v2.0.1

User-visible outcomes

This patch makes the public install path work out of the box and publishes the Azhou skin so a fresh Agent Skin clone builds without a private overlay.

• npx skills add TeFuirnever/azhou-ai-hub can now install all 12 public skills. The unquoted YAML description values in super-archify, super-asd-ste100 and super-ask-azhou that contained : are quoted; scripts/check_repository.py now strictly parses every public SKILL.md frontmatter.
• migrate-receipts prints nothing to migrate and exits 0 when no legacy receipts directory exists, instead of crashing.
• The 21 project-local .agents/skills packages set metadata.internal: true, so they stay out of the public npx skills add --list catalog. They remain installable by name, and local Cursor/Codex discovery is unchanged.
• A fresh Agent Skin clone builds the Azhou skin without a private overlay: the four build assets (working.png, wallpaper-scene-q-v3.png, entry-avatar.png, app-icon/source.png, about 6.5 MB) are tracked under packages/agent-skin/assets/azhou/ as CC BY-NC-ND 4.0, separate from the MIT code, with a root TRADEMARKS.md. The resolver prefers an ignored assets/private/<name> overlay when present. Bundles built from the library art may only be shared non-commercially, unmodified and attributed.
• Agent Skin's code license is recorded as the repository root MIT license after the independent rewrite in #441. The historical legacy bytes from the unlicensed upstream remain unlicensed and are not redistributable.
• The maintained pstack marketplace identifier is now pstack-azhouzhou@azhou-hub (renamed from the fork-era pstack-claude). Two Hub-authored pstack overlays were corrected to match the shipped setup steps; the pstack distribution version stays 0.15.37.
• Documentation, README layout/hero, governance wording, and a docs-only CI fast lane were brought in line with current repository reality.

Affected skills

• Fixed install metadata: super-archify, super-asd-ste100, super-ask-azhou. No skill runtime behavior or schema change.
• Docs only: Agent Skin Pet commands now state the Pillow prerequisite.
• Package: packages/agent-skin (published Azhou library assets and license record).
• Overlays only: pstack Hub-authored setup references. No pstack runtime, manifest or patch change; version stays 0.15.37.
• No canonical skill was added or removed; the catalog remains 12 public skills.

Upgrade / migration action

• If you installed from v2.0.0 or from main before #465 and could not add those three skills, reinstall from this tag or from current main:
  npx skills add TeFuirnever/azhou-ai-hub --skill <skill-name>
• Compare an installed skill directory against this tag with the snapshot in the installation guide:
  git clone --depth 1 --branch v2.0.1 https://github.com/TeFuirnever/azhou-ai-hub /tmp/azhou-tag
• If you use the maintained pstack marketplace, the identifier is pstack-azhouzhou@azhou-hub. Existing installs that still show pstack-azhouzhou@pstack-claude need a marketplace-root update; see the support matrix and the pstack closeout notes in the changelog.
• Agent Skin users who want the published Azhou art can build from a fresh clone. A local assets/private/ overlay, when present, still wins; overlay-built bundles are not part of this public release.

Verification summary

• python3 scripts/verify.py passed on the release-prep tree and again on a clean checkout in the Draft release workflow (run 43421454114).
• Release-prep PR #466: Required, CodeQL and dependency review passed. The docs-only fast lane skipped the heavy platform legs because this cut changes only CHANGELOG.md and the installation-guide tag example. The 16 preceding commits already landed through #450–#465 with their own checks.

Security / privacy

• Agent Skin code is MIT; Azhou artwork under packages/agent-skin/assets/azhou/ is CC BY-NC-ND 4.0; the Azhou name, character and art are reserved by TRADEMARKS.md. Provenance records, prompts, raw source material, the Codex Pet spritesheet and any assets/private/ overlay stay out of this release.
• No secrets, private paths or unpublished raw assets were added to Git for this cut.

Known holds

• Pet real-machine visual acceptance remains open.
• Release artifacts ship unsigned. This tag has no downloadable assets; signing stays on the rehearsed keyless sigstore path and will apply once downloadable artifacts exist. Git tags remain unsigned.
• pstack 0.15.37 remains a candidate, not an accepted pstack product release (inventory, mirror-gap and patch-debt receipts record holds H1–H5). This Hub patch does not change that status.

Full changelog

v2.0.0...v2.0.1

v2.0.0

User-visible outcomes

Three surfaces are physically retired in this cut (spec #442, PR #449):

• agent-orchestration-kit (model router) — deleted with its entire Hub integration surface; never released, so no shipped version is affected.
• Legacy codex-startup-animation root — deleted after the maintainer-reviewed receipt-ownership handoff; the maintained Pet toolchain lives in agent-skin.
• agent-skin's ZCode host layer — removed; Codex and DSH hosts and the shared core keep their behavior. Last supported ZCode build: 0.9.0/build 41.

A permanent retired-surface reverse-scan gate is now part of scripts/verify.py.

