The first OUT-OF-PROCESS subagent backend, proving the seam generalizes past the
in-process backends. @deepseek-ai/dsh-subagent-acp runs each child agent in a
spawned subprocess, driven over the Agent Client Protocol as the CLIENT — the
direction-inverted twin of the dsh-acp server bridge. Point the configured
command at the acp-agent example and the harness talks to its own process.
- Fresh process per run: start spawns, runs one ACP session (initialize →
newSession → prompt), dispose kills the subprocess and awaits its exit.
- Minimal client stub: advertises no fs/terminal; accumulates agent_message_chunk
text as the result output; auto-answers session/request_permission by a
configured policy (reject default / allow). No start-time capabilities (an
out-of-process child can't enforce the parent's depth/tool-filter); ignores
request.parent; injects only `subagents`.
- StopReason mapping (end_turn→completed, cancelled→aborted, …); result resolves
error/aborted on a child failure, never rejects (seam contract).
- Security: credential-shaped ambient env vars are scrubbed; the child's own key
is forwarded only via explicit config.env. A spawn-level error (ENOENT) is
captured and raced against the ACP drive so a bad command settles error rather
than crashing the parent.
Testing designed at every tier: keyless integration drives a scripted mock ACP
server subprocess (cancellation incl. the pre-newSession race and a
torn-pipe-after-cancel, permission auto-answer, non-message updates, spawn
failure, HMR, export shape) at 100% coverage; a with-key e2e drives the REAL
acp-agent example process (PONG + real file write, verified on disk) — the
harness driving itself. Snapshot coverage of an ACP child is deferred as
TODO(acp-subagent-replay) (each child is its own process with its own replay).
Stayed on @agentclientprotocol/sdk 0.25.1: the proposed 0.28.x bump only
deprecates the stable ClientSideConnection/AgentSideConnection API this layer
uses (33 sites incl. the server bridge), turning no-deprecated red across code
this PR shouldn't rewrite — that fluent-API migration is its own follow-up. The
backend needs nothing 0.28.x adds.
This completes the subagent seam stack (PR1 interface → PR2 in-process → PR2.5
snapshot infra → PR3 ACP); the seam RFC moves to implemented/, amended.
The snapshot tier was built single-session: dsh-llm-replay served calls from
one global positional cursor, and the harness harvested one session log. A
subagent runs as a second agent with its own session, so a parent→child
scenario could neither replay deterministically nor harvest the child's log.
This resolves the TODO(subagent-snapshots) deferral from the subagent RFC.
- Stamp the calling session id onto the model request: GenerateOptions.sessionId
(typed Branded<'SessionId'> to avoid the dsh-llm↔dsh-session cycle), set by the
agent loop from agent.session.id. Adapters ignore it; an llm/stream listener
routes by it.
- Key replay per session: dsh-llm-replay loads the parent log plus one per child
(childFiles / $DSH_SNAPSHOT_CHILD_FILES), derives a script per recorded session,
and binds each live (freshly-random) session to a recorded script by first-call
order — parent first (earliest createdAt, first to stream). Keys by WHO calls,
so it survives a future concurrent/backgrounded subagent; a global cursor would
not. An unrecorded extra session fails loud.
- Harvest every log: the harness collects all .jsonl across cwd buckets, ordered
primary-first (top-level, then children by createdAt), and RunResult exposes the
plural sessionLogs. The spec writes each back on record (session.jsonl +
session.<n>.jsonl) and diffs each against its fixture on replay.
- Wire the subagent seam + spawn + fork + tool into the acp-agent example (both
cordis configs) and add two nested scenarios recorded against the real API:
subagent-spawn (parent + 1 child) and subagent-multi (parent + 2 children, 3
sessions). Both replay keyless in the default gate.
A new RFC documents the design (docs/rfc/implemented/testing/). Single-session
replay is unchanged (a call with no sessionId is one anonymous primary session).
TODO follow-up: a dedicated branded-ids package could own the SessionId brand and
dissolve the cross-package cycle note; out of scope for this testing PR.
Two merge-blocking bugs in the shared in-process run driver, both rooted in
`readResult` scanning the whole child session and deriving the stop reason only
from `turn/end`:
- A pre-turn `cancel()` cleared the queued prompt before any `turn/end` was
logged, so the run settled `error` instead of `aborted`, violating the
`SubagentRun.cancel()` contract. The driver now tracks that a cancel was
requested and maps the no-turn case to `aborted`.
- A fork child whose own turn produced no `assistant/message` returned the
SEEDED parent's last message as a `completed` success. `readResult` now scopes
to the child's OWN events (after the seed prefix), so a message-less child
yields empty output.
