Skip to content

CHANGELOG.md

Latest commit

 

History

History
961 lines (812 loc) · 75.3 KB

CHANGELOG.md

File metadata and controls

961 lines (812 loc) · 75.3 KB
Raw

Changelog

All notable changes to harness-anchor are documented here. Format follows Keep a Changelog.

Unreleased

0.18.0 - 2026-07-31

MINOR: watchdog coverage on two hooks that never had it is a new backward-compatible capability, and it ships with a new gate check and a new contract test — an ### Added section, which per the versioning rule means MINOR rather than PATCH.

Added

  • hooks/stop and hooks/user-prompt-submit now carry the R1 5s watchdog. Both fork a JSON engine — ha_json_engine_init probes by running python3 -c 'print(1)', and both then call ha_flist_active — so a wedged interpreter (antivirus scan, network-mounted PATH entry, broken python3 shim) hung them indefinitely with nothing to intervene. Measured against the wedged engine before the fix: 25s and still running when the harness cut it off, on both hooks. After: 5s, emitting nothing, exit 0. The three peers have had this since v0.10.0; these two were missed and nothing noticed for eight releases. ha_json_engine_init is deliberately inside main() — outside it, the probe's own hang would sit in front of the watchdog rather than behind it.

  • tests/hook-contracts/stop-prompt-timeout.sh — wedges every engine in the chain (python3/python/py/node stubs that sleep 30) against an anchored fixture, then asserts each hook exits 0 within 8s emitting nothing. Verified red on the pre-fix hooks before being accepted as green on the fixed ones.

  • scripts/validate-anchor.sh [11/12] now asserts the watchdog's presence in every hook hooks/hooks.json registers, so a newly added hook cannot ship without one. Structural check: it catches absence, not misbehaviour — proving the watchdog fires is the contract test's job. (163 assertions, was 158.)

  • Invariant #5 is now enforced, and reworded to describe what actually gates what. It read "C/C++ skills are gated by cpp-detect.sh. They must include a frontmatter trigger condition referencing the detected build system" — wrong on both counts. cpp-detect.sh gates the banner's project type, the meta-skill's cpp-only regions and the two C/C++ commands, but not skill loading: skills are chosen by description matching, and nothing consults it at that moment. And three of the four skills name no build system, correctly — clang-format has no reason to. The rewritten invariant states both gates and requires what actually does the work: every skills/cpp-* description must scope itself to C/C++ inside the first 80 characters, the window invariant #6 reserves for trigger load. validate-anchor [3-4/12] now asserts exactly that, with a non-vacuity guard for the day the glob stops matching. (168 assertions, was 163.)

    The skills were left alone deliberately. Editing four descriptions to name build systems would spend the scarcest trigger real estate on a term irrelevant to three of them — and CLAUDE.md forbids rewriting skill descriptions without evals showing improvement. It also would not have made the invariant's first sentence true. An unenforced entry in that list contradicts what the section below it promises: the list is the part that a test mechanically enforces.

  • Invariant #9 is now enforced too — its wording needed no change. Auditing the other nine invariants after fixing #5 found that #9 ("docs-lookup is the canonical procedure; new skills MUST reference it rather than inlining Context7 → WebSearch waterfalls") was accurate but unchecked. validate-anchor [3-4/12] now asserts that any skill naming Context7 or WebSearch also references docs-lookup. Naming them is normal and expected — all seven current cases use the shape "Invoke docs-lookup … (Context7 → WebSearch fallback)"; what the check catches is that sentence with the reference removed, which is how the failure-mode detection and the calibrated-uncertainty fallback drift out of sync between copies. Non-vacuity guarded; both directions verified by mutation.

    Invariant #6 was first written up in this release as "the remaining unenforced entry". That was wrong, and the correction is the more useful finding: #6 is enforced by the skill-triggering harness — check-coverage.sh asserts on every CI run that each skill has an adversarial prompt that never names it, and run-all.sh puts those prompts through a live session before a release. The error came from equating "no grep for it" with "not enforced"; a behavioural eval is the stronger form of enforcement, not the absence of one. #6 now says where it is enforced, in what tier, and warns against replacing that eval with a static check on the description text. With #5, #7 and #9 closed in this release, all ten invariants are enforced, which is what the section below the list has always claimed of it.

