The sandbox Service Definition adds SandboxEscalationError with
SANDBOX_ESCALATION_INVALID_ARGUMENT, _UNAVAILABLE, _REJECTED, and
_CANCELLED, thrown by validateEscalationArgs() and approveEscalation()
with their existing messages. A cancelled approval for an aborted call
rethrows the call's cancellation instead.
It also adds SandboxPreparationError. sandbox-local wraps Windows ACL
SID, grant, temp-directory, and path-resolution failures as
SANDBOX_PREPARATION_FAILED with the Win32Error or Node error (or an
AggregateError with cleanup failures) as cause, and reports a temp root
inside the workspace as SANDBOX_INVALID_TEMP_ROOT. sandbox-windows-acl
exports isTempRootInsideWorkspace() for that check.
Refs #7869
JobError and TerminalError only need the error base, so dsh-jobs and
dsh-terminal depend on dsh-error instead of the LLM runtime. The Python
SDK runtime closure includes the new package.
Refs #7869
dsh-llm depends on dsh-attachment and dsh-typert-protocol, so classes in
those packages could not extend a base defined in dsh-llm without a
project-reference cycle. HarnessError now lives in @deepseek-ai/dsh-error,
which has no dependencies; dsh-llm declares it as a peer and re-exports
the same class, so existing imports and instanceof checks keep working.
The Agent Note records why the original leaf-package placement no longer
holds.
Refs #7869
Type the approval denial helper with a closed ApprovalErrorCode, state
that a cancellation is recognized by the signal reason or an Error named
AbortError, note that ApprovalError names a denial rather than a thrown
class, and separate JOB_CONTROLLER_UNAVAILABLE from a missing jobs
service. The jobs README describes JobError as the seam's export rather
than as current caller behavior, and the terminal test asserts one
concurrent send failure.
Refs #7869
The plugin entry may export only Loader fields and Service or Error
classes, so the ApprovalError and ToolContextUnavailableError code
constants move to an internal module; the recorded strings are
unchanged.
Refs #7869
Add pi-ai's google-generative-ai protocol to hand-declared routes, read
Gemini's native model listing during discovery, and offer it as "Gemini
API" in the Custom model API form.
Refs #7978
Co-authored-by: LegGasai <[email protected]>
* feat(worktrees): clean up safe UI checkouts when sessions are archived
* refactor(workspace): share archive notice registration
* refactor(session-query): return observation sessions directly
* docs(worktrees): align archive guide and client catalog after merge
* refactor(worktrees): move archive cleanup into the worktree service
The worktree service now records each checkout it creates as a
`plugin:worktree/created` record, folds them in its own `worktree`
projection with fork-aware ownership, and reclaims a clean, unreferenced
checkout when the Session that created it is archived. The Forge panel
reads the service's record and forwards its cleanup state; it no longer
writes a `created` record, owns a retention projection, or gates every
Session's directory access.
Admission registers the attempt inside the Workspace write queue and
returns; inspection runs afterwards, reading cold Sessions through the
projection cache when mounted and from their log otherwise, with subagent
children of an archived ancestor counted as archived. Only a path being
removed holds `working-directory/access`, whose request now carries the
access's `signal`, and listeners withdraw before the drain. Ignored files
no longer retain a checkout, matching Git's non-force removal; processes
and terminals are not consulted.
The Forge bundle mounts one `worktree` service: the `forge-worktree` row
and the `optional-forge-panel` group that isolated a second service are
gone, and `forge-panel` is a top-level row. `canRemove`, `remove`, the
`forgeWorktreeRetention` projection, and `ForgeWorktreeCleanup` are
removed; `WorktreeService.cleanupOf`, `WorktreeCleanup`, and the
`worktree/cleanup` Host event replace them. The upgrade guide moves to
the current version directory and covers the id removals.
* docs(worktrees): move the archive upgrade guide to v0.2.2-rc.1 and wrap a test line
* test(worktrees): satisfy the static and consumer gates
Move the removal step into `removal.ts` as `removeCheckout`, so the specs
drive the real function and intercept it through a module mock instead of
reaching a private method through an assertion to unknown. Classify
`CREATED_RECORD` as a peer-required Host export for the Forge panel, ship
the compiled `lib/types/*.js` that the `./types` subpath names, refresh the
worktree recorded session with the creation record the service now writes,
and regenerate the module graph.
* refactor(worktrees): share the creation record schema with the Forge panel
The worktree package owns `createdWorktreeSchema`, the validator of its
`plugin:worktree/created` payload; the Forge projection imports it instead
of redeclaring the same object.
* test(web): refresh the Forge plugin card golden for the single worktree row
The card lists six components: the isolated `forge-worktree` row is gone.
