2026.08.02.1 — latent-bug fixes (IL-1/2/4/7) + five-suite quality

Model behavior change — model type version bumped to 2026.08.02.1 (first release since the 2026.07.16.2 five-suite quality backfill). Real fixes for four of the seven latent bugs triaged in the LOCAL issue-lifecycle model issue-lifecycle-latent-bugs; three are explicitly re-affirmed as intentional behavior, not fixed. No globalArguments or resource-schema change — the version bump ships as an identity upgrades[] migration.

Fixed

• IL-1 — start used to overwrite whatever current held with no read-before-write — approved plans, review history, everything — with no guard or confirmation. It now reads existing state first: a fresh instance (no current yet) or one already in a terminal state (complete/closed) proceeds as before; anything else throws unless the new force: boolean argument (default false) is passed.
• IL-2 — approve_plan/tests_approved gated only on hasBlockingFindings (open CRITICAL/HIGH count), so a reviewer could post verdict: "FAIL" with zero findings (or all findings resolved/non-blocking) and approval would still succeed. Both methods now also call the new failingReviewers() helper and block if any reviewer's verdict is FAIL, after the existing blocking-findings check (gate order preserved so "N CRITICAL and M HIGH" fires first when both conditions hold). tests_approved's override_reason now bypasses both gates together, as before. SUGGEST_CHANGES remains non-blocking by design — only FAIL hard blocks, so the autonomous zero-CRITICAL/zero-HIGH loop can still converge without a human override.
• IL-4 — resolutions is a flat Record<string, string> keyed by finding description text; two different reviewers whose findings happened to share description text collapsed into one entry. resolve_findings now expands each supplied key against the current round's findings: every reviewer whose finding matches that description gets its own `${reviewer} :: ${description}` composite key. A key matching no finding is stored verbatim (legacy-safe; unaffected: empty-map callers).
• IL-7 — record_review appended every submission unconditionally, so recording the same reviewer twice in one round double-counted their open findings in the blocking gate. It now replaces (last-write-wins) the reviewer's earlier entry in place instead of appending a duplicate.

Kept as designed (re-affirmed, not fixed)

• IL-3 — no model-enforced iteration cap on iterate/iterate_tests. MAX_CODE_ITERATIONS/MAX_TEST_ITERATIONS stay skill-layer policy; enforcing a cap in the pure model would couple it to skill policy and could break the human override_reason escape hatch.
• IL-5 — close has no guardState call and works from any state, including terminal ones. This is intentional: close is the abandon/escape hatch (manifest.yaml has always documented "works from any state") and an escape hatch must never itself be blockable.
• IL-6 — hydrate's summary.snapshotAt is stamped fresh via now() on every call, so two calls produce two different values even though current is never mutated. snapshotAt is a wall-clock capture stamp — it should differ per call; the property suite already proves non-mutation of current by freezing the clock (@std/testing FakeTime), which is the invariant that actually matters here.

Five-suite quality (carried from the 2026.07.16.2 backfill, updated)

• extensions/models/issue_lifecycle_methods_test.ts — success + exact guardState-throw-message regression for each of the 20 model methods, a sweep pinning "No issue state found — run 'start' first" on every method but start, and a sweep pinning the REAL (not assumed) unknown-key behavior of every method's zod arguments schema: swamp model type describe --json renders additionalProperties: false (its own JSON-Schema view), but none of the 20 methods call .strict(), so a bare .parse() silently strips an unrecognized key rather than throwing.
• extensions/models/issue_lifecycle_adversarial_test.ts — illegal out-of-order transitions from varied source states, malformed reviewer input (bad severity/verdict enums rejected by zod), hostile approve_plan gate combinations (missing-matrix-reviewer, combined CRITICAL+HIGH counts), corrupted-stored-state pins for the "no plan found" branches in approve_plan/tests_approved, whitespace override_reason still gated, and pins asserting the FIXED IL-1/IL-2/IL-4/IL-7 behavior above plus the re-affirmed IL-3/IL-5 by-design behavior.
• extensions/models/issue_lifecycle_coverage_test.ts — branch fill for allMatrixReviewersRecorded across the security/ux/skill matrix dimensions,

Release 2026.07.16.2 — align model versions with manifests

Maintenance release across the @magistr extensions. For most packages this carries no functional change: the only edit is the model's version: field, brought back in line with its manifest version so the published model type version and the package version no longer drift.