Affected packages

• Removed: packages/agent-orchestration-kit (unreleased), packages/codex-startup-animation (last 0.7.1/build 8).
• Changed: packages/agent-skin (ZCode host layer removed; frozen persisted identifiers kept verbatim for installed-state migration). No canonical Skill changed.

Upgrade / migration action

• An installed ZCode skin keeps working as a frozen install; to retire one, follow the exit statement in the README (restore the retained original launcher or remove the entry with its receipts).
• If you ever installed the model router, uninstall it before installing pstack.
• All pre-deletion releases and tags are retained as the recovery/uninstall source.

Verification summary

Per-ticket verify.py gates, final verify on the branch tip, an independent review of the whole branch diff (14 findings, all low, each with evidence; residuals swept in 3befab0), CI 14/14 green, the release workflow's clean-checkout gate, and a maintainer-checkout gate run after archiving the retired root's untracked remnants.

Security / privacy

No redistribution rights changed: THIRD_PARTY_NOTICES rows for the retired sources are rewritten as historical statements; the no-license source remains non-distributable; private artwork and receipts stay outside Git.

Known holds

• The squash merge skipped the standard no-mistakes pass under the maintainer's explicit waiver (recorded in commit e937b48): its claude agent shares a provider quota capped until 2026-11-17.
• Pet real-machine visual acceptance (the #441 tail) remains open.
• Release artifacts ship unsigned; signing activates via the rehearsed keyless sigstore path once downloadable artifacts exist.

Full changelog

v1.0.0...v2.0.0

v1.0.0

What's Changed

🧰 Repository

• feat(hub): brand-prefix all skills and unfreeze super-caveman metadata by @TeFuirnever in #326
• docs(changelog): record 1.0.0 brand-prefix and metadata-unfreeze release by @TeFuirnever in #328
• chore(super-caveman): rebind beb0568c Git-external approval and review by @TeFuirnever in #329

Full Changelog: v0.12.0...v1.0.0

Breaking changes

• Every canonical package path moved to skills/super-<name>/. Relink or copy installs with azhou-setup.
• Old canonical names remain description/compatibility triggers; check residual host hooks with azhou-doctor --host-root.

Upgrade

git pull
python3 scripts/azhou_hub.py setup --target "$SKILLS_HOME"
python3 scripts/azhou_hub.py doctor --target "$SKILLS_HOME"

Verification

• python3 scripts/verify.py and --promotion-evidence passed on the release commit in the Draft release workflow (run 39939809945).
• Windows full-flow and required status checks passed on PRs #326, #328, #329.

Known holds

• Release artifacts are not yet attached; tag signing remains unsigned per docs/releasing.md until a tag-signing step is rehearsed.

v0.12.0

What's Changed

🔐 Security and dependencies

• chore(deps): Bump source-map-js from 1.2.1 to 1.2.2 in /skills/excalidraw-diagram/scripts by @dependabot[bot] in #279

🧰 Repository

• fix(ci): run the signing rehearsal with the pinned sigstore CLI (T7) by @TeFuirnever in #281
• fix(ci): point the signing bundle at its own output path by @TeFuirnever in #282
• docs(evidence): record the keyless signed-release rehearsal (T7) by @TeFuirnever in #283
• fix(skills): pin skill CLIs' stdio to UTF-8 against console locale (#236) by @TeFuirnever in #284
• fix(skills): pin the remaining UTF-8 console surfaces and harness pipes (#236) by @TeFuirnever in #285
• docs(evidence): windows full-flow fully green after the #236 console fix by @TeFuirnever in #286
• chore(super-lavish): bump the locked lavish-axi CLI baseline to 0.1.89 by @TeFuirnever in #289
• docs(research): record the 2026-10-17 upstream freshness re-scan by @TeFuirnever in #290
• feat(skills): add the asd-ste100 controlled-English rewrite skill (#288) by @TeFuirnever in #291
• chore(super-repo-pedant): re-baseline the neat-freak reference onto upstream v3.0.0 by @TeFuirnever in #293
• feat(skills): add the archify presentation-diagram skill (#292) by @TeFuirnever in #294
• docs(research): record the super-caveman rule-9 adoption ride attempt-1 rejection by @TeFuirnever in #295
• docs(research): record the super-caveman rule-9 ride-2 rejection (revision-12b5552a) by @TeFuirnever in #296
• fix(state): enforce private boundaries and modes at every default write site (F01) by @TeFuirnever in #308
• fix(super-repo-pedant): require majority of all judges and reject unresolved holds (F03) by @TeFuirnever in #310
• fix(benchmarks): share one first-pass criterion and require the exact attempt-1 matrix (F04) by @TeFuirnever in #311
• fix(super-llm-wiki): validate the archive lock before freezing the page (F05) by @TeFuirnever in #312
• fix(super-llm-wiki): isolate bad MCP requests and align the archive page contract (F06) by @TeFuirnever in #313
• fix(azhou-hub): judge host-migration residue by rename record and dead references (F08) by @TeFuirnever in #315
• ci(windows): propagate receipt-step exit codes instead of echoing past them (F09) by @TeFuirnever in #316
• docs(setup): publish per-command apply forms and correct stale lavish paths (F11) by @TeFuirnever in #317
• fix(super-repo-pedant): serialize append against the whole archive transaction (F02) by @TeFuirnever in #319
• fix(session-insights): treat a damaged cache hit as a miss, not a crash (F07) by @TeFuirnever in #314
• docs(audit): record the 2026-10-18 repository audit report by @TeFuirnever in #318
• docs(support): distinguish the recorded caveman run from the invariance rebind (F10) by @TeFuirnever in #320
• docs(reconcile): record the transaction lock, pre-write lock validation, and cache miss semantics by @TeFuirnever in #323
• feat(brand-layer): paren-field broadcast grammar with memory credit protocol by @TeFuirnever in #324
• feat(brand-layer): S1 chat presentation + brand-effect audit by @TeFuirnever in #325