  • check-coverage.sh now verifies the prompts are actually adversarial, and self-tests its own matcher first. It had asserted only that a case was registered and its file existed — so a prompt reading "use the docs-lookup skill" would have satisfied it while testing nothing, since the run would pass on the user naming the skill rather than on the description doing its job. Prompts are now rejected if they contain their own skill name in either spelling.

    The first cut of that check was itself vacuous, which is the more useful half of the story: grep -qi -e A -e B aborts on the MSYS2 grep (exit 134, nothing on stderr) — the same crash class tests/windows-compat.sh documents for -i combined with -F. A crash returns non-zero, non-zero read as "no match", and all fifteen prompts were reported clean by a grep that never ran. The loop-counter guard did not notice, because the loop had run — it was the judge that was broken. Single -E alternation is crash-free, and a matcher self-test against a known-positive now runs before any prompt is believed clean. Both directions verified by mutation. (44 assertions, was 28.)

  • tests/unit/doc-align.sh — integrity of the doc-align markers, which were a standing unverified claim. Per marker: exactly one 40-hex sha, it resolves to a real commit, that commit is an ancestor of HEAD, and any abbreviated sha in the prose agrees with it. Markers are discovered by glob with a non-vacuity guard. Two real failures in one session motivated it: a 40-character sha typed out from a 7-character one (the second time that has happened in this repo), and a marker left pointing at a commit an --amend had rewritten. Both look exactly like a correct marker when read. Sha resolution needs real history, so it SKIPs — loudly, on its own line — where the clone is shallow; the CI fast-tests job now checks out with fetch-depth: 0 so that the two assertions that matter actually run there rather than self-skipping past the exact case they exist to catch.

  • tests/bench/hook-timing.sh covers pre-compact, and guards its own list. The benchmark for the 5s budget enumerated four hooks and had silently omitted pre-compact since v0.15.0 — the test for invariant #7 was not measuring one of the hooks the invariant governs. The hooks stay enumerated, because each needs its own env/stdin to reach its active path, but a completeness check now fails if any hook on disk is not benchmarked.

  • docs/troubleshooting.md #16 — "a hook goes completely silent". With every hook now bounded at 5s, a wedged JSON engine produces silence indistinguishable from a legitimately quiet hook. The entry names the likely causes (antivirus on first interpreter run, the Microsoft Store python3 alias, an unreachable network share on PATH), gives the two timing commands that tell them apart, and states that any one engine of the four is enough.

Fixed

  • docs/commands.md had drifted for six releases and said so nowhere. Its doc-align marker still pointed at 74a06eb (v0.12.0, 2026-07-14) — an assertion of "verified" that had not been true since v0.13.0. Re-read against every file under commands/ plus the scripts and agent that implement them. What the stale marker had been hiding: /verify's report was documented with five sections when verification-runner emits nine and had gained ### Integrity; /status was documented as parsing JSON with python3→node when the chain has been python3 → python → py -3 → node → pure-bash since v0.13.0; /session-end's fact-gathering omitted the secrets scan and state-hygiene facts, and its steps omitted the consent-gated golden-rules consolidation and the secrets-before-commit gate; /gc never mentioned that doc-drift scanning is bounded and can return PARTIAL; the discoverability table still described PostToolUse as firing only on Edit/Write, which stopped being true in v0.15.0. The new marker carries an explicit scope note naming what was checked against implementation versus merely read.
  • tests/README.md listed five hook-contract tests while sixteen shipped, and its "quick test" block enumerated them by hand directly below a comment explaining that enumeration rots. The per-file table is now indexed by hook — five rows that change when the hook set does, not when a test is added — and the quick-test block globs. Also corrected there: the hook-contract timing (two tests deliberately wait out a 5s watchdog, so <5s was never true of the directory) and the CI matrix, which has included Windows since v0.13.0.
  • docs/design.md claimed cpp-detect.sh gates which C/C++ skills load. It does not. Skills are selected by description matching — all four C/C++ ones open with "Use in C/C++ projects", which is the actual gate — while cpp-detect.sh gates the injected meta-skill's cpp-only regions and makes /cpp-init and /sanitize refuse outright. The sentence had restated CLAUDE.md's normative wording ("must include a frontmatter trigger condition") as a description of what the files do; only one of the four actually names a build system.
  • tests/windows-compat.sh [1/5] punished documenting the hazard it polices. Its own comment promised that "prose mentions and comments must NOT trip this", but the grep had no comment filter — so a comment explaining that ha_json_engine_init probes by running python3 -c was reported as a bare python3 invocation. It fired on both hooks touched in this release. Comment lines are now dropped exactly as [2/5] already did it. Coverage is unchanged: an executable line is never comment-leading, and a trailing comment after real code still matches. Both directions verified by mutation.