Both fixes carry a regression test proven to go red on the pre-fix driver.
Also: correct the `SubagentRun.id` / event-payload docs (it is the child AGENT
id, not a session id — the backend mints distinct tokens); refresh the stale
`coding-agent` welcome string (subagent is now a tool); and replace the stale
`TODO(sub-agents)` "deferred" prose in the Agent interface, core.md, and
architecture.md with an accurate pointer to the realized seam.
The second PR of the subagent seam: the two in-process backends that run a
child agent on the same cordis context, reusing the agent factory's quiescent
AgentHandle teardown. Both register on ctx.subagents (PR1's named-provider
registry) and share one run driver.
- dsh-subagent-spawn: a FRESH child via ctx.agents.create — own session, the
parent's model by default (overridable), zero inherited conversation. Also
exports the shared in-process run driver (startInProcessRun): mint ids, stamp
cwd/parentSession-lineage/depth, drive the one-shot (send → whenIdle), read
the last assistant/message + turn/end reason, dispose to quiescence.
- dsh-subagent-fork: a child SEEDED with the parent's balanced completed-turn
prefix (the log up to and including its last turn/end), so the child inherits
context. The in-flight unbalanced turn is excluded — a raw seed would fail the
invariants replay. Proven: a regression test goes red if the boundary seeds
the open turn.
- Seam extension: CreateAgentOptions.seed, threaded through AgentLoop.createAgent
→ ctx.sessions.prepare({ seed }) (the primitive resume already used). This is
the fork-lineage path the TODO(sub-agents) markers anticipated.
- Depth: a merge-extensible AgentOptions.subagentDepth (0 top-level, parent+1 for
a child); the depthLimit capability refuses a spawn past request.maxDepth.
Tests: real-loop unit tests for both backends (mock MODEL only, real loop +
invariants), a multi-subagent test (one parent drives a fork AND a spawn child
then keeps working), and a with-key e2e (a real parent delegates via the
`subagent` tool to a real child that writes a file on disk — world-verified).
100% per-file coverage. The coding-agent demo wires the spawn backend + tool.
Snapshot coverage of nested agents is deferred to a stacked follow-up
(TODO(subagent-snapshots)): dsh-llm-replay is a single global positional cursor
that cannot route calls to a parent vs. a child on one context. Recorded in the
RFC's deferrals and a new AGENTS.md rule: designing a subsystem must design its
test infrastructure END TO END up front, verifying the snapshot/e2e harness can
express the new shape — a gap this plan hit.
A single try/catch around ctx.emit prevented a thrown subagent/start or
subagent/end listener from propagating, but cordis emit dispatches listeners in
a `.map(cb => cb())` that HALTS on the first throw — so a bad subscriber still
starved the listeners registered after it, violating the AGENTS.md
callback-boundary rule ("one bad subscriber must not starve the listeners after
it"). Resolve the listener callbacks via ctx.events.dispatch and contain each
call individually, the same per-listener guarantee BashExecutor.notifyTaskDone
gives its own listener set.
The two containment tests now register TWO listeners where the first throws and
assert the second still observes the event (start) and the settle (end) — a
regression that fails on the per-emit code (verified: reverted, watched both go
red, restored).
Address four findings from the first Codex review round:
- Contain subagent/start|end listener throws (emitContainedStart/End): a
thrown lifecycle listener could escape SubagentService.start() before the
caller received the live run to dispose it (a leaked child), and a thrown
subagent/end listener could surface as an unhandled rejection on the detached
result-settle hook. Both emits now log-and-contain, mirroring the agent
registry's agent/created|disposed containment.
- Make the model-facing tool name configurable (Config.toolName, default
subagent). The docs say to load dsh-tool-subagent once per provider to expose
multiple transports, but the hardcoded name made the second load throw a
duplicate-tool-name error; a distinct toolName per load is now required and
documented.
- Reach the per-file 100% coverage gate: tests for the subagent/end error
branch, lifecycle-listener containment, every stopReasonError arm + the
merge-extensible default, the multi-provider toolName path, agentOptions
forwarding, and the direct-apply schema-bypass fallbacks.
- Document the seam vocabulary in docs/core-data-structures/subagent.md with
verbatim type-equiv blocks + manifest entries, and link it from core.md (a
brand-new core/seam type the doc-sync gate cannot detect on its own).
Codex review of the trace-event fold found two merge-blockers.
