orchestrate run prd352-20260803-015333 — the slice-executor delegation layer (PRD #352)

[rodrigorjsf]

Autonomous orchestrate run of PRD #352, ADR-0017's slice-executor delegation layer. 12 of 12 slices passed, none failed, none skipped. 49 files, +7931/−1008, across six dependency waves.

Left unmerged for review. Ships orchestrate 1.7.0.

What changed

The orchestrator no longer runs slices — it delegates them. Its per-slice section is five steps: create the worktree, freeze routing, spawn orchestrate:slice-executor-<variant> with a briefing, validate the returned envelope, act on it. Investigation, implementation, the continue-in-place loop, review, the changeset check and the capability gate all moved into a skill the executor preloads.

Wave	Slice	PR
0	#353 run-discovery extracted from the watchdog	#366
0	#354 slice-executor envelope role + failure classes	#367
0	#355 slice progress record + recover_slice_progress	#368
0	#356 routing entry for the executor	#369
0	#364 complete a partial .orchestrate/ config	#370
1	#358 spawn-budget awareness in the watchdog	#371
1	#359 the slice-pipeline skill	#372
1	#365 pre-flight capability probe	#373
2	#360 the two slice-executor definitions + variant parity test	#376
3	#361 read guard hook + hook-matcher consistency test	#377
4	#362 orchestrator spine restructure — the keystone	#379
5	#363 release closure — version cascade and README	#380

Premises the run falsified

Four acceptance criteria rested on claims that turned out to be wrong. Each was corrected rather than implemented as written:

• orchestrate: complete a partial .orchestrate/ config instead of running with a vacuous capability gate #364's bug was in the prose, not the code. bootstrapConfig() was already idempotent; the defect was that its trigger gated on .orchestrate/ being wholly absent.
• orchestrate: read guard hook and hook-matcher consistency test #361's AC8 asked for a test over "a hook matcher that names an agent type" — but only SubagentStart/SubagentStop match on agent type, and this slice adds a PreToolUse, which matches on tool name. Inventing a matcher to have something to test would have been an unrequested hook. It shipped as a general invariant over hooks.json whose load-bearing layer is the bare-name trap.
• orchestrate: orchestrator spine restructure — delegate the slice, map the failure class #362's AC12 rested on a cap that does not reach the file. The 500-line SKILL.md cap's paths: covers only two other plugins — corroborated behaviourally, since the same rule mandates semantic tags neither orchestrate skill has, so if it reached them both would be in permanent hard-fail. The cap binds by self-declaration; the spine's own claim to be "within" it was false at 501 lines.
• orchestrate: Slice progress record and its structured-recovery tool #355's AC5 targeted cross-run isolation that resolveRunDir already provides for free; the real hazard was intra-run, so the filename had to carry the issue number.

Defects found by review that the implementation would have shipped

• orchestrate: spawn-budget awareness — wave-width cap, backpressure, watchdog threshold #358's first review failed the slice on a per-run spawn numerator measured against a per-session denominator — every successor session would have re-raised backpressure on its first spawn, degrading the run to one slice per session.
• orchestrate: slice-executor agent definitions (standard and deep) with variant parity test #360's parity suite compared skills: by nothing, so a misspelled skill name in the deep variant passed 27/27 green — precisely the silent-preload failure the slice exists to prevent.
• orchestrate: orchestrator spine restructure — delegate the slice, map the failure class #362's spine declared its failure-class mapping lived in exactly one place, then restated one of its rows eighteen lines later. A grep could not see it because the spine's copy carried no label literal.
• orchestrate: release closure — version cascade and README for the delegation layer #363's condensed routing.json example dropped a required tier with a note implying it was inferable. Anyone copying it would have produced a config that fails validation.
• orchestrate: routing entry for the slice-executor role #356's test claiming to cover "label beats default" used a fixture where no default was ever derived — it passed under a reversed order.

Verification

Every slice passed an orchestrator-run capability gate independent of the reviewer's self-report. The suite went 698 → 808 tests; the single failure throughout is one known pre-existing baseline (resolveMergeConflict RE-ENTRANCY), confirmed by knownFailureMatches on every run. dist/ is committed and CI's bundle check gated the three slices that touched source.

#361 was proven red from both directions — a permissive stub turned all 29 deny cases red, and a deliberately over-broad implementation turned all 27 allow cases red together. A one-sided proof would have accepted a guard that blocks everything.

Known limitations, filed rather than hidden

• orchestrate: the watchdog cannot count a slice-executor's nested spawns when the hook fires from a worktree #374 — the watchdog resolves the active run from cwd, which a slice worktree cannot satisfy, so nested executor spawns may go uncounted. Becomes live now that the executor actually spawns workers.
• orchestrate: the slice-executor's skills: preload is unverified, and a miss leaves it with no procedure and no error #375 — the skills: preload for a plugin subagent is unverified. A probe series could not confirm it fires; the plugin-subagent path could not be tested in-session because plugin definitions register only at startup. orchestrate: slice-executor agent definitions (standard and deep) with variant parity test #360 ships a body guard that turns a silent miss into a blocked envelope. First real run of this layer verifies it.
• orchestrate: the slice progress record loses the continuation handoff note across an executor resume #378 — the progress record does not carry remainingWork, so a resumed executor loses the continuation handoff note. Deliberate closed set; filed so the decision is explicit.

Closes #353
Closes #354
Closes #355
Closes #356
Closes #358
Closes #359
Closes #360
Closes #361
Closes #362
Closes #363
Closes #364
Closes #365