Changed

  • README.md rewritten to lead with the failure modes and a real hooks/session-start banner rather than a component inventory; the long-form rationale moved to the new docs/design.md. Also corrects a stale hook count (four claimed, five shipping), a false "Subagents (5, read-only)" heading (index-curator carries Write — it has to), and a description of done_criteria as booleans (it is an array of strings; the enforced rule is evidence: null ⇒ status cannot be pass, per feature_list.schema.json).
  • docs/design.md added — the design rationale at full depth, with the alternative rejected in each case. Its warn-only section corrects the previous README's claim that all four warn hooks inject additionalContext: Stop has no such channel and uses systemMessage, and PreCompact reaches only the user.

0.17.1 - 2026-07-31

PATCH, not MINOR: both entries are repairs to components that already shipped — nothing belongs in an ### Added section.

Fixed

  • doc-drift-scan.sh flooded the consumer it reports to. HARD_CAP bounded the symbol set; nothing bounded the candidate rows, which are the unit the reader actually pays for. Measured on this repository's own v0.16.0..v0.17.0 range: 54 symbols → 4765 rows / 759 KB. The 3000-token SessionStart budget invariant is this repo's scale for what injected context may cost; 759 KB from one sensor is not in that world. Wherever the consuming tool's output limit falls, a payload that size is past it — so drift-analyst was adjudicating a list the harness had already cut, silently, with no note either end could see (an inference from the size, not a measurement of a particular tool; the budget argument stands without it). The script announced PARTIAL for the one truncation it performed and was blind to the larger one it caused. Rows are now capped per symbol (12) and in total (400), each with a PARTIAL marker, and the summary reports matched-vs-shown (K candidate(s), S shown). Same range after: 379 rows / 70 KB (−91%).
  • doc-drift-scan.sh harvested comment prose as if it were code. Symbol extraction reads raw diff lines, comments included, so English containing word ( becomes a "symbol": O (from O(files)), b (from the regex \b(${ALT})) and it (from prefixes it () produced 3100 of those 4765 rows — 65% of the output, none of it naming any code symbol. Tokens under three characters are no longer searched at all: prefix matching is case-insensitive by design, so their rows are undecidable at any count. They are named individually on stderr — a symbol the scan chose not to look for must never read as one it looked for and found nothing about. This does not touch the common-word noise the header documents (read, get, check are real identifiers whose rows a reader can judge; they stay).
  • tests/windows-compat.sh stopped checking the newest hook two releases ago. Its HOOKS list was hard-coded and never learned about hooks/pre-compact (v0.15.0), so checks [1/5] (no bare python3), [3/5] (eol=lf) and [5/5] (sources portable.sh + ha_platform_init) skipped it — and [1/5] has no other coverage anywhere in the suite. Nothing was broken (the hook complies with all three), but invariant #10's only mechanical gate was not enforcing it. Hooks are now globbed, with a non-vacuity guard so an empty discovery fails loudly instead of passing every loop below it. Same defect class tests/README.md and the CI workflow already name in their own comments: enumeration rots.
  • Two troubleshooting entries gave pre-0.13.0 / pre-0.16.0 advice. #6 named a missing python3 as a cause of hook-contract failures — false since v0.13.0 made the engine chain python3 → python → py -3 → node → pure-bash and made an engine-less machine emit SKIP, never FAIL, so the entry sent Windows users chasing an interpreter that cannot be the problem. #5 told the reader to install a build tool on the strength of init.sh's command -v check, which is PATH-only — the exact inference 0.16.0's discovery chain exists to prevent, in the guide that is supposed to teach it.
  • The meta-skill's eol=lf pin had never been verified. .gitattributes pins skills/using-harness-anchor/SKILL.md because hooks/session-start awk-consumes it for the conditional-region filter and the injection length count — and tests/windows-compat.sh [3/5] is the only place in the repo that reads an eol attribute at all. It did not check that file. It does now.
  • A hook registered in hooks.json but missing on disk was invisible to every gate. validate-anchor [1/12] names two of the five hooks; [11/12] enumerated all five by hand but only to check their wiring; hooks.json itself is validated for JSON syntax alone. [11/12] now derives its list from hooks.json — the file Claude Code actually executes — and runs registry → disk: registered implies exists, sources portable.sh, calls ha_platform_init. Complementary to windows-compat [5/5], which globs hooks/ and runs disk → wiring; neither direction alone catches both failures. An empty parse, or a registration in a shape the parser does not recognise (parsed count ≠ declared "type": "command" count), fails loudly instead of skipping the loop.