Full Changelog: v0.11.0...v0.12.0

v0.11.0

This cycle closes the 2026-10-07 upstream freshness evaluation batch at all three decision gates (#272).

User-visible outcomes

• super-lavish moves its locked npm baseline lavish-axi@0.1.47 → 0.1.83 (#274): the relay-contract regression (optimistic revision guards, stale-write rejection, exact visible-ledger projection, responsive layout, fuzz, entry-budget suites) and a two-baseline isolated-prefix artifact-mode smoke (visible-state block byte-identical across both CLIs) passed on the new CLI. Documented 0.1.83 command-surface deltas: an additive reply command, an 8th explanation playbook, share --private/--site/--unpublish, and hand-authored inline SVG as the diagram default.
• super-repo-pedant's neat-freak capability account is re-audited against upstream v3 (pin bab17831 → upstream HEAD 322346de, 50 commits / 117 files) and judged an equivalent evolution — kept 18 / changed 10 / removed 0 across the 28 baseline rows. neat-freak-compatibility.md moves to contract v2 with a "Baseline and upstream v3 delta" section; parity tests gain positive and negative controls (#275).
• super-caveman's pinned i-have-adhd contract stays byte-identical by decision (T1, #273): the upstream rule-9 rewrite is a presentation-completeness strengthening that contradicts no current benchmark criterion. The corrected decision, rationale and a maintainer-triggered promotion-ride adoption path are recorded in docs/research/2026-10-07-upstream-freshness-scan.md; no skill byte or pin was touched.

Affected skills

super-lavish (locked baseline), super-repo-pedant (contract docs + tests), super-caveman (decision record only). Dependency bumps touch only excalidraw-diagram's local script tooling (jsdom 30.1.2, undici 8.11.2) and pinned CI actions.

Upgrade or migration action

None required. Re-run the super-lavish setup step to pick up the re-locked baseline on existing installs; no schema, receipt or trigger changes ride this release.

Verification summary

python3 scripts/verify.py passed all 8 gates and verify.py --promotion-evidence passed with the Git-external record pair; the release candidate re-passed the full promotion-evidence gate on CI (run 39014807574). The 0.1.83 landing carried its relay regression, fuzz, entry-budget and two-baseline byte-identity smoke; the contract v2 landing carries positive and negative parity controls.

Security and privacy changes

Dependency bumps only. No receipt, permission or data-handling behavior changes.

Known holds

• Tags remain unsigned pending the signed-release rehearsal (#225).
• The neat-freak 28/28 parity claim is scoped to the pinned implementation baseline (bab17831), re-audited against upstream 322346de — not a claim about upstream HEAD.
• super-caveman's pin b42a45a0 is immutable; upstream adoption rides only a maintainer-triggered promotion.

---

What's Changed

🔐 Security and dependencies

• chore(deps): Bump the actions-minor-patch group across 1 directory with 4 updates by @dependabot[bot] in #278
• chore(deps): Bump jsdom from 30.0.1 to 30.1.2 in /skills/excalidraw-diagram/scripts in the npm-minor-patch group across 1 directory by @dependabot[bot] in #248
• chore(deps): Bump undici from 8.10.0 to 8.11.2 in /skills/excalidraw-diagram/scripts by @dependabot[bot] in #249

🧰 Repository

• docs(research): upstream freshness scan with first-pass evaluations by @TeFuirnever in #271
• docs(research): record the t1 decision to keep the pinned i-have-adhd hard cap by @TeFuirnever in #273
• chore(super-lavish): bump the locked lavish-axi CLI baseline to 0.1.83 by @TeFuirnever in #274
• docs(super-repo-pedant): audit upstream neat-freak v3 delta into contract v2 by @TeFuirnever in #275
• chore(changelog): record the landed upstream freshness evaluation batch in Unreleased by @TeFuirnever in #276
• docs(changelog): align the evaluation-batch entry with the review-corrected records by @TeFuirnever in #277
• chore(release): cut 0.11.0 by @TeFuirnever in #280

Full Changelog: v0.10.0...v0.11.0

v0.10.0

v0.10.0 — cross-harness evidence batch and gate restoration

Repository-level minor release (SemVer): lands the accepted u1–u14 upgrade batch (#252) from the 2026-10-06 cross-harness execution-evidence research, plus the promotion-receipt restoration that keeps the public gate honest.

