Review bot findings: DSH_TOOLS_MODE values, the wire-replacement
wording, 'code-only' mode references, and the spill README's
dispatch-log waterfall name all still named the removed 'code' value.
The review bot found stale 'code' configuration values in the zh tools
and agent-tool-presentation READMEs and the zh tool catalog, which the
runtime schema already rejects in favor of 'ptc'.
Master removed the ApiProxy package; its former test paths now live in
the client connection, API gateway, and settings controller test
directories. Update the three notes' references so verify-package-paths
resolves.
The rebase merged master's SDK-example and telemetry prose with the
ptc renames in the tools and CLI reference READMEs and the executor
collapse note; re-record their pair hashes.
The DSH_SNAPSHOT refresh recorded the internal seq-range form; the
canonical fixture layout expands those ranges. Apply the mechanical
migration instead.
Master replaced the session-log-version-mechanism note with the
fail-closed-session-event-vocabulary note; update the rename note's links
(en/zh) so cross-links resolve.
The subagent review found stale code mode-value prose the mechanical pass
missed: tools and agent-tool-presentation READMEs (en/zh), the CLI reference
page (DSH_TOOLS_MODE and the preset roster), the zh tool catalog, the
execute JSDoc and collapse comments in dsh-tools, the codeModeHarness
helper in agent-loop tests, and stale code-mode titles in ptc.spec.
Apply the rename pass to READMEs and docs the master sweep rewrote, fix
PTC mode anchors and the renamed-note links in the spill READMEs, and
regenerate the doc graphs.
The split kept the session-persistent vocabulary (tool/code-dispatch*,
tools-code-mode, :code:) on this PR, but several prose surfaces still named
the new values: the zh persistence/tool catalogs, the renamed Agent Notes'
event mentions, spill-policy comments, and a garbled 're-enPTC mode'
replacement. Also rename the mode value to ptc in the places the rename
missed (tools and agent-tool-presentation READMEs, the Config JSDoc, note
mode unions) and the codeModeHarness* e2e helpers.
Rename the tool-presentation transport from code-mode to ptc everywhere
that is not written into session logs: the mode config value becomes 'ptc',
the preset directory/id becomes ptc, the demo becomes demo:ptc, the
dispatch waterfall becomes tools/ptc-dispatch-log (types PtcDispatch*), the
prompt rule becomes tools:ptc-only, source/test files become ptc.ts etc.,
and prose says PTC mode / PTC 模式. The session-persistent vocabulary
(durable events tool/code-dispatch*, logged plugin name tools-code-mode,
sub-call id segment :code:) intentionally stays and moves in the stacked
persistence PR, which is blocked until the SESSION_FORMAT_VERSION v0→v1
migration lands with it. run_code, its code parameter, CodeSdkLanguage,
CodeRunFailedError, the dsh-code-runtime family, third-party codex names,
and frozen archived notes keep their names.
chmod 0 can only deny directory reads on POSIX to a non-root owner:
Windows exposes no directory permission bits for readdir, and root
bypasses them. Where the fixture stays readable the sealed candidate is
indexed, so the unreadable-branch behavior is pinned on POSIX non-root.
An injected readdir failure keeps that branch covered on every platform.
collectPackageNames repeated collectPackageAliases' directory walk
verbatim, which jscpd reported as a 7-line clone and which failed the
lint-and-duplication gate on both the Linux and Windows lanes. Both now
read one workspacePackages() iterator: the alias collector keeps only
packages named after their directory, and the coverage collector keeps
every one of them.
Review found the CLI entry guard compared import.meta.url against a
concatenated file:// URL. That fails whenever import.meta.url encodes
something process.argv[1] does not — a repository path containing a
space, or any Windows drive path — and it fails silently: the script
exits 0 having done nothing, so verify-tsconfig-paths would pass
without checking exactly where it is needed. Running a copy from a
directory whose name contains a space reproduces it. The guard now uses
the comparison the repository's seven other generators already use.
Deleting the group wildcards also removed the fallback that resolved a
package nobody had aliased, so the generator now asserts coverage:
every workspace package with a src directory must be mapped, and
--check names any that is not. That assertion immediately found four
packages named dsh-<group>-<directory> whose siblings carry hand-written
aliases while they did not, so they reached built lib/ output through
the workspace symlink. They now carry aliases too, which takes the
resolution differences in this branch from seven to eleven.
Two comments in tsconfig.base.json still pointed at the deleted
wildcards; they now state why those aliases stay hand-written. The
package-inventory proposal recorded the wildcard collapse as current,
so both notes are cross-linked as partial supersession.
* feat(ui-conversation): fold turn process before final answer
* fix(ui-chat): polish turn-process control row from review
* test(web): drive preset slash catalog with gestures
* fix(ui-chat): keep turn process order stable
* fix(ui-chat): preserve prompt order after pagination
Co-authored-by: Yif <[email protected]>
tsconfig.base.json resolved @deepseek-ai/dsh-* through 49 candidate
globs and @deepseek-ai/dsh-*/invariant through 45, one per package
group. Resolution tries candidates in order, so a package late in the
list paid for every earlier miss — and under the dsh source launch each
miss is an ERR_MODULE_NOT_FOUND that Node decorates with a full
CommonJS resolution walk. A boot profile attributed 934.6 ms, 35% of
startup, to that decoration path across 60,942 failed resolutions. The
cost landed hardest on packages/util/*, which sits at position 44 of 49
and holds the leaf utilities nearly every plugin imports.
gen-tsconfig-paths writes one explicit alias per package into a marked
region and both group wildcards are gone; verify-tsconfig-paths reports
drift and runs in the ci-static lane. Booting the headless profile from
source drops from ~2,157 ms to ~1,055 ms with --help output unchanged.
All 1,022 dsh specifiers in repository sources resolve to the same
target as before, except seven /invariant specifiers in the lsp,
terminal, and runtime-diagnostics groups that the deleted wildcard
never listed: those reached built lib/types instead of src, against the
rule that static gates resolve through paths to src on a clean tree.
Progressive enhancement on body in the shell base sheet so mixed
Chinese/English copy gets consistent spacing without content edits;
engines without support ignore the property.
A title now comes from an attached session's live projection cut or a cold
one's durable checkpoint, and from nothing else. Attachment is decided by the
session store at read time, so a session that attached after the listing is no
longer answered from a checkpoint its log has moved past — the stale-title case
`api-session.list` already handles this way.
The log fold and its per-log memo are gone. Folding one title costs a whole
log, and this call sits under every keystroke; a session no projection answers
for is labeled by its id and regains its title the first time it is opened.
`lib` leaves the default exclusions: Ruby gems and many npm packages keep
sources there, and the miss would be silent and total. A traversal whose root
is unreadable now rejects instead of publishing an empty index over entries
that are still good, which is what the stale-while-revalidate path claimed but
could not do while every readdir error was swallowed. A drill marks the menu
drilled only when its edit actually reached the draft.
Refs #3154
Refs #3180
Review found the shared observation defeated the very fence it fed:
it carries the original PID-to-start-time pairing forward, so a
recycled PID still matches it and takes a signal meant for the process
that exited. Capturing it outside the per-member try also let one
failed read abort a whole teardown round, breaking the synchronous
host-exit contract, and an empty round paid a read for no members.
signalProcess now reads ProcessInspector.isAlive immediately before
delivering, from the narrowest per-identity source each platform
offers; signalMembers and waitForMembers return before capturing when
a round has no members. snapshot() keeps serving the readiness poll,
whose per-poll table read stays at one.
Windows enumerates Toolhelp32 lazily on the first tree question, so a
snapshot asked only for liveness — the 25 ms teardown poll — performs
no table walk at all.
Replace the debounce-timer polling with Include.stop(), which flushes the
file-backed write queue deterministically. This avoids Windows coverage
flakes where the self-dispose write could land after the expect.poll
timeout.
TurnNavigator was an unmemoized component rendering one div and one
button per loaded Turn, so every ChatView render rebuilt the whole rail:
143 button rebuilds per commit in a 300-Turn session against 8.4 in a
4-Turn one, while a streaming answer commits dozens of times.
memo alone would not have helped, because navigateToTurn was rebuilt on
every render and broke prop identity, so it moves into useCallback.
MacProcessInspector answered the descendant tree and every member's
liveness with its own `/bin/ps` fork, so one readiness poll cost N+1
full table reads for N tracked descendants. With execFileSync on that
path and a 50 ms poll interval, any command spawning two or more
children saturated the host event loop until it exited.
ProcessInspector.snapshot() now returns one ProcessSnapshot that
answers tree, session, and alive from a single observation, and
signalProcess takes the caller's observation so its PID-reuse fence
does not re-read the table per member.
Collapsing every relative-time bucket to `{{age}}` reached the session-tree
goldens, where a literal age is the assertion: a fresh row reads `now` and an
older one does not. Six e2e files failed, two of them by aborting mid-scenario
and leaving their replay fixtures half-consumed.
`captureStableAria` now takes the rule as an opt-in, and only the reference
menu — whose rows are dated from the live Host list — asks for it.
Refs #3154
The Host session list already carries each session's `updatedAt`, which is
the number its own rows show; reading it there keeps the two surfaces from
disagreeing and avoids making a context capability depend on the BFF
assembly that owns the `sessionListMetadata` projection. A session the list
does not carry falls back to the candidate's creation time.
Refs #3154
Session candidates labelled from projection checkpoints instead of a full
log fold per keystroke, with the uncheckpointed remainder folded once and
memoized while its log stays cold. The file index keeps answering while an
invalidated traversal rebuilds behind the caret, and its default exclusions
now cover build outputs so deep sources stay reachable.
Rows carry only what distinguishes them: a file names its parent directory,
a session names its workspace only when that workspace is not the current
one, and a drilled listing names none because its new breadcrumb does.
Resolves#3180
Related to #3154
Stacked on #3115: keep its windows-* setup-pnpm-js-<run_id>-<run_attempt>-<job>
destination, and extend the same isolation to non-Windows jobs in ci.yml
and ci-master.yml with setup-pnpm-<run_id>-<run_attempt>. This prevents
sequential self-hosted Windows jobs from tripping over a stale locked
pnpm.exe/reflink native module.
The anchors-a-relative-add-spec case serializes two subprocesses (plugin
add + remove) under a hardcoded 90s budget, which the 2x60s worst case
exhausts; derive it from SPAWN_TIMEOUT_MS * 2 + 30s like the other
dual-call cases. The pnpm setup isolation note now also records the python
SDK exe build's suffixed destination and its regression-test coverage.
The merge resolution had reverted THIRD_PARTY_NOTICES.md to the SDK 0.3.220
rows; the lockfile pins 0.3.241, so CI regenerated a mismatch. The apply spec
now drives the injected onHover face, closing the per-file coverage gap on
src/client/index.ts.
The plugin add and dump-default-config cases serialize two runBuiltBin
calls, each with a 60s execa cap; the 90s outer budget could be exhausted
before the second call. Raise them to SPAWN_TIMEOUT_MS * 2 + 30s, matching
the multi-call treatment.
Candidate rows lead with domain icons instead of localized text
prefixes; pointer and keyboard share one reducer-owned highlight (last
input wins); drillable folder rows reveal a localized Browse-folder +
Tab keycap hint with the library chevron; pending sources render
skeleton bars; the menu spans the composer card. The editable @dir/
text decorates color-only — the domain icon now belongs exclusively to
the settled reference chip. Composer placeholders advertise / and @,
and the zh copy for commands is unified to 指令.
The Agent Note (both languages), the PR prose, and the commit message
claimed a 13px floor for the table variants; the formula has none —
max(13px, setting − 2px) selects the −1 branch at low settings rather
than clamping the result, so the tier bottoms out at 11px at the 12px
setting, matching think text. Rewrite the claim, say so in the axis
comment, and split the README sentence that lumped body-pair and
secondary-pair consumers together.
Assert the engine-resolved secondary size in the settings-chrome e2e
(13px at the default, 13px at the 15px boundary, 14px at 16px,
unchanged across reload), sync the StatsLine and workflow-panel spec
headers with the tier they now pin, and note why memberLabel stays at
the body size.
Derive --dsh-content-font-size-secondary (setting -1 at <=14, setting -2
above; 13px at the default) with --dsh-content-font-delta-secondary in
gradient-shadow-text.css, and move every one-step-under-the-body text
onto it: think text and reasoning summaries, the shared DisclosureRow
title, ToolRow and bash-row summaries and file links, compaction/
context/command/retry/error rows, StatsLine, the workflow-run panel
tiers, reference summaries, the turn-status clock, and the feedback
note trigger. The markdown table variants join the same tier instead of
staying fixed; its 13px floor keeps them legible at the 12px setting.
At the default setting the flow-row titles and summaries render at 13px
(previously 14px) so they match think text at every setting instead of
sitting 2px above it.
The requires-profile and routes-help cases serialize 4-6 runBuiltBin calls,
each with a 60s execa cap; under the loaded pool the 90s outer budget was
exhausted before the last call and vitest truncated the run without the
execa diagnostics. Raise both to SPAWN_TIMEOUT_MS * 3 + 30s.
* ci(build-exe): drop pull_request label trigger to avoid skipped checks
* docs(build-exe): sync agent note and pin event set in workflow spec
* test(ci): type-safe event key assertion for build-exe workflow
* ci(build-exe): use present-tense trigger comment and drop label-run note
The built-bin help/usage case still used a win32-conditional 60/30s outer
budget while serializing six runBuiltBin calls, and startProfileLifecycle
lacked the execa timeout/killSignal the sibling helper has; the tool-ralph
cases pinned 20-30s explicit timeouts that the 90s lane default cannot
override. Align all of them to the SPAWN_TIMEOUT_MS + 30s (or 90s) pattern.
The per-case 15-30s budgets on the Windows native and coverage lanes fire
before oxlint, workflow-worker-thread, and other subprocess-spawning cases
finish under the loaded self-hosted pool; the failures rotate across cases
as load shifts, so per-case widening only moved the flake. Raise the lane
defaults (DSH_COVERAGE_TEST_TIMEOUT_MS and the native --testTimeout) to 90s,
align the oxlint and workflow-worker-thread case budgets, and keep the
built-bin SPAWN_TIMEOUT_MS at 60s under a 90s outer budget.
The provider-card seat's keyConfigured now derives from the reference the
page would use — the profile's apiKeyEnv, or the page's derived
<ROUTE>_API_KEY while the profile names none — so the add-provider draft
agrees with its own editor about an existing conventional credential (the
store joins the derived describe in the same batched call, as
ProviderRow.derivedCredential). ModelsSectionProps makes the renderSlot seat
required so a direct render that forgets it fails to compile; the one such
render in provider-form.client.spec regained a real mount and the test
boilerplate collapsed to renderSlot={() => null}. The extension-slots Agent
Note now states the keyed cell's real override rule (same priority throws,
a different priority shadows), and docs/subsystems/slots.md carries the two
new seats in its hierarchy, both languages.
The windows-* jobs now install pnpm under a run/attempt/job-suffixed
destination, so the pre-install step that cleared the old fixed
setup-pnpm-js path no longer touches the actual destination and its
comment claims stale state. The suffix already gives every job a fresh
directory, so remove the four cleanup steps.
The execa timeout was widened to 60s but the outer vitest case budgets stayed
at 30s, so a cold-starting built bin would trip the vitest budget first and
the execa SIGKILL cleanup could not run inside it. Extract SPAWN_TIMEOUT_MS,
share it across the execa deadline, its error text, waitForFile, and the
outer case budgets (60s spawn + 30s headroom), so the widening is coherent.
Both suites spawn real subprocesses (oxlint probes; the dsh built bin) that
cold-start slowly on the contended self-hosted Windows pool, so their 20-25s
timeouts fire before the child finishes. Raise the oxlint contract case
timeouts to 60s and the built-bin execa timeouts to 60s, matching the
tool-ralph budget treatment.
The windows-* jobs keep a separate standalone pnpm executable under
runner.temp/setup-pnpm-js. A previous job on the same self-hosted runner
can leave a locked @reflink native module there, so the next job's
pnpm/action-setup fails with EPERM during unlink before any test runs.
Suffix the destination with run_id, run_attempt, and job so every job
gets a fresh directory even when sequential jobs land on the same
runner; apply the same to the python SDK exe build. Update the pnpm
setup isolation note to record the Windows-specific destination.
pi-ai credentials carry optional members as explicit undefined (a github.com
Copilot grant holds enterpriseUrl: undefined), and the store bridge committed
the object verbatim, so the credential store's strict validator refused the
write and sign-in failed after the provider had already authorized it.
toRecord now drops explicitly-undefined members and renders undefined array
entries null, exactly as JSON.stringify would; everything else passes through
untouched so genuinely unstorable values still fail loud at the store.
Review follow-ups: wait on the caller's own locator instead of assuming
it is the page's first composer, describe the actual read-only window
(submit adjudication and locked states — a running turn stays editable
for queueing), drop the pre-Lexical narration from the JSDoc, and record
the gesture-semantics trap as an Agent Note.
The Models section now declares two SlotMap seats for out-of-tree plugins:
settings.models.provider-card (keyed by the row's settingsNs, dispatched on
saved cards, the first-run setup posture, and the add-provider draft, with
the row view, configured join, and confirmed api-key state as owner props)
and settings.models.footer (ordered list after the rows and add controls).
Without registrants both seats render nothing. First consumer: the
llm-pi-ai-oauth companion plugin's sign-in surface.
`import.meta.resolve`'s `parentURL` argument takes effect only under
`--experimental-import-meta-resolve`, which no launch passes, so the
fallback resolved from this module rather than from the harness — the one
question it existed to answer. The disk walk is the whole answer now, and
the refusal memo it needed goes with it. A `file:` URL joins the file
branch rather than the package one, where a resolver would only normalize
it and report a missing target as present, and a row is skipped on the
Loader's own `Boolean(disabled)` so `disabled: 0` is checked like the
Loader checks it.
