aesop 0.3.1 — Multi-core waves

aesop 0.3.1 — Multi-core waves

0.3.1 ships as the release tag for the 0.3.0 milestone (a defective v0.3.0 tag was burned by an automation error and superseded).

0.2.0 shipped the seams; 0.3.0 ships the proof: a non-Claude model core ran a full
supervised wave — intake → build → verify → ship — through the same engine, with the
same gates, and the release was preceded by a fresh adversarial hardening loop that
exited clean.

Headline: the wave engine is core-agnostic (WS3)

• wave_scheduler.py (WS3a pilot): deterministic single-cycle orchestration — tracker
  intake with fail-closed validation (empty/missing ownership rejected; paths normalized
  platform-independently; absolute/traversal paths rejected), HALT + cost-ceiling gates
  that abort on module failure (never fail open), manifest via the driver bridge, one
  run_wave call, stop-before-merge Report. Atomic tracker claim (mkstemp + os.replace,
  content-hash conflict abort) prevents double-dispatch across runs.
• Gate-1 handoff kit: --driver claude|codex CLI injection; per-item Report
  observability {slug, backend, tier, verified, testExit}; documented orchestrator
  REPORT-CONTRACT; offline FakeTransport codex route proven in CI.
• LIVE PROOF (gate 1, DONE): a supervised codex wave (gpt-4o-mini via CodexDriver)
  took a real backlog item (wave_templates validate --json), implemented it, passed the
  real 25-test suite (testExit 0, tier 2), and the ship phase committed and pushed —
  human-reviewed and merged as PR #325. Two supervised corrections were applied (unicode
  glyphs ASCII-coerced by full-file replacement), and four scheduler Report-plumbing
  defects the live run exposed were fixed with real-shape regression tests.
• Survived two adversarial review rounds pre-merge (12 verified defects fixed, including
  a dead-code tracker write and a symlink TOCTOU) — see the hardening section.

Measured, not asserted

• Live structured-output accuracy: gpt-4o-mini 32/32 (100%) composite
  (valid-JSON / schema-exact / ownership-respect) under the driver-faithful payload
  (bench/results/accuracy-live-2026-07-22.json). Single run, N=32 curated tasks —
  supports the probe's conservative 0.92 assertion; not a transfer claim. The path to
  this number (33% → 0% → 4% → 100%, each step a real harness defect fixed and
  regression-guarded) is documented in the bench history.
• Frontier discrimination slice: 20 hard judgment tasks with per-task discrimination
  rationales, deterministic scoring, live runs cost-gated behind --confirm-spend
  (exit 2 USER-GATED otherwise).
• Transcript-sampled judgment set: N=150 sanitized tasks from real fleet transcripts.
• Cross-OS drift measurement: tools/crossos_drift.py quantifies windows-vs-ubuntu CI
  divergence from real run history. Baseline at introduction: windows 0/6 where present;
  after the parity campaign (env-tunable child timeouts, eod_sweep repo-delimiter root
  cause, 8.3 containment fixes) the windows job went GREEN on main — the promote-to-
  required streak is counting from run 29955999466.

State consolidation (WS4)

• ReadAPI facade (state_store/read_api.py): one read seam over tracker / orchestrator
  status / heartbeats / ledger — delegates to existing parsers, never forks logic.
• WriteAPI seam (state_store/write_api.py): event-append + atomic projection with
  conflict detection; first two tracker write ops behind one facade (caller migration
  is the 0.4 track).
• StateAPI ratchet in CI: stateapi_lint gate live — new direct state reads outside
  the facade fail CI against a committed, posix-normalized baseline (currently 33
  entries: a visible migration worklist that can only shrink).
• Agent lifecycle events: dispatched/working/done/stalled event types + projection
  with transition history, feeding the Activity view live.

Cost: observed, projected, bounded, unified

• cost_projection.py: burn-rate from a ledger window, end-of-wave projection,
  idempotent 70%/90% ceiling alerts (honest fired_alert semantics under partial failure).
• One window contract: projection and ceiling share a single window helper — they can
  no longer disagree about "spent".
• Cost Analytics dashboard panel: spend per wave, per-model split with the all-Opus
  counterfactual, burn vs ceiling — with honest DATA-UNAVAILABLE states and a Playwright
  proof (verify_cost_panel.py).