Affected skills

• azhou-doctor: new read-only --host-root probe for stale hook paths and residual installs left by canonical renames, with a host-migration runbook (u2).
• ask-azhou: the routing catalog is now generated from every canonical SKILL.md description with a fail-closed drift gate, and unrouted requests can be recorded as hash-only telemetry (u4).
• eli5: paired MUST-trigger / anti-over-trigger phrase list, and the receipt now carries an audience read-back assertion (u4, u14).
• excalidraw-diagram: GUI-free sandbox fallback for render verification; "sandbox blocked preview" is now an explicit hold instead of a silent skip (u5).
• azhou-setup: same-source conflicts are separated from real conflicts and exit 3 (u6).
• session-insights: zcode 3.14.4 store scout concludes a bounded go (metadata-only, read-only); the Kimi Code support-matrix column lands behind a fresh receipt (u1, u9).
• Super Caveman promotion machinery: validators name the exact stale field on re-derive, raw record shapes are versioned, and two receipt rebinds (#253, #269) restore the --promotion-evidence replay (u7).
• Repository gate: per-skill platform view derived 1:1 from the support matrix with a consistency check (u3), reference-doc script-path check with negative control (u13), verify.py ends on a single aggregate verdict with a machine-readable summary (u8); docs additions cover the interrupt/quota resume protocol (u10), treehouse lease-return cwd step (u11) and the offline dependency recovery runbook (u12).

Upgrade / migration action

• Re-run azhou-setup / azhou-doctor against your install; if you scripted against setup exit code 1, match conflict-same-source via exit 3 or the JSON status.
• If your host hooks still reference skills/llm-wiki/, skills/repo-pedant/ or other pre-rename paths, run the new doctor --host-root probe and follow skills/azhou-doctor/references/host-migration.md.

Verification summary
Every landing rode the full deterministic gate (python3 scripts/verify.py) locally and 12/12 CI checks before merge; the release cut ran scripts/verify.py --promotion-evidence against the receipt-only rebind locally and in this release workflow.

Security / privacy
Unrouted telemetry stores SHA-256 hashes, counts and day stamps only; the zcode scout read metadata and schema, never transcript content.

Known holds
Tags are not signed yet (signed-release rehearsal pending, #225); the pre-rebind review record remains missing from Git-external storage (#170); the session-insights zcode adapter stays fail-closed pending an adapter ticket (scout verdict: bounded go).