A broken card says so through `aria-disabled` rather than `disabled`, and
refuses the pick in its own handler. `disabled` took it out of the tab
order, which with the reason moved onto the badge left it unreachable
without a pointer — reachable before this change, so hiding it was a
regression rather than a path that never existed.
Both notes this decision partly supersedes are updated in place and
cross-linked, one README pair loses an editing residue that repeated a
sentence, and the single-row diagnostic no longer reads "row row 1".
A running turn disables the composer by flipping contenteditable to
false on the same element. fill() throws there immediately — a disabled
textarea used to hold it back through actionability — and isEnabled()
reports true for a div regardless, so the permission-policy scenario's
post-settle wait never waited and its next gesture raced the re-enable
render. The window is a few frames wide; #3083's Remote-routed subagent
control stretches settle enough to hit it on CI.
writeComposerDraft now waits for contenteditable="true" before acting,
and the permission-policy scenario drives all four sends through it with
the settle wait pinned to the attribute.
Discovery proved only that a composition parsed, so a preset naming a
package a later rename took away kept a healthy card and its place in
every picker until a person switched to it. It now resolves each row it
can prove will start, reading the package off disk and falling back to
the resolver only for names that look absent — the resolver costs a
synchronous hooks-thread round-trip under the source launch's tsx hook,
which the walk avoids for every row it clears.
The mount diagnostic followed `AggregateError.errors` but never a cause,
so a group that failed on two rows named neither. It now follows a cause
that carries more than its own message.
A refused switch left the chip's label snapping back with no account of
why, which is the only account there can be for a preset that resolves
and then refuses. It announces through the shared Toast, which gained a
caller-set hold for a cause that names packages and rows.
The tenth master merge staged the correct 0.3.241 notices, but the
pre-commit regenerator ran against a local pnpm store that still held a
stale 0.3.220 SDK directory alphabetically ahead of it and silently
committed the old payload table. Regenerated after removing the stale
store entries; the lexical rows this branch adds stay.
The generator names the first matching virtual-store directory, so a local
store still holding an older SDK payload alongside the locked one renders
that older version into the notices.
The full-corpus transform gate spawns one child that transforms and imports
every built bundle, so it runs for 8-25 minutes as a single case and
dominates one native Windows coverage partition, blowing its 900s budget
under load. Move it to the coverage-exempt heavy gate, which runs it with
its own worker budget instead of competing with the instrumented
partitions. The package's src is threshold-excluded in vitest.config.ts, so
the exemption carries no coverage; the exempt-heavy roster note records the
entry.
The self-hosted Windows coverage pool (16 shards x 12 workers on 192
threads) pushes real subprocess boots past their vitest deadlines: the
tool-pwsh Loader smoke reaches ~40s against a 30s process cap, and the
tool-ralph worker-thread cases exceed the 5s default. Give the pwsh
smoke a 90s process deadline (the subprocess keeps the assembled boot,
the vitest deadline stays at 120s), and give the two un-budgeted ralph
cases 30s each, matching the existing 20s quiescence case.
On Linux, pwsh may trap SIGTERM and exit cleanly when the subprocess service
is disposed, so the handle status is 'completed' rather than 'killed'. The
test already proved the process tree is gone via kill(pid,0); both statuses
satisfy the contract.
Nothing under packages/host/apiproxy imports @deepseek-ai/dsh-goal after the
unary domain deletion, so the dependency and its project reference go too.
The goal domain has been served by GoalService's @Remote namespace since it
shipped; the API Proxy copy was a second implementation of the same six
mutations. Remove the goals contract, schemas, route rows, IApiClient stub,
host implementation, and the fixture's compatibility face, leaving
ctx.remote.goals as the only path.
The fixture's goal fold keeps its coverage through the Goal Remotes: its
lifecycle case moves out of the unary-dispatch test, which no longer has
goal rows to cover.
Drop the width handle's double-click reset and tooltip — the handle is now
drag-only and a stored preference is only replaced by another drag — and
add a tertiary description line under the Settings font-size title stating
the size only affects conversation content. Update both Agent Notes and
the settings dialog goldens.
Adopt --dsh-content-font-size / --dsh-content-font-delta across the flow
rows around the transcript body: DisclosureRow header (row height, title,
leading box, and registered glyphs, with StateDot exempt), ToolRow and
bash-row summaries and file links, think text, compaction/context/retry/
error rows, message clock and icon actions, the workflow-run panel, and
the workspace browser. Update the font-size Agent Note and add CSS-text
specs for the new adoptions.
The scenario landed on master with a textarea locator; the composer is a
Lexical contenteditable surface here, so the fill waited 30s for a node
that never exists. Use the shared per-key draft helper against
[data-composer-input] like the other composer scenarios.
Other PRs are blocked by Windows ACP half-close tests timing out. Keep the
coverage job running for signal, but remove it from all-checks-passed.needs
until the Windows skip fix is validated.
Windows anonymous pipes do not surface a child stdout EOF while the child
process stays alive. The three tests that simulate 'child closes protocol but
stays alive' therefore cannot be reproduced on Windows and hang until the
test timeout. Skip them on win32.
These tests spawn real ACP child subprocesses. On contended self-hosted
Windows runners the default 30s budget times out. Instead of raising the
global coverage timeout, give this file 2x the configured default
(DSH_COVERAGE_TEST_TIMEOUT_MS) so it follows future default changes.
After the 4-partition split, other PRs' windows coverage now fails on
process-bound subagent-acp tests timing out at 30s under self-hosted
concurrency. Give the coverage lane the same 60s per-test budget that the
earlier failover profile used.
The rail's items now ride the Chat snapshot: a structural upsert re-derives
the loaded Turn set, a content-only upsert re-derives only the Turns whose
nodes changed, and each preview is capped so navigation state never holds a
copy of the transcript. The published array keeps its identity until an item
changes, so ChatView selects it as both data and change signal — and a
streaming reply's preview follows the in-place node update instead of the
last structural publication.
A scroll frame resolves the active mark with one hit test at the reading
line, falling back to a single row scan, rather than a DOM query per mark.
Flow-height changes resync through the existing column observer, navigating
during a pending page keeps the paging anchor, and the rail height no longer
holds a floor taller than the band it centers in.
Render reconstructable system prompts at each request-series boundary, preserve series declarations through pre-step wrappers, and keep the presentation and replay snapshots aligned across clients.
Keep five non-blank rows stable while the selected blank New Session is provisional, and derive the overflow count from the rows still hidden.
Fixes#2841
The rail is a landmark on every Chat wide enough to show it, so each
recorded conversation with at least two loaded Turns now carries the
navigation node and its marks.
The concurrent-migrator test wrote the winner document with a plain
writeFile, whose truncate-then-write window lets the boot's unlocked
initial read observe an empty file and boot an empty store under load.
Seed fixtures through writeFileAtomic instead, matching how the provider
itself persists, so a reader sees either the old or the new complete
document.
ChatView derives one navigation mark per currently loaded Turn, keyed by
Turn number and anchored on that Turn's first loaded user node. The rail
sits against the scrollport's right edge, centered in the band the sticky
composer leaves visible; hover and keyboard focus preview the Turn's
prompt and response, and activating a mark moves the shared scrollport
and records the resulting restoration anchor.
ConversationRoot publishes --dsh-conversation-viewport-height beside the
composer height it already measures on the scrollport, so floating View
chrome can center in that band without assuming a Session header height.
The PR changes native Windows coverage partitions from 8 to 4 to reduce
vitest worker startup pressure under high self-hosted concurrency. Sync the
implemented Agent Note (EN/ZH) so the decision record no longer says Windows
is fixed at 8, and revise the same-partition-count alternative accordingly.
Keep recognized code fences syntax-highlighted while assistant text streams. Preserve completed Shiki token lines across chunks, mirror token styles and CRLF handling, and retain plain rendering for unsupported or math-like fences.
Add unit, DOM-parity, and keyless assembled-Web coverage for the streaming-to-settled transition.
Closes#1499
Keep the 4-partition coverage profile, split the monolithic windows-native
job into smaller required jobs (build, coverage, native-tests) plus a
non-blocking observational job. Update ci-workflow.spec for the new topology.
Under high self-hosted concurrency, 8 partitions per Windows native job
triggered vitest fork worker startup timeouts. This branch lowers Windows
coverage to the same 4 partitions Linux uses, trading some single-job
coverage wall time for lower process-creation pressure.
CI hit the same dropped-fill race in lifecycle-chrome's slash-menu
sequence that the folder scenario hit: fill()'s single-task select-all +
edit lands on a Lexical selection that has not absorbed the DOM
selection after a trigger-menu interaction, so the previous draft
survives and poisons the next test. Promote the per-key gesture to
support.ts and use it at both proven-fragile sites.
The base-mounted projection cache's write-behind forces session-log flushes
at cache-chosen times, splitting packed chunk rows and changing the durable
JSONL batching the keyless headless and sdk replay fixtures pin. Neither
profile exposes a session-listing surface — the one-shot runner and the SDK
protocol — so both bundles disable the base row, restoring the pre-base
behavior the fixtures were recorded against (the same pattern acp-app uses).
The pinned prompts were recorded on master before this branch's
file-reference guidance rewrite; refresh them so the pin carries the
directory-aware wording the assembly now produces.
Directly after a chip deletion, fill()'s single-task select-all +
insertText lands on a Lexical selection that has not absorbed the DOM
selection yet and is dropped, leaving the previous draft in place. Real
keystrokes leave room for selectionchange between keys, matching what a
user's typing does.
The Windows x64 installed-wheel job timed out while waiting for initialize even though the same head passed on rerun. Exact packaged-runtime VM evidence showed a 6.47-second first cold handshake and 2.69-2.94-second warm fresh-home handshakes, leaving too little variance below the public 10-second default.\n\nRaise the independent initialize default to 30 seconds in both Python SDK configuration layers. Ordinary turn and shutdown timeouts remain unchanged, callers retain an explicit override, and tests plus paired documentation pin the public behavior.
Run the GitHub Windows runtime leg under the runner’s native PowerShell instead of inheriting the POSIX Bash body. POSIX and Windows now own explicit output resolution, virtual-environment setup, environment scrubbing, and keyless/live black-box commands, while portable build commands continue to use each runner’s default shell.
Put the pinned uv installation on the GitLab Windows job PATH before either the smoke or release builder invokes it. Reject a runtime executable whose basename does not match the selected platform manifest, and reject Intel macOS at platform selection instead of reporting a misleading missing artifact.
Add a complete PowerShell path to the published Python tutorial and record the three-phase shutdown-time bound in the Windows runtime decision. Workflow, Python, and bilingual documentation tests pin the resulting behavior.
Record win-x64 as the sole Windows Python carrier: node24-win-x64 builds a py3-none-win_amd64 wheel with dsh.exe, rg.exe, and both ConPTY addons; Windows arm64 remains explicitly unsupported. The note also pins native build ownership, shell-free pnpm launch, installed-wheel keyless/live gates, and the PowerShell-specific minimal snapshot.
Update the active SEA, sole-launcher, profile-runtime, installed-wheel, and publication decisions from three runtime wheels to four, preserving their existing rationale while linking the Windows extension. Contributor and runtime references now state the exact target, filenames, sidecars, snapshot ownership, and five-wheel release set in both languages.
Make the checked-in minimal SDK overlay disable both one-shot shell rows and mount exactly one persistent PTY stack: Bash on Linux/macOS and PowerShell on Windows. The SDK server, explicit dsh home, persistence, editor, timeout, and reduced tool catalog remain unchanged.
Update the runnable example and tutorial to list Windows x64 as supported, describe the platform-selected shell, and remove the obsolete POSIX-only restriction. This keeps the documented first Python task executable through the packaged Windows dsh profile instead of advertising a Linux-only overlay on a Windows-capable SDK.
Add node24-win-x64 to the required pull-request and public-release matrices on a native windows-2025 runner, and publish the same win_amd64 artifact from the GitLab tag pipeline. GitHub uses Git Bash for the shared release script while selecting the Windows venv's Scripts/python.exe explicitly; the Linux and macOS legs retain their existing commands and native checks.
Run the complete installed-wheel keyless suite and the trusted two-turn DeepSeek smoke on Windows exactly as on the existing targets. Make the minimal blackbox choose persistent PowerShell on Windows, keep advanced and restart snapshots platform-stable by disabling both one-shot shell variants, locate the generated dsh.exe console command, and validate text lines without assuming POSIX newlines.
Workflow tests pin the four-target matrix, Windows runner and wheel tag, cross-platform venv selection, GitLab publication dependency, and full blackbox invocation. The existing POSIX minimal snapshot changes only its platform-neutral prompt wording; Windows owns a separate model-visible snapshot.
Add node24-win-x64 as the only supported Windows runtime target and publish it as a py3-none-win_amd64 wheel containing the conventional dsh and ripgrep .exe payload names. Keep Windows ARM64 rejected explicitly so Python cannot claim a carrier that CI and release automation do not build.
Teach the pkg builder to require a native x64 Windows host, validate both node-pty ConPTY addons, copy @vscode's win32 ripgrep executable, and recognize pkg's .exe output. Extend runtime resolution, wheel staging, payload validation, and the preset closure check so the Windows-specific PowerShell plugins and sidecars fail loud when omitted.
The sidecar resolver now maps a packaged main.exe to main-rg.exe; focused TypeScript and Python tests cover that name, the win_amd64 manifest, x64-only host selection, complete wheel payload, ConPTY inventory, and platform-conditioned plugin closure.
The keyless max-token smoke carried a disable overlay that reproduced the removed private carrier's reduced tool roster. That legacy roster is now owned by the standalone sdk-minimal profile, so mutating the full sdk profile hides the application the SDK actually ships.\n\nLaunch sdk without a patch for both the successful and invalid-config paths. The separate sdk-minimal process test continues to pin its exact platform-selected two-tool request.
The startup test helper always constructed an explicit sdk profile, so coverage never exercised apply's supported default-config path even though runtime behavior depended on it.\n\nPass Config objects through the helper and default them to an empty config. The existing startup and help tests now prove sdk defaulting, while the explicit sdk-minimal case remains covered.
Stop narrowing the standalone DeepSeek adapter to a DSH_MODEL-derived one-entry catalog. The direct adapter already accepts model ids outside its advisory catalog, so retain only the DSH_CONTEXT_WINDOW fallback and let the JSON-RPC initialize model be the single runtime selection.
Remove model mirroring from minimal.py and the packaged smoke. The keyless process now initializes deepseek-v4-pro without DSH_MODEL, while the packaged scenario continues to use its unlisted smoke-model; together they prove both cataloged and arbitrary SDK model arguments reach the adapter directly.
Update the bundle, tutorial, SDK/example references, and owning Agent Notes to keep DSH_MODEL only as minimal.py's optional default input, never as a second value callers must synchronize.
Record sdk-minimal as the narrow repository-owned exception to base-first profile composition: callers still launch only dsh and cannot provide an arbitrary Cordis tree, while the shipped bundle may own a complete explicit roster. Cross-link the launcher, profile-bundle, Python-runtime, minimal-agent, snapshot, and telemetry decisions; the supersession audit keeps each older note active because its remaining rationale is independent.
Update the CLI, architecture, Python tutorial/reference, example, runtime-wheel reference, and bundle documentation. The docs distinguish the full sdk profile from sdk-minimal, explain explicit-home/plugin/patch customization, state the minimal permission and persistence choices, and retain the separately packaged web profile and frontend assets for direct dsh use.
Correct dsh-base descriptions to cover base-backed profiles, make SDK startup configuration visible in the generated config catalog, add sdk-minimal to the module graph, and regenerate the base-composition graph. English and Chinese pairs are re-recorded at the exact reviewed contents.
Make minimal.py select the shipped sdk-minimal profile directly and pass its selected model into the profile-owned adapter catalog. The Python SDK still starts only the bundled dsh CLI with an explicit Harness home; it no longer supplies an invocation overlay for this mode.
Drive both the source keyless process test and installed-wheel smoke through the same named profile. The keyless test pins the generated profile manifest and exact two-tool model request, while the packaged smoke keeps the persistent-shell, editor, session-log, and model-visible snapshot evidence.
Delete minimal.patch.yml and the unused complete-config/replay fixtures. Their composition now has one owner in @deepseek-ai/dsh-sdk-minimal, so the example and tests cannot drift into separate launch trees.
Add a startup-only sdk-minimal template whose sole bundle inserts the complete JSON-RPC agent tree over the empty profile root. The roster is an explicit composition allowlist: it contains one DeepSeek adapter, the minimal agent spine, persistent Bash, the string-replace editor, local execution, and JSONL persistence, while dsh-base and Web remain absent.
Reuse the SDK app startup provider so the new profile retains help, stdin EOF, and bounded launcher shutdown semantics. Make that provider render its configured profile name, which keeps both sdk and sdk-minimal help truthful without duplicating process lifecycle code.
Register the package in the CLI closure, TypeScript graph, lockfile, Knip policy, and bilingual bundle references. Exact manifest, row-roster, profile-template, config-dump, and HMR tests make later additions visible instead of relying on a blacklist.
Resolve the installation fallback generation before locking and return immediately when every required symlink or packaged proxy is complete. Parallel SDK rollouts sharing an initialized DSH_HOME therefore do not queue on profiles/node_modules.lock.
Missing or stale entries still acquire the cross-process writer lock, recheck the generation, and repair under exclusive ownership. Tests hold the lock to prove the steady-state bypass and verify that a partial repair retains already-correct siblings.
Give the SDK JSON-RPC server a per-root tool filter and let deployments mark the configured persona as the complete system prompt. The checked-in minimal overlay now names only bash and str_replace_editor, so later global tools and unrelated guidance from dsh-base cannot appear implicitly.
Keep the shared SDK host services and packaged Web capability intact. Only workspace instructions, compaction, and the conflicting one-shot Bash row remain disabled. Unit coverage pins the configuration paths, and a real dsh profile smoke proves the assembled prompt and exact two-tool request.