Operability

• aesop reproduce: offline verification suite from a clean clone/install; doctor
  failure classification is exact-match (a real missing dependency can no longer be
  mistaken for a pre-init condition).
• docs/PORTING.md: step-by-step adopter port with the 10 likeliest failure modes and
  recoveries, sourced from this repo's real incident history.
• Windows CI job (non-required): node+python on windows-latest; parity fixes for
  file-locking, SSE disconnect noise (WinError 10053/10054 as normal lifecycle), and
  eod_sweep failing CLOSED on git errors (the 8.3 short-path fail-open root cause). Promotion to required tracked at
  5 green merges.
• Monitor stall detection: stall_check.py active-task predicate + advisory recovery
  emission, surfaced as a monitor signal.
• Wave preflight: backlog validation flags (missing ownership, stale refs, overlaps,
  ledger-aggregate retry rate with DATA-UNAVAILABLE honesty).

Security

• Redaction hardening: URL-credential patterns are scheme-agnostic, consume
  embedded-@ userinfo to the last @, handle IPv6 hosts, and refuse to over-redact
  letterless ratios; over-redaction is the documented failure direction.
• Scanner exemption, done in the open: connection_string stays fatal everywhere;
  ONE file (the redaction-pattern source) downgrades to a reported, never-silent
  ALLOWED-REDACTION-SOURCE — a user-approved, single-rule, test-pinned exemption.
  Notable property it surfaced: the pre-push hook runs main's scanner, so a branch
  cannot weaken its own gate.
• Ship-phase hygiene: git-add failures unstage their residue; per-repo ship errors
  carry stderr detail in the Report.

Hardening (the release gate)

• 0.3.0's release condition was a full /refinesystem loop: expert + adversarial lens
  fleets with regression re-verification, every P1 deterministically verified by the
  orchestrator before any fix was paid for. This cycle's honest ledger:
  • Round 1 (7 lenses): ~14 verified defects fixed pre-merge (incl. a symlink TOCTOU,
    a dead-code double-dispatch guard, and 4 redaction under/over-redaction defects);
    5 findings refuted with evidence.
  • Round 2 (6 lenses, full): 12 verified defects fixed (incl. codex broken-by-default,
    write_api OCC contract lie, 16-site dead-client 500 discipline, the eod_sweep ':'
    delimiter root cause) + a LIVE incident caught by the regression lens's README
    canary (fixture escape into the working tree — contained, guarded, two long-lived
    identity polluters eliminated); 4 severities corrected downward.
  • Round 3 (3 lenses): 3 small findings fixed; the new identity tripwire caught a
    polluter predating the entire cycle (hook self-test rewriting git identity on every
    run — active for months, invisible until instrumented).
  • Round 4: exit verification (fix re-attack + tripwires) — clean.
    Net: ~30 verified defects fixed across 4 rounds, ~10 lens claims refuted with
    evidence, 3 integration trains + 2 solo ships, ending at a fully-green main
    including windows for the first time in the repo's history.
• Test-infrastructure classes fixed this cycle: zero-collection test classes (a gate now
  fails baseless Test* classes), scaffold-test load-sensitivity (shared fixtures,
  env-tunable child timeouts), local-server timeout starvation, stdin-inheritance hangs.

Breaking / behavior changes

• Mixed git-ship manifests (some items with explicit repo, some without) are rejected
  at preflight; pure-legacy and fully-explicit manifests are unchanged.
• npm run test:sh no longer invokes reconstitute.sh --test directly (its wrapper suite
  exercises --test internally); the pre-push hook's own self-test remains an explicit
  invocation because its wrapper suite does not run it.
• NEW CLI surfaces: frontier_slice.py exits 2 (USER-GATED) without --confirm-spend;
  cost_ceiling.py gains --window (backward-compatible); stall_check.py gains
  --active-from, --emit-recovery, --recovery-dir; wave_scheduler.py gains
  --driver claude|codex; test_battery.py added (parallel local union battery).
• eod_sweep: repo list delimiters are now os.pathsep (';' on Windows); nonexistent or
  non-git explicitly-listed repos are AT-RISK findings (exit 1), never silent skips.

Honest residuals

• Windows job remains non-required until 5 consecutive green merges post-parity.
• StateAPI baseline: 33 direct-read sites remain; burn-down is the 0.4 track alongside
  caller migration to WriteAPI and validation-ownership consolidation.
• Codex live proof is one supervised wave on one small item — the unsupervised loop,
  failure-recovery ownership (WS3b), and multi-item non-Claude waves remain future work.
• Benchmark discrimination slice is authored but not yet live-run (spend-gated).