What's Changed

🧰 Repository

• docs(agents): repoint the promotion-record bookkeeping at the 0.9.0 cut receipt by @TeFuirnever in #205
• docs(readme): regenerate bilingual READMEs with azhou-covers hero image by @TeFuirnever in #206
• feat(hub): retire the super-ci-test-reliability and super-prose-standard canonical skills by @TeFuirnever in #207
• docs(support-matrix): correct the neutral availability count to fourteen by @TeFuirnever in #208
• feat(session-insights): roast tone over the same numbers (#164) by @TeFuirnever in #209
• feat(session-insights): offline single-file HTML artifact (#165) by @TeFuirnever in #210
• feat(session-insights): Codex adapter for the shared pipeline (#166) by @TeFuirnever in #211
• feat(session-insights): deliver the zcode fail-closed hold as final (#168) by @TeFuirnever in #212
• docs(research): OpenSSF Best Practices pre-enrollment draft (#27) by @TeFuirnever in #213
• docs: reconcile stale facts after the session-insights M2/M3 landings by @TeFuirnever in #214
• docs(research): production-readiness gap audit + autoresearch non-NVIDIA feasibility by @TeFuirnever in #215
• docs(readme): correct the benchmark-integrity suite count to four by @TeFuirnever in #216
• docs(agents): scaffold the matt-pocock skills configuration by @TeFuirnever in #217
• docs(installation): teach consumers to verify what they installed (#220) by @TeFuirnever in #227
• docs: minimal behavior eval-case contract for the no-benchmark tier (#222) by @TeFuirnever in #228
• feat(autoresearch): named unsupported holds for non-NVIDIA accelerators (#219) by @TeFuirnever in #226
• docs(evidence): named threat-review of distribution path and receipt parsing (#221) by @TeFuirnever in #229
• feat(ci): coverage and regression-rate baseline job (#223) by @TeFuirnever in #231
• feat(ci): windows full-flow receipt workflow (#224) by @TeFuirnever in #233
• fix(ci): use RUNNER_TEMP in the windows full-flow steps by @TeFuirnever in #234
• fix(super-llm-wiki): never render store-controlled text into SessionStart context (#230) by @TeFuirnever in #232
• fix(ci): windows receipt steps tee output and never abort mid-flow by @TeFuirnever in #235
• docs(evidence): windows full-flow receipt with encoding divergence (#224) by @TeFuirnever in #237
• feat(ci): keyless signed-release rehearsal workflow (#225) by @TeFuirnever in #238
• fix(ci): sign via the sigstore CLI (composite action outside the allowlist) (#225) by @TeFuirnever in #240
• chore(assets): ignore azhou-scenes generation scratch by @TeFuirnever in #242
• docs(readme): unify skill effect previews as azhou-scenes showcase sections by @TeFuirnever in #243
• docs(skill-standard): bind readme showcase sections to scene figures 1:1 by @TeFuirnever in #244
• docs(closeout): sync promotion pointer and changelog for readme showcase landing by @TeFuirnever in #245
• fix(policy): separate local tool installations from public packages by @TeFuirnever in #246
• docs(closeout): record local tool installation boundary by @TeFuirnever in #247
• docs(readme): embed promo video under hero by @TeFuirnever in #251
• fix(promotion): rebind super-caveman receipt for post-#251 protected-blob drift by @TeFuirnever in #253
• docs(research): scout zcode 3.14.4 sqlite session store as bounded go by @TeFuirnever in #254
• feat(azhou-doctor): detect stale host hook paths and residual installs after renames by @TeFuirnever in #255
• docs(platform-view): derive per-skill platform view from the support matrix by @TeFuirnever in #257
• feat(router): generate the routing catalog and record unrouted misses (u4) by @TeFuirnever in #256
• fix(setup): separate conflict-same-source from real conflicts with exit 3 (u6) by @TeFuirnever in #258
• fix(verify): end on a single aggregate verdict with a machine-readable summary by @TeFuirnever in #260
• feat(eli5): assert audience read-back in the receipt judgment layer (u14) by @TeFuirnever in #259
• feat(excalidraw-diagram): add GUI-free sandbox fallback and explicit preview holds (u5) by @TeFuirnever in #261
• feat(promotion): name stale fields on re-derive and version raw record shapes (u7) by @TeFuirnever in #265
• docs(kimi): add Kimi Code support-matrix column behind a fresh receipt (u9) by @TeFuirnever in #264
• docs(skill-standard): document the interrupt and quota resume protocol (u10) by @TeFuirnever in #263
• docs(worktree-policy): cd out of the worktree before returning a lease by @TeFuirnever in #262
• docs(installation): add offline dependency recovery runbook (u12) by @TeFuirnever in #266
• feat(policy-gate): verify reference docs cite existing script paths (u13) by @TeFuirnever in #267
• chore(changelog): record the ...

v0.9.0

v0.9.0 (2026-09-12)

User-visible outcomes — the hub grows to sixteen canonical skills and Windows becomes an enforced platform rather than a sensor:

• New skill session-insights (#161/#163): fact-bound agent-usage insight reports from local session stores, with a Claude Code adapter, a deterministic aggregate/report pipeline, and an incremental metadata cache (#161/#167) whose cold/warm/deleted/corrupt states are byte-identical under benchmark and unit negative controls.
• Fidelity axis (#176/#177–#183): every skill now carries an evidence-cited classification; the five adapted skills were renamed under the super- prefix (ci-test-reliability → super-ci-test-reliability, prose-standard → super-prose-standard, repo-pedant → super-repo-pedant, lavish → super-lavish, llm-wiki → super-llm-wiki); the repository gate fails prefix-fidelity contradictions and rename residue.
• Windows enforced (win-01..win-13): Verify on Windows flipped into the Required gate (#188) after the POSIX-only os.fchmod root-cause fix (#184) and two further Windows-only divergences were closed; docs/support-matrix.md and both READMEs claim Windows as enforced.
• Runtime state: chained migrations no longer travel a prior migration receipt as content (#199), with a regression test across all three azhou_runtime_state.py copies.

Affected skills — session-insights (new), super-caveman, super-repo-pedant, super-lavish, super-llm-wiki, super-prose-standard, super-ci-test-reliability, plus the repository gate and runtime-state machinery.

Upgrade / migration action — reinstall the five renamed skills under their new canonical names. Existing runtime state under .azhou/<old-name>/ is migrated by migrate_state.py under the documented dry-run plan / atomic apply / source-preserved protocol; legacy receipt schema spellings remain dual-accepted; old names survive only as compatibility triggers.

Verification summary — python3 scripts/verify.py exit 0 (951-file repository policy, 467 unit tests, six benchmark-integrity suites, whitespace); CI green on Python 3.11/3.14, the Windows Required leg, CodeQL, Scorecard and Fuzz; super-caveman revision-8f493670 re-bound through the producer-context invariance pathway with workspace-owner exact-diff approval (PR #204).

Security/privacy changes — none new. The session-insights metadata cache stores no transcript text (first-prompt repetition is counted on SHA-256 only) and the skill remains a read-only observer that never writes inside a session store.

