The transport schema requires both `code` and `description`, but the tool
description and both SDK instruction flavors described the call as passing a
program. `description` was reachable only through the parameter schema, so a
model following the prose emitted `{code}` alone and lost the whole written
program to an INVALID_ARGS rejection.
The length and format guidance stays in RUN_CODE_DESCRIPTION_PARAM_DESCRIPTION
alone, so the schema and the prompt cannot drift.
Fixes#2426
`collapses()` read `defaultMode`, so the collapse only applied when the
DEPLOYMENT was `code`. An agent handed `code` by an agent preset under a
native default announced `[run_code]` on the wire and still executed a
model-direct native call -- the bypass this collapse exists to close,
reopened for exactly the composition `dsh-agent-tool-mode` produces.
`modeFor(scope)` is the same resolution `wireSchemas` and the SDK section
already use, so presentation and execution cannot disagree, and a mode
inherited from a standing preset scope collapses like a declared one.
The per-agent and preset tests asserted only the wire, which is why the
regression passed them. They now assert through the executor: the body
never runs, the call resolves UNKNOWN_TOOL, and the native sibling beside
it still executes.
The executor collapse landed without telling the model it exists. Every
tool contributes its own guidance section naming its tool, none of them
qualify how that tool is reached, and they all render before the SDK
(orders 100-199 against SDK_SECTION_ORDER 150), so the prompt said "Use
the read tool" eleven times and never said only run_code is callable.
A real session shows the consequence: the model emitted a native call,
read `unknown tool "read"` for a tool the same prompt declares, and
concluded the deployment was inconsistent rather than routing through
run_code.
The registry now contributes `tools:code-only` at order 99 -- ahead of
the guidance band -- stating the rule, registered wherever `tools:sdk`
is and rendering empty outside an effective `code`. `both` renders it
empty because its native calls do execute, which is also why
both-mode-turn no longer shares code-mode-turn's expected prompt. The
denial itself now names the route back, since a bare UNKNOWN_TOOL for a
declared tool is what misled the model.
Machine-produced by `pnpm run rescope-vendor --apply` plus the regeneration it
prints: `pnpm install` for the lockfile, `pnpm run gen-third-party-notices`,
`verify-translation-pairing --write` for the touched bilingual pairs,
`gen-doc-graphs`, and one typert snapshot whose ids embed character offsets.
`pnpm run rescope-vendor --check` verifies the result.
Renames nine vendored packages (cordis, cosmokit, schemastery and the six
@cordisjs plugins) and every reference that resolves them: manifest names and
dependency keys, module specifiers including declare-module merges, cordis.yml
plugin names, tsconfig paths, every Markdown fence, and `docs/` prose.
Directory names, upstream versions, and dependency ranges are unchanged, so
vendor/README.md still reads as an upstream snapshot; its manifest table gains
an upstream-name column so THIRD_PARTY_NOTICES keeps MIT attribution pointed
at each fork's origin.
The tutorial tier follows the rename end to end: its yaml fences named plugins
the Loader can no longer resolve, its `ts ignore-check` fences disagreed with
the compiled fences beside them, and its prose quoted both. The contracts that
told readers to keep upstream names — the root convention and the vendoring
cookbook's tree comment and manifest invariant — now say to rescope instead.
Two rules read `@deepseek-ai/` as "another workspace plugin": the client bundle
purity gate now names the vendored libraries a browser bundle inlines, and the
files where a bare `cordis` is an agent-preset id keep that product data.
modeFor read only the exact scope's layer, so a `code`-preset session
advertised the native catalog: the mode is declared on the preset's STANDING
scope, and the agent only parents to it. Nearest scope wins along the chain —
the mode decides what the model SEES, which is the class of fact the chain
inherits. Caught live (the model politely computed with bash and said
run_code was not in its list); the chain test pins it.
The client half of the label fix: the create echo and the session-added
frame's agentPreset now reach the session list (newest wins in the upsert —
every producer of the field reports the CURRENT composition), and a confirmed
blank-session switch publishes through the new ISessions.noteAgentPreset, so
the header label moves with the composition instead of waiting for a reload.
Code Mode was a deployment-wide field on the host `tools` row: a
deployment ran every session that way or none. The obvious product
shape — 代码模式 beside 标准/极简/创造 in the preset picker — had
nothing to hang on.
The registry itself cannot move into a preset; the agent loop's
scheduler, the api-proxy's presenters, and every tool plugin are its
consumers. So split the registry from its projection: `presentAs(mode)`
writes one cell on the calling agent's scope layer, exactly as
`restrict()` does, and the three reads that decided presentation take
that scope's mode instead of the service's. The config `mode` becomes
the default agents shadow rather than a process-wide fact.
