fix: a subagent must not run its own 0→1 phase machine

[linhdmn]

Summary

A subagent the model spawns is not the run. It was getting its own 0→1 phase machine, burning the research phase's step ceiling, and being rejected — which blocked the whole run in research.

The measurement

Live headless run, 2026-10-07 (harness 0.2.0-rc.2, fl-live profile). The model spawned four research subagents. Every one of them has its own run record and its own phase table:

67770243-…  Verdict: blocked   research | current | 24 | $0.0638 | $0.2000
3b8306ef-…  Verdict: blocked   research | current | 24 | $0.0531 | $0.2000
f3ab6708-…  Verdict: blocked   research | current | 24 | $0.0700 | $0.2000
120412bc-…  Verdict: blocked   research | current | 24 | $0.1093 | $0.2000

Four subagent sessions, each stopped at exactly 24 steps — the research phase's ceiling (maxSteps: 240 × research share 0.10 = 24). Each subagent's own session log confirms it:

• session.origin: 'subagent', delegationDepth: 1, parentSession: <the run>
• the subagent was told "0→1 PIPELINE — phase 1 of 5: RESEARCH" — the parent's pipeline notice, in a session that has no goal
• and each ended turn/end {"reason":{"kind":"blocked"}} after its 24th step

The parent then received, four times:

Background subagent 67770243-… declined the task.It left no closing message.

declined is stopReason: 'refusal' (packages/subagent/tool-subagent/src/index.ts). The parent's own reject is the same string: research step ceiling reached (24 steps).

Result: the run blocked in research, 24 of 240 steps, $0.0471 of $2.00 (2%), with no research note written.

Why

policyFor builds one policy per agent (policies = new WeakMap<Agent, FeatureLoopPolicy>()), and createPolicy builds a pipeline for each:

const pipeline = spec === undefined || options.pipeline?.enabled !== true
  ? undefined
  : buildPipelineRuntime(spec, options.pipeline, options)   // → startPipeline()

startPipeline() returns { state: 'research', … }. So every helper the model delegates to inherits the run's goal, its phase machine, and — via phaseJustEntered — a phase-1-of-5 briefing, then dies at that phase's ceiling.

The fix

A session that is a subagent child — origin: 'subagent' or a non-zero delegationDepth, both read structurally from session.header — now runs under the same spec, budget, ladder and review gate, with only policy.pipeline cleared:

if (isSubagentSession(sessionHeaderOf(agent))) policy.pipeline = undefined

Three deliberate choices:

• The gate stays ON. A subagent's write and bash are exactly as irreversible as the parent's, and the gate is the thing that asks a human. Only the phase machine — meaningless without a goal — is switched off.
• Both signals are read. origin and delegationDepth are set by different parts of the harness, and a release that moves one should not silently re-enable the phase machine on every helper.
• Malformed signals fail closed to "not a subagent." An unclassified session keeps the phase machine it would have had before this existed.

The predicate is its own import-free module (src/subagent.ts), so the CI job that installs nothing can test it — the same reason watcher-ttl.ts is.

Verification

• test/subagent.test.ts fails with the policyFor branch reverted — both not ok 6 (the predicate is wired) and not ok 7 (the branch clears only the pipeline) — and passes with it. Confirmed, then restored.
• Test 7 asserts the whole branch body is that single assignment, so a future edit that also clears gate/spec/budget/worktreeRoot fails rather than quietly disarming a subagent's gate.
• Full suite 829 pass / 0 fail / 1 skipped (830 tests); tsc --noEmit clean; pre-commit scan clean.
• src/subagent.ts added to the typecheck list and test/subagent.test.ts to the no-install test job in ci.yml — with both lists re-checked so every path they name exists.

Context

Third bug from the same headless run, and the second one that only became observable after the previous was fixed:

1. fix: an open dashboard tab must stay a watcher, or a headless run hangs #46 (merged, b4c3f44) — an open dashboard tab was a watcher only 60% of the time, so the ask hung.
2. fix: /product and /loop must start a turn, not just return text #47 (merged, 18a94e3) — /product and /loop produced a command row and no turn.
3. this PR — a subagent ran its own phase machine.

Each was found by driving the real thing and reading what it actually did, not by reading the code.