Known holds — session-insights ships wiring-integrity benchmarks only (no behavior benchmark yet); its Codex and zcode adapters fail closed pending format verification; super-caveman's invocation-class declaration remains held by the gate declaration table until its next whole-tree promotion ride.

What's Changed

🧰 Repository

• feat(skill-standard): hold super-caveman's invocation class in the gate, require the rest by @TeFuirnever in #169
• feat(skills): windows/macos cross-platform compatibility groundwork (win-01..09) by @TeFuirnever in #171
• feat(ci): collect windows verification evidence as a non-blocking job by @TeFuirnever in #172
• test(gates): skip posix-bound check and benchmark tests on windows by @TeFuirnever in #173
• feat(super-caveman): promote revision 74f8d585 (windows atomic-write hardening + command surface) by @TeFuirnever in #174
• fix(super-caveman): land Windows fchmod guard via producer-context invariance pathway (revision-8f493670) by @TeFuirnever in #184
• feat(ci): enforce the deterministic gate on Windows with OS support declarations (win-10) by @TeFuirnever in #185
• test(super-caveman): preserve PATH in cleared test environments (win-13) by @TeFuirnever in #186
• docs(evidence): win-13 windows CI root-cause diagnosis and green receipt by @TeFuirnever in #187
• feat(ci): enforce the deterministic gate on Windows (win-13 required flip) by @TeFuirnever in #188
• docs(specs): land session-insights and super-prefix specs with backing research by @TeFuirnever in #189
• fix(super-caveman): authenticate invariance-pathway landings in promotion-evidence replay (#170) by @TeFuirnever in #190
• docs(github): add approved proof visuals by @TeFuirnever in #191
• chore(treehouse): raise the worktree pool to 8 by @TeFuirnever in #192
• docs(research): zcode session store format scout (session-insights #162) by @TeFuirnever in #193
• feat(session-insights): sixteenth canonical skill with Claude Code aggregate/report pipeline (#161, #163) by @TeFuirnever in #194
• docs(skill-standard): fidelity axis with evidence-cited classification for all sixteen skills (#176, #177) by @TeFuirnever in #195
• refactor(super-ci-test-reliability): rename from ci-test-reliability under the fidelity axis (#176, #178) by @TeFuirnever in #196
• refactor(super-prose-standard): rename from prose-standard under the fidelity axis (#176, #179) by @TeFuirnever in #197
• refactor(super-repo-pedant): rename from repo-pedant under the fidelity axis (#176, #182) by @TeFuirnever in #198
• fix(runtime-state): chained migrations no longer travel the prior receipt as content by @TeFuirnever in #199
• refactor(super-lavish): rename from lavish under the fidelity axis (#176, #181) by @TeFuirnever in #200
• refactor(super-llm-wiki): rename from llm-wiki under the fidelity axis (#176, #180) by @TeFuirnever in #201
• feat(gate): enforce prefix-fidelity and rename-residue in the repository gate (#176, #183) by @TeFuirnever in #202
• feat(session-insights): incremental metadata cache for aggregate runs (#161, #167) by @TeFuirnever in #203
• docs(changelog): cut 0.9.0 by @TeFuirnever in #204

Full Changelog: v0.8.0...v0.9.0

v0.8.0

What's Changed

🧰 Repository

• test: pin docs-to-code contracts and completion criteria by @TeFuirnever in #138
• feat(llm-wiki): optional decision lifecycle with alternatives gate by @TeFuirnever in #139
• feat(llm-wiki): archive freeze with content-hash tamper detection by @TeFuirnever in #140
• feat(llm-wiki): supersession hint, receipt v3, migration half-state pin by @TeFuirnever in #141
• feat(repo-pedant): consumer classification for simplification candidates by @TeFuirnever in #142
• feat(skills): add ci-test-reliability canonical skill by @TeFuirnever in #143
• feat(skills): add prose-standard canonical skill with recall batteries by @TeFuirnever in #144
• feat(skills): add ask-azhou router with gate-enforced coverage by @TeFuirnever in #145
• docs(changelog): land the absorption batch under promotion record evolution by @TeFuirnever in #146
• docs(changelog): cut 0.7.0 by @TeFuirnever in #147
• feat(skill-standard): standardize the invocation-class axis by @TeFuirnever in #151
• bench(repo-pedant): consumer-classification end-to-end case and first run by @TeFuirnever in #152
• feat(llm-wiki): surface decision lifecycle on the MCP tools by @TeFuirnever in #154
• docs(changelog): land the invocation-axis, eval-case, and MCP-lifecycle entries by @TeFuirnever in #155
• docs(agents): align the promotion-record parenthetical with the re-bound receipt by @TeFuirnever in #157
• docs(readme): refresh catalog facts and add the invocation-axis overview by @TeFuirnever in #158
• feat(skill-standard): declare the invocation class for fourteen canonical skills by @TeFuirnever in #159
• docs(changelog): cut 0.8.0 by @TeFuirnever in #160