* build(worktrees): pin zod to the unified 4.6.6
master unified every zod declaration on 4.6.6, so the worktree package's new
dependency follows it and the lockfile resolves again under --frozen-lockfile.
* test(worktrees): cover every reclamation path for the per-file gate
Drop the dead branches CI's coverage lane found: the placeholder trailing
slash in `within`, the aborted-lifetime check in `schedule`, the rejection
arm of a task that never rejects, and the undefined arms of the cold reads.
`cleanupOf` re-reads the attempt after waiting, so a cold Session without
creations settles to null instead of a stale placeholder. Removals of one
path never overlap, so `remove` keeps one entry per path without chaining.
New cases: a cold Session without creations, a borrower that appears
between inspection and the locked recheck, a subagent child with an unknown
parent, a creation record whose payload fails validation, a live fork that
moved out of the checkout, a cached cold Session without a directory, the
Forge projection ignoring a creation without an open request, the Remote
with and without the service, and the client's removed follow-up.
* feat(worktrees): add the retain-on-archive switch
`retainOnArchive`, a volatile field of the worktree service's Config, skips
cleanup admission while it is on; it is read at each archive, so a change
applies to later archives only. The service keeps its instance off the
generated Settings pages, and the Forge client draws the switch as the
`plugins.row.config` page of the bundle's `worktree` row, staged through
the shared Plugins form until one explicit save.
Docs: the worktree, Forge, and bundle READMEs, the archive-worktrees
upgrade guide, the config and client catalogs, and the
explicit-worktree-creation note.
* docs(module-graph): record the worktree service's settings dependency
`gen-module-graph --check` failed on CI after the retain-on-archive switch
declared `@deepseek-ai/dsh-settings`; regenerate the graph.
* test(snapshots): refresh terminal-panel-tools for master's runtime-context reminder
Master's #7321 wrapped the runtime-context snapshot in a `<system-reminder>`
after this fixture was last refreshed, so `test:snapshot` fails on master's
own tree; refresh the one user message so the replay matches.
* feat(release): publish package source, declaration maps, and documentation
Every release member publishes its TypeScript source and the declaration
maps that resolve into it, and @deepseek-ai/dsh publishes the repository
Markdown documentation staged from docs/ at prepack time.
The ./src/* export most members already declared now resolves for an
installed consumer, and a consumer editor follows a declaration to the
implementation. Bundled JavaScript maps stay unpublished; the vendored
family and the Desktop runtime file policy are unchanged.
Rationale: .agents/notes/implemented/process/2026-10-08-published-source-and-docs.md
* feat(release): publish JavaScript maps and compress bundles
Every emitted JavaScript map ships beside the bundle it describes, and the
workspace build and the client preset compress bundles without mangling
identifiers.
The documentation payload now mirrors the website publication manifest and
is removed again by `postpack`; Desktop and the Python executable exclude
the published source, documentation, and maps; `clean` removes a staged
copy. Typert's publication contract gains the Remote declaration map.
* fix(release): keep node bundles readable for the worker image
The browser worker image lowers `yield*` only as a whole statement and
rewrites `for await` loops by their source layout, so compressed node-side
entries either fail that lowering or lower to unparseable code. Only the
browser bundles keep compression, which removes about a third of their text.
The bundle-purity assertion now matches the module-loader call regardless of
the quote style compression picks, and the release-pack fixture exercises
the installed package's new documentation requirement.
* feat(release): publish the whole documentation tree
`@deepseek-ai/dsh` copies `docs/` in full instead of the website's page
set: the maintainer documents, the generated migration records, the
translation guidelines, and the `*.i18n.yaml` consistency records ship
with the installed package, so the documentation's own links resolve
without a checkout.
* feat(release): publish the upgrade guides with the documentation
The staged documentation keeps the website's page set and adds every
`docs/upgrade-guide/` document, which is what an installed reader consults
when a released change breaks something. The rest of the maintainer tree
stays out.
* feat(agent-preset): ship plugin artwork skill with creator mode
* docs(skills): clarify artwork icon lookup and official review steps
* refactor(agent-preset): fold plugin artwork into cordis plugin development
* feat(agent-preset): require detail documents for creator bundles
* test(release): expect source and maps in master's new payload cases
* refactor(agent-preset): consolidate creator skills into dsh-doc
* feat(release): keep installed docs and source readable in Desktop
Desktop keeps the published src/ and docs/ trees inside app.asar, and
every model shell call receives DSH_INSTALL_DIR and DSH_NODE_EXECUTABLE
so an agent can read them through Electron's Node mode. Staging also
generates docs/upgrade-guide/README.md, an index of the upgrade guides
grouped by the release they upgrade from.
