Replace send/steer/inject with one Agent.send primitive over the
(target × wakeup) matrix; followup/steer/inject become fixed-preset
alias methods on the now-abstract Agent class. Coalesce context/message
into user/message (injected context is a non-user source). Replace
agent/queued with agent/inbox/enqueue/dequeue/discard, add cancel
keepInbox, and add a FIFO-conservation invariant.
Drain idle injection flushes before agent teardown, snapshot approval and subagent provider inputs, and gate subagent lifecycle events on real child readiness. Align the RFCs and generated contracts with the hardened behavior.
Every subject-extractor row of the invariants carrier table is exercised
with a matching and a foreign-keyed carrier; the HMR re-apply seed path
(sessions of agents that predate the plugin are marked started) is pinned;
the scoped tool-provider disposal, plural restrict() validation, singular
scopeHost absentee, tool-subagent passthrough, stale-stage drop, and
disposing-parent spawn (INACTIVE_EFFECT, no orphan) each gain their test.
Two genuinely defensive branches carry justified v8-ignore markers.
SubagentStartRequest gains persona (capability-gated like toolFilter); the
in-process driver composes the child's scoped world in the factory's setup
window — persona as a scoped shadowing deployment:persona section,
toolFilter as a scoped tools.restrict() (loud unknown-name validation),
outputSchema as the scoped structured runtime. spawn/fork now advertise
every start-time capability; ACP stays all-false. A parent-scope teardown
effect links each child to its parent through the memoized handle, so a
disposed parent reaches its whole subtree even if the delegating tool's
finally never runs; subagent/start|end dispatch in the delegating parent's
scope.
structured.ts loses the placeholder schema, the final-assembly swap/strip,
the refcounted root runtime, and the WeakMap state: each child registers
its OWN capture tool (real schema), instruction section, and enforcement
listeners on child.ctx, riding the child's fiber. The commit listener is
call-keyed (a stale stage from a short-circuited post-execute chain is
dropped, never promoted on a later call), and one scoped prepend re-assert
listener preserves the final-assembly guarantee against a stripping global
listener.
tool-subagent gains persona/toolFilter/maxDepth passthrough config —
deny-listing the delegation tool (or maxDepth) is how a deployment bounds
recursion; the omitted-toolFilter schema key is forced absent (a
materialized {} would mean an empty allow-list, i.e. deny-everything).
The Codex simplification concern plus the duplication comment on the spawn
apply, resolved by deletion: the backend-lifetime holds are gone, so the
runtime registers at the first structured run and disposes when the last
settles — a deployment that never passes outputSchema carries no always-on
global state, and there is no per-backend acquisition block left to extract.
The driver spec now drives an INLINE spawn-shaped provider over
startInProcessRun, which removes the spawn/fork devDependencies (the
test-only workspace cycle); plugin-level structured coverage moves to the
backends' own specs (capture through the shipped plugin, mid-run backend
unload, seeded fork capture). tools.md, the driver README, and both backend
READMEs describe the run-scoped lifetime; the module-graph regenerates
without the cycle edges.
Two human review directives:
- No re-prompt. A structured child that finishes a turn cleanly without
calling structured_output settles error to the parent immediately —
readResult already carried that mapping; the nudge loop only delayed it.
Deletes the loop, its cancellation-window guard, STRUCTURED_OUTPUT_NUDGE,
and the structuredNudgeRetries Config on both backends.
- FIXME in the structured module doc: per-agent/per-session tool registry and
prompt assembly would dissolve the final-assembly enforcement dance (the
placeholder tool, the swap, the strip, the global-registration lifetime).
Carved out of #170 per review feedback — the foundation the workflow tool
builds on, now standing alone on master:
- dsh-tools: the structured-output JSON Schema subset (StructuredOutputSchema,
assertSupportedOutputSchema, validateStructuredValue) — rejects loud outside
the enforced subset, listing every violation
- dsh-subagent: SubagentStartRequest.outputSchema / SubagentResult.structured
become a real capability; the service rejects a schema'd request whose
provider lacks it
- dsh-subagent-inprocess: the shared structured runtime — one global
structured_output capture tool, a prepend final-assembly listener that
strips the placeholder for plain agents and swaps in the run's own schema
(plus the calling instruction as a trailing section) for structured
children, an agent/turn-continuation veto once captured, and the
capture/nudge loop in the run driver (structuredNudgeRetries, cancellation
honored mid-nudge); lifetime refcounted by backends and live runs
- subagent-spawn / subagent-fork flip outputSchema: true
One deliberate divergence from the #170 revision: the backends do NOT add
'tools' to their plugin inject. Doing so deferred their apply past the todo
plugin, and the delegation tool mirrors provider lifecycle — so the
model-visible tool order of every existing prompt changed, invalidating every
recorded snapshot fixture. The runtime now gates its capture-tool registration
on tools availability itself (sync when live, a scoped inject fiber when the
Loader starts the backend first), keeping this PR byte-invisible to existing
transcripts: all 35 snapshot scenarios pass against master's fixtures
unchanged.
Review round 2 (tianyicui inline comments):
- dsh-system-prompt itself registers the harness:identity (-100) and
deployment:persona (0) sections — they must survive a swapped loop
plugin, so they leave dsh-agent-loop; the persona text is the plugin's
own validated 'persona' config. The model/cwd variables STAY on the
loop: runtime facts of the agents it drives.
- AgentOptions.systemPrompt is deleted with all its forwarding plumbing:
the app configs' systemPrompt keys become 'persona' routed through
dsh-agent-core (schema = z.intersect of the owners'), the ACP bridge
and tool-subagent stop carrying persona configuration, and subagent
children now render the deployment persona like every other agent.
- Example personas drop transport/interface trivia (ACP, CLI) — facts
irrelevant to the model.
- Root CONTEXT.md removed (not idiomatic); its persona definition was
wrong under the new ownership anyway.
- Docs, READMEs, the prompt-variables RFC, and generated catalogs
updated; new loop test pins the assemble-waterfall escape valve
(an emptied assembly sends NO system field).
The shared run driver lived inside dsh-subagent-spawn, so the spawn package
carried fork-aware seeding logic and dsh-subagent-fork depended backward on
dsh-subagent-spawn — the two in-process backends were not independent.
Move the driver (startInProcessRun, depthOf, SubagentDepthError,
InProcessRunOptions) into a new pure-library package
@deepseek-ai/dsh-subagent-inprocess that registers nothing. spawn and fork now
both depend only on that driver and neither knows about the other; spawn no
longer re-exports it and fork no longer imports from spawn.
Also wire BOTH backends in examples/coding-agent/cordis.yml (config-only): load
dsh-subagent-spawn + dsh-subagent-fork + two dsh-tool-subagent instances with
distinct toolNames (subagent → spawn, subagent_fork → fork), demonstrating that
exposing multiple transports needs no code change.
A request signal aborted BEFORE the run starts never fires an `abort` event
(`addEventListener` only fires on the transition), so the backend-level bridge
missed it and ran the child to `completed`. The driver now checks
`request.signal?.aborted` at the top of the result path and settles `aborted`
without running the child. Regression test proven red on the pre-fix code.
Also refresh two stale RFC prose blocks the round-1 fix left behind: the
subagent RFC's Problem statement (cited the removed `TODO(sub-agents)` markers
and claimed nothing existed yet) and the unify-id RFC's fork/spawn risk bullet
(described the seam as "explicitly deferred" via `AgentLoop.create`'s old TODO),
now pointing at the realized seam.
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.
Turtle
Tianyi Cui
Yichen Jiang