Changed

  • agents/drift-analyst.md reads the scan's stderr through a six-row state table (up from three prose states), including the two new cap markers and the never-searched token list. Its known-blind-spot header records the new three-character floor: a clean doc-drift section now means "no doc claim about a changed, 3+ character, function-shaped symbol in a scanned language looks stale".
  • docs/troubleshooting.md gains entries 14 and 15, covering the two sensors 0.16.0/0.17.0 added — both of which are built to report things that look like failures and are not (doc-drift-scan's six stderr states and its candidates-not-violations contract; cpp-tool-discovery's NOT_FOUND for a tool in a non-standard prefix, and the dated phrasing that lets init-verification re-check it next session).
  • docs/troubleshooting.md's doc-align marker is re-verified and bumped, v0.9.0 → v0.17.1. It had sat at v0.9.0 through eight releases — two of which added entries to that very file — while its own text said "re-verify and bump this marker if they change". All 15 entries were re-checked against hooks/ and scripts/, every mechanically checkable claim run or grepped rather than read, and the marker now states that scope instead of asserting bare "verified".
  • tests/unit/doc-drift-scan.sh 22 → 32 assertions; tests/windows-compat.sh 19 → 24; scripts/validate-anchor.sh 157 → 158.

0.17.0 - 2026-07-29

Added

  • The Default-FAIL contract now runs both ways. anti-hallucination-gates bound only positive claims ("done", "fixed", "passing"); a negative one — "clang-tidy isn't installed here", "there's no such function" — was bound by nothing, and the skill's own description carried no negative trigger, so it was not even loaded at the moment it was needed. Adds a two-class contract (capability / search) with an executable probe for each, the mirror of the "should" detector, and an explicit exclusion for judgement-shaped negatives.
  • Negative capability conclusions carry an observation date. 0.16.0 made "not found" state its search scope; it still did not state when. The mandated form is now searched <scope>, not found (as of <YYYY-MM-DD>) at all seven sites (init-verification additionally re-checks inherited negative conclusions written in that form at session start — only negative ones, because those fail silently while positive ones fail loudly at the next invocation).
  • tests/unit/mandated-phrasing.sh — the wording rule spans seven files and was held together by instruction alone through 0.16.0, whose own review found it already drifted, and whose release then drifted again between the command and the docs page restating it. Now mechanical.
  • doc-drift-scan.sh announces what it did on stderr: skipped — <reason> for every early-exit path, scanned N symbol(s) x M doc(s) on completion. stdout remains a pure candidate list.