Functional changes in this release are limited to:

• anime-cron: normalizeTitle now strips a ": subtitle" suffix and a trailing parenthesized year before comparison, fixing dedup false-misses where the torrent title carries a subtitle or year that the AniList romaji does not.

• arckit: first publish. Standalone ArcKit port — a 12-phase architecture governance state machine with 65 bundled templates, driven by a bundled skill.

Also tracks three extensions (kaiten, observability-agent, music-library) that previously existed only as untracked working-tree directories, recovered from stashes.

Merge pull request #14 from umag/feat/issue-lifecycle-scenario-9-resume-eval

feat(issue-lifecycle): eval scenario-9 — resume-dispatch from the TDD sub-cycle (2026.06.12.3)

Merge pull request #13 from umag/feat/issue-lifecycle-record-reproduction

feat(issue-lifecycle): record_reproduction method (2026.06.12.2)

Merge feat/issue-lifecycle-tdd-subcycle-docs: skills catch up with the TDD test-review sub-cycle (2026.06.12.1)

issue-lifecycle: release 2026.05.25.3

Re-trigger after the -y publish fix (631455c). Same content as the tagged- but-unpublished .25.1/.25.2 (the 2026.05.24.x skill changes); model type version stays 2026.04.30.1.

Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com

2026.05.24 release — plan presentation + implementation discipline (skill-only; model type version unchanged at 2026.04.30.1)

Plan presentation:

• New skimmable BLUF plan format: Goal / Approach / Domain impact (4 lines) / Scope table with a DDD-role column / conditional Risks / numbered one-line Steps / Review coverage / Non-goals / Open questions.
• Front-loaded "Plan output format (always)" pointer in SKILL.md so the format applies even when a plan is produced outside the full lifecycle flow.
• Opt-in HTML-artifact escalation, Wardley maps (strategic build-vs-buy only), and DDD diagram conventions; diagrams off by default.
• Planning Step 9 and the autonomous-loop approval gate render in this format.

Implementation discipline (implementation.md Step 2):

• Anchor changes in existing code: map integration points first, reuse/extend, no parallel code paths.
• Right-size backward compatibility: no compat shims/migrations/version flags for unreleased code; new adversarial-review check enforces it at plan time.
• Explicit RED/GREEN/REFACTOR; refactor in-place, never deferred.

Validated with tessl evals (BLUF format 100/100; integrate + no-compat 100%).

Note: 2026.05.24.2 == 2026.05.24.1 content; this version adds the release notes/changelog that the .1 push omitted.

2026.04.30.5 — TDD test sub-loop with human escalation + explicit plan display

This release wraps up the work introduced across 2026.04.30.3 / .4 (which went out without proper release notes) and adds the CI publish-flow fix that prevents future versions from shipping with empty changelog entries.

What's new since 2026.04.30.2

Phase 5 (Implementation) is now a TDD test-first state machine

implement no longer transitions straight to implementing. It enters writing_tests, and the lifecycle enforces the TDD discipline at the state-machine level: tests must be authored, reviewed, and cleared of blocking findings before any production code is written.

New states:

• writing_tests — author failing TDD tests (RED).
• reviewing_tests — tests under review.
• implementing (existing) — tests are clean; now write code (GREEN) and refactor.

New ReviewPhase value: test_review (alongside plan_review and code_review). Each test-review round is snapshotted into reviewHistory just like plan/code rounds.

New methods

• review_tests — fan out reviewers from reviewMatrix against the tests (writing_tests → reviewing_tests). Resets the round and stamps reviewRoundStartedAt.
• iterate_tests — return to writing_tests because findings remain; bumps testReviewIteration, snapshots round with outcome: "rejected_auto" or "rejected_human" per source.
• tests_approved — gated transition reviewing_tests → implementing. Default autonomous gate (full matrix coverage AND zero open CRITICAL AND zero open HIGH). Accepts an optional override_reason that bypasses the blocking-findings gate as an explicit human override after the iteration cap; matrix coverage is still enforced. The override path snapshots outcome: "human_override" with the reason stored in rejectReason for audit.

Human escalation path (cap-reached safeguard)

After 5 test-review iterations (or signature-loop detection), the autonomous loop must stop and escalate to the human. The skill presents the full iteration history and open blocking findings, then offers two explicit, audited paths:

• Human correction — iterate_tests --input source=human --input reason="<guidance>". Counter bumps; the round snapshots rejected_human; tests get rewritten per the human's guidance.
• Human override — tests_approved --input override_reason="<justification>". Bypasses the blocking-findings gate; matrix coverage still required; the round snapshots human_override with the reason recorded.