The packaged dsh launcher must expose installation modules to profile-local plugins without writing symlinks into pkg's virtual filesystem. The first review fix selected ESM exports correctly in ordinary Node, but real carrier execution exposed package metadata and VFS behavior that a synthetic tree did not cover: executable and declaration packages have no import entry, legacy main fields rely on Node probing, and pkg's Windows VFS prevents filesystem package-scope resolution from seeing exports such as zod/mini and @google/genai/web.
Resolve explicit exports directly from each installed manifest with the maintained resolve.exports package under Node import conditions. Publish only package-local candidate files that exist, reject escaping or malformed targets, preserve the package installation URL without realpath, and keep Node's legacy resolver only for exports-less packages. This avoids pkg filesystem package lookup entirely while retaining fail-loud behavior for broken runtime entries.
Add regression coverage for import-only, nested, symlinked, zod-style, and genai-style condition maps; unavailable and types-only entries; invalid and escaping targets; executable/declaration packages; extensionless main; and legacy index fallback. profile.ts remains at 100% statements, branches, functions, and lines. Update the bilingual package and Agent Note contracts, replace the runtime dependency and generated notice, and regenerate the lockfile through pnpm.
Resolve packaged profile proxies with Node ESM import conditions from each package installation, and fail loud when an explicit runtime export or legacy main entry is missing. Serialize the shared profile fallback under the existing cross-process writer lock so concurrent dsh processes cannot observe partial proxies; either carrier now replaces the other carrier’s managed entry without manual cleanup.
Give Python initialize its own 10-second default bound and name the selected profile in timeout diagnostics, while leaving ordinary agent turns unbounded by default. Package the dynamically resolved web frontend and skill-badge assets so the runtime wheel’s normal dsh profiles do not depend on pkg static-discovery accidents.
Rewrite the root launch rule and every active stale SDK-runtime note to the shipped dsh profile architecture in both languages. Focused tests prove import-only and transitive package exports, lock contention, cross-carrier transitions, missing-entry failures, asset inventory, and bounded initialization.
Document dsh as the only application launcher across architecture, CLI, SDK, app-boot, Python package, contributor, tutorial, and example references. Explain explicit home selection, profile and patch precedence, persistent external plugin installation, the Node-free runtime path, and the absence of complete-config or ~/.dsh fallbacks.
Record the Python profile-runtime decision and update the active naming, installed-wheel, and SEA packaging notes with precise supersession. Regenerate the configuration catalog and module graph after deleting the carrier, update both reviewed languages and pairing records, and classify the retained standalone Cordis files as lower-level test fixtures rather than launch interfaces.
Migrate the packaged-runtime smoke inventory from complete Cordis trees to the sdk profile plus ordered patches. Preserve the focused minimal and advanced behaviors, update the generated durable snapshots for explicit permission events and the smaller RunResult, and keep worker, MCP, ripgrep, PTY/editor, direct JSON-RPC, and real-provider coverage.
Add an installed-only external bundle scenario that invokes the wheel's dsh plugin command with a local file package, verifies profile manifest reconciliation, imports @deepseek-ai/cordis as a peer, asserts the packaged proxy returns the exact host Context instance, and proves its system-prompt contribution reaches the model. Migrate the repository source e2e and runnable minimal example to the same profile grammar.
Replace complete-config, session_root, runtime-bin, bridge-bin, and public argv override options with dsh_bin, profile, ordered patches, and dsh_home. Resolve executable/home/patch/cwd paths before spawn, select the sdk profile by default, and fail before launch unless dsh_home or non-empty DSH_HOME is explicit; Python never inherits ~/.dsh silently.
Remove Python-owned DSH_CORDIS_CONFIG, DSH_SESSION_ROOT, and DSH_CWD injection and drop session_root from RunResult. Keep arbitrary argv only as an underscore-prefixed fake-runtime adapter, retain provider/model/token and process controls, and append subprocess stderr to initialization JSON-RPC errors so profile boot failures name their actual plugin cause. Unit and carrier tests cover both exe and Node modes.
Make the zero-code dsh-python-runtime-closure depend on the real @deepseek-ai/dsh application and every required profile peer, then package apps/cli's built bin instead of the deleted Python carrier. Rename executables to deepseek-harness-sdk-runtime-<platform>-<arch>, update wheel/platform/build workflow discovery, and install a Python dsh console command that requires explicit DSH_HOME before exec.
Include profile, bundle, preset, native addon, and shared-library assets needed by the full CLI. Remove the checked-in default cordis.yml and preserve the existing wheel distribution names, Python module names, sidecar validation, and wire identity. Runtime resolution and release tests pin the new artifacts and dev Node carrier.
Teach the profile installation fallback to use normal symlinks under Node and real ESM proxy packages under pkg. Each proxy records the source package version, mirrors its explicit runtime subpath exports, and re-exports the virtual /snapshot URLs, so built-in Loader rows and external plugin peers resolve one shared Cordis/module instance from an on-disk profile.
Keep proxy healing idempotent, reject foreign real directories, cover root and subpath imports in packaged mode, and expand AggregateError startup diagnostics so concurrent Loader failures retain their individual import causes. This is the reusable packaged-profile mechanism; Python-specific artifact wiring remains in the next commit.
Delete @deepseek-ai/dsh-sdk-python-runtime and its packaged-bin entry now that Python uses the repository's dsh application launcher. Remove the package's project reference, Knip entry, workspace-policy exception, executable allowlist entry, and README-model classification together so no tooling preserves the old exception.
This commit is deliberately mechanical: it removes the obsolete package and gate accommodations without introducing the replacement launch behavior. The following commits add the packaged dsh runtime and Python profile API, keeping the architecture change separate from deletion noise.
The read-image-reencode lane feeds a 16-bit gradient PNG through the
shipped app: pass-through is impossible, the master converts down the
opaque JPEG ladder, and the 640,000-pixel request budget re-encodes a
downscaled request version — the projection the byte-identical tiny
fixtures never exercised.
Master dimensions move from a 2048 long-edge rule to a total-pixel
budget (normalizedImageMaxPixels, default 2048x2048) with an 8192
long-edge cap, so extreme aspect ratios keep short-edge resolution.
The shared quality ladder and lazy execution move to encoding.ts,
review-round doc fixes land across attachment and llm packages, and
the superseded facts in the unified-image-pipeline note now describe
the shipped routing.
Field test: the model received a bare '@niulai/' and guessed a user
mention — the guidance section was present but said only that @-prefixed
paths are files, never covering the trailing-slash directory form, the
workspace-relative root, or the quoted spelling. Rewrite the section to
name all three; the section's presence conditions and every consumer
pin the constant, so nothing else moves.
CI's snapshots lane timed out waiting for the folderx candidate: the
fixture directory was created after connectFreshWorkspace, racing the
Host's file index on a loaded runner. Land every fixture (and the
workspace directory itself) before the connect, and give the first
folder query a cold-start allowance.
A directory row in the @ menu had one verb doing two jobs: picking it
inserted literal @dir/ text and kept the menu open, so a user wanting
the folder itself never got a settled entity — the token kept its
trigger character and stayed editable, nothing like a file's atomic
chip.
Split the intents on the same row, mapped to shell-completion instincts:
row click / Enter settles the directory as an atomic folder chip (the
file chip's exact language; canonical @dir/ mention as its serialized
form — the { insert } arm the folder path never took), while Tab or the
row's trailing chevron drills: literal editable text, menu open on the
children. One new dimension carries it: candidate.drill advertises the
verb, InputTriggerPick.action reports it, ArbitrateKey gains 'tab', and
the keymap intercepts Tab only while a drill row is highlighted.
Covered by controller arbitration, MenuView chevron routing, the
ui-reference verb split, a keymap Tab-passthrough spec, and a real-
browser e2e driving all three gestures; menu golden refreshed for the
chevron and the fixture directory.
foldSurfaceTokens rebuilt the priced surface on every surface event: an
append allocated [...nodes, node] and a replacement copied the whole
array before splicing, charging every well-formed event O(surface) for
an atomicity property only malformed events need. Benchmarks put the
copy at ~99.9% of an append's cost (100µs at a 50k-node surface vs
0.1µs for pricing) with O(S²) accumulation over a session, inside the
synchronous session/event publication path.
Split the fold into the session core's planSurfaceEvent/applySurfacePlan
shape: planSurfaceTokens performs every fallible step against the
read-only surface, commitSurfaceTokens applies the plan in place and is
infallible by construction. _foldEvent plans first, runs the remaining
fallible anchor validation, and only then commits, so retry identity is
preserved by ordering instead of by allocation. Appends drop to
amortized O(1) (100.3µs -> 1.9µs at 50k nodes); replacements keep their
O(surface) findIndex but stop paying the extra full copy (21µs -> 4.2µs).
A new regression test pins the one hazard this introduces: an event
whose surface plan is valid but whose later anchor validation throws
must leave the priced surface and running total uncommitted across
repeated failures.
- The base-mounted projection cache's write-behind forces session-log
flushes at cache-chosen times, splitting packed chunk rows and collapsing
the persistence window the keyless acp transcripts pin. ACP exposes no
session-listing surface, so the automation profile disables the base row,
restoring the pre-base behavior the fixtures were recorded against.
- The windows coverage lane runs suites that read built lib/ output (the
webworker-packer image tests) and raced the build gate's tsdown writes
against a partial tree; every coverage gate now waits for the build gate.
Delete the 5-bit colour-count classifier and palette PNG branch that
misrouted high-frequency photographic JPEGs (issue #2885 images 23/24)
into an encoder 100x slower with 4x larger output. Both normalization
and request-image encoding now route by the decoded alpha fact alone:
opaque sources down a JPEG ladder and alpha sources down a WebP
effort-0 ladder, each at qualities 85/75/60. Byte budgets become ladder
targets: the downscale retry loop is gone and a ladder-exhausted encode
keeps its smallest output, while provider byte caps stay enforced at
the transmitting route. Request transforms move to request-image-v5.
The exact-shape fix added two filesystem-failure catch branches that only
fire on a race/permission fault the caller already guards against (the
session-dir readdir after an isDirectory() check, and the discovered-root
rmdir after the root was observed empty). Neither is deterministically
reproducible in-process, so tag both with the same reasoned v8 ignore the
sibling catch blocks already use, restoring per-file 100% coverage and the
symmetry between the parallel rmdir handlers.
Tighten the startup sweep to backend-generated name shapes and fix the tests
that had drifted from the SweepRoot-based API:
- Match roots by the exact `dsh-spill-<6>` mkdtemp shape and session dirs by
`session-<12 hex>` (DEFAULT_ROOT_RE / SESSION_DIR_RE), replacing loose
startsWith checks so foreign or fixture-shaped directories are never swept.
- Carry `SweepRoot { path, pruneWhenEmpty }` through SweepOptions so a discovered
prior-default root is removed once emptied while the active root is never
pruned; lstat each session entry so a symlinked session dir is not followed.
- Fix the tests to the SweepRoot API: import SweepRoot, correct the gatherRoots
override return shapes, build discovery fixtures with the real mkdtemp shape,
and route the warn-wiring test through a deterministic failure path.
The local spill backend never reclaimed its files, so configured roots
grew without bound and default per-process dsh-spill-* temp roots piled
up across runs. Immediate deletion is unsafe because persisted, resumed,
and forked sessions may still reference an older locator.
Add a fiber-owned, best-effort sweep that runs once after activation
(never delaying availability, awaited on disposal): it deletes regular
files older than cleanupPeriodDays (default 30; 0 disables) across the
configured root and prior default temp roots, prunes emptied dirs, and
skips symlinks/unknown entries. Every filesystem failure is contained
and logged, so the sweep cannot fail activation or a concurrent write.
preview-boot's hero wait was master's new textarea:enabled locator (the
conv refactor predates the Lexical composer); migrate it to the
data-composer-input surface like every other lane. Drop the unnecessary
DOMRect assertion oxlint flagged on the new Range stub.
- user-text.tsx enters ui-primitives' per-file 100% gate: replace three
regex-guaranteed impossible ?? fallbacks with asserted captures, make
the precedence sort a branch-free rank comparator, and pin the nested
recall-label ordering and the no-basename quoted-path fallback with
tests (100% statements/branches/functions locally).
- assembled-boot stubs Range.prototype.getBoundingClientRect: jsdom has
no Range geometry, and Lexical's selection reveal now reaches it in the
built-graph lane after the architecture merge (the unhandled TypeError
behind command-image-envelope and preview-boot).
The worker-preview pack landed on master declaring dsh.configTrees
against apps/cli/config/agent-presets, which this branch moved into
packages/preset/agent-presets/presets. The VFS mount and the worker-side
roster patch keep their paths; only the source directory follows the
move.
The spec landed on master before the roster gained the plugin-bundled
shipped root; its empty-roots harness now opts out explicitly, matching
every other exact-roster suite.
The Landlock packed-install rehearsal packed a hand-maintained list of
workspace tarballs. When dsh-llm gained the dsh-util-crypto runtime
dependency, the list stayed stale and npm tried to fetch the unpublished
release candidate from the public registry, failing both Linux master jobs
with E404 before confinement ran.
Read the current pnpm workspace inventory and traverse dependencies,
optionalDependencies, and required peerDependencies from the packed test
roots. Verify package identities, fail loudly on unresolved workspace names,
sort the closure deterministically, and leave native-family packages to the
existing mode-preserving native packer.
Cover runtime traversal, optional-peer exclusion, native filtering, and
invalid workspace metadata. Remove the obsolete vendoring exact edit for the
deleted manual list so future runtime workspace additions are included by
their manifests instead of becoming post-merge CI failures.
Unix PowerShell emits cursor-position requests while PSReadLine starts and
redraws prompts. The subprocess PTY is only a transport, so those requests
went unanswered. Startup could then accept the dsh> literal echoed from its
setup source as a rendered prompt, and later sends were lost or clipped.
Feed raw PTY output into a zero-scrollback @xterm/headless state machine and
write generated replies through the provider-owned terminal handle. Drain
replies before caller input, repeat foreground inspection when terminal
activity races the sample, and retain send ownership until parser and reply
work quiesce. Coalesce raw chunks behind one active parser write so large
Windows output cannot create thousands of queued parse callbacks.
Publish pwsh only from backend stdin_read evidence and start one timeoutMs
deadline before the complete startup retry loop, so inferred-idle follow-ups
cannot reset the bound. Dispose the emulator when the
terminal or cleanup fails. Document the fail-loud ConstrainedLanguage path
and add focused coverage for split queries, reply ordering, foreground
resampling, failure containment, batching, timeout, and disposal.
* test(python): exercise installed wheels as black boxes
Add an installed-wheel mode that refuses source/editable imports, repository working directories, mismatched SDK/runtime versions, unpinned runtime dependencies, and executables outside the installed runtime distribution. The mode resolves the wheel-owned executable itself, so callers cannot accidentally prove an explicit checkout artifact.
Add a real-API scenario that drives two tool-using turns through the public synchronous SDK, verifies the file bytes outside the agent, checks completed turn/tool events and persistence, and projects provider failures without retaining credential-bearing error text. The existing deterministic scenario set remains the keyless behavior oracle.
Refs #2952.
* ci(python): require installed-wheel checks on every release target
Move the complete deterministic runtime scenarios behind construction and clean installation of the SDK and matching runtime wheels. Each native leg runs outside the checkout with source-resolution environment variables removed; Linux manylinux smokes assert the same installed provenance.
Expand the required pull-request call from Linux x64 to Linux x64, Linux arm64, and macOS arm64. Trusted heads receive only DEEPSEEK_API_KEY_EXTERNAL for a fail-loud live two-turn smoke on each carrier, while fork and Dependabot heads retain the full keyless path without exposing secrets.
Pin the reusable secret declaration, matrix call, aggregate dependency, untrusted-head condition, and live/keyless commands in the workflow contract test.
Refs #2952.
* docs(testing): make installed wheels the Python CI authority
Record the clean-wheel provenance boundary, complete keyless scenario set, trusted real-API contract, secret handling, and three-target required topology in a new implemented testing decision. Update the SEA distribution and portable-CI authorities plus the Python contributor reference to describe the same current state.
Archive the fully superseded Linux-x64-only decision after consolidating its rationale and alternatives into the new owner. Preserve its bilingual triplet as a sealed historical snapshot and redirect every active current-state reference.
Refs #2952.
Move the shared module-reload policy into dsh-base by inserting its HMR row disabled, then remove the redundant disabled overrides from Web, headless, SDK, and ACP. No shipped profile enables server module reload; live profile patch watching continues through the launcher-owned config-only fallback, and browser client HMR remains a separate mechanism.
A later profile layer can opt into source-module reload explicitly with disabled: false while retaining the base root configuration. Composition tests cover every shipped mode and the explicit enable path, and the bundle references plus launcher Agent Notes document the resulting ownership and safety rationale.
Set DSH_E2E_MAX_WORKERS=4 for the credentialed e2e workflow and pin that environment contract in the workflow test. Profile-launched SDK and ACP scenarios each boot a complete subprocess tree, so the previous file-level fan-out could multiply process and provider pressure far beyond the runner's useful concurrency.
The bound changes scheduling only: every e2e file still runs, the Vitest configuration retains its explicit override knob, and local callers can choose a different positive worker count when their resources allow it.
Regenerate the configuration catalog and module graph after replacing standalone application packages with dsh profile bundles and renaming the private Python carrier. Update built-bin coverage to launch the shipped sdk and acp profiles, assert retired bins stay absent, and keep the Web golden text aligned with the same assembled runtime.
Clarify that headless is a startup profile whose patches freeze after boot. This commit is projection and verification work: it contains no application implementation, and every generated document is produced from source committed earlier in the series.
Update workspace manifests, the lockfile, Host project references, Knip inputs, package constraints, vendoring rewrites, and Python runtime build/smoke scripts for sdk-app, acp-app, and @deepseek-ai/dsh-sdk-python-runtime. Add the ACP hook packages to the dsh dependency closure so installed profile materialization resolves the same plugins as source workspaces.