Full Changelog: v0.6.0...v0.8.0

v0.6.0

• Promoted the super-caveman promotion ride (revision 93f38a6b) through a fresh-producer evaluation after the landing simulation proved the package's promotion digest binds its whole tree: the brand layer gains the Startup value discipline section (the English twin of the promoted startup-line rule) and a pass/scope consistency bullet, the SKILL.md Persistence paragraph is deduplicated against Runtime boundary, and the duplicated setup.md heading is fixed. The byte-reuse chain is retired for this revision: all 19 response and 19 lifecycle cases were re-evaluated by GLM through isolated zcode subagents (criteria panel 19/19 three times; blind paired A/B against a baseline set produced from the pre-change tree under the same producer voted 3:0 candidate with two judges finding the candidate strictly better), a pre-flight run without the contracted fixtures is disclosed and discarded in the summary limitations, and the shared brand-startup discipline test now covers all four brand layers. Raw output sets, the review record, and the exact-diff approval records are Git-external; the in-repo approval record binds seven reviewed blobs including the extended test, manifest inputs, bookkeeping, and this entry.
• Promoted the second same-day evolution batch from the maintainer exact-diff checkpoint (all candidates 3:0 in independent paired reviews with reversed A/B order; the landing simulation passed the public gate before apply): the excalidraw-diagram style verdict rewrites references/design-system.md Modern Aesthetics as a user-named opt-in aligned with the live hand-drawn gate — roughness: 1 / fontFamily: 1 (Virgil) stay the defaults, roughness: 0 / Cascadia exist only inside the named style, and the dangling "Deliverables & Export" pointer is gone — pinned by a new docs-alignment regression test that also fixes the entry, json-schema.md and element-templates.md default statements; the startup-line value discipline rolls out to the llm-wiki, repo-pedant, and arch-doc brand layers as purely additive sections (concrete values, unresolved as the only legal pending marker backfilled at each skill's scope-lock point, angle brackets contract-only) with per-skill status-consistency rules and a shared public-gate regression test that pins every startup contract line verbatim and guards the promoted excalidraw-diagram discipline; and a nine-file wording-prune batch (nine skills) removes the duplicated "emit the startup protocol" workflow step in six single-stage skills (timing moves into each Brand protocol line), collapses arch-doc's duplicated note-format, Planned-synonym and parameter-drift rules into pointers at the canonical references, deletes one redundant repo-pedant prohibition that restates the body verbatim, and fixes two dangling pointers in diagram-types.md (an orphan-period defect a judge flagged was fixed and re-reviewed before landing).
• Approved but not landed: the super-caveman variant of the startup-line discipline (English brand layer, approved_deferred) rides the next super-caveman promotion — the landing simulation proved that package's promotion digest binds its whole tree including references/brand-layer.md, so any text change there requires a fresh benchmark attempt and approval record; the same constraint moved two prune items (SKILL.md adapter-overview dedupe, duplicated setup.md heading) onto the super-caveman-promotion-ride ticket.
• Promoted four evidence-gated skill evolutions from recorded zcode/Claude/Codex usage (signals + isolated regressions + 3-judge paired reviews, all 3:0, maintainer approval on green gates): repo-pedant routes real audit phrasings (tech-debt/repo-health-check) with an anti-over-trigger guard and its description is pruned to body-covered content; azhou-verify derives its verdict mechanically (green requires exit 0, the gate's passing tail and no failure verdict line anywhere) with a stdlib check-gate-report.py validator; excalidraw-diagram brand-layer gains a startup-line value discipline (unresolved as the only legal pending value, template text rejected in real broadcasts) with a stdlib check-brand-startup.py validator. Regression tests pin all validator fixtures in the public gate.
• Amended docs/skill-standard.md section 3 with the startup-line value discipline promoted by the excalidraw-diagram evolution (real broadcasts carry concrete values, the machine-stable literal unresolved is the only legal pending marker, angle-bracket templates stay contract-only); fixed the arch-doc entry and provenance template pointing at a repo-root scripts/verify_doc.py that never existed (scripts live under the skill directory) with a new regression test pinning every referenced script path; published the cross-harness trigger-phrase inventory under docs/research/.
• Promoted the lavish entry slimming evolution: the artifact-mode workflow, visual guidance, playbooks, and command rules moved verbatim into a new references/artifact-mode.md (relay workflow, authorization boundaries, receipts, mode selection, and the locked npx baseline stay in the entry), cutting the always-loaded entry from 15,479 to 8,512 bytes against the frozen 12,000-byte budget derived from the recorded zcode context-overflow sessions; a public-gate regression test pins the byte budget and the reference wiring. 3:0 paired judge majority; pointer issues flagged by judges (relay polling-rule scoping, relay <skill-dir> resolution, in-reference relative link) fixed before landing.
• Completed the arch-doc skill backlog: brand layer, setup contract, controlled-evolution files, benchmark runner with evidence receipt, demo document, read-only template provenance and deterministic gate refinements (--plantuml-cli, --trace).