Blocker #1 — format version. Folding usage onto assistant/message and removing
the standalone usage/error events changed the persisted SessionEventMap shape,
which per the AGENTS.md "bump the version and reject — don't migrate" policy
requires a backend to reject any non-current log. Centralize the version in an
exported SESSION_FORMAT_VERSION constant (dsh-session), read by both write sites
(Session constructor default, SessionStore.prepare header) and the coordinator's
load-time assertVersion check. The constant is pinned at 0: while unreleased the
on-disk format is unstable/pre-release, so breaking shape churn is absorbed at v0
(no monotonic bump until the first tagged release) and any non-0 log is rejected
on load — no migration. Update every test/fixture/doc that stamps a
currently-written header to the constant, bump the ACP snapshot fixture + golden
headers to v0, and keep the version-rejection test meaningful by switching its
bad value to a clearly non-current 99. AGENTS.md documents both the monotonic
(SQLite SCHEMA_VERSION) and pinned-0 (session log) pre-release stances.
Blocker #2 — restore the late turn-end warn. failTurn now sets the error reason
only while the turn is still open; once turn/end is appended (a throwing
agent/turn-end listener after closeTurn) the reason can no longer reach the
durable log, so the late throw is logged via ctx.logger.warn instead of
vanishing into a futile post-close assignment. A regression test asserts the
warn fires.
Also guard the normal-step assistant/message append with the same
content-or-usage condition as the max-tokens branch (a content-less, usage-less
step records no trace-only row), with a covering test.
The whenIdle() JSDoc cited "a closing ACP connection" as a non-owner that
awaits whenIdle(). That is false against the code: ACP OWNS its agent handles
and tears them down via rec.dispose()/handle.dispose() (quiesce() at
packages/ui/acp/src/index.ts:666-686), never whenIdle(). The only whenIdle()
consumers are tests (acp dispose/turns/edges specs, agent specs) — which is
genuinely why the primitive stays (a test harness programs against the seam),
but the contract doc must not claim a production ACP path uses it.
Replace the parenthetical with truthful non-owning observers (a test awaiting a
turn to settle, a monitor) and state explicitly that an OWNER does not need
whenIdle() because AgentHandle.dispose() already awaits the loop-exit promise.
- packages/core/agent/src/types.ts: the Agent.whenIdle() contract JSDoc.
- docs/core-data-structures/core.md: the type-equiv mirror (re-copied verbatim).
- Regenerate the cordis catalog (whenIdle source line shifted).
Codex's confirmation pass found the teardown-framing error went deeper than the
three prose spots already fixed: the whenIdle() JSDoc itself (and its mirrors)
claimed "the quiescence signal a teardown awaits ... a lifecycle owner disposes
the agent through its AgentHandle which ... awaits THIS". The disposer does not
call whenIdle() — it does `stop(); await agent.done` directly
(packages/core/agent-loop/src/index.ts:271). whenIdle() is the NON-OWNER
observation hook; owner teardown awaits the loop-exit promise (done) through
AgentHandle.dispose(). Reframe every copy accordingly:
- packages/core/agent/src/types.ts: the Agent.whenIdle() contract JSDoc.
- packages/core/agent-loop/src/agent.ts: the impl JSDoc.
- packages/core/agent/README.md and docs/core-data-structures/core.md (the
type-equiv mirror of the types.ts JSDoc — re-copied verbatim).
- docs/rfc/proposed/feature/2026-06-14-acp-agent-client-protocol.md:40 and :70:
owner teardown via AgentHandle.dispose(); a non-owner observing quiescence
uses the interface-level agent.whenIdle(), not hand-rolled agent/status.
- Regenerate the cordis catalog (whenIdle source line moved).
The session event vocabulary carried two standalone trace-only events that
were not load-bearing as separate records. Fold their facts into nearby
load-bearing events and delete the standalone variants.
- Token usage now rides on `assistant/message` as an optional `usage` field —
the assembled model output and its accounting travel together. The loop folds
`assembler.usage` onto the append instead of emitting a separate `usage`
event.
- The max-tokens path is the no-data-loss host: a step cut off with usage but
EMPTY content (e.g. only a dropped tool call) previously emitted a standalone
`usage`; it now records an empty-content `assistant/message { content: [],
usage }`. `deriveMessages()` skips empty-content assistant messages, so the
usage host never injects a spurious content-less assistant turn into the
provider transcript. A step with neither content nor usage appends nothing.
- An operational error's step number now rides on `turn/end.reason` for
`kind: 'error'` (`{ kind: 'error', step, message, code? }`) — the durable
turn outcome ACP and resume already consume. `failTurn` sets the reason
directly (no separate session `error` event). `agent/error` + logging are
unchanged for live diagnostics.
- No format-version bump: pre-release, no persisted data, so per the format
policy there is nothing to migrate or reject (the RFC's "refresh the format
version" criterion over-reached). `version` stays 1.