Keep Python distribution outputs deliberately unchanged: the wheel modules, executable names, and smoke targets retain their public identities even though their private npm carrier moved. Constraint fixtures pin the new package locations and catch missing application dependencies on every platform.
Add verify-application-entrypoints to the top-level gate graph. It inventories executable sources and package bins across apps, packages, and examples; rejects unclassified launchers including root-level js/mjs/cjs/ts files; and permits only the dsh CLI plus the explicitly private Python runtime carrier exception.
Update repository, architecture, CLI, and naming records to state the same rule: Node consumers select a dsh profile instead of invoking application-package bins, and no compatibility aliases remain. Fixtures prove both allowed classifications and representative escape attempts, making the architectural rule mechanically enforceable.
Rename the relocated npm workspace to @deepseek-ai/dsh-sdk-python-runtime and the runnable example to python-sdk-agent, then update every owning English/Chinese document, test, and packaged-runtime reference. Remove the obsolete standalone bin while retaining the carrier entrypoints used to assemble the wheel runtime.
This is intentionally a naming and ownership change, not a Python SDK migration. The public deepseek_harness_sdk and deepseek_harness_runtime module, wheel, executable, environment, and wire behavior remain unchanged. Documentation records that this private carrier is the temporary sole non-dsh application exception and will move to profile launch later.
Regenerate the four TypeScript SDK replay scenarios after switching their subprocess to dsh --profile sdk. The fixtures now project profile-owned runtime context, tool composition, nested-child persistence, and the opt-in DeepSeek session-log acceptance event while preserving each scenario's final response and file assertions.
This commit contains only committed snapshot outputs. Separating them from the SDK implementation keeps protocol/API review focused and makes the model-visible consequences auditable as generated evidence.
Replace the TypeScript SDK's public arbitrary command/argv launch surface with the same-version dsh CLI, a named profile, ordered per-launch patches, optional process cwd, and explicit Harness home selection. Installed consumers use the built CLI; clean source checkouts use the package's src/bin.ts through an absolute tsx/esm loader and a source-only patch that omits build-generated Typert loading.
Materialize explicit or inherited environments at spawn time, keep profile-internal patches below caller patches, and resolve every caller-relative path before the child starts. The SDK subagent provider validates its optional CLI and patch files at plugin load and requires an isolated absolute child home.
Treat JSON-RPC initialize as the Loader-owned readiness point: Loader settlement joins entry imports, fiber lifecycle work, and synchronous effect registration, so the server needs no scheduler tick. A delayed profile entry registers a private adapter before initialize resolves, proving caller-supplied plugin routes are visible without fallback.
Update sdk-app ownership, server diagnostics, TypeScript SDK examples, nested-loader coverage, and session-upload replay layering together. The following commit contains only the regenerated SDK transcripts, keeping this API and lifecycle change directly reviewable.
Move examples/jsonrpc-agent to examples/python-sdk-agent and packages/examples/jsonrpc-demo to packages/sdk/python-runtime while preserving every file byte-for-byte. The directory placement now reflects the example's Python-distribution role and the private runtime carrier's SDK ownership.
This commit deliberately keeps the old package names, commands, configuration, and snapshots inside their new directories. All 42 files are 100% renames; API/profile migration and naming changes follow separately so reviewers do not have to disentangle behavior from filesystem movement.
Regenerate the ACP composition graph and committed replay outputs from the assembled dsh base plus acp-app profile. The updated logs, system prompts, tool schemas, and stdout projections capture the same scenarios after shared plugins move out of the retired demo package and into profile-owned composition.
There is no runtime source in this commit. Keeping generator-owned artifacts separate lets reviewers validate the behavioral migration first, then inspect expected wire/model-visible consequences as a mechanical projection update.
Replace the standalone @deepseek-ai/dsh-acp-demo application with dsh --profile acp plus ordered example patches. The shipped acp-app bundle owns only the protocol bridge; every example overlay now targets shared dsh-base rows instead of copying a complete application tree.
Move launcher responsibilities into the ACP snapshot harness: it materializes profile patches, links required packages, reserves stdout for JSON-RPC, observes spawn and drain failures, and escalates process teardown deterministically. The relocated control-surface fixture and the ACP/subagent integration tests now exercise the real CLI/profile path.
This commit contains authored runtime, configuration, and test changes only. Generated transcript and projection churn is deliberately left for the next commit so reviewers can inspect the migration logic without hundreds of expected-output edits.
Move the ACP control-surface e2e, scripted LLM, and Cordis fixture out of the application package that later commits retire and into the runnable examples/acp-agent test tree. This gives the assembled dsh profile example ownership of its integration fixture.
This commit is intentionally mechanical: all three files retain their exact blobs and appear as 100% renames. Imports and launch behavior are updated only in the subsequent ACP migration commit, keeping reviewer-visible code changes separate from file movement.
Introduce @deepseek-ai/dsh-acp-app as the thin application layer for the built-in acp profile. It contributes only the ACP protocol bridge and profile metadata; dsh-base remains the single owner of shared agent composition, providers, persistence, permissions, and tools.
Wire the bundle into CLI resolution, catalogs, workspace configuration, and built-bin coverage. The focused bundle and startup tests prove that base plus acp-app exposes automation sessions while keeping stdout reserved for ACP JSON-RPC.
Introduce @deepseek-ai/dsh-sdk-app as the thin application layer for the built-in sdk profile. The bundle contributes the JSON-RPC server and startup-only profile metadata, while dsh-base continues to own the shared agent, provider, persistence, and tool composition.
Publish ctx.appReady from the launcher only after the Loader tree and launcher-owned setup succeed. The stdio lifetime binding leaves stdin unread until the protocol transport claims it and defers EOF exit 0 until readiness commits, so early protocol frames remain buffered and a racing startup failure remains the nonzero process outcome. Fiber disposal cancels both pending lifecycle listeners.
Register the bundle in the CLI resolver closure, generated configuration catalog, workspace graph, and built-bin smoke. Startup tests prove that base plus sdk-app exposes the SDK server without taking ownership of shared runtime plugins; focused and built-bin regressions cover early input, EOF readiness, and startup-error precedence.
Add a patchReload field to built-in profile metadata and carry it through CLI profile resolution into app boot. Live profiles install the existing patch watcher; startup profiles freeze every layer after boot and apply later edits only on the next launch. Missing or invalid metadata fails before the plugin tree starts.
The implementation keeps reload policy with the profile that owns it instead of inferring behavior from an entrypoint. Unit tests cover metadata validation and both lifecycle modes, while the CLI and app-boot references document which built-ins are live versus startup.
Record the final architecture before any runtime or file-layout changes. The decision makes named dsh profiles the only supported Node application launch path, gives TypeScript SDK callers ordered profile patches for per-launch customization, and retains the packaged Python runtime as an explicitly temporary exception with a later migration obligation.
Keeping this decision in a documentation-only commit gives every following commit one stable naming, lifecycle, and compatibility reference. The English and Chinese notes and their pairing record enter together.
- stop attributing the 45-case helper/Node divergence to the path port spec,
which pins only the Node-facing port to Node
- keep the divergence measurement as provenance, without a dead owner
- fix the path corpus comment so it does not claim divergence from the spec
that asserts equality
- drop the v2 generation stamp from the cordis mount fallback test
- restate the watcher refusal rationale in current-state terms
- re-record the ui-workspace bilingual pair after rebase
Remove dead design-session citations, change narration, indexical
stamps, and review-adjacent justification found by the
dsh-trim-cot-leakage recall batteries in prose that landed after the
last purge. Bilingual README pairs are re-recorded.
Config catalog, module graph, event producer-consumer tables, and
third-party notices regenerate over the webworker surface; the oxlint
rule fingerprint and the ui-renderer NodeNext import face follow.
One Vite build emits dist/index.html and dist/preview.html sharing every
chunk; the only difference is one prepended bootstrap entry whose module
connects the worker host, so the page from the stock entry onward is the
served startup chain verbatim. The dist moves to a relative base so the
preview mounts under any static directory, and the served form anchors
deep SPA-fallback paths with a rendered <base href="/">. The preview-boot
e2e serves the real built pages, packs the VFS image when absent, and
holds the boot line's lowering contract, the interactive hero, and a
clean page-error channel in headless Chromium.
The webserver renders a boot-readiness tail after the injection rows and
AppWebEntry.run awaits the __DSH_BOOT_READY__ deferred before reading any
injected state. Whichever bootstrap applies the injection table settles
the deferred - the served renderer resolves it inline, an asynchronous
bootstrap installs it ahead of the entry module and settles it with the
handshake - so both deployments run one startup chain and a failed
handshake surfaces on the boot page instead of proceeding on missing
globals.
Buy the grammar, own the execution: @yarnpkg/parsers parses the command
line - aliased at bundle time to its shell entry so the root barrel's
syml/js-yaml closure stays out of the worker - and a VFS-backed evaluator
with a coreutils command table runs it inside the worker host. Each shell
process is a real child WebWorker spawned from the same bundle (the first
frame decides the role), so the TERM-then-KILL ladder is real - TERM
requests, KILL terminates the worker - and the file face stays
asynchronous end to end, since the deployment target serves no COOP/COEP
headers and SharedArrayBuffer never exists there. node:child_process
reports through the ChildProcess surface the subprocess service consumes;
execSync, execFileSync and fork refuse, and node-pty stays stubbed.
Two private experimental packages run the whole harness tree inside one
dedicated Web Worker. dsh-experimental-webworker-runtime owns the in-memory
VFS (BigInt stats with per-path identity and strictly increasing mtimes),
the CommonJS wrapper loader over a lazily-evaluated builtin table whose
shims typecheck against Node's own module types, the postMessage tunnel
speaking plain HTTP, the AsyncLocalStorage runtime, and the worker
assembly. dsh-experimental-webworker-packer lowers every module body at
pack time against the shared wrapper contract, sweeps the profile closure
by static reachability, and writes a deterministically gzip-compressed tar
the worker inflates through the browser's native DecompressionStream while
it downloads.
The old backdrop overpainted a folder reference's trigger character with
IconFolderClose16; the Lexical rewrite kept the data-ref-appearance
attribute but nothing consumed it, so the composer showed a bare blue
token while the sent bubble carried the icon. A Lexical text node cannot
split out its trigger character for overpainting, so the glyph renders
as an icon prefix instead: a currentcolor mask of the same asset before
the intact literal token. The Lexical note's behavior line follows.
Keep the active implemented and proposed Agent Notes current with the
renamed ToolCallId brand: branded-ids, content-block-vocabulary,
approval-seam, tool-output-spill-files, and task-surface, in both
languages, with re-recorded .i18n.yaml pairing records.
Review-round hardening of canonical admission:
- passthrough now requires a single-frame source free of EXIF/XMP/IPTC
metadata, so location/device metadata never enters durable storage and
stored dimensions always describe the perceived pixels; animated WebP joins
GIF on the always-re-encode path (first frame only)
- SourceImageInfo records orientation-applied dimensions, keeping source and
stored raster on shared axes for coordinate mapping
- validateImage runs a canonical-encoding dry run, so a validated batch can no
longer be refused mid-write by the byte target (no partial writes)
- read_image names per-axis multipliers when rounding splits the two ratios
and maps IMAGE_TOO_LARGE to actionable downscale guidance
READMEs (both languages) describe the wide source envelope, the canonical
encoding and its fixed encoder parameters, and read_image's downscale
envelope; tool/config catalogs regenerate for the new schema and Config
fields. The read-image-dimension scenario now pins the acceptance the old
2000px admission cap refused: the 2001x1 source is admitted and stored
byte-identically, so the fixture stays platform-independent.
The canonical ref serializes mediaType, width, height, bytes in the order the
pre-canonicalization store used, so existing session-log fixtures and logged
histories keep byte-identical reference JSON.
When the attachment store's canonical encoding shrinks the file on disk, the
read_image envelope names the original dimensions and the multiplier that
maps coordinates measured on the attached image back onto the file, and the
output schema carries sourceWidth/sourceHeight for programmatic callers.
Admission now validates a wide source envelope (32MiB, 100MP, 16384px per
side) and persists a canonical encoding instead of refusing large sources:
EXIF orientation baked in, metadata stripped, long edge downscaled to the
configured canonical target (default 2048px), PNG palette for alpha/PNG/GIF
sources and a fixed JPEG quality ladder (85/75/60/45) until the canonical
byte target holds (default 1MiB). In-budget PNG/JPEG/WebP passes through
byte-identically so equal sources keep deduplicating to the same content
address; GIF always re-encodes to the PNG of its first frame, pinning the
first-frame meaning providers apply. Encoder parameters are fixed by design;
only the canonical budget is deployment configuration.
AttachmentStore.saveImage now resolves SavedImageAttachment: the durable
reference paired with the submitted raster's intrinsic facts, so a store may
persist a canonical re-encoding while callers keep the source dimensions for
coordinate mapping. saveImages keeps returning refs; every fake store and the
cordis API catalog follow the new signature.
The user-bubble decorator rendered every plain run through the
block-level MessageText div, so a decorated single-line message broke
into one line per run and the space between two tokens rendered as a
blank line. The queue dock's read-only row printed row.preview verbatim,
showing the wire session form (@[label](dsh-session:...)) instead of a
readable label. Both predate the Lexical composer; the logged model text
was correct in both cases.
One shared inline projection (reference/user-text.tsx) now owns sent
user text for the bubble and the queue row: plain runs are spans with
white-space policy left to the consumer (bubble pre-wrap, queue nowrap),
and a highest-precedence rule folds the wire session form to its label
chip, shielding the URI from the bare-token scan. The queue edit field
keeps the literal sent text. user-text.client.spec pins the inline
guarantee and every fold rule; queue-actions.e2e locators move to
row-container matching (the projection adds one span layer).
Exposed by this branch touching rescope-vendor.ts, which runs the full
rescope check: the knip-logger-console exact edit targeted the
packages/util/home knip section that #2758 deleted (drop the edit), the
zh vendoring-cookbook anchor predates the rescope.zh.md link
localization (follow it), and the new shipped-root.spec.ts joins the
files whose bare 'cordis' tokens are preset ids.
The files constraint tables gained per-package expectations on master
while this branch changed two files lists: apps/cli no longer ships
config/, and dsh-agent-presets ships presets/ (ordered where the
expected-files derivation places extras).
The documentation site deployed on every master push, with no reviewer and
no version check, while npm, PyPI, and the public source repository all
advance only at a release tag. The Pages site is reachable without
authentication, so a merge published documentation ahead of every artifact
readers could obtain.
docs-pages.yml now declares workflow_dispatch alone and verifies the ref
through the gate npm publication already runs, so the site and the npm
sequence share one definition of a released version.
Review asked why the launcher special-cases one plugin's row. It no
longer does: the four shipped compositions move into the package
(presets/, in files), dsh-agent-presets resolves its own shipped root
and prepends it before configured roots (includeShippedRoot, default
true, opt-out for bare-machinery embedders), and the per-composition
derived patch, its spec, and the dump layer are deleted — profile-boot
and dump-config return to plain layer stacking. The always-load
guarantee now rides the schema default instead of patch ordering, so a
whole-config replacement keeps the shipped set and the squash, reload
freeze, and dump divergence stop being possible.
Gate globs, the web scaffold, and both preset browser lanes drop their
hand-fed shipped roots; the roster e2e keeps asserting configured roots
beside the shipped four against the built lib.
Fixes#2863.
The Windows native coverage lane was hitting Vitest's 30s per-test
ceiling on slow subprocess/ACP fixtures. Doubling the per-test budget
absorbs cold-start and process-teardown jitter without changing the
assertions or the job-level 120-minute cap.
python-release.yml only triggers on workflow_dispatch, so build.if:
github.event_name == 'workflow_dispatch' is always true and redundant; remove it
(the exact event set is already pinned in the spec). Update the spec assertion
accordingly.
Address PR #2875 review:
- Re-record python/development.i18n.yaml (corpus verify-translation-pairing was
out of sync after editing development.md/zh.md) and the 2026-08-11
python-publication-workflow pair after the dry-run wording tweak.
- Tighten the python-release spec assertion to the exact event set
(['workflow_dispatch']) instead of not.toHaveProperty('pull_request').
- Fix the 'dry-run run' wording in development.md and the note.
Corpus-wide verify-translation-pairing (1001 pairs) and note-format (594) pass;
ci-workflow.spec.ts 14/14.
The static gate (translation pairing) failed on a pre-existing master note:
2026-08-20-composer-edit-range-from-selection.zh.md:17 linked the zh target with
the en .md path. Point it at the .zh.md target and re-record the i18n hash. This
unblocks the required node 24 / static gate (it is not part of the python-release
gray-check change but sits on the same PR's CI path).
An intermediate cache iteration made cachedSnapshot async, which forced
listProjectionsFor onto a multi-line async signature; the final design
kept the sync read but the formatting residue stayed. No behavior
change.
The per-record contract test forced readFile to fail via chmod 0o000,
which is a no-op on win32, so the readRecord catch stayed uncovered on
the windows native coverage gate. Route record documents without an
isFile pre-filter: a directory where the document should be throws
EISDIR on every platform, and readRecord's existing contract already
reads an unreadable document as absent. Drop the win32 skip.
Review follow-ups: enumerate the derived shipped agent-preset root in
apps/cli README/reference dumps and the profile-boot module JSDoc
(bilingual pairs re-recorded), correct the stale AppCLIEntry/distIndex
analogy in the web scaffold and the shipped-root cross-reference in the
web preset e2e, drop the write-only ComposedProfile.rows field, and make
the Agent Note describe the dump path as sharing the derivation rather
than the builder.
Master's #2844 asserted composer recovery through a textarea locator the
Lexical composer no longer renders; the merge carried it in silently and
the CI web-snapshot lane timed out waiting for it. Use the same
data-composer-input wait the file's other cases already migrated to.
The master merge brought the doc-sync regenerations (locale-specific
paired document paths). Re-apply the zh-side link fixes to the agent
note and the session-projection intro, re-record translation pairing,
and keep the e2e event arrays as asserted literals like master.