Explicit full plan display before approve_plan (Phase 4)

The plan must be presented to the human in full — verbatim, no compression — before approve_plan. SKILL.md now mandates an ordered display:

1. Plan summary
2. Every step (description, files, risks)
3. Full DDD analysis
4. Full TDD strategy
5. Review matrix
6. Potential challenges
7. Aggregated review findings (every reviewer, every finding)
8. Explicit approval prompt with the accepted phrases ("approve" / "approved" / "LGTM" / "ship it" / "go").

This closes a gap where the model could ask for approval after only summarizing the plan, undermining informed consent at the only true human-gated transition.

Other surface changes

• record_review guard extended to allow reviewing_tests.
• hydrate summary now surfaces testReviewIteration alongside codeReviewIteration (so the autonomous loop can detect the cap and escalate cheaply without parsing full state).
• New ReviewOutcome value: human_override.
• Model TypeVersion bumped to 2026.04.30.1; package version 2026.04.30.5.

Backwards compatibility

Existing recorded state parses through schema defaults (testReviewIteration defaults to 1). Issues already past approved (in implementing, code_reviewing, etc.) keep working; the new sub-loop applies to any issue that re-enters via implement.

Tests

47 model tests pass (33 prior + 14 new covering the test sub-loop, the override gate, escalation paths, and the new state machine guards).

Why .3 / .4 changelogs were sparse

2026.04.30.3 was auto-published by CI immediately after the code push, and the CI publish step did not pass --release-notes — so the swamp.club changelog entry for .3 was empty. 2026.04.30.4 was a manual republish to attach notes, but the page still shows .3's empty entry in the "Previous Versions" section because per-version notes are immutable once published.

The CI workflow has now been updated to read HEAD's commit message and pass it as --release-notes on every auto-publish, so future versions will always carry notes derived from the version-bump commit.

TDD test sub-loop in implementation phase, with human escalation after the 5-iteration cap.

Highlights since 2026.04.30.2:

• New states: writing_tests, reviewing_tests. New ReviewPhase: test_review.
• New methods:
  • review_tests — fan-out reviewers against TDD tests (writing_tests → reviewing_tests).
  • iterate_tests — loop back on test-review findings (reviewing_tests → writing_tests); bumps testReviewIteration; snapshots outcome rejected_auto/rejected_human.
  • tests_approved — gate transitioning reviewing_tests → implementing. Default autonomous gate (matrix coverage + zero CRITICAL + zero HIGH). New optional override_reason bypasses the blocking-findings gate as an explicit human override after the 5-iteration cap; snapshots outcome=human_override with the reason in rejectReason for audit.
• implement now enters writing_tests (not implementing) so production code is only written after the test-review round comes back clean.
• record_review guard extended to allow reviewing_tests.
• hydrate summary now surfaces testReviewIteration alongside codeReviewIteration.
• New ReviewOutcome value: human_override.

Skill instructions (issue-lifecycle/SKILL.md):

• Phase 4 now mandates explicit, full-content plan display before approve_plan: summary, every step (with files/risks), DDD analysis, TDD strategy, review matrix, potentialChallenges, aggregated review findings — all verbatim, no compression — followed by an explicit approval prompt.
• Phase 5 documents the autonomous TDD sub-loop and the cap-reached escalation path: when testReviewIteration >= 5 (or signature loop detected), the skill must surface the iteration history and open findings to the human and offer two explicit, audited paths — iterate_tests source=human (correction) or tests_approved override_reason=… (force-approve).

Tests: 47 passing (33 prior + 14 new covering the test sub-loop, override gate, and escalation paths).

Backwards compatibility: existing recorded state parses through schema defaults (testReviewIteration default = 1). Any in-flight issue still in implementing/code_reviewing continues to work; the new sub-loop applies to issues that re-enter via implement.

Restructure to monorepo layout. Add repository field, root README as additionalFile, improved manifest description with state machine and method docs.

Initial publish: issue lifecycle model (v2026.04.09.1) with prior-art lookup, DDD+TDD planning, autonomous code-review iteration, UAT/KB harvest, and hydrate summary resource. Bundled with 9 user skills: issue-lifecycle, ddd, tdd, moldable-dev, review-code/adversarial/security/ux/skill.