- ACP fixtures + goldens re-recorded (keyless replay): dropped standalone
usage/error lines, usage folded onto assistant/message, error step on
turn/end.reason.
RFC moved proposed -> implemented with an implementation note recording the two
scope refinements.
The public Agent handle exposed abort() (step-only) and cancel() (queue-aware).
No production caller used abort() — ACP maps session/cancel to cancel(), and
lifecycle owners tear down via AgentHandle.dispose(); the loop's own stop paths
abort their per-step AbortController directly. So abort() is latent generality
that keeps a private loop mechanic public.
RFC-premise correction: the public-agent-stop-surface RFC proposed removing
whenIdle() too. Implementation found whenIdle() load-bearing — a real
quiescence primitive with a deliberate loop contract (settle-without-transition,
the replacement-turn race) and ACP test consumers; its proposed replacement
("observe the running->idle transition") is exactly the async-state race
AGENTS.md warns against. So only abort() is removed; whenIdle() stays. The RFC
is amended on the way to implemented/ to record the narrowed scope, and the new
AGENTS.md "RFCs are proposals, not golden truth" principle (PR1) gets its
worked example.
- Remove Agent.abort() from the interface + the ReactLoopAgent impl; the no-arg
'aborted' default goes with it (cancel() keeps its 'cancelled' default).
- Migrate tests: empty-queue abort() -> cancel(reason); the two review-fixes
tests whose subject is the in-flight step's AbortController drive that
controller directly via the private currentAbort field (cancel() would clear
the inbox and destroy the queued steering one of them proves survives a step
abort). The no-arg-default test is dropped (cancel()'s default is already
covered in cancel.spec.ts).
- Resulting public stop surface: cancel() + whenIdle(). Update agent/agent-loop
READMEs, architecture.md, core.md type-equiv, the extension cookbook, the
lifecycle RFC (short note), and the proposed ACP RFC.
Implements docs/rfc/implemented/simplification/2026-06-20-public-agent-stop-surface.md
Type-only change (brands are zero-cost casts; no runtime/wire impact). Closes
the two gaps in the "brand ids that cross package boundaries" policy and fixes
the dependency direction so a capability package never pulls in an unrelated one.
- Extract the `Branded<B>` primitive into a new standalone type-only package
`@deepseek-ai/dsh-brand` (packages/util/brand) with no harness-package deps.
dsh-llm keeps its owned CallId but imports Branded from dsh-brand; dsh-session,
dsh-agent, and dsh-bash all import Branded from there. dsh-bash depends on
dsh-brand ALONE — never on dsh-llm or dsh-session (the architectural fix: a
generic execution backend must not couple to the LLM or session vocabulary).
- Mint BashTaskId + OwnerToken in dsh-bash and thread them through BashTask.id,
the get/ownerOf/list/readOutput/kill seam, the bash-local generation site, and
the dsh-tool-bash validate/access surface. OwnerToken is a DISTINCT brand from
SessionId so the seam stays decoupled; dsh-tool-bash is the single boundary
that casts SessionId -> OwnerToken.
- Brand at the SOURCE, not via mid-pipeline casts: agent-loop's Config types
agents[].id as AgentId and resumeSessionId as SessionId, so the brand enters
at the config boundary and the inner create()/resume casts disappear (only the
genuinely-new per-run session-id string is cast).
- Stop brand erosion: propagate CallId/SessionId/AgentId to the registry/store
Map keys and public params/exports (SessionStore, AgentRegistry + factory
options, the ACP session-id surface + ToolPresenter CallId map, the
persistence coordinator, invariants pendingCalls, the pi-ai tool-call maps).
- Docs: document BashTaskId/OwnerToken in bash.md (type-equiv re-pasted), point
the Branded type-equiv at dsh-brand, fix stale param types in the session/
agent/bash READMEs, regenerate the cordis catalog + module graph.
Implements docs/rfc/proposed/architecture/2026-06-20-branded-ids.md
The original prune removed BashExecutor.get()/.list() too, but each is a
one-line accessor over the executor's already-tracked tasks map, and removing
them forced dsh-tool-bash's tests onto a ~35-line onTaskDone completion-tracking
harness just to replace the one-line ctx.bash.get(id) lookup. Per the AGENTS.md
"RFCs are proposals, not golden truth" principle, that disproportionate
migration cost is evidence the methods earn their keep — a test harness IS a
consumer programming against the seam.
Restore get()/list() (seam + LocalBashExecutor impl + the bash tests that used
them, dropping the doneFor/trackCompletions scaffolding). The persistence
has()/delete()/deleteStored removal stands — it had only contract-test callers
and no test-ergonomics cost. The RFC is retitled persistence-only with an
implementation note recording the bash revert.