The per-record rm in the cache spec hits an EPERM flake on windows
native (directory still draining); retry the recursive rm like the
subagent spec does. The unreadable-record probe is meaningless on
windows, where chmod 0o000 is a no-op; skip it there.
The reworked cache moved from the awaited coldSnapshot(id) to a sync
coldSnapshot(meta, events) whose write-back is fire-and-forget, and the
e2e fixture dropped the seeding call with it. The web bundle mounts the
cache again, so list rows read projection columns from stored rows only;
the one-shot and grandchild fixtures therefore lost their projections
column and the tree golden drifted. Re-seed both through the new API and
poll for the write-back to land.
Remove the pull_request:[labeled] trigger from python-release.yml so the
workflow no longer fires (and shows a gray skipped check) when a PR gets any
non-dry-run label. The credential-free dry-run validation is now manual-only
(workflow_dispatch with publish=false), preserving the validation capability
without a PR gray segment.
- python-release.yml: on is workflow_dispatch only; build.if is
github.event_name == 'workflow_dispatch'.
- ci-workflow.spec.ts: assert python-release has no pull_request event and the
simplified build.if.
- python/development.(md,zh.md) and 2026-08-11-python-publication-workflow note
(en/zh/i18n): describe the manual dispatch-only dry-run path.
Verification: ci-workflow.spec.ts 14/14, typecheck clean, note-format 585,
verify-translation-pairing consistent.
The boot-time agent-presets overlay replaced the composed roots with the
shipped root alone, so roots configured in a profile's cordis.patch.yml
vanished from the roster (externally reported in
deepseek-ai/deepseek-harness#3636). The overlay also froze the row's
boot-time config above every live reload and never reached the config
dump, which therefore showed roots the boot dropped.
Derive the roster patch from the current layers instead: prepend the
shipped root (system trust, wins duplicate ids) to configured roots,
share one builder across boot, live user-layer reloads, and
--dump-config, and fail loud on a roots value the launcher cannot
statically rewrite.
Fixes#2863.
isKeyboardSelectable() defaulted to true, so an arrow key at a chip edge
created a NodeSelection whose DOM projection collapses to an element
point; the plain-text binding's arrow/delete/insert handlers all bail on
non-Range selections, deadlocking arrows, typing, and Backspace at the
chip until a pointer click. False restores the placeholder semantics:
arrows cross the chip in one move and Backspace/Delete remove it whole.
Reproduced and verified with real-key Playwright probes (CDP raw
keydowns carry no engine default and cannot reproduce it); the
reference-composer e2e pins the gesture in the browser lane.
The localized-link pairing rule that landed on master requires Chinese
notes to link Chinese counterparts; fix this PR's note and the
composer-edit-range note the rule landed after, so the merge ref
passes the gate. The full-manifest closure walk gets an explicit
timeout: coverage instrumentation on a loaded runner pushes it past
vitest's 5s default. Refs #2846.
- drop the space-key debug probe from the production keymap
- give pastes their own undo boundary (PASTE_TAG via $addUpdateTag on the
nested dispatch path), with an input-bar regression test
- claim decoration outranks text-ref entities on the leading-token seat:
the entity transform skips the active claim token, restoring the warn
color for lexicon-listed command names (probe-confirmed regression test)
- caret-only commits no longer advance draftRev or re-publish InputState;
content changes still do (snapshot-built CAS spans stay valid)
- retire stale JSDoc/contract wording (paste-upgrade, set-invalid; the
invalid bit only promises the render treatment)
- rename the keydown probe spec to keymap-routing and drop test dead code
- reference-composer e2e gains the #2813 type-ahead-of-chip gesture
- archive three superseded composer notes (Safari soft-wrap, text layers,
decoration keys), rewrite the input-machine note's superseded half in
place, and record the new behavior decisions in the Lexical note
Each index route also emits a parent-level alias twin, projected over
the alias route so its relative links stay correct; llms.txt and the
docs now state the drop-trailing-slash convention exactly. Frontmatter
failures name their page, the twin pass refuses to overwrite existing
build files, the dev middleware documents the deliberate in-page
fetch() divergence, and the per-locale collection order moves to one
shared export. Refs #2846.
Append .md to any published route for the page as plain Markdown: the
build emits a raw-Markdown twin of every route (frontmatter dropped,
home bodies kept, images beside pages) plus a manifest-generated
llms.txt, the dev server serves both per request, and the post-build
gate fails when either is missing. Fixes#2846.
Review follow-up: the consumer-facing contract still described the removed
suppression. TurnErrorNode's JSDoc, the client-runtime retry-projection
README section, and the ui-conversation chat-flow section (both languages)
now state that a terminal turn/end error always projects the node, beside
the settled retry chain when retries exhausted, and that only intermediate
failures that scheduled another retry stay retry-notice-only. The recovery
note sheds its remaining cross-turn wording (closed-step boundary, retry
turns, closed failed step/turn) for the same-turn reality.
Every consecutive composer send waits out the submit round-trip's
read-only span; the end-of-document caret gesture becomes select-all +
ArrowRight (Cmd/Ctrl+End moves no caret in mac contenteditable), which
lets Lexical's own ancestor scroll walk prove the typing reveal.
textContent/data-placeholder probes replace inputValue/placeholder reads,
evaluate-string selectors move to the composer anchor, queued fills wait
out the submit round-trip's read-only span, and the refreshed aria goldens
drop the hover tooltip the old interaction order happened to capture.
Paste-command text entry (awaiting its microtask commit), data-placeholder
lookup, a scrollIntoView stub for the menu the settled-caret re-track now
opens, and a trailing separator on the bare /plan paste so Enter submits
instead of picking from that menu.
The client-bundle preset pins the production/development exports condition
(lexical's node-condition file selects its flavor with a top-level await a
CJS bundle cannot carry); the composer carries an explicit aria-label (a
div's data-placeholder does not name it the way a textarea placeholder
did); READMEs, the composer.bar slot doc, and the Agent Note record the
editor architecture; web e2e drives the contenteditable surface.
The turn-error Definition suppressed its node permanently once the owning
turn carried any llm/retry event — a rule from the retired model where a
retry opened a new numbered turn. Retries now run inside the failing turn,
so the suppression hid exactly the exhausted terminal failure it existed
to defer to: spending every transient retry left the conversation with a
neutral collapsed retry row and no error row at all.
Delete the suppression: turn-error matches only turn/start and error-reason
turn/end, and renders whenever its turn recorded a terminal error; the
settled retry chain renders beside it through the separate model-retry
node. The Definition suite now asserts the exhausted-retry error node in
full history, tail-only windows, and after prepending the chain, and a new
keyless live e2e scenario exhausts a scenario-owned two-retry policy and
pins the terminal error row beside the settled retry row (the scaffold
gains replayRetryPolicy so exhaustion runs in milliseconds).
Both stale notes are corrected to the same-turn retry reality, and the new
bug-fix note owns the removal rationale.
Drive keyboard gestures as real KeyboardEvents at the contenteditable
(Lexical routes them through the command layer), write drafts through the
shell, and probe decorations at their new DOM (chip decorators, styled
claim leaf, hint CSS variable, text-ref entity nodes). jsdom lacks
Selection.modify, so the Backspace-deletes-chip gesture moves to the
browser lane.
The editor (shell-owned, per-session) is the draft + chip truth; the
machine slims to the submit plane. Chips are atomic decorator nodes with
NodeKey identity; TokenSpan coordinates ride the detect projection (chip =
one U+FFFC), persistence and InputState.draft ride the clipboard
projection. The mirror/backdrop layers, Safari soft-wrap repair, manual
undo log, boundary occurrence deletion, and clipboard expansion all
retire; the producerless paste-attempt and set-invalid planes go with
them.
Internal builds before the versioned layout wrote .credentials.yaml as a flat
mapping; refusing it outright would strand every key stored through the Models
page and fail the next model request. Boot now recognizes exactly that layout
- addressable names over non-empty string scalars, no directives - and
rewrites it under the writer lock, nesting the original lines verbatim under
refs: with the values byte for byte unchanged. Everything the recognizer
declines keeps the loud by-name refusal, a live reload still never migrates,
and the parser continues to read exactly one layout. The migration step
retires with the pre-release stance at the first tagged release.
os.homedir() reads HOME on POSIX and USERPROFILE on Windows; stubbing
only HOME left the Windows lane expanding ~ into the real profile
directory and failing the fileExists assertions.
- dsh-authorization's manifest becomes a release member on current
master's terms: version matches the root 0.1.0-rc.6 and
publishConfig.access is public, which the constraints gate and the
tarball pack's version verify both enforce after the master merge.
- Regenerate docs/module-graph (zh mirror included) for the
authorization package and the llm-pi-ai -> authorization edge; the
graph gate lives outside doc-sync and was never regenerated when the
package was added.
- The built-bin smoke seeds the versioned credentials document; this
branch's provider refuses the pre-release flat layout by design, and
the master-side test still wrote the old shape.
Review findings on #2509, all confirmed:
- Every writer of .credentials.yaml now waits out the record-mutation
lock (DOCUMENT_LOCK_WAIT_MS): refs and records share one file and one
lock, so a reference write or record delete contending with an OAuth
refresh must not fail at the 2s file-work default.
- api-key records are admitted before they are rendered: an empty key,
a non-POSIX env name, or an empty env value is refused at the write
instead of persisting a document the next boot rejects wholesale.
- llm-pi-ai no longer lets the credential-key grammar reject legal
route ids: reads answer "nothing stored" via isCredentialKeySegment
(new dsh-credentials export), deletes have nothing to remove, and only
a write refuses, as LlmError UNSTORABLE_PROVIDER_ID; flow registration
skips a future catalog id outside the grammar instead of failing the
mount.
- authorization/settled fans out with contained listener failures on
the credentials seam's terms (INVARIANT still rethrows), so a broken
watcher can never turn a finished attempt into a failure.
- notify() is fire-and-forget at the seam: a surface that cannot render
a notice loses the notice, never the attempt.
- A declined prompt is an outcome: interactions reject with the new
AuthorizationDeclinedError and the attempt settles cancelled instead
of failed.
- NOT_COMMITTED now confirms a commit observed during the attempt
(credentials/record-updated for the flow's key), so a re-auth cannot
pass a stale record off as fresh; a flow that deletes its record is
refused on the same code.
READMEs, the subsystem/event/config catalogs, and the Agent Note follow
the shipped behavior; memory.ts carries the dedup TODO.
Why the record union is one step more abstract than pi-ai's credential
in exactly two places, why a record key names the owning plugin rather
than the provider, why the flow owns the write, and why the interaction
travels with the request instead of a registry. Cross-links the release
fix it supersedes, and states what is still missing: the wire contract
and the Models-page control that would let a human start a login.
The two web e2e goldens regain exactly the openai-codex option line
they lost when the provider was withheld.
pi-ai's auth model reaches this adapter through three translations, all
of which live here: a CredentialStore over the harness credential
records, an AuthContext over the credential plane and the host
filesystem, and one authorization flow per installed provider that
ships a login. The seams they consume name nothing from pi-ai, so a
second adapter family can arrive with a different auth model and share
them.
Every collection is now built with the store and the context rather
than with nothing, which is what makes a signed-in provider stay signed
in across the collection rebuild a configuration change causes. With a
posture that works, the configurable-provider directory no longer
withholds OAuth-only routes and `openai-codex` is offered again; the
predicate that withheld it is gone.
The credential plane stays optional. Reads answer "nothing stored"
without a credentials service because such a composition genuinely
holds no credential, while writes refuse by name — a login whose grant
evaporated would report success and then fail every request. Flow
registration is scoped to the authorization seam, so a headless or ACP
composition mounts with no sign-in and everything else unchanged.
Two fixes found while wiring this up: pre-release credential fixtures
in the llm suites still used the flat document the record work
replaced, and a flow that ignores its cancellation signal would have
held its key for the life of the process — withdrawal now settles the
attempt either way.
Some credentials cannot be configured, only obtained: getting one means
a conversation — open this page, paste that code, pick an account. The
new seam owns that conversation and the one-attempt-per-key lifecycle,
and never the protocol, so a second authorization protocol arrives as
another flow rather than as another seam.
A flow is registered under the CredentialKey it writes, which is also
how the seam knows which plugin answers for the format inside that
record. The flow owns the write: run() resolving means the record is
already committed through ctx.credentials, and the seam confirms it.
That keeps a library persisting through its own store adapter the
single writer instead of being copied back out and written twice.
The interaction travels with the request rather than a registry,
because whoever starts an authorization is the one who can talk to the
human about it. A request already withdrawn never claims the key and
never starts the flow — relying on each flow to check its signal before
the first await would let one that does not hang holding the key.
The seam answered one question — what is behind this environment-variable
name — and that shape cannot hold what an authorization grant is: a
multi-field, rotating value keyed by a provider id rather than by a POSIX
identifier. The Models page already works around the gap by inventing a
synthetic environment name (`MINIMAX_CN_API_KEY`) for a route the user added
by hand, because the store's key must look like one.
`CredentialKey` is `<scope>/<id>`, where the scope is the owning plugin's
registered name. The owner is in the key because a `grant` payload is written
in its owner's format: two plugins serving the same provider name would
otherwise read each other's payload, and a record left by an uninstalled
plugin could not be told from a live one. The `/` also keeps the grammar
disjoint from `CredentialRef`, so the key spaces cannot collide.
`CredentialRecord` is `api-key` (key and/or provider environment values) or
`grant` (an opaque, owner-owned payload). The asymmetry is deliberate: an api
key is the harness's own data, a grant is a package it carries for someone
else. `modifyRecord` is the only write path because a correct write depends
on the current value — a token refresh is read-decide-replace under one
cross-process lock, without which two processes rotating one refresh token
lose whichever wrote first.
`.credentials.yaml` becomes a versioned two-section document. The pre-release
flat layout is refused by name, with the entry count and the one edit needed,
rather than read as an empty store — which would surface as an authentication
failure on the first request instead of at load. A grant payload is admitted
in both directions, so a value the document could not read back exactly as
written is refused rather than stored lossily.
How long a contender waits is a property of the operation the lock holder
runs, not of the write protocol. The 2s default was sized for the
render-and-rename cycle every call site had; a credential mutation that
refreshes an expired token performs a network round trip while holding the
lock, and leaving the default in place would fail every other writer of that
file for the duration.
`withFileLock` takes an optional `waitMs`; the retry cadence stays fixed
because it governs how often a contender asks, which no caller varies. Every
existing call site keeps the default.
The two standalone unit classes deliberately mirror the KvUnit drain/guard
lifecycle (close + assertOpen); mark the block with a reason-carrying
jscpd ignore so the duplication gate stays green.
Regenerate config/cordis catalogs and doc graphs (the master merge changed
configs and consumers), add the DomainSpec layout field to the storage
type-equiv block, record session-projection-cache as a session/created
consumer, and drop the leftover experimental/team ghost directories from
the master rename. All 37 static gates pass.
Keep the base class's relative method order (write before coldSnapshot) so
the diff against the base shows the cold-read methods as a pure insertion
instead of a reorder of existing methods.
A caret Backspace or Delete replaces no selection, so `beforeinput` reports
the bare caret and the previous derivation produced a negative inserted
length and fell back to the ambiguous scan. The delete half of the defect
survived, and the component test missed it by pre-expanding the selection to
the span the engines never report.
The range for a caret delete now comes from the direction `inputType` names
and the number of characters the draft actually lost, measured rather than
assumed to be one, so a grapheme, word, or line deletion sizes correctly.
Only the insert and delete families are recorded; a history replay reports
wherever the caret sits and would name a wrong span while passing every check.
Component tests cover the caret Backspace, Delete, and word-delete gestures,
and an assembled browser scenario drives them as real key presses, which is
the only place an engine's reported range is observable.
The card-wide `border-box` reached `.fieldMirror`, so the growth cap counted
text plus padding. The inline variant carries none and landed on the declared
line count, while the optionless variant's 16px inset spent two thirds of a
line and delivered its last one as an 8px sliver. The e2e measured only the
inline shape, so nothing caught the drift.
The mirror now takes `box-sizing: content-box`, which states the cap in the
units it is written in, and the cap moves to six lines. The e2e asserts whole
text lines rather than a box height, and covers the optionless shape — asked
straight through the user-questions seam, since a layout metric needs no model
round — including that the reserved empty field is fully covered by its own
control.
A session that never talks — a forked child seeded with its ancestor's
title, say — previously got its first cache row only at detach; a crash,
or a fork held live in the store, left the seed-derived values (the
title) unreadable on the cold list. Session creation is now a third
mandatory write point: the creation checkpoint folds the seed and
persists immediately. Write-policy docs (README + catalogs) updated.
Opening a per-record unit splits a legacy `<root>/<name>.json` (the
pre-per-record single-file layout) into per-record documents; an
already-present new record wins, and the legacy file is deleted once
every record migrated. The migration also runs when the new tree is
absent — the fresh-upgrade shape — and foreign, shapeless, or malformed
legacy files are left alone. This preserves previously cached session
titles, list metadata, stats, and subagent identity across the medium
change.
A detached history read still traverses the complete log, but each unit's
fold is now seeded from its cached checkpoint: the registry's restore
slices off the already-folded prefix (events at or below the row's seq)
and applies only the tail. The first cold read writes the refreshed
checkpoint back (fail-soft), so the cache row is created on first read
and kept current afterwards. The recipe lives on the cache
(cachedCheckpoint, coldSnapshot, writeBack); the api-proxy carrier only
supplies the stored header and the full log.
Replace per-plugin tapIndex regex edits with pure-data IndexInjection rows
collected fresh per render over one webserver/index-inject event. One table,
two renderers: the served form renders rows into index.html; a static worker
form ships the same rows over its boot payload. tapIndex survives as the
raw-HTML escape hatch, applied after row rendering; client-modules and
ui-theme move to the event, and the manifest global renders as
globalThis["__DSH_BOOT__"].