Two consequences are load-bearing. `run_code` now enters a view only
for scopes whose own mode presents it — a native agent must not find it
dispatchable because another agent in the process does — and the
reserved name holds whatever the configured mode, since any agent may
select a code mode later.
`dsh-agent-tool-mode` is the row a preset carries to declare this. A
code mode waits for the host's `codeRuntime` rather than assuming it,
so a runtime-less deployment fails the preset at mount, naming the
row, instead of at the session's first request.
The shipped `code` preset is `standard` plus that row, ordered second.
"defineTool compiles an object root, so the annotation is a bare
TypedDict class name that opens nothing" is a false universal:
parameterSchemaSpecToJsonSchema compiles an OPEN object root, so an
empty parameter table and one with unrepresentable field names both
degrade to dict[str, Any], which opens one bracket. The conclusion the
sentence carries is unaffected -- 1 or 2 against a 182 cap -- so say
"a bare TypedDict class name or dict[str, Any], neither of which
carries a chain", in the JSDoc and the test comment that copied it.
pyScalar's docstring said the subscript tool-name comment quotes
"through the same call". It quotes through its own JSON.stringify call
site in renderToolsSdkPy and never reaches pyScalar, which only takes
const/enum scalars. Same function, different call site.
The mode-'both' test attributed assembly.tools to the public schemas().
That projection is wireSchemas, wired at ctx.systemPrompt.tools.
pyScalar's docstring named only the two code points CPython refuses
anywhere in source. A bare quote, a trailing odd backslash, and a bare
LF/CR break the Literal line just as fatally, and JSON.stringify is what
covers those too. The argument also leaned on an unstated coincidence:
every escape JSON.stringify can emit is a Python escape for the same
character, which is why the emitted text both parses and decodes back to
the declared value. Say both, and assert the second class.
"statement head" does not describe `case`, whose clause block is not a
statement. Split the positions three ways.
Add the mode 'both' by python assembly, pinning the mode-by-language
matrix rather than leaving it to the shared code path.
A description was emitted above the `async def`, where Python treats the
first string as the `Tools` class docstring and every later one as a dead
expression — leaving each method undocumented in the model's only source of
tool semantics. Emit it as the first statement of the method body instead.
Also names the known languages in the run_code flavor guard (the reachable
rejection, symmetric with the SDK_RENDERERS guard) and corrects three doc
claims: the code-runtime group README no longer calls the generated SDK
TypeScript, the base Code Mode note states its serial dispatch in past
tense, and the tools README points at the rationale the language-dispatch
note actually carries.
Address ds-review-bot v5/v6 review on the Python SDK renderer:
- resolveFlavor now takes a peekRuntime() reader: undefined (no runtime,
the doc-catalog harvest) degrades to the TS flavor, but a mounted
runtime whose language is absent from RUN_CODE_FLAVORS fails loud. This
removes the try/catch that silently swallowed the invalid-language path
and drops the /* v8 ignore */ that hid the flavor guard from coverage;
wireSchemas validates the runtime before projecting schemas so the
renderer-table rejection stays the canonical assembly error.
- py-types RESERVED drops the soft keywords match/case: they are legal as
TypedDict fields and methods, so keeping them needlessly degraded
common search/regex arg objects to dict[str, Any].
- py-types treats an object with omitted properties as {} like the unified
validator and TS renderer do, so a closed empty object declares an empty
TypedDict instead of a permissive dict[str, Any].
- README: symmetric jsonSchemaToPy->Any note; a stale zh SDK bullet and
limitation corrected; link the service-wide-language limitation to its
Agent Note.
Code Mode generated only a TypeScript SDK and rejected any runtime whose
language was not "typescript". Add py-types.ts (jsonSchemaToPy /
renderToolsSdkPy) and select the SDK-section renderer and the run_code
schema flavor by ctx.codeRuntime.language through two parallel tables
(SDK_RENDERERS, RUN_CODE_FLAVORS), read with Object.hasOwn and failing
loud on a language with no renderer. The tool layer depends only on the
code-runtime seam's language field, so it lands independently of the
Python protocol and backend.
The staged-promotion machinery (concludingParents keyed by parent token
plus a pendingParentConclusions staging map) spread one fact — this call
concluded the turn — across three registry-side collections with manual
cleanup. Align it with how additionalContexts already crosses the same
boundary: concludeTurn() marks only its own execution, the marker rides
that execution's successful result (ToolExecutionFailure types
concludesTurn as never, so a policy-converted failure sheds it with the
type), and the composite that owns the nested dispatch forwards it —
Code Mode's binding does so beside its existing context forwarding.
The registry loses both parent-keyed collections and the promotion
block; the propagation decision moves to the owning boundary; the
structured-output consumer's own two-phase commit is untouched.