Changed

  • doc-drift-scan.sh scans a maintained language whitelist (C/C++, Python, JS/TS, Go, Rust, Ruby, Java, Kotlin, C#, shell) instead of C/C++ only.
  • doc-drift-scan.sh is now O(files) per chunk end to end, not O(symbols x files) — but the search-phase rewrite alone (one alternation grep per document per chunk instead of one grep per (symbol, document) pair) was not what delivered that. Measured on --base v0.15.0: the search-phase change alone regressed to over 10 minutes, killed by timeout (the pre-branch nested loop completed there in 8m25.794s), because attribution still spawned two greps per CANDIDATE LINE, and candidates outnumber symbol-file pairs. The fix that actually shipped (ac440ff) replaced attribution with one awk pass per document. Measured on the identical --base v0.16.0 range (44 symbols x 53 docs, 4363 candidates): pre-branch nested loop 5m44.427s, shipped version 3.917s — byte-identical output apart from a trailing-colon truncation the old IFS=: read silently caused. Chunking above 400 symbols and truncation above 2000 are both announced.
  • The contract is restated in both directions everywhere it appears: CLAUDE.md design invariant #8, the using-harness-anchor meta-skill's Hard Rule 1 (injected at every SessionStart), the README skill table and Default-FAIL section, and the scaffolded AGENTS.md template's Definition of Done.

Fixed

  • doc-drift-scan.sh could not see a single one of 0.16.0's own 21 changed files — it was pathspec-limited to C/C++ in a bash-and-markdown repository — and reported that by returning exactly what a clean scan returns. "Did not scan" and "scanned, found nothing" no longer share a channel.
  • cpp-tool-discovery.sh searched versioned tool variants against a hard-coded ladder ending at 22, giving it a roughly twelve-month fuse: an installed clang-tidy-23 would have reported NOT_FOUND, recreating the exact bug the script was written to fix. Versioned variants are now glob-enumerated.
  • tests/README.md's "Quick test" block enumerated 7 of the 18 unit tests that actually exist — the list had gone stale across several releases, so anyone following the documented steps ran under 40% of the suite while believing they had run it. Replaced with a glob, matching what CI already does for the reason its own comment gives: enumeration rots.
  • tests/windows-compat.sh blamed a MSYS2 grep crash on backslashes. The actual trigger is the -i+-F flag pair alone: it aborts during matcher construction, so there is no safe input, and grep writes nothing to stderr — inside $(...) or an if condition a SIGABRT is indistinguishable from a clean "no match".

0.16.0 - 2026-07-28

Added

  • scripts/cpp-tool-discovery.sh — resolves a C/C++ tool through PATH and the platform's known install locations (VS-bundled LLVM/Ninja on Windows, keg-only Homebrew llvm on macOS, versioned /usr/lib/llvm-* on Linux). An empty command -v is no longer treated as proof a tool is absent.
  • scripts/doc-drift-scan.sh — reverse-associates symbols touched by a change to *.md lines that mention them, surfacing documentation claims that should have changed and didn't. Attributes body-only changes to the enclosing symbol.
  • Adversarial trigger prompt for cpp-static-analysis covering the "tool looks missing, let's skip the check" failure shape.

Changed

  • cpp-static-analysis / cpp-formatting / /cpp-init now resolve tool availability via the discovery script, and are required to report absence as "searched PATH + , not found" rather than "not installed on this machine".
  • drift-analyst doc-drift now covers stale claims (symbol still exists, its contract changed) in addition to dangling references, and pulls unchanged *.md into scope when they mention a touched symbol. Its symbol-keyed blind spot is stated in the agent header.
  • cpp-build-systems escalates to cpp-build-doctor after a second failed attempt at the same build failure, not only on "anything cryptic".
  • /anchor closes by recommending /cpp-init when the project is C/C++ and .clang-format/.clang-tidy are absent.
  • /cpp-init records resolved tools portably: a tool already on PATH keeps its bare name, and one found off PATH is written as a PATH-first lookup with the resolved path as fallback. scripts/lint.sh and AGENTS.md are git-tracked, so a bare absolute path would have broken the next machine and CI.

Fixed

  • tests/skill-triggering/run-test.sh passed -p --output-format stream-json without --verbose, which newer claude CLI versions reject at argument-parse time — before any prompt is read. Every triggering case failed identically regardless of content, which made tests/README.md's pre-tag ritual unexecutable and left the "an adversarial prompt must PROVE the skill triggers" authoring rule unsatisfiable.
  • scripts/doc-drift-scan.sh reported a silent clean in the two most common /gc contexts: on main (where merge-base HEAD main is HEAD, and the HEAD~1 fallback was guarded on the base being empty rather than useless), and on uncommitted work (a BASE..HEAD range excludes the working tree, but /gc runs on uncommitted batches by design). Both fixed and covered by regression tests that exercise default base resolution — the path the original suite never touched.

