The group's convention is package suffix == provider default
(subagent-acp/'acp', subagent-spawn/'spawn', subagent-fork/'fork'), and the
provider default became dsh-sdk in the last review round — so the package
follows: @deepseek-ai/dsh-subagent-dsh-sdk at
packages/subagent/subagent-dsh-sdk, plugin name subagent-dsh-sdk,
diagnostics prefixed subagent-dsh-sdk:. The dsh echo has precedent
(dsh-llm-deepseek). Directory, fixture path, knip/tsconfig/examples
registrations, catalogs, READMEs (en+zh), and the Agent Note follow; the
sdk-client dispose ladder moves to its own module (src/dispose.ts) with the
deterministic FakeChild tier tests restored alongside it.
Master's #660 replaced dsh-subagent-subprocess with the dsh-subprocess
capability seam (ctx.subprocess + scrubbedParentEnv, tree-scoped teardown)
and moved subagent-acp onto it. Convergence for this branch's packages:
- The shared out-of-process provider vocabulary this branch had grown in
the deleted library (NO_START_CAPABILITIES, assertPositiveFinite, cwd
resolution, settleRunResult, subprocessRunHandle) moves into the subagent
seam package as out-of-process.ts — it enforces subagent-seam contracts,
not process mechanics, and both out-of-process backends now import it
from there (subagent-acp keeps master's shape otherwise).
- subagent-sdk spawns THROUGH the SDK client (the subprocess README's
documented exception for SDK-managed transports) and now applies the
seam's scrubbedParentEnv() + explicit-env merge in place of the deleted
buildChildEnv.
- sdk-client inlines the EOF→SIGTERM→SIGKILL ladder as private helpers (it
runs outside any harness context, so it cannot ride ctx.subprocess).
- The child harness fixture gains the now-required dsh-subprocess-local
entry for bash-local; the fixture cordis.yml keeps exercising the
shipped provider default.
The live-link E2E intentionally creates a project without a lockfile. Prevent ambient CI=true from making Yarn 4's first install immutable, while preserving the CI environment for every other behavior under test.
Keep restore-only and self-hosted stores outside pnpm/action-setup's replaceable install directory, resolving the configured path for every paired cache. Pin Yarn 4 for the generated-project e2e so runner images without Corepack do not fall back to Yarn Classic, and update the bilingual decision record.
- subagent-sdk: the default registry name becomes `dsh-sdk` (the bare
`sdk` read ambiguously in configs); READMEs, config catalog, fixture,
and suites follow. The Loader fixture now omits providerName to exercise
the shipped default end to end.
- loader-composition.e2e: two full harness runtimes boot in sequence, so
the default 30s loader-smoke window times out under host load; raise the
subprocess deadline to 120s with matching vitest headroom (the
real-model.e2e precedent).
sha256sum-or-shasum selection happens in preflight (macOS ships shasum,
not GNU coreutils) with a clear missing-tool hint; provisioning now waits
for every child and reports each named status before exiting, so a fast
failure can neither hand the EXIT trap a scratch tree that a still-running
install is writing into nor bury the root cause behind a downstream
symptom.
- api: resolve a relative workspace cwd to absolute before the handshake —
the child spawns relative to the parent cwd, but the wire cwd is resolved
again inside the child, so a relative value double-resolved
(worker -> worker/worker).
- api: make the documented handshake retry real — HarnessClient.close() is
permanent, so a failed initialize now reaps the runtime and swaps in a
fresh client; DeepSeekHarness.close() is terminal and stops the respawns.
- api: validate session.event envelopes, assistant/message content, and
session.finished reasons at the wire boundary — a malformed runtime
surfaces as SdkProtocolError instead of type-invalid TurnResult data or a
TypeError out of finalResponse.
- client: a throwing subscribe() filter fails and detaches only its own
subscription (normalized to Error); sibling fan-out and the transport read
loop are undisturbed.
- client: NotificationSubscription.close() drops its queued notifications,
matching its documented contract; runtime-death fail() still leaves
already-delivered items drainable.
- client: subscribe() after close()/runtime death returns a born-failed
subscription so next() rejects instead of parking forever.
- client/transport: bounded requests abandon via AbortSignal — the transport
drops the pending entry at timeout, so repeated bounded calls against a
hung method retain no per-call state.
One test per finding; per-file coverage stays 100% on both packages.
AGENTS.md and the Wine Agent Note pair state the run policy: CI owns the
everyday win32 signal; run check:windows-wine only while investigating a
known Windows-related failure, and pre-push selection never includes it.
scripts/wine-windows-gates.sh owns the whole lane — workspace snapshot,
concurrent Node/wineboot/install provisioning, entrypoint resolution, the
vue link, smoke, and the two concurrent blocking gates — so the ci.yml
windows job shrinks to runner provisioning plus one script call, and
pnpm run check:windows-wine runs the identical script locally. The local
path never mutates the working tree, persists the Wine prefix and the
checksum-verified Windows Node zip under .cache/wine-windows/, and falls
back to the cached zip offline. AGENTS.md ceiling 1680 -> 1695 for the
one-line command entry.
- The clone exclusion is now '.runner*' (catches .runner_migrated /
.credentials_migrated, which GitHub writes on migrated runners and
which equally trip the already-configured refusal — hit live during
the 2026-07-27 org-group migration).
- The problem statement now states the verdict runs on ubuntu-latest,
scopes the switch to enterprise Linux-pool outages, and names the
residual dependency: node-compat/python-sdk/windows stay standard-
hosted, so a broader hosted-capacity failure still blocks the
verdict, with Windows the only leg lacking any in-house substitute.
Static gate green locally: 32 passed, 0 failed.