Conflict resolution: drainDispatches composes both sides — the reworked
single-lane driver's drive() (ordered stages, mid-flight commit drain)
followed by this branch's logWork drain, so every shaped settle append
still lands inside the open run_code turn.
Responding to ds-review-bot round 2 on #658 (three critical findings, one
warning — all rooted in the pump/commit split racing ordered stages):
- ONE driver lane now owns every ordered stage: the start append, prepare
(pre-execute/guards), and the head-of-line commit (post-execute, context
deferral, settle append). start() is awaited before the next entry can
start, so concurrent submissions can no longer run pre-execute pipelines
concurrently; only the around-dispatch/body stage overlaps, matching the
native loop's fillPool sequencing.
- An exclusive call's barrier now holds through its COMMIT: later starts
wait for the exclusive pipeline (post-execute included) to finish, the
native exclusive-group semantics.
- drainDispatches() awaits the driver run itself, so a commit already
mid-flight when the program returns is drained before run_code closes
the turn — the settle event and deferred contexts land inside it.
- maxParallelSubCalls is resolved and validated at construction (positive
integer), so direct construction can no longer wedge the pool with 0.
New tests: overlapping-submission ordered-prepare, barrier-through-commit,
drain-mid-commit, cap rejection. 96 keyless snapshots replay unchanged;
Agent Note updated (both languages).
CI's full-tree coverage flagged three untaken paths this PR introduced:
- shapeDispatchLog's catch (a throwing tools/code-dispatch-log listener
must be contained — the settle event logs the unshaped content);
- the spill listener's flatten-decline arm (non-text sub-result content
passes through unchanged);
- the generated scope-key extractor row for tools/code-dispatch-log
(registered in the scope invariant matrix like the other tools events).
The coverage gate flagged three untaken paths in the bridge scheduler:
- the exclusive-head inFlight re-check was dead (the shared guard above
already returns for an exclusive head with any in-flight sibling) — removed;
- the commit-cursor undefined-dispatched break was structurally unreachable
once entries join commitQueue only after start() ran synchronously —
reordered the pump so the invariant holds by construction, annotated;
- the finish (final-result) commit arm and the pump re-entry guard gain a
covering test (throwing tools/pre-execute listener) and a defensive
annotation respectively; mid-run unregistration test renamed to match its
actual post-result settlement path.
Also covers the direct-construction maxParallelSubCalls default (index.ts)
and commits the regenerated persistence catalog for the new dispatch pair.
ds-review-bot findings on the parallel bridge: sub-dispatches now run
through the registry's staged scheduler view (the loop's own seam) —
ordered prepare (pre-execute/guards) at submission-ordered start time,
only the dispatch/body stage overlapping, and a head-of-line commit
cursor running post-execute, context deferral, and the settle event in
submission order (new spec pins post order + context order under
out-of-order completion). Queued dispatches reclassify via
executionMode() immediately before starting, so a registry mutation
while queued flips them exclusive (native lazy-reclassification
semantics). Config and tool catalogs regenerated; the tool-catalog
metadata now names the start/settle pair and the scheduling contract.
The bridge replaces its serialization queue with a pool that reuses the
native concurrency contract: submissions classify through
registry.executionMode (fail-closed isConcurrencySafe), start strictly in
submission order, overlap up to the validated maxParallelSubCalls config
(default 10; 1 restores serial), and exclusive calls drain the pool, run
alone, and bar later calls. Each started sub-call logs a
tool/code-dispatch-start event at pool entry; the existing
tool/code-dispatch settles the pair (started ⇔ settles exactly once;
abandoned queued calls log neither). SDK prompt guidance now states the
true Promise.all contract — re-recorded across every code/both-mode
snapshot (plus the stale cordis-dynamic-toolchain fixture gaining the
required description arg).
Client: CodeSubCall widens to RunningToolCall | ToolResultNode — starts
land the running shape (rows wear the native running ring), settles
replace in place preserving start order, callTime pairs to the start
time. Fixture emits start/settle pairs; jsdom pins the running sub-row;
runtime specs pin in-place settlement and out-of-order completion.
run_code gains a required bash-style description parameter: presentCall
titles the card with it and moves the program to rawInput, so every
surface gets a readable label. tool/code-dispatch now logs each
sub-call's complete content/isError (the tool/result vocabulary),
replacing the bounded resultSummary and deleting the summarize/cwd
machinery — a UI renders sub-calls through the identical path as native
results. The dsh config tree mounts the worker code runtime and reads
DSH_TOOLS_MODE (temporary seam until per-session mode selection lands).
Session format stays v0 (pre-release churn). Code-mode ACP/TUI fixtures
re-recorded; TUI presenter pin refreshed; catalogs regenerated. Keyless
web smoke pins the code-mode wire contract (tools=[run_code] + SDK
prompt section).
