akm docs

Review corpus — task/workflow format unification

Reference record for ../../task-workflow-format-unification.md. Every revision of that spec from v8 onward was shaped by these reviews; the spec's §11 cross-references findings by the ids used here.

Round 0 — three independent adversarial reviews of spec v7

Three reviewers, identical briefs, no shared context — overlap between reports is therefore independent confirmation, and unanimous findings (3/3) were treated as settled facts.

File Contents
opus-review-1.md Reviewer 1 full report
opus-review-2.md Reviewer 2 full report
opus-review-3.md Reviewer 3 full report
findings-consolidated.md Deduplicated findings (C1–C5, M1–M13, minors) with per-reviewer attribution — the ids the spec's §11 uses

Outcome: v7 → v8 full redesign (IR v4 invocation union, hashVersion 5, two output schemas, cascade built-not-exposed, with:, env provenance split, migration rebuilt).

Round 1 — four-lens panel on spec v8

File Lens
panel-r1-coverage.md Findings-coverage audit: does §11 resolve each round-0 finding, or just name it
panel-r1-architecture.md Module boundaries and seams against the codebase's own idioms
panel-r1-adversarial.md Fresh defect hunt on what v8 itself introduced
panel-r1-security.md Script-execution policy, env/redaction contracts, migration mechanics

Outcome: v8 → v9. Headline: the per-bundle script-execution grant was withdrawn (contradicted activation-policy.ts's no-new-trust-machinery decision) in favor of exact 07 P1-D ceiling inheritance; persona snapshot added to the hash preimage; with: merge defined; tombstone read-only carve-out and interruption-safe conversion ordering.

Round 2 — verification on spec v9

File Lens
panel-r2-verification.md Every round-1 critical/major re-checked against v9 and code — verdict: all fixes hold
panel-r2-adversarial.md Cold pass, no prior context — verdict: no design flaws; two evidence corrections (the real placeholder filler; indexer-classifier recognition), both applied

Library evaluations — buy-vs-build investigation

Three independent investigations into replacing the hand-rolled engine with a third-party library, all verified against live artifacts rather than training knowledge. Decision record: ../../workflow-engine-buy-vs-build.md.

File Scope
library-eval-bunqueue.md bunqueue — full rubric + source read
library-eval-openworkflow.md OpenWorkflow — full rubric, executed (SIGKILL durability probe, reproduced schema collision)
library-eval-alternatives.md Landscape survey — 25+ candidates against the no-server / SQLite / Bun+Windows / license constraints

Outcome: build. ~6% of exec+ir is the generic durable-execution concern libraries sell; the best candidate deletes 8–12% while removing cancellation, concurrency caps, and transactional atomicity.

Reading notes