The client boot chain gains the seams a pre-injected transport needs: the
module loader takes loadBundle from the transport global by default, HTTP
prefetch stands down when a transport owns bundle bytes, the web-app bundle
can decline frontend serving, the gateway client installs a namespace's
whole method group inside its fiber apply so a parked dependent never
observes the service without its methods, and the dynamic-code precheck
gates through new Function so hosts without a real node:vm keep the
define-time parse gate.
The composition tests hardcoded shared /tmp/dsh-acp-demo-* persistence
roots. On the shared self-hosted CI VM, concurrent jobs running the same
master-level test file opened the same SQLite session-query.db and one
failed with "database is locked". Use mkdtemp-unique directories so each
run owns its SQLite index while still exercising explicit
persistenceRoot forwarding.
Address the latest review pass on PR #2798:
- client-build-environment.client.spec.ts: add release-publish.yml to
dshBuildWorkflows so the 'workflow env must not set DSH_CLIENT_*' gate covers
the new dsh publish path (it runs build:official and writes a dsh client
build record). vendor-publish runs only build:lib:host, so it is not added.
- 2026-08-10-npm-release-sequences note (en/zh/i18n): the Release-publish group
is carried by the publish job (job-level concurrency), not the whole
workflow; corrected the wording.
Address the fresh review pass on PR #2798:
- ci-workflow.spec.ts: restore the subscription-type assertions the rewrite had
dropped (issue-lifecycle pull_request types omit ready_for_review and include
review_requested; pull_request_review types == ['submitted']) alongside the
new step-level gate checks, so the 'tests pin the subscribed events' gate
holds.
- 2026-08-10-npm-release-sequences note (en/zh/i18n): the three-sequence table's
Workflow column now lists the pack and publish workflows for dsh/vendor, and
line 114 no longer claims release.yml publishes (the dsh pack job packs the
vendored family for verification; release-publish.yml repacks and publishes).
Follow-up for the pack-job copy drift is filed as #2816.
Ordinary typing supplied no edit range, so the machine recovered one by
scanning the two drafts for a common prefix and suffix. That reading is
ambiguous whenever the inserted text repeats what it lands against, and the
greedy scan always slides the edit as late as the characters allow. A
reference renders as '@' plus its label, so typing '@' immediately before one
reads as an insertion inside the reference; reconcile then applies its
intersect rule and drops the occurrence. Submission takes the occurrence-free
path and sends the human-facing label instead of the owner's model form, past
the serialization guard that only fires when an occurrence survives.
InputBar now records the textarea's selection during beforeinput and passes
the resulting range to setDraft, which the machine already prefers over its
own scan. A record that cannot describe the reported change yields no range
and the scan still runs.
The pending-question registry holds the awaiting tool call's own
resolve/reject, so it is host-process memory. events.mux replays every
still-pending question on reopen, which covers a browser reload or a
reconnect, but a host restart takes the awaiting turn with it and the
reopened Session offers no composer for that question.
A question that carried options collected its free-text answer in a
single-line input: a long sentence scrolled sideways inside one line and
Shift+Enter was inert, so an answer with structure could not be typed.
The optionless question already used a textarea, but a fixed 64-140px box
that never followed the draft.
Both shapes now answer into one AnswerField: a `textarea rows=1` sharing a
grid cell with a hidden mirror that renders the draft plus a trailing
newline and so owns the height. Soft wraps are invisible to a '\n' count,
so the mirror is what grows the box; growth stops at eight lines and the
textarea scrolls from there, keeping the choices the answer belongs to in
view. Enter still continues and submits, Shift+Enter breaks the line, and
the IME guard is unchanged.
Address ds-review-bot findings on PR #2798:
- release-publish.yml: use pnpm run build:official (not build) so the dsh
pack step's verifyBuildArtifacts (families.ts:327, readClientBuildRecord with
officialClientBuildEnvironment) finds the official client-build record; build
would fail Pack release tarballs on a clean runner.
- issue-lifecycle.yml: move the previous job-level if to step level on
Create project token and Handle repository event, so approved/commented
reviews pass (job reported success, no gray segment) without minting a
write-capable App token or touching the board — preserving the original
least-privilege property.
- ci-workflow.spec.ts: lock the step-level gate on the two lifecycle steps, and
add a release-workflow invariant test (release.yml/vendor are pack-only;
release-publish.yml/vendor-publish.yml are workflow_dispatch-only with the
npm-publish environment and Release-publish group) to prevent #2797 recurrence.
- Update 2026-08-10-event-directed-pr-review-status and 2026-08-10-npm-release-
sequences notes (en/zh/i18n) to the new split and step-level behavior.
Verification: ci-workflow.spec.ts 14/14, typecheck clean, all five workflows
YAML-parse, verify-translation-pairing consistent, note-format 582.
The themed WebKit scrollbar skin keeps a wide table's horizontal bar
permanently painted. Chromium never repaints state-conditioned scrollbar
styles (hover-conditioned ::-webkit-scrollbar* rules and :hover
scrollbar-color changes both compute but never reach the painted bar,
measured headed and headless), so the hover reveal toggles overflow-x
itself: hidden at rest with a padding-bottom matching the themed bar
height, auto on hover or keyboard focus with the padding released — the
appearing bar exactly replaces the padding and nothing below shifts.
Resting hidden overflow drops Chromium's implicit scroller focusability,
so wide wrappers carry an explicit tabindex for arrow-key scrolling.
Cache README (EN/ZH): the medium is the session_projcache domain in
per-record layout (one version-stamped document per session under the
json backend root), reads are synchronous from the domain's in-memory
tables, and the storage stack rides in base. storage-json README
documents both layouts and their contracts; web-app README notes that
storage and the projection cache live in the shared base. Regenerated:
session-projection subsystem catalog (sync cachedSnapshot, domain
medium), config-catalog (Config.root gone), module-graph (cache now
depends on storage-domain, not session-persistence), cli composition
(storage rows in base), and the projection-cache Agent Note — which now
records the file-root revision and its revert as rejected alternatives.
cache.spec now boots the real storage stack (storage, storage-json,
storage-domain) and asserts the per-record medium directly:
<root>/session_projcache/sessions/<id>.json carries a version-stamped
{version, record} document, cachedSnapshot is synchronous (zero-I/O from
the domain's in-memory tables), and the write-policy / fail-soft / listing
coverage is preserved at 100%. json-backend.spec gains a per-record layout
block (per-record documents, overwrite/delete/reopen, unsafe keys and
undeclared tables rejecting, foreign-document discard on open, closed
guard, close drain, unreadable-as-absent); storage-domain domain.spec
covers layout validation and descriptorOf projection. list-children.spec
mounts the storage stack for its projectionCache cases and its
cachedSnapshot mocks and reads go synchronous; the api-proxy specs' cache
mocks go synchronous too. devDeps and tsconfig references updated for the
storage stack.
- document the exported SqliteStore prefix/suffix loaders (verify-export-jsdoc)
- share createStoredEventRead from session-persistence so the standalone
SQLite store stops duplicating the service helper (duplication gate)
- route replaceStored header upserts through writeRow (duplication gate)
- move replacement/conflict test SQL into closed test resources so the
SQLite SQL resource boundary test passes
The storage hub, json backend, and domain form are general infrastructure,
and the projection cache is a session-layer service that depends on them —
both belong in the shared base, not in the web overlay. Base now provides
storage / storage-json / storage-domain / session-projection-cache; the
web-app overlay keeps its surface consumers (workspace, message-feedback),
which inherit storageDomain from base (a child layer sees parent services).
This reverts the storage stack's historical web-app-only placement and the
file-root design's base mount of the cache.
The cache now opens its domain through ctx.storageDomain, which the
web-app overlay provides (storage-json + storage-domain, backend json).
A parent layer cannot see a child layer's services, so the base-layer
mount from the file-root design is reverted: the cache mount moves back
to web-app next to its storage dependencies, and Config.root is gone.
Restore the storage-domain medium the file-root design replaced: the cache
opens the session_projcache domain (per-record layout — one document per
session under the json backend root) and checkpoint writes land through
the domain's write chain. Reads and writes now share ONE coherent state:
cachedSnapshot reads synchronously from the domain's in-memory tables,
and every write is durability-first-then-memory, so a read can never go
around the write chain to the medium. The hand-rolled write chains,
in-flight tracking, per-session file paths, owner-only file modes, and
the sqlite no-path special case are gone; Config.root is removed and the
domain's version stamp makes a checkpointRecord bump discard stale
sessions per record instead of rejecting the whole medium. The async
ripple of the old file read is reverted: api-proxy's listing column and
subagent's cold identity read go back to synchronous cachedSnapshot.
The json backend now serves two layouts. single (the default) keeps the
whole unit as one document at <root>/<name>.json; per-record keeps one
version-stamped document per record at <root>/<name>/<table>/<key>.json
(plus global.json), so one write rewrites one record instead of the whole
unit. The per-record unit is stateless — the directory is the state,
loadAll re-reads the tree, and every write is a single durable file
operation — while single keeps its authoritative in-memory state and
whole-file publish. Records keys must be path-safe ([a-zA-Z0-9_-]+);
an unsafe key rejects. A record document that is malformed or stamped
with another version reads as an absent record: one bad or stale file
never bricks the unit, and a version bump discards stale records instead
of migrating them. DomainSpec and KvUnitDescriptor gain the optional
layout field (defineDomain validates it, descriptorOf projects it).
Review follow-up on #2794: the pin was applied inside nextSessionOrderAccount,
whose output syncSessionOrderAccount persists into sessionOrderByAccount, so a
reused blank stayed first forever after turning non-blank (even across
reloads). Move the pin to the render-derived layer (orderedWorkspaces,
orderedUngroupedSessionIds, ungrouped order, flat rows) so the persisted
account order is never touched, and skip drags the pin would mask (dragging
the pinned blank, or parking another row into the pinned slot) entirely,
including the Host account write.
Adds regression tests for the masked-drag skip and for the pin releasing when
the blank turns real; updates the sidebar-order Agent Note (EN + ZH) and its
pairing hash.
Address ds-review-bot re-review: collect per-cleanup rejections and rethrow
as AggregateError so real flush failures surface instead of being swallowed;
register the cold-resume context's second persistence handle in cleanups so
it closes even on a failure path; drop the unverifiable flake-history aside.
Remove the three skipped (gray) checks from the PR check panel without changing
functional semantics:
- issue-lifecycle: remove the job-level 'if' that skipped the lifecycle job on
non-changes-requested pull_request_review events, so it now runs and reports
success (the lifecycle handler already no-ops for approved/commented reviews).
The changes-requested board transition is unchanged.
- release.yml / release-vendor.yml: drop the publish job (and its
workflow_dispatch 'publish' input + RELEASE_PUBLISH pass-through) so it no
longer appears as a skipped Publish-to-npm check on PRs; the files keep the
pack job that validates tarballs on PR/push.
- new release-publish.yml / release-vendor-publish.yml: manual workflow_dispatch
only, repack on the current tree then publish, so publication behaves exactly
as the old publish job (explicit dispatch, uses the packed bytes) but never
shows as a PR check.
Update the 2026-08-10 review-status note (en/zh/i18n) and the issue-lifecycle
spec assertion to match the unconditional lifecycle job.
Verification: ci-workflow.spec.ts 19/19, all five workflows YAML-parse, typecheck
clean, verify-translation-pairing consistent, note-format 582.
The backdrop keyed each plain-text reference mark by its draft offset, so
any character typed ahead of a reference changed the key and made React
unmount the mark with its nested spans and inline glyph. Every keystroke
before a reference rebuilt every reference after the caret.
A scan derives these ranges fresh on every render, so a range carries no
identity past the render slot it occupies. The mark now keys by its index
in the offset-sorted textRefs list. Structured chips keep occurrenceId,
the identity their occurrence table owns.
The New Session being created renders first in its account while selected,
in both order modes; opening or reusing a blank never advances its
updatedAt, so the pin keeps the reused row at the front.
Clicking New Session reuses the workspace's existing blank session when one
exists; reuse only opens it, so its updatedAt (host: max(createdAt,
lastPromptAt)) never advances and the row kept its old creation-time
position in the sidebar. nextSessionOrderAccount now pins the currently
selected blank session to the front of its account in both Manual and Last
updated modes; non-blank navigation stays untouched.
Fixes#2788
Trim knip.json to 655 lines by removing 15 stale or glob-duplicate workspace entries. Behavior-neutral under knip's specificity-based workspace config selection; CI green, issue policy green, review threads resolved.
ProjectionDefinition lost its persist?: boolean opt-in when every
projection unit became uniformly checkpointed, but the subsystem doc's
type-equiv paste still carried the field, so verify-type-equiv drifted
from packages/session/session-projection/src/index.ts. Remove the field
and its JSDoc from both language sides and re-record the bilingual
pairing record.
- subsystem docs: drop the removed persist flag from the ProjectionDefinition
type block (both languages).
- config-catalog regenerated (cache requires sessionProjections/sessions,
config gains root) and the zh side synced by hand; doc graphs regenerated.
- Agent Notes: the per-session cache note records the owned root tree and
no-persistence design; the storage-root proposal's link to it is corrected
(two levels up).
The web e2e seeds no longer warm the deleted coldSnapshot; assertions do not
depend on the cache path. list-children mounts the cache with a scratch
root instead of a storage-domain backend, and drops the now-unused storage
devDependencies.
The cache now owns its storage root (dshHomePath('projections')), so the
mount moves from the web-app overlay to base with that root declared, next
to the other base session layers. web-app inherits it; its overlay mount is
removed.
Store each session's projection_cache.json under the cache's own root tree
(<root>/<session-id>/projection_cache.json, wired to dshHomePath('projections')
in the base bundle) instead of beside the session log via
sessionPersistence.locate(). The cache owns its directory layout, keys
directories by the code-generated session id, and never consults the
persistence layer; the service now injects only sessionProjections and
sessions.
Drop the coldSnapshot method and its readFrom-tail fold ladder: every cold
consumer refolds from the log itself, so the cache only serves the listing
read (cachedSnapshot, one async file read per session) and the write side.
Fail-soft durability, per-path write serialization, in-flight drain, and
atomic 0600 writes are unchanged; the chain cleanup now observes its own
rejection so a failed write cannot surface as an unhandled error.
dsh-session-persistence leaves peer/dev dependencies and the tsconfig
reference; dsh-atomic-write moves to peerDependencies. Config gains a
required root.
- Write through @deepseek-ai/dsh-atomic-write with { mode: 0o600,
dirMode: 0o700 } instead of exporting a second atomic-write primitive
from dsh-storage-json; the session tree stays owner-only like the jsonl
backend's own directories.
- Serialize atomic replacements per cache path so an older cut can never
overwrite a newer one; track in-flight writes and drain them on
disposal so a late flush cannot land after teardown.
- Detect the absent per-session directory before the checkpoint cut and
durability flush: sqlite-style backends no-op the write entirely.
- Cold-read write-back path and identity both come from the stored log
header (tail.meta), so a stale caller header cannot mint an orphan
cache file.
- Add no-path coverage (write no-op, cachedSnapshot undefined, cold
fallback to the full-log rung) and a concurrent-write ordering test;
the package now holds 100% statement/branch/function/line coverage.
- Sync README.md/zh (inject list, coldSnapshot signature, per-session
file read wording), package description, the Agent Note alternatives,
and the superseded proposed/implemented notes (EN/ZH); add the
concurrent-checkpoint Known Limitation.
Replace the single global session_projcache domain with a per-session
cache file inside the session's own persistence directory, resolved
through sessionPersistence.locate(meta) — the persistence backend owns
the session-directory layout, the cache service keeps every checkpoint
and cold-read responsibility.
- cachedSnapshot(meta) becomes async (one file read per session);
coldSnapshot takes the session header so it can locate the file, with
the stored log header remaining the identity witness.
- Backends without a per-session directory (sqlite) disable the durable
cache: writes no-op and cold reads fall to the full-log rung. An
obsolete global cache is never read — derived data refolds on first
cold read (no migration).
- writeAtomic is exported from dsh-storage-json as the shared atomic
whole-file replace primitive; api-proxy listing and subagent cold
reads await the now-async cachedSnapshot.
- READMEs and a new Agent Note document the per-session medium.
The product model has no cross-process writer exclusion (the coordinator
serializes per-session operations in-process; the README documents one
live writer per session), so the wx-created .lock sibling only guarded
byte-level races while adding two failure modes: a crash leaves a stale
lock that permanently wedges that log's appends/repairs/replacements, and
a post-commit lock cleanup failure makes a committed append look failed,
so the retained write-behind batch retries into duplicate seqs.
Remove withLogLock and keep replaceStored's revision compare-and-swap at
the commit boundary (recheck immediately before the atomic rename).
buildStepIndex rejected the whole decoder at initialization whenever any
registered step could not reach the current version, so one retired old
upgrader blocked every session, including later versions whose path to
the current version is complete. planSteps already refuses a specific
stored version when a needed step is missing; initialization now checks
only step legality and duplicates.
Markdown tables were always rendered at natural width, so anything wider
than the 748px message column could only be read through horizontal
scrolling. Following the deepsuite chat TableWrapper treatment: tables
under four columns (and any table inside a blockquote) now fill the column
and wrap cell text down to the cells' minimum readable width, while
four-or-more-column tables keep their natural width behind the wrapper's
horizontal scroll and carry the stable md-table-wide hook. The chat
transcript widens hooked tables past the message column with a pure-CSS
container-query breakout (100cqw against ChatView's scroll container
standing in for deepsuite's JS-measured list width), keeping the table
content aligned with the message column and clamping to neutral when the
transcript is narrower than the column.
The exact-event-set assertion called Object.keys on a Record<string, unknown>'s
on field without narrowing, failing typecheck (TS2769). Guard with isRecord
before asserting the full event sets.
Address the 7 remaining review threads on PR #2768:
- spec: assert the exact full event sets for both workflows (ci-master =
[push, workflow_dispatch], ci.yml = [pull_request]) instead of only the
negated checks, so losing/adding a wrong event fails.
