, and renders today. Removing it would drop a DSH-only layer for anyone who shares instructions across tools, and sessions that recorded `user-global` state would lose their referent.
**Load the shared root only when the harness-home file is absent.** The two files then alternate as the session-visible user-global file, which makes a DSH-specific file silently suppress the shared one instead of composing with it.
**Load the shared file as a second candidate name in the harness home.** `instructionFileCandidates` names files within one directory; the two roots are different directories, so the option cannot express this chain without redefining candidate semantics.
**Rely on path identity instead of content for the duplicate case.** The common layout is `$DSH_HOME/AGENTS.md` symlinked to `~/.agents/AGENTS.md`, which is two paths for one file. Path identity drops neither, so the same text would render twice under two display labels.
## Consequences
- Shared instructions load without a copy or symlink, and the symlinked layout still renders once, from the harness home.
- A session that started before this change receives one transition when the shared file exists with content that differs from the harness-home file.
- `dsh-agent-instructions` exposes no `dshHome` or `agentsHome` config field; its two roots are process policy. `dsh-skill-filesystem` retains its explicit home fields, which select that provider's own skill roots.
- The two user-global files share one fate on probe failure: the group preserves its last-good state together, because one digest decides which member renders.
- `packages/context/agent-instructions/tests/agent-instructions.spec.ts` pins the two-root order, content deduplication, duplicate promotion after the retained candidate changes or disappears, the global delta wording, the `replace` and `remove` transitions, `$DSH_AGENTS_HOME` resolution, the `~/.agents` display label, and the escaped project directory name.
这两个文件是同一个候选组中的候选：按去除首尾空白后的内容去重，完全相同的副本只从 harness home 渲染一次，候选不可观测时保留该组最后一次成功状态。预算保留内容的重复候选仍参与对账，因此保留的候选被删除或改变时它会被提升为可见；字节预算省略会让该候选及其内容重复项不再被探测。harness home 文件保留更早会话记录过的 `user-global` scope key；共享文件使用 `agents-global`。对账在恢复会话时会探测两个根目录，因此外部编辑、移除或新建文件都会表现为变更通知，而不会重新生成基线。
用户全局 scope 目录是对账用的内部 key，因此不会进入模型可见文本：全局增量声明 `These user-global instructions apply to all work.`，而不指名 scope 目录；首段路径与 scope 目录同名的项目路径会带前导 `.` 渲染，从而保留自己的 scope。
根目录顺序由 `USER_GLOBAL_DIRECTORIES` 固定，发现、基线对账与去重都按它迭代。因此由 harness home 决定内容重复项，并在基线中保持自己的位置；共享根目录中内容不同的文件排在其后。
## 备选方案
**用共享根目录替换 harness home 文件。** harness home 存放 DSH 专有配置，其 `AGENTS.md` 是该包当前记录、测试并渲染的用户全局文件。移除它会剥夺跨工具共享指令者的 DSH 专有层，也会让记录了 `user-global` 状态的会话失去指涉对象。
**仅当 harness home 文件不存在时才加载共享根目录。** 这样两个文件会轮流充当会话可见的用户全局文件，使 DSH 专有文件静默压制共享文件，而不是与之组合。
**把共享文件作为 harness home 中的第二个候选名加载。** `instructionFileCandidates` 命名的是同一个目录内的文件；这两个根目录是不同目录，因此该选项无法表达这条指令链，除非重新定义候选语义。
# Agent Note: Preserve Web and Desktop plugin identities across experimental consolidation
Status: implemented
English | [中文](2026-10-06-web-desktop-plugin-identity-compatibility.zh.md)
## Problem
Moving a capability into the experimental directory changes its ownership and support classification. Renaming its published package or an existing configuration row additionally changes how installed profiles refer to that capability. Those identity changes can make a working Web or Desktop customization require edits unrelated to the capability's behavior.
## Decision
The moved packages retain the published names declared by the [experimental naming policy](../../../../scripts/experimental-package-policy.ts). That typed directory-to-name map is the sole maintained inventory of naming exceptions. Package manifests, consumers, installation checks, and bundler checks agree with it; a missing, moved, or duplicated exception is rejected. [Experimental status](../../../../packages/experimental/README.md#status) remains independent of spelling, Official discovery, and explicit selection.
Existing badge, Ralph, and native delegation row IDs remain valid Host row selectors. New isolation groups and newly introduced capabilities retain their new IDs. Later user layers still replace configuration or disable rows. Selecting an optional bundle is still explicit: a patch targeting an absent row does not select its bundle.
Compatibility protection covers Web and Desktop profiles, including existing native provider package identities, selections, configuration, and authentication. Native bundles insert ordinary Host rows, so Headless, SDK, and ACP compositions can select them too; separate provider-only compatibility wrappers are unnecessary. Public APIs remain pre-stable, and new output metadata stays required unless an actual consumer needs an accommodation.
## Deferred naming cleanup
| Retained identity | Unconstrained cleanup | Why it is deferred |
|---|---|---|
| Published names in the naming-exception map | Give every experimental package the `@deepseek-ai/dsh-experimental-` prefix | Existing dependencies, imports, module declarations, and profile files use the established names |
| Existing badge, Ralph, and native-tool row IDs | Give optional contributions consistently distinct `optional-*` selectors | Existing user overrides use those IDs; the ownership move does not require a new selector |
These naming changes are deferred, not prerequisites for experimental classification. A later cleanup requires an explicit release and migration decision, an inventory of affected Web/Desktop configurations, and installed-profile upgrade evidence. Directory moves or pre-stable API status alone do not retire these exceptions. The package list stays in the typed policy rather than being copied into this note or package READMEs.
## Alternatives considered
**Rename every package and row during consolidation.** Uniform spelling makes the new ownership visible, but adds dependency and configuration migration to a change that can preserve those references.
**Publish aliases or keep duplicate compatibility rows.** Parallel identities need continuing resolution and removal rules and can register the same capability twice. Retaining the established identity avoids that additional mechanism.
**Keep provider-only bundles through additional wrapper packages.** Native bundles insert ordinary Host rows that compose into any profile. Extra delivery packages would preserve compositions outside the selected compatibility scope and increase catalog and release maintenance.
## Consequences
The repository carries a bounded naming-exception map and tests that classify retained names through direct imports, aliases, transitive dependencies, and packed or bundled artifacts. An unprefixed name does not grant admission to the default product. Existing row identities have Web/Desktop composition coverage for user overrides and disabled state. No general alias or automatic profile migration mechanism is introduced.
The [classification decision](../process/2026-10-05-product-package-classification.md), [optional-composition decision](2026-09-21-experimental-capabilities-as-optional-bundles.md), and [on-demand discovery decision](2026-10-05-official-on-demand-bundles.md) retain their separate responsibilities. Released Session generations and persisted hook event names remain protected independently of this profile compatibility scope.
Loaded 100 of 3226 files, more files were not shown because too many files have changed in this diff.
Show more
Reference in new issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
07akioni