Added

• Completed the Super Caveman Codex lifecycle adapter (spec #115, implementation #116): codex_adapter.py now owns one UserPromptSubmit registration beside SessionStart in the same explicit scope and adds prompt/enable/disable/status; render and prompt delegate to the canonical claude_adapter handlers so the capsule builder, mode hierarchy, stop phrases, one-shot routes and persistent defaults stay single-source — the Codex SessionStart capsule becomes the mode-aware canonical capsule (disclosed codex-host-visible change), covered by the existing contract tests plus a new lifecycle suite; live firing stays unclaimed until a trusted real-host receipt exists. Landed via the byte-reuse promotion chain with two isolated response-output-neutral delta reviews and fresh exact-diff approval.
• Added the opt-in zcode lifecycle adapter to super-caveman (scripts/zcode_adapter.py): two proven events (SessionStart plus UserPromptSubmit) managed in .zcode/cli/config.json for one explicit scope, reusing the canonical Claude adapter handlers so the capsule builder, mode hierarchy and session state machine stay single-source; a defensive camelCase-to-snake_case payload bridge, a 15-case deterministic contract test file, a config-registration smoke receipt carrying the verified headless -p no-hooks negative, and a re-derived evidence/zcode-hook-surface-2026-09-04.md replacing the 2026-09-02 receipt found committed empty. Promotion followed the byte-reuse chain re-bind: the 19-case output set is byte-identical, two isolated delta reviews confirmed response-output neutrality, and the superseded revision-ed8a4b80 evaluation is re-bound to the new tree digest.
• Added the relay mode to lavish by merging Spec Relay into it as the single canonical package with two modes: artifact remains the unchanged general rich-HTML review loop while relay serves the portable handoff packet with embedded spec-relay.html-state.v1 state, stable data-review-ids, feedback disposition, next-owner handoff, optimistic revision guards and spec-relay.receipt.v1 receipts; relay_state.py moved byte-identical into skills/lavish/scripts/ alongside the Spec Relay reference, machine-stable names (the state schema, spec-relay.receipt.v1, the spec-relay: CLI error prefix and the pinned lavish-axi@0.1.47 baseline) are preserved, receipts and the fuzz workflow re-point at the single package, and the canonical package count drops from twelve to eleven.

Fixed

• Fixed the llm-wiki hook-wiring renderer emitting a bare "*" matcher: runtimes that compile hook matchers as regular expressions reject it as invalid ("nothing to repeat"), so render-hooks now emits the valid match-all ".*" and a regression test pins all three event matchers as compilable regular expressions ending at the canonical event name, and the SessionEnd wiring now requests 3 seconds, matching the short session-end budget hook hosts enforce instead of triggering a clamp at load time.

Changed

• Unified the Azhou identity anchor across all canonical packages: every SKILL.md now carries the 🦊 阿舟 · <Skill> identity and motto in its own body, Super Caveman included — its stage-event protocol stays byte-identical in references/brand-layer.md and ordinary terse replies still add no lifecycle emoji. check_skill_brand_contract enforces identity and motto against SKILL.md itself instead of the combined skill+brand surface, a regression test proves identity removal from SKILL.md alone fails the gate even when the brand layer retains it, and docs/skill-standard.md states the uniform rule without the previous Super Caveman exemption. The Super Caveman package tree digest changes, so the passing evaluation re-binds only through fresh promotion evidence.

Planned

• First complete cross-harness evidence set.

v0.5.0

What's Changed

🧰 Repository

• fix(spec-relay): add bounded fuzzing for the relay state parser (#27 fuzzing slice) by @TeFuirnever in #91
• docs(evidence): canonical SKILL.md package load receipt (zcode) by @TeFuirnever in #93
• docs(evidence): zcode hook-surface receipt by @TeFuirnever in #94
• docs(evidence): repo-pedant manual invocation receipt (zcode) by @TeFuirnever in #95
• docs(evidence): lavish rich-HTML review-loop receipt (zcode) by @TeFuirnever in #96
• docs(research): zcode lifecycle adapter feasibility (two-event design on receipt-backed contract) by @TeFuirnever in #97
• docs(evidence): super-caveman compact delegation receipt (zcode) by @TeFuirnever in #98
• docs(evidence): spec-relay review-loop receipt (zcode) by @TeFuirnever in #99
• docs(support-matrix): receipt-back zcode cells via fresh SC promotion (b829d7c5) by @TeFuirnever in #100
• feat(skills): absorb eli5 and autoresearch as canonical Azhou skills by @TeFuirnever in #101

Full Changelog: v0.4.1...v0.5.0

Note on the tagged tree: the v0.5.0 tag is immutable under this repository's tag-protection ruleset and was cut from cebe9dd, minutes before the [0.5.0] CHANGELOG section itself landed in 4ad46db (#102). The section on main is authoritative for this release; the tagged tree carries the same code delta with the changelog section arriving one commit later.