- 2026-07-21 note line 29: 'The ci-master and Sandbox workflows keep their
cross-platform references on master pushes' (en+zh) — ci.yml is PR-only.
- failover-runbook line 19 (en+zh): the push-reachable classification now names
ci-master.yml for the drills/wine cache and ci.yml for the PR jobs, dropping
the stale pull-request-gated phrasing.
i18n hashes re-recorded for both notes; spec 13/13, pairs consistent,
note-format 575.
Address the second review pass on PR #2768:
- Restore the serial-linux-selfhosted drill comment in ci-master.yml (it was
lost in the split boundary, leaving serial-linux-selfhosted as the only
comment-less job); adapted to the current-state fact that this workflow never
listens to pull_request while keeping the no-cache-steps reason, the
DSH_CI_FAILOVER_LINUX switch guidance, and push-triggered semantics.
- Update 2026-07-29-pnpm-setup-runner-isolation (en/zh + i18n): the regression
test and setup steps now span ci.yml and ci-master.yml.
- failover-runbook line 19: the push-reachable classification now names
ci-master.yml for the drills/wine cache and ci.yml for the pull-request jobs,
dropping the stale pull-request-gated phrasing.
- all-checks-passed comment in ci.yml notes that needs cannot reach across
workflow files (ci-master job are deliberately outside this PR verdict).
Verification: ci-workflow.spec.ts 13/13, both workflows YAML-parse with correct
concurrency, 3 translation pairs consistent, note-format 575.
- The pnpm-caching note placed node-compat in ci-master.yml, but it stays a PR
job in ci.yml; only the two runner benchmarks moved. Correct en/zh line 15 and
re-record the pair.
- The pnpm/action-setup destination test only iterated ci.yml; extend it to also
cover ci-master.yml so its five pnpm setups stay regress-tested.
Per review on PR #2768:
- Restore a concurrency block in ci.yml (cancel-in-progress: true) so a fresh
PR push cancels the superseded run; GitHub has no default PR auto-cancel,
so removing it would stack a second full 9-job run on every push.
Update the ci-workflow.spec assertion accordingly.
- Fix comment misplacements from the split boundary: ci.yml's all-checks-passed
now carries its own branch-protection comment (the drill comment that was
orphaned above it is gone), and ci-master.yml no longer ends with an orphaned
all-checks-passed comment.
- Restore the DeepSeek e2e workflow describe-block in ci-workflow.spec.ts that
the rewrite had silently deleted (e2e.yml is unchanged).
- serial-windows comment no longer claims non-blocking-for-PR/absent-from-needs
since ci-master never listens to pull_request.
- DSH_TELEMETRY_DISABLED safe-use comment restored in ci-master.yml env; split
rationale (mi gray segments in PR check) documented in .github/AGENTS.md.
Verification: ci-workflow.spec.ts 13/13, both workflows YAML-parse with correct
concurrency, git diff --check clean.
Split the single ci.yml into two workflows so the PR check panel stops listing
master-only obs jobs that skip (gray) and block the aggregate green-check:
- ci.yml is now pull_request-only, holding only the 9 PR jobs (node-24,
node-24-coverage, node-24-consumers, node-compat, python-sdk, python-runtime,
windows, windows-native, all-checks-passed). It drops the workflow-level
concurrency block since master carve-outs no longer apply.
- ci-master.yml (new) carries the six master/dispatch jobs (wine-apt-cache,
serial-linux-selfhosted, serial-macos, serial-windows, and both runner
benchmarks) with the push-exempt cancel-in-progress block and suite input.
It does not listen to pull_request, so its jobs never appear in PR checks.
ci.yml keeps the wine apt cache seed/restore (repo-scoped by key), so moving the
producer to ci-master.yml does not break the PR windows job's restore.
Update ci-workflow.spec.ts to assert the two-workflow split: ci-master owns the
drills/benchmarks and the push-exemption concurrency, ci.yml is PR-only without
concurrency. Update the serial-reference, portable-required, failover-runbook,
and pnpm-caching notes (en/zh + i18n) and .github/AGENTS.md to reflect ci-master.
Verification: scripts/ci-workflow.spec.ts 12/12, both workflows YAML-parse,
verify-translation-pairing consistent, verify-agent-note-format passes.
Drop the persist?: boolean opt-in: every unit's state — client-visible and
host-only alike — is now written to the projection cache. A unit can no
longer silently skip the durable cache, host-only units no longer need an
explicit flag to participate in cold restore, and the persist-sharing
conflict check disappears with the field.
- ProjectionDefinition/ErasedDefinition lose persist; register overloads
simplify; checkpoint/restoreFloor/restore fold every registered unit.
- Registry and cache tests drop the persist:true fixtures and the
persistence-policy sharing test.
- READMEs and the state-and-client-views note record the uniform rule;
cordis API catalog and subsystem signatures regenerated.
- Keep cleaning remaining roots even if one cleanup rejects.
- Collapse the duplicated Windows-EPERM rationale to one comment.
- Dispose the cold-resume context's second JsonlSessionPersistence handle on
the shared root before afterEach removes it.
Remove 15 workspaces entries from knip.json: 2 keys naming packages that no
longer exist (packages/util/home, packages/client/web-ui) and 13 entries
whose entry/project equal the packages/*/* glob default. knip selects one
config per matched key by specificity, so a removed entry either lost an
unresolvable target or fell back to an identical glob config; knip still
runs clean (0 issues). 655 lines, down from 790.
The continuation spec cleaned each mkdtemp root with rmSync in afterEach
without closing the JsonlSessionPersistence handle first. On Windows,
rmSync recursive over a directory whose file handle is still open fails
with EPERM, surfacing intermittently in native complete as
dsh-subagent-continuation-* cleanup failures. Close the persistence
fiber before rmSync (mirrors jsonl.spec.ts), preserving per-root
dispose-then-delete order.
The snapshot embedded five live bilingual document pairs as reviewed
examples, so editing any of them (README, development guide, i18n docs)
churned the snapshot. Replace them with three synthetic fixture pairs
(product, rules, agent-note shapes) under scripts/fixtures; the prompt
examples stay representative without tracking real document content.
# Agent Note: GUI layering and the RPC protocol — host/client layering by capability provider, the four-quadrant message model, and the fetch carrier
Status: implemented
Archived: 2026-08-27
English | [中文](2026-07-19-gui-layering-and-rpc-protocol.zh.md)
@@ -209,7 +210,7 @@ The same domain tree as `ApiProxy`, but unary methods **take the business payloa
### The instance-level envelope observation aspect
All four quadrant full forms pass through `onEnvelope`; the base implementation is an **instance-owned microtask-batched buffer** (frame storms must not disturb consumers per frame; module-level state would leak across instances/tests, hence instance-owned). Observers subscribe via `subscribeEnvelopes(listener)` (receiving whole batches as `readonly RpcMessage[]`, returning an unsubscribe function); a listener throw is isolated (observation must never bite the carrier). With no subscribers the buffering costs nothing. No shipped consumer subscribes today — the aspect is the designated seat for wire diagnostics (the retired RPC debug panel was its first consumer, and a future one plugs in without touching the carrier).
All four quadrant full forms pass through `onEnvelope`; the base implementation is an **instance-owned microtask-batched buffer** (frame storms must not disturb consumers per frame; module-level state would leak across instances/tests, hence instance-owned). Observers subscribe via `subscribeEnvelopes(listener)` (receiving whole batches as `readonly RpcMessage[]`, returning an unsubscribe function); a listener throw is isolated (observation must never bite the carrier). With no subscribers the buffering costs nothing. No shipped consumer subscribes — the aspect is the designated seat for wire diagnostics (the retired RPC debug panel was its first consumer, and a future one plugs in without touching the carrier).
# Agent Note: WebSocket carrier for browser downlinks
Status: implemented
Archived: 2026-08-27
English | [中文](2026-08-04-websocket-downlink-carrier.zh.md)
@@ -16,7 +17,7 @@ WebSocket carries only the host→browser downlink. All client→host unary call
## Upgrade and lifecycle boundaries
`dsh-host-webserver` provides an exact upgrade-route registration point alongside ordinary routes, dispatches Node upgrade sockets by pathname only, contains raw-socket errors, and waits for surviving upgraded connections to close during server teardown; it knows nothing about Harness frames or WebSocket messages. `dsh-client-connection` owns the WebSocket handshake, frame output, and stream cancellation, and reuses the `/api` Host/Origin trust fence before upgrade. An untrusted authority or cross-origin Origin is rejected before `ctx.apiProxy.events.*` starts.
`dsh-host-webserver` provides an exact upgrade-route registration point alongside ordinary routes, dispatches Node upgrade sockets by pathname only, contains raw-socket errors, and waits for surviving upgraded connections to close during server teardown; it knows nothing about Harness frames or WebSocket messages. `dsh-client-connection` owns the WebSocket handshake, frame output, and stream cancellation. Before upgrade it applies the `/api` Host/Origin checks followed by the same signed browser-cookie authentication as unary HTTP. An untrusted authority or cross-origin Origin receives 403; a trusted but unauthenticated request receives 401; neither starts a Remote stream.
A browser abort or socket close cancels the corresponding host stream; plugin teardown also waits for that source iterator's cleanup. If a host stream throws midway, the carrier sends one existing `stream/error` frame and then closes the socket; the client treats that frame as connection loss rather than delivering it to a business sink. Each WebSocket reports open independently, and the existing readiness handshake still waits until mux and host are both open and the `host.describe` HTTP call has succeeded before publishing connected.
外部报告 dsh-external/issues#107（内部聚类为 deepseek-harness#1406）测得视口宽度在 760px 到 850px 之间时 Plan 控件与模型选择器发生重叠，模型选择器覆盖 Plan 控件的点击区域，导致在 800×720 下无法用鼠标退出 Plan 模式。其验收清单要求增加浏览器回归测试，断言 Plan 中心命中 Plan 按钮。
# Agent Note: Composer edits carry the range they applied to
Status: implemented
Archived: 2026-08-20
English | [中文](2026-08-20-composer-edit-range-from-selection.zh.md)
## Problem
The input machine keeps its reference occurrences aligned by reconciling them against one edit range: entries before the range shift, entries after it hold, and an entry the range intersects loses its structured identity and stays behind as ordinary draft text. That last rule is the deliberate meaning of editing inside a reference.
Ordinary typing supplied no range. A controlled textarea's change event carries only the resulting string, so the machine recovered the range by scanning the two drafts for a common prefix and suffix. That recovery is ambiguous whenever the inserted text repeats the text it lands against, and the greedy scan always resolves the ambiguity the same way: it slides the edit as late as the characters allow.
A reference renders as `@` followed by its label, so typing `@` immediately before one produces exactly that collision. The user inserts at the reference's own offset; the scan reports an insertion one character later, inside the reference; reconcile applies the intersect rule and drops the occurrence. Deleting a character in front of such a reference slides the same way.
The draft then still reads correctly to the eye while carrying no structured reference, and submission takes the occurrence-free path that sends the draft verbatim. The host receives the human-facing label instead of the owner's model form and resolves nothing. The serialization guard that exists to prevent exactly this downgrade never runs, because it only fires when an occurrence survives to be serialized.
This became reachable when references [became literal inline text](../feature/2026-07-27-web-file-and-session-references.md). A reference previously occupied one `U+FFFC`, a character no keystroke produces, so the scan had nothing to collide with.
## Decision
`InputBar` records the textarea's selection and `inputType` during `beforeinput` and passes the resulting range to `setDraft`, which the machine already accepts and prefers over its own scan. A textarea exposes the edit no other way — `getTargetRanges()` is empty for form controls.
An edit that replaces a selection reports that selection, and it is the range outright; the inserted length is whatever the draft grew by once the replaced range is accounted for. A caret delete replaces nothing and reports the bare caret, so its range comes from the direction `inputType` names and the number of characters the draft actually lost. The count is measured rather than assumed to be one, because a single caret gesture removes a multi-unit grapheme, a word, or a line just as readily. Chromium, WebKit, and Firefox all report the collapsed caret for `deleteContentBackward` and `deleteContentForward`, and all three derive the same range from it.
Only the `insert` and `delete` families are recorded. A history replay reports wherever the caret happens to sit, which would survive every check while naming the wrong span; ignoring it leaves that path on the scan.
The record is consumed once and cleared. A record whose draft length disagrees with the draft the change reports, a selection past the draft, a shrinking edit over a selection the caret cannot explain, or an undirected delete all yield no range, and the machine falls back to its scan. Paste and the boundary Backspace and Delete gestures already supplied their own ranges and are untouched.
## Testing
Component tests cover the trigger character typed in front of a reference, a caret Backspace, a caret Delete, a caret word delete, and a delete over a selection, asserting in each case that the occurrence survives at the shifted offset. The caret cases fail against the scan-recovered range, and the word case fails against a fixed one-character step.
An assembled browser scenario drives the same gestures as real key presses against the shipped composition, which is the only place the range an engine reports for them can be observed; its golden projects the backdrop's segments, since the decoration layer is aria-hidden and the accessibility tree cannot see the chip. A real composition driven through the browser's IME path reports the composing segment as the selection at every intermediate state, and the reference survives each one.
## Alternatives considered
**Disambiguate the scan with the post-edit caret.** The caret pins which of the textually equivalent readings happened, and the change event already carries it. Rejected because it keeps a reconstruction where an exact fact is available, and it cannot separate the deleted and inserted halves of a replaced selection at all.
**Give references a leading marker no keystroke produces.** A private-use character in place of the literal `@` removes the collision at the representation level, and the reject list for pasted text already names that range. Rejected because it re-adds a character that every serialization, selection, and accessibility path has to strip, to buy what an exact range buys directly, and it would leave ordinary typing reconstructing its range for every other reason.
**Widen reconcile to keep an occurrence when the range only touches its edge.** Rejected because the misattributed offset lands strictly inside the reference, not on its boundary, so the rule change would not reach this defect while making the intersect rule vaguer.
## Consequences
Every native textarea edit that reaches `onChange` now names the range it applied to, so occurrence offsets follow the edit that actually happened rather than one merely consistent with the resulting characters. Draft writes that originate in the facade rather than the DOM — `insertText` and the command-token splices among them — still carry no range and keep the scan, as do the fallbacks above.
The composer now depends on `beforeinput` preceding each value change, and on `inputType` naming the direction of a caret delete. Any future edit path that mutates the value without either silently returns to the scan rather than breaking, which keeps the failure mode the old behavior instead of a wrong range.
# Agent Note: Composer reference decorations key by draft-order ordinal
Status: implemented
Archived: 2026-08-20
English | [中文](2026-08-20-composer-reference-decoration-keys.zh.md)
## Problem
The composer backdrop renders the draft as an array of segments: plain strings, a leading claim-token mark, one element per structured reference, and one mark per plain-text reference range. React reconciles that array by key.
Structured references carry an identity — the occurrence table mints an `occurrenceId` that survives every edit — so their chips key by it. Plain-text reference ranges have no such identity: `scanTextRefs` re-derives them from the draft on every render, and nothing outside that scan remembers a range between two keystrokes.
Keying those ranges by their draft offset made the key change whenever earlier text changed length. React then treated the range as a different element, unmounted the mark with its nested spans and inline glyph, and mounted a replacement. Every character typed or deleted ahead of a reference rebuilt every reference after the caret, and the work grew with the reference count. [Directory-syntax ranges](../feature/2026-07-27-web-file-and-session-references.md) made that path routine: they match on `@path/` syntax without a lexicon, and each one renders an icon.
## Decision
A plain-text reference mark keys by its index in the offset-sorted `textRefs` list, computed where the boundary list is assembled so a skipped boundary cannot shift it. The scan already returns the ranges in draft order, so the ordinal names the render slot a range occupies, which is the only identity a scan-derived range has.
Structured chips keep `occurrenceId`. The two key strategies differ because the two range kinds differ in identity, not by oversight: a range the occurrence table owns keeps its node across reordering, and a range only a scan knows keeps its node across offset shifts.
A range that stops matching the scan still loses its decoration, because it disappears from `textRefs` and the ordinal it held no longer exists.
## Testing
A component test holds the mark element and its glyph, types a character ahead of the range, and asserts the same nodes are still mounted; it then edits the token out of match shape and asserts the decoration is gone. The test fails against an offset-derived key.
## Alternatives considered
**Key by the range text.** Rejected: duplicate references collide on one key, and editing inside a range changes its key, which reintroduces the remount this fixes.
**Give scan-derived ranges an identity table.** Rejected: it adds mutable state whose only consumer is a render key, and the scan would have to diff against the previous draft to maintain it. An edit that breaks a match simply dropping the range on the next scan is what keeps `scanTextRefs` a pure derivation.
**Drop the keys and let React match by position.** Rejected: React requires keys on elements inside an array, and the plain string segments between them already match by index, so an unkeyed element warns without changing the outcome.
## Consequences
Typing ahead of a reference updates text nodes only; the mark and its icon stay mounted. The backdrop's per-keystroke DOM work no longer scales with the number of references in the draft.
Because the key names a position, inserting a reference ahead of existing ones reuses the earlier nodes with new content instead of re-creating them. That is correct for these marks, which hold no focus, selection, or animation state, and it is the condition any future decoration on this layer meets before it keys by ordinal.
# Agent Note: Equal-order system-prompt sections render in activation order
Status: implemented
Archived: 2026-08-25
English | [中文](2026-08-24-system-prompt-section-order-ties.zh.md)
## Problem
`SystemPromptRegistry` sorts sections by `order` with a stable sort, so equal orders render in plugin-activation order. `tool:cordis` and `tool:workflow` both declared `order: 115`, while their activation order varies between clean platform compositions. ACP and SDK snapshot replays could therefore assemble the same sections in a different order from their committed `system-prompt.expected.md` files.
