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
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.
LegGasai noted that filtering prompt sections by tool:* prefix is a poor
heuristic: it conflates section naming convention with presentation
semantics and would incorrectly drop tool:structured_output. The executor
collapse already enforces the boundary — a model-direct native call is
rejected as UNKNOWN_TOOL regardless of what the prompt says, so filtering
the prompt adds no security and only risks losing useful guidance.
The tool:read/tool:bash/etc sections describe capability usage patterns
that apply to both native and code presentations, and keeping them does
not reopen the native direct-call path because the executor blocks it.
Only `docs/module-graph.*` conflicted, from master's new
`@deepseek-ai/dsh-message-feedback` package; generated, so taken from master
and regenerated.
Hook bypassed as before: the staged-pairing check hands an archived note path
to `verify-translation-pairing`. The full-corpus gate passes.
Generated docs (`module-graph`, `event-producer-consumer`) taken from master
and regenerated. The `cordis-inspect-jsdoc` golden likewise: master's copy is
the base, and this branch's `presentAs` per-scope rewording is re-applied on
top, since `cordis_inspect` renders that JSDoc into model-visible output.
Hook bypassed as before: the staged-pairing check hands an archived note path
to `verify-translation-pairing`. The full-corpus gate passes.
One assistant message carrying several subagent calls now overlaps them
under the rolling pool (maxParallelToolCalls) instead of serializing each
foreground delegation behind an exclusive barrier; results still commit
in model order.
tool-subagent declares isConcurrencySafe: () => true for every call form:
children work in their own sessions, a run never mutates the parent
session, and sibling workspace coordination already belongs to the model
for background, continuable, and workflow children. The former
serial-execution pin flips to a parallel pin plus a gated genuine-overlap
test, and the authored subagent-parallel snapshot pins the assembled
transcript (tool/call, tool/call, tool/result, tool/result) with
interchangeable twin children, which replay race-free under the existing
first-call binding and createdAt harvest ordering.
The provider seam now requires concurrent starts and continuable
preparations for distinct children to isolate operation-local state,
cancellation, settlement, and cleanup; a continuable gate test proves a
cancelled preparation leaves no Agent or durable Session while its
sibling persists independently.
Closesdeepseek-harness/deepseek-harness#1489
list_agents called a storage-only continuable child 'complete', inviting
parents to poll it for a result that will never appear there. Rename the
status to 'ready' — resumable, not terminal — and rewrite the tool
description to say a parent is told when a child finishes, so listing is
for recall, not polling.
A continuable child that stopped without reporting — an error, a token
ceiling, cancellation, teardown — left its parent nothing to act on.
The continuation manager now delivers an unconditional settlement
notice to the durable direct parent before releasing ownership, folding
consumed work (foldConsumedWork supersedes findLastMessageTurnEnd) so a
claimed-but-unrun prompt reads as aborted rather than completed, waking
an idle parent, steering a busy one, and never waking a closing tree.
A continuable child could finish without giving its parent a usable
result: the report tool was described as optional and its default
delivery stayed quiet. Install the child-scoped report obligation (tool
description plus the tool:report prompt section), default reportDelivery
to wakeup, bind shipped subagent_fork delegations to one-shot so fork
keeps its prefix reuse, and extend the ACP snapshot harness to pin child
system prompts alongside child tool schemas.
Yichen Jiang
imccyu
Tianyi Cui
pku-xht
Chinesezjc
Dudu-0223
Turtle
Hypatia May
creatixchu