0.15.0 - 2026-07-19

Added

  • Session Pulse (PostToolUse fast lane, all tools). Sliding-window self-supervision: duplicate-call nudge (3× identical input), error-streak nudge (3× same-tool failures), periodic feature checkpoint quoting the new optional out_of_scope ledger field — one nudge max per call, 10-call cooldown, pure bash on the hot path (no JSON-engine spawn outside the 1-in-25 checkpoint).
  • Two-stage context watermark. The v0.12.0 flush reminder (T1) migrates into the fast lane — observing every tool, not just Edit/Write — and gains a T2 stage advising /session-end + a fresh session over automatic compaction.
  • PreCompact forensics. New warn-only hook records .harness-anchor/last-compact.meta (trigger, transcript size, branch, dirty count, handoff age) and notifies the user when the handoff is stale; the SessionStart compact notice consumes the marker into concrete recovery anchors.
  • Evidence integrity. /verify gains a ### Integrity (tests-touched) report section; /session-end's precheck scans state files for credential patterns before the commit offer (labels only — matched values never echoed) and reports golden-rules hygiene with a consent-gated consolidation flow ([user]-tagged rules are never touched).

Changed

  • PostToolUse registration drops its Edit|Write matcher (fires on all tools); the Edit/Write slow lane is behaviorally unchanged, and the JSON-engine pre-warm moved into it so the fast path stays spawn-free.

0.14.0 - 2026-07-18

Added

  • Cross-platform content modularization. Platform-specific agent-facing content now has two dedicated channels: skills/<skill>/platform/<os>.md sidecars for operational depth (loaded on demand behind an inline same-skill pointer; decision-shaping facts — availability, verdict rules — stay inline), and <!-- os-<name>-start/end --> regions in the injected meta-skill, mechanically dropped by SessionStart unless <name> matches the runtime HA_OS (fail-slim: unknown names never fatten the injection). First sidecar: cpp-sanitizers/platform/windows.md (substitute-tool preference table). Adding a platform = adding content; no mechanism change.
  • SessionStart banner Platform: line (HA_OS taxonomy: windows (Git-Bash) | darwin | linux; unknown pre-set values pass through verbatim).
  • ha_platform_init respects a pre-set HA_OS (tests inject platform states; users may override classification) — the Windows PATH shield stays keyed on the real uname.
  • validate-anchor: joint flat sequencing across both conditional-region families, os-name taxonomy whitelist, inert-marker detection outside the meta-skill, and [12/12] platform-sidecar ↔ SKILL.md pointer integrity (bidirectional).

Changed

  • cpp-sanitizers Windows substitute detail moved to platform/windows.md (SKILL.md keeps the availability matrix and the never-CLEAN verdict rule inline); /sanitize pointer updated accordingly. CLAUDE.md invariant #2 wording generalized to conditional regions; the skill-authoring rules gain the platform decision-weight split. The meta-skill is pinned eol=lf (it is awk-consumed by the injection filter).

0.13.0 - 2026-07-16

Added

  • Windows support (Git-Bash baseline). scripts/lib/portable.sh — shared platform layer sourced by all four hooks and runtime scripts: JSON engine chain (python3 → python → py -3 → node → narrow pure-bash) with run-validated detection (immune to the Windows-Store python stub), C:\ path normalization at hook entry, fixed-point project-root walk (the old != "/" loop spun the 5s watchdog on drive-letter paths), Windows PATH shield (System32's incompatible find/sort/timeout lose to /usr/bin), portable mtime. Plugin-controlled formats (plugin.json version, cpp-detect output) parse even with ZERO engines — fixes the vunknown banner and C++-projects-typed-generic on Windows.
  • .gitattributes eol rules: every bash-consumed file checks out LF on all platforms (run-hook.cmd deliberately stays text=auto for the cmd/bash polyglot).
  • hooks/run-hook.cmd: %ProgramFiles%-based + user-scope Git discovery; WSL's System32\bash.exe excluded (hooks must see Windows paths).
  • C/C++ Windows counterpart mapping: sanitizer-build.sh.tpl turns detect_leaks off on MINGW*/MSYS*/CYGWIN* (LSan unsupported — same abort class as the v0.8.0 macOS incident); /sanitize reports TSan-on-Windows as INFRA-FAIL with substitutes; cpp-sanitizers gains Windows platform notes with a substitute-tool table (WSL2/Linux-CI TSan, Intel Inspector, Dr. Memory, CRT debug heap, UMDH, /RTC1); cpp-static-analysis gains Windows compile_commands/driver-mode notes.
  • tests/windows-compat.sh — static Windows invariants on every platform; new contract tests: session-start-engine-degradation.sh, post-tool-use-windows-paths.sh (Windows path bugs are string bugs — Linux CI catches them); windows-latest CI arm (curated core subset).