## Decision
Give the affected sequence distinct values without changing its established relative order: `tool:cordis` stays at 115, `tool:workflow` uses 115.5, `tool:ralph` stays at 116, continuable subagent guidance stays at 116.5, and child-report guidance stays at 117. Prompt text and tool schemas remain unchanged.
## Alternatives considered
**Normalize section order in the snapshot harness.** Rejected because the runtime, request header, and model prompt would remain sensitive to activation timing while only the fixture comparison hid the difference.
**Tie-break equal orders by section name in the registry.** Rejected because it would silently reorder every existing tie. Explicit orders keep each model-visible placement local to the contributing plugin.
## Consequences
The Cordis and workflow guidance has a platform-independent order while Ralph remains before continuable subagent and child-report guidance. Prompt-section placements that require a stable relative position need distinct `order` values; other equal-order sections retain activation-order semantics and are outside this decision.
## Testing
The keyless ACP and SDK snapshot replays pin Cordis before workflow and preserve the workflow, Ralph, continuable-subagent, and child-report sequence. The full snapshot suite verifies the refreshed fixtures.
English | [中文](2026-08-18-product-subagent-failure-facts.zh.md)
## Problem
The [Claude Code and Codex product providers](2026-08-04-claude-code-and-codex-subagent-backends.md) receive structured product failures, but a published run historically flattened most of them to the shared `error` stop reason. Product logs retained detail that the foreground parent and a [one-shot background Job](2026-08-12-product-subagent-one-shot-background-tasks.md) could not use to distinguish a product limit, an execution failure, or an early process exit.
Copying SDK error text, app-server payloads, or stderr into the result would expose task text, paths, environment values, credentials, or product internals. Adding shared error fields would also make the provider-neutral [subagent seam](2026-06-21-subagent-capability-seam.md) own product version vocabularies that change independently.
## Decision
Each product Provider owns the mapping from its pinned official structured failures, current operation, and managed process outcome to one fixed safe diagnostic line. `SubagentResult` remains unchanged: consumers receive the existing bounded `diagnostic` string and do not parse its product-private fields. The [minimal-diagnostics decision](../simplification/2026-08-21-product-subagent-minimal-diagnostics.md) supersedes this note's complete Claude Code subtype mirror; this note continues to own the current detailed Codex categories until that provider adopts the same simplification.
The Provider omits unavailable optional fields. Exit code and signal are independent facts and are each retained when observed. A contributing permission decision from the [non-interactive permissions decision](2026-08-15-product-subagent-noninteractive-permissions.md) follows the structured line; the latest safe permission fact remains operation-local. The shared result boundary limits the complete text to 4096 UTF-8 bytes.
Successful results and local cancellation expose no failure fact. Raw product errors, stderr, tool input, paths, environment values, credentials, and protocol payloads never enter the diagnostic. Startup and cleanup rejections use the same safe line in their Error message. Original failures remain on internal cause chains; Provider Host logs and forwarded stderr remain product-local observation only.
### Claude Code facts
The [minimal-diagnostics decision](../simplification/2026-08-21-product-subagent-minimal-diagnostics.md) exclusively owns Claude Code categories, stages, process facts, permission ordering, and verification for Agent SDK 0.3.241 and Claude Code 2.1.241. This note carries no separate Claude category contract.
### Codex facts
Codex app-server 0.147.0 defines eleven string categories and five object variants. The Provider preserves `contextWindowExceeded`, `sessionBudgetExceeded`, `usageLimitExceeded`, `serverOverloaded`, `cyberPolicy`, `internalServerError`, `unauthorized`, `badRequest`, `threadRollbackFailed`, `sandboxError`, and `other`. It also preserves `httpConnectionFailed`, `responseStreamConnectionFailed`, `responseStreamDisconnected`, `responseTooManyFailedAttempts`, and `activeTurnNotSteerable`; the four connection/stream variants retain numeric `httpStatusCode`, while the active-turn variant does not expose `turnKind`. Unknown strings, objects with another variant set, malformed values, and unclassified exceptions use `unknown`.
| Stage | Owned operation | Observable failure |
| --- | --- | --- |
| `initialize` | App-server spawn and initialize/initialized handshake | `start()` rejects with fixed safe facts and any process outcome already observed |
| `thread-start` | Ephemeral `thread/start` request and response validation | `start()` rejects with the thread stage and any available process outcome |
| `turn-start` | Published `turn/start` request, provisional ids, and early frames | The run resolves as `error` with a safe unknown fallback when no structured category exists |
| `turn` | Terminal notification, final-answer selection, and error-info mapping | The complete category and optional HTTP status reach the non-completed result |
| `process` | Managed app-server exits before another terminal path settles | The run resolves as `error` with `process-exit` and any available code and signal |
| `teardown` | Wire close and process-tree release | `dispose()` rejects independently; startup rollback aggregation exposes both startup and teardown lines |
`contextWindowExceeded` remains `max-tokens`; every other known or unknown Codex category remains `error`, and `cyberPolicy` does not become `refusal`.
### Ownership and lifecycle
| Fact or resource | Owner | Consumer behavior |
| --- | --- | --- |
| Codex error category | Codex Provider over its pinned official app-server | The Provider preserves its current structured category and uses `unknown` outside the recognized set |
| Current failure stage | Product Provider operation | Derived at the failure site; never persisted or used as a recovery state |
| Exit code and signal | `dsh-subprocess` process handle | The Provider displays observed values without inferring missing ones |
| Diagnostic bytes and delivery | `dsh-subagent`, foreground tool, and Job runtime | The same bounded text is presented separately from assistant output in both scheduling modes |
| Raw product failure | Product runtime, internal cause chain, and Host observation | It remains internal and never becomes model-visible result text |
## Verification
Claude Code verification is owned by the [minimal-diagnostics decision](../simplification/2026-08-21-product-subagent-minimal-diagnostics.md). Codex package tests pin all sixteen current error-info variants, HTTP status presence and absence, all six stages, unknown fallback, stop-reason preservation, permission ordering, sanitization, cancellation, concurrency, and cleanup aggregation. The real app-server fixture produces an actual Codex `internalServerError` and covers process/protocol failure and whole-tree quiescence. The keyless ACP snapshot records the Codex diagnostic in foreground error output, a background completion notice, and `job_output`.
## Alternatives considered
**Return raw SDK errors, app-server payloads, or stderr.** These values can contain commands, paths, workspace content, environment values, credentials, or upstream prose. A fixed allowlisted mapping preserves actionable facts without expanding the model-visible trust boundary.
**Add a shared product-error enum or structured result fields.** Claude Code and Codex version their error unions independently. A shared enum would duplicate those authorities and force unrelated Providers and consumers to track product releases.
**Parse generic stderr and exception messages.** Free-form text is neither stable nor safe. Only pinned structured product fields and the managed process outcome qualify as diagnostic input.
**Persist stages or add a recovery controller.** The stage is derived from the current call site only when a failure is reported. Persistence, retries, resume, and remediation need separate ownership and user contracts.
**Map product limits to new shared stop reasons.** Claude Code turn and budget limits are not token-window exhaustion, and an error category does not establish refusal semantics. Existing stop reasons remain unchanged.
## Consequences
The parent can distinguish the current Codex budget, usage, service, policy, request, connection, stream, rollback, sandbox, and active-turn categories without receiving raw product text. The [minimal-diagnostics decision](../simplification/2026-08-21-product-subagent-minimal-diagnostics.md) owns the corresponding Claude result. Foreground and background scheduling preserve the same fact because both consume one `SubagentResult`.
The diagnostic is display text rather than a new public protocol. Callers may present it but must not branch on its punctuation or product-private category names. A pinned product-version upgrade revalidates the Provider mapping and evidence without requiring every official error member to remain model-visible.
This decision adds no product session persistence, retry policy, recovery state, stderr classifier, authentication or configuration taxonomy, progress stream, or human interaction path.
[最小诊断决策](../simplification/2026-08-21-product-subagent-minimal-diagnostics.zh.md)独占负责 Agent SDK 0.3.241 与 Claude Code 2.1.241 的 Claude Code 类别、阶段、进程事实、权限顺序与验证。本说明不再承载独立的 Claude 类别约定。
English | [中文](2026-08-19-high-cache-hit-decimal-display.zh.md)
## Problem
The Web conversation stats line rounded every non-empty cache-hit ratio to an integer. Once the actual ratio passed 99%, the display hid further progress, and a ratio of at least 99.5% appeared as 100% even while uncached input or cache writes remained.
Users therefore could not distinguish a nearly complete cache hit from a true full hit.
## Decision
`StatsLine` continues to derive the ratio from the whole-session `tokenUsage` projection owned by `@deepseek-ai/dsh-token-meter`; the projection remains the only owner of the uncached-input, cache-read, cache-write, and output counts ([projection decision](../architecture/2026-07-29-projected-token-usage-and-request-context.md)). The presentation layer changes only the text inserted into the existing `stats.cacheHit` locale template.
| Actual ratio | Display |
|---|---|
| No billed input | Cache-hit group omitted |
| Integer rounding is below 100% | Rounded integer |
| Non-full ratio whose current rounding is 100% | Minimum decimal precision whose rounded result is below 100% |
| 100% | `100%` |
Every non-empty ratio starts at zero decimal places. A non-full ratio increases precision one place at a time only while rounding would produce 100%, so `99.1%` and `99.49%` remain `99%`, while `99.5%`, `99.95%`, and `99.995%` retain one, two, and three decimal places respectively. `StatsLine` uses exact small-factor comparisons over the safe-integer token counts, then scales the near-full gap only while the intermediate remains within that range. This avoids floating-point tie errors without imposing a precision cap or substitute label. A full hit does not carry a redundant decimal. The same derived string feeds the inline row and its overflow tooltip.
## Ownership and lifecycle
Token-meter continues to fold usage from the complete durable session log. `StatsLine` performs a synchronous display derivation whenever the standard projection value changes. It introduces no setting, stored percentage, event, wire field, client state, or recovery path.
Live updates, reload replay, and reconnect recovery all restore the same `tokenUsage` counts and run the same display function. A missing projection still omits every token group, and a zero input denominator still omits only the cache-hit group.
## Verification
The component spec pins the zero denominator, ordinary integer rounding, half-step rounding at several decimal precisions, each precision boundary through three decimal places, a near-full cumulative sample that needs fourteen decimal places, the true `100%` result, both locales, and equality between inline and tooltip values. The assembled `lifecycle-chrome` replay sidecar selects `9,950 / 10,000 = 99.5%` as a deterministic ratio that integer rounding would misreport as 100% while the base session fixture remains recordable; the live assertion and post-reload browser snapshot both display `99.5%` without another model call.
## Alternatives considered
**Keep integer rounding for every ratio.** Rejected because it hides all movement above 99% and still reports some non-full hits as 100%.
**Truncate the high band to one decimal.** Rejected because `99.95%`, `99.995%`, and still closer ratios all collapse to `99.9%` instead of retaining the minimum precision that distinguishes them from a full hit.
**Cap precision and use a substitute such as `<100%`.** Rejected because the exact cumulative counts can produce the required numeric result, and a cap would make display behavior depend on an arbitrary presentation limit.
**Show one decimal at every ratio.** Rejected because the additional low-band motion adds noise and changes the established display where integer precision is sufficient.
**Persist a display percentage in token-meter.** Rejected because the projection already carries the exact counts, while presentation precision belongs to the Web stats line. A second stored value would duplicate derivable state and expand replay and wire responsibilities.
## Consequences
High cache-hit sessions remain visually stable until integer rounding would falsely report a full hit, then expose only the decimal places needed to preserve that distinction. Extremely close non-full ratios can therefore produce long decimal strings; this is the accepted cost of having no arbitrary precision cap or nonnumeric fallback. Every delivery and recovery path stays on the existing durable projection lifecycle.
# Agent Note: Deleted stale and duplicative knip.json workspace entries
Status: implemented
Archived: 2026-08-22
English | [中文](2026-08-19-knip-config-cleanup.zh.md)
## Problem
`knip.json` carried workspace entries that did no work. Some pointed at packages that no longer exist, and some duplicated the `packages/*/*` glob default exactly. Both kinds made the file larger — 790 lines — and signaled a config that had outgrown the packages it described, so a reader could not tell which entries protected real behavior and which were inert.
## Decision
Deleted 15 `workspaces` entries: 2 stale keys naming packages absent from the working tree and from `HEAD`, and 13 entries whose `entry`/`project` values were byte-identical to the `packages/*/*` glob default.
- Stale keys: `packages/util/home` (removed in `4a09d9b34d`, the harness-home resolver collapse) and `packages/client/web-ui` (no directory and no git history, an orphan key). knip 6.16 does not flag stale workspace keys — that stability check arrived in knip 6.18 — so these were inert config that only deleted when their packages disappeared.
- Glob-duplicate entries: `packages/host/webserver`, `packages/client/runtime`, `packages/core/tools`, `packages/context/tmux-context`, `packages/util/timeout`, `packages/util/output-retention`, `packages/goal/goal-round-driver`, `packages/goal/tool-goal`, `packages/util/home-paths`, `packages/fs/tool-fs-search`, `packages/client/ui-settings`, `packages/client/modules`, `packages/client/hmr`. Each declared exactly `entry: ["tests/**/*.spec.ts"]` and `project: ["src/**/*.ts", "tests/**/*.ts"]`, which equals the `packages/*/*` glob, and each package still exists, so the glob now covers it identically.
The change is a deletion only: `knip.json` went from 790 to 655 lines with no behavioral change. `pnpm run knip` runs clean (zero issues, exit 0) before and after, because knip selects one workspace config per matched key (`getConfigKeyForWorkspace` uses specificity, not array merge), so a removed entry either lost an unresolvable target or fell back to an identical glob config.
## Alternatives considered
- Fold `zod` and other workspace-level `ignoreDependencies` up to the root. Rejected: the root `ignoreDependencies` is a repository-wide fallback, and these exemptions are deliberately workspace-scoped (the README of `cordis-host-runner` records why `src` cannot import the flagged dependency while the generated TypeRT face in `lib` needs it). Widening scope would mask a genuinely misplaced dependency in any future package.
- Upgrade knip to 6.18+ to get an automatic stale-workspace check. Deferred: 6.32.2 (latest at the time) re-flags many `@deepseek-ai/...` test dependencies as unused, i.e. it changes analysis semantics, not just adds hints. That is a separate dependency-upgrade decision with its own CI blast radius, not part of this cleanup.
- Keep the entries as documentation of intent. Rejected: an entry identical to the glob it sits under documents nothing beyond the glob itself, and a key naming an absent package actively misleads.
## Consequences
-`knip.json` is 135 lines shorter and names only packages that exist with config that differs from the glob default.
- Still-explicit entries (54) all carry a real reason to differ — an `e2e`/fixture/tsx `entry`, a `project` outside the default, or a workspace-scoped `ignoreDependencies`.
- knip 6.16 cannot itself detect the next stale key, so a package removal must still remember to drop its `knip.json` key; upgrading to 6.18+ (after the analysis-semantics change is separately assessed) restores that guard.
- This realizes the "never a restatement of the default stanza" criterion of the package-inventory proposal ([topic](../../proposed/process/2026-06-20-discover-package-inventory.md)); its remaining items — the e2e entry folding and the generated inventory — stay open there.
- Multimodal blocks return only with coordinated adapter, UI, and compaction support; see [the drop-image Agent Note](../../archived/simplification/2026-07-04-drop-image-content-block.md).
- Cache hints and assistant prefill remain absent until a shipping adapter can honor them; see the [producer-less variants](../../archived/simplification/2026-07-04-prune-producerless-vocabulary-variants.md) and [inert request knobs](../../archived/simplification/2026-07-04-drop-inert-request-knobs.md) Agent Notes.
- Every adapter pays a translation cost; the first real adapters have since validated the streaming protocol, and new adapters should continue proving their provider-specific mapping in adapter-local tests.
- IDs that cross package boundaries are branded (`CallId`, the shared agent/session `SessionId`) — nominal typing at zero runtime cost.
- IDs that cross package boundaries are branded (`ToolCallId`, the shared agent/session `SessionId`) — nominal typing at zero runtime cost.
@@ -6,7 +6,7 @@ English | [中文](2026-06-13-capability-seams.zh.md)
## Problem
The harness has swappable capabilities — bash execution today, sandboxed/remote executors and alternative model providers tomorrow. A capability has three concerns that change at different rates and for different reasons: the *contract* (what the capability is), the *implementation* (how it runs), and the *consumer API* (what the model and other plugins program against). Bundling them in one package couples those rates of change — swapping a local executor for a sandboxed one would churn the tool schemas the model sees, even though the model-facing contract never changed.
The harness has swappable capabilities, including shell execution and model providers. A capability has three concerns that change at different rates and for different reasons: the *contract* (what the capability is), the *implementation* (how it runs), and the *consumer API* (what the model and other plugins program against). Bundling them in one package couples those rates of change — swapping a local executor for a sandboxed one would churn the tool schemas the model sees, even though the model-facing contract never changed.
This is distinct from "who provides vs. needs a capability at runtime", which Cordis already answers with services + `inject` (a provider registers `ctx.shell`; a consumer declares `inject: ['bash']` and its fiber pends until the service exists). That mechanism is necessary but doesn't dictate package boundaries; this Agent Note does.
持久化是一个具有抽象 Service Definition 的**能力 seam**（[能力 seam](2026-06-13-capability-seams.md)，`dsh-shell` 模板），而非循环或核心逻辑：
持久化是一个具有抽象 Service Definition 的**能力 seam**（[能力 seam](2026-06-13-capability-seams.zh.md)，`dsh-shell` 模板），而非循环或核心逻辑：
1.**接口**（`dsh-session-persistence`，`ctx.sessionPersistence`）：一个抽象的 `SessionPersistence` 服务，提供 `locate`/`create`/`append`/`prepare`/`load`/`inspect`/`readFrom`/`list`/`listSnapshots`。其持久化单元就是现有的 `SessionEvent`（`{ type, seq, time, data }`），原样复用，无转换类型。
Loaded 100 of 7075 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.
imccyu