Minified client bundles print string literals with backticks: the
worker image packer now recognizes those chunk requests, client license
banners are prepended after minification so their text survives, and
the affected artifact tests accept either quote.
* docs(agent-preset): add dsh-doc install, upgrade, and fallback guidance
* docs(agent-preset): align dsh-doc archive reads and upgrade start versions
* docs(agent-preset): bound dsh-doc upgrade index reads and test documented commands
* test(desktop): read archived docs through the Windows PowerShell form
Regenerate the config catalog for the shifted shell-env source line.
* docs(upgrade-guide): correct dsh-doc session verification step
* test(agent-preset): use a block body for dsh-doc fixture cleanup
* test(desktop): wait for the GUI-subsystem Electron process from PowerShell
* docs(agent-preset): read the running DSH version from the installation
* build(release): publish source and maps from the message-quotes package
* docs(agent-preset): document the verified PowerShell runtime invocation
* docs(agent-preset): keep PowerShell runtime output streams and exit code
* docs(tool-cordis): describe the shipped dsh-doc skill
* docs(skills): load configuration guidance separately
* fix(skills): declare shared template details in every locale
* fix(skills): verify dsh-doc guidance through recorded sessions
* test(skills): retain current runtime reminder in dsh-doc replay
* test(snapshot): refresh terminal panel runtime context
Tool bodies that lack a calling Agent, an injected service, or an open
turn get one shared HarnessError with the TOOL_CONTEXT_UNAVAILABLE
code, instead of each tool throwing a plain Error. The stacked
per-area refactors adopt it at their call sites.
Refs #7869
A ToolGuard may return a ToolGuardDenial with optional ToolErrorInfo,
and a post-execute block may supply info; the registry records it on
the failed result. checkGuards now returns the normalized denial.
String guards and blocks without info keep their existing results.
Refs #7869
An ask that approval does not grant now denies with ApprovalError and
APPROVAL_REJECTED, APPROVAL_CANCELLED, or APPROVAL_UNAVAILABLE, so a
user rejection is routable on tool/result like other structured
failures. Model-facing reasons are unchanged.
Refs #7869
A started tool body that throws the signal's reason or an AbortError
without HarnessError identity now keeps its message and carries
AbortError / ABORTED, so the cancellation is routable like the
registry's own cancellation results. HarnessErrors and distinct
failures that settle during cancellation stay unchanged.
Refs #7869
`retainOnArchive`, a volatile field of the worktree service's Config, skips
cleanup admission while it is on; it is read at each archive, so a change
applies to later archives only. The service keeps its instance off the
generated Settings pages, and the Forge client draws the switch as the
`plugins.row.config` page of the bundle's `worktree` row, staged through
the shared Plugins form until one explicit save.
Docs: the worktree, Forge, and bundle READMEs, the archive-worktrees
upgrade guide, the config and client catalogs, and the
explicit-worktree-creation note.
user_terminal_list and user_terminal_read list terminals the user
controls, including Agent terminals the user took over, so their
descriptions now say so. The visible-terminal guidance states that
terminal_close on a taken-over terminal only removes it from the Agent's
list while its shell and processes keep running. Subagents are refused
by both tools, because their Sessions hold none of the user's terminals,
and the lines parameter is an integer schema. The README records the
sandbox policy refusal, the subagent refusal and the release behavior;
the tool catalog and the terminal-panel-tools snapshot carry the new
model-visible text.
* chore(deps): unify zod to 4.6.6
Raise all 88 workspace zod declarations to exact 4.6.6 and move @browserbasehq/stagehand to 4.4.0, which pins the same version, so the lock keeps a single zod copy. Test expectations and the Typert consumer's node types follow the new declarations; pnpm-workspace.yaml records the release-age exemption for [email protected].
* fix(stagehand): drop the deprecated extensionId option
Stagehand 4.4.0 ignores extensionId and discovers the installed extension itself, so the option no longer selects anything. Remove it from the provider config and worker plumbing, update the READMEs and regenerated config catalog, and adjust the attach test fixture.
* test(bench): widen the browser stream overlap window
The input-overlap probe can start after the paced synthetic reply finishes on a loaded runner, so double the per-chunk pace and re-record the README translation hash after the extensionId removal.
Move the removal step into `removal.ts` as `removeCheckout`, so the specs
drive the real function and intercept it through a module mock instead of
reaching a private method through an assertion to unknown. Classify
`CREATED_RECORD` as a peer-required Host export for the Forge panel, ship
the compiled `lib/types/*.js` that the `./types` subpath names, refresh the
worktree recorded session with the creation record the service now writes,
and regenerate the module graph.
sjw1231