Codex review of the PR2 diff caught doc/comment sites doc-sync does not gate
(core-data-structures prose) and a latent test-helper race:
- docs/core-data-structures/persistence.md + bash.md, sqlite README, and two
source comments (coordinator.ts, jsonl.spec.ts) still listed the removed
has/delete/get/list methods — updated to the surviving four-method
persistence surface and the get/list-free bash seam.
- doneFor(ctx, id) attached its onTaskDone listener lazily, after the task
could already have closed (e.g. `true`), so it could miss the completion and
hang. Replaced with trackCompletions(ctx): one eagerly-installed listener
(mounted in setup() before any task starts) records every completion, and
doneFor resolves immediately for an already-finished task or on completion
otherwise. Race-free, and there is no get-by-id seam left to poll instead.
The LLM service exposed three call surfaces (stream/streamBlocks/generate) but
the only production consumer — the agent loop — uses stream() exclusively,
feeding raw chunks through its own BlockAssembler for replay fidelity. Drop the
speculative convenience surfaces and the registry-change event that no listener
consumed, leaving stream() as the single model-call contract for both
production and tests.
- Remove LlmService.streamBlocks() and generate(), the llm/generate waterfall,
and GenerateResult.
- Remove the llm/adapter-change event (declaration + emits) and the
listener-throw rollback ordering that existed only to protect it; keep the
HMR rollback disposer.
- Remove BlockAssembler.flushReady()/flushRemaining()/result() and the flushed
cursor — the streaming-flush slice existed only for streamBlocks().
- Adapter tests drive a stream()+BlockAssembler helper (tests/assemble.ts)
instead of generate(), exercising the same path production uses.
- Land the AGENTS.md "RFCs are proposals, not golden truth" principle and move
both RFCs proposed -> implemented.
Implements:
- docs/rfc/implemented/simplification/2026-06-20-drop-unconsumed-llm-adapter-change-event.md
- docs/rfc/implemented/simplification/2026-06-20-drop-unconsumed-llm-assembled-surfaces.md
Merge brought in the RFC-classification reorg and two new doc gates;
rewrite every drifted packages/<name> cross-link (Markdown link targets,
moved-README relative depths, and .ts comment paths) to the grouped paths.
Add two doc-sync/hygiene gates so the manual checks this restructure
needed become automated:
- verify-package-paths.ts: flags a packages/<path> reference (in Markdown
or a .ts comment/string) that does not resolve AND names a real package
in a segment — i.e. a stale path to a MOVED package. A path naming a
non-existent package (a forward-looking proposal) is left alone, so it
applies uniformly across proposed/implemented/rejected.
- check-workspace-constraints: assert the packages/<group>/<pkg> depth-2
shape (group dirs carry no package.json; no flat or over-nested
packages). Group names stay open; only the shape is fixed.
Add a second axis to every RFC — its class (feature, bug-fix,
simplification, architecture, process, testing) — encoded in the path
as docs/rfc/{lifecycle}/{class}/file.md. The folder is the label, so
the closed set is enforced by structure rather than a parsed field.
Two new doc-sync gates back it:
- verify-rfc-classification: every RFC sits in a valid class folder and
the README index lists it under the matching lifecycle→class heading.
- verify-doc-refs: every docs/*.md path cited in a packages|examples TS
comment resolves — closes a drift class verify-md-links can't see, and
catches the four comment refs this reorg moved.
The README gains a Classification section explaining the taxonomy and
per-class index sub-sections. A self-referential process RFC records why
the scheme is path-encoded and gated.
Review found verify-type-equiv only scanned docs the manifest already named, so
a type-equiv block in an unmanifested doc was silently skipped — defeating the
1:1 guarantee. Scan all docs in the markdown glob scope instead, so an orphan
block in any doc is caught. Also parse `abstract class` in blockSymbol (matches
sourceDeclaration's class support).
persistence.md listed the SessionPersistence surface as create/append/load/list;
the abstract service also exposes has/delete. AGENTS.md's doc-sync command
summary omitted verify-md-links and verify-type-equiv.
A new docs/core-data-structures/ folder: a self-contained core.md defining what
counts as a "core" data structure (the agent-loop spine) and covering the spine
vocabulary, plus per-seam sub-pages (llm-streaming, session, persistence, tools,
bash). Type definitions are pasted verbatim via `ts type-equiv` blocks and
drift-checked by verify-type-equiv. Cross-linked from architecture.md; the
`ts type-equiv` mechanics are documented in development.md.
Tianyi Cui