Changed

  • Hooks and status-report.sh/session-end-precheck.sh widen command -v python3 gates to the shared engine chain; the four duplicated escape_for_json copies collapse into ha_json_escape; hooks/stop staleness checks use mtime arithmetic instead of find -mmin (BSD/GNU/System32-neutral). Dev-surface scripts (validate-anchor / validate-manifests / measure-context) discover any python via ha_python; the test suite SKIPs honestly (visible, never silent) where an assert's engine is missing. CLAUDE.md gains design invariant #10 (Windows surface).

0.12.0 - 2026-07-14

Added

  • Write-at-realization contract for durable memory: capturing-golden-rules now mandates capture in the turn the signal appears, with a legitimate rough-stub form (origin = pasted evidence at hand; Check defaults manual review → /gc's [MANUAL] tier); template note synced.
  • hooks/post-tool-use Check 1d — context-fill flush reminder: transcript-size threshold (6 MiB const), warn-once per session via a .harness-anchor/flush-warned-<session_id> marker; X-vs-Y blind spots documented in the check header.
  • hooks/session-start — compact caution line: fired with stdin source=="compact", the banner warns that memory-from-recall is unreliable and points at on-disk evidence; stdin read is -t 0-guarded (non-blocking), regular startups zero-increment.
  • Contract tests: post-tool-use-flush-sentinel.sh, session-start-compact-caution.sh.

Changed

  • hooks/stop — the stale-progress nudge now also reminds to flush chat-only durable memory; observation-point header documents the mtime-proxy blind spot (cannot see chat content).
  • hooks/post-tool-use — stdin capture bounded (1s read -t; callers holding stdin open no longer hang the hook) and watchdog hardened to the SIGKILL/stdio-detached idiom (session-start's v0.10.0 lesson): the TERM version added ~5s wall to every command-substitution consumer and could not actually kill a runaway main.
  • Same-turn flush cross-links: self-correction-loop (capture hop after a recurrent fix), context-budget-discipline (flush-before-compress ritual + rebuild-from-disk after compaction), feature-state-keeper (mid-session milestone progress.md prepends), /session-end step 7 reframed as the safety net rather than the capture moment.

0.11.0 - 2026-07-14

Added

  • scripts/golden-rules-check.sh — mechanical Check runner for golden rules: parses ### GR-<n> blocks, executes the first backtick-quoted command per Check line (5s SIGKILL watchdog each, per-check isolation), three-state verdicts — CLEAN / FINDINGS(n) / CHECK-ERROR — so "found nothing" is never conflated with "didn't look"; --count feeds /status. Check convention documented in the golden-rules template + capturing-golden-rules skill: output = candidate violations, empty = clean, "manual review" in the line wins over backticks.
  • scripts/status-report.sh — the whole 7-section /status snapshot in one deterministic run (python3→node JSON engine chain; if both are missing only the JSON-derived lines degrade, the rest still reports; reuses toc-freshness.sh and golden-rules-check --count).
  • scripts/scaffold.sh — template placement for /anchor and /cpp-init (--cpp): placeholder substitution, chmod, skip-by-default for feature_list.json/golden-rules.md, conflicts (need decision) reporting, --render for diffs, --overwrite <allowlist> as the only write path over non-empty files — the default path physically has no overwrite branch.
  • scripts/session-end-precheck.sh — one-call fact block