fix: address codex review round 3

- Strict steer additionally requires an OPEN STEP: between steps the
  loop may be awaiting its continuation/turn-stop checkpoints, where
  pending steering was already folded and a terminal stop discards a
  later arrival. A message accepted during an open step is drained and
  recorded at that step's settlement before any terminal decision, so
  the acknowledged-then-discarded window is closed. New keyless test
  holds agent/turn-stop open and pins the rejection.
- tool-subagent-control README: distinguish synchronous not-delivered
  errors from started-Task failures (unknown/foreign/descriptor-less
  ids settle the started Task as failed), and drop the claim that the
  completion notice carries the child's response.
This commit is contained in:
Dudu-0223
2026-07-23 18:47:19 +08:00
committed by imccyu
parent 4eda48d002
commit 9e5ae0d12e
7 changed files with 49 additions and 7 deletions
@@ -49,7 +49,7 @@ For a continuable initial activation, the control service allocates the stable c
Every continuable child turn is admitted through this Task-backed path. A non-terminal Task is the only supported live activation; when no activation exists, its run has already been disposed and the durable child is resumable. Before routing any by-id operation, the control service synchronously compares its association with `ctx.agents.get(childId)`. A registry Agent with no association, or a registry Agent different from the associated `run.localAgent`, is an ownership conflict: the control service fails rather than adopting an idle Agent or attaching an untracked turn. When neither exists, cold resume may proceed; a competing publication after that check still loses at the Agent registry collision boundary.
Routing follows the Task association. A running Task accepts live delivery through the run's optional strict `SubagentRun.steer` capability. An absent Task starts a fresh Task and cold-resumes the child. In-process spawn and fork implement this capability with synchronous checks that share one frame with the `Agent.steer()` call: the child must be `running`, its turn must still be open in the log (status stays `running` through a closed turn's durability flush, where the loop strands drained steering), and no structured capture may have committed (its terminal stop makes the loop discard late steering). Providers must not expose the Agent-level idle fallback as strict steering, because that fallback may start an untracked turn after the observed run has ended. If the Task settles between association lookup and this strict check, `steer()` fails, `send_message` reports the message as not delivered, and that call does not fall through to cold resume; a later retry after Task terminal may start the next activation.
Routing follows the Task association. A running Task accepts live delivery through the run's optional strict `SubagentRun.steer` capability. An absent Task starts a fresh Task and cold-resumes the child. In-process spawn and fork implement this capability with synchronous checks that share one frame with the `Agent.steer()` call: the child must be `running`, its turn must still be open in the log (status stays `running` through a closed turn's durability flush, where the loop strands drained steering), a step must be open (between steps the loop may sit at its continuation/turn-stop checkpoints, where steering was already folded and a terminal stop discards a later arrival), and no structured capture may have committed (its terminal stop makes the loop discard late steering). Providers must not expose the Agent-level idle fallback as strict steering, because that fallback may start an untracked turn after the observed run has ended. If the Task settles between association lookup and this strict check, `steer()` fails, `send_message` reports the message as not delivered, and that call does not fall through to cold resume; a later retry after Task terminal may start the next activation.
The control service does not serialize two callers that race a stopped child through paths outside it, nor does it model a separate settling phase between result production and disposal. The synchronous association install before the producer's first await admits one activation per child in this process — a competing `sendMessage` during resume load observes the pending activation and fails explicitly — while a bypassing publication still loses at the Agent registry's same-session collision boundary. Delivery racing startup, cancellation, completion, or cleanup may also fail. These limitations are explicit rather than hidden behind a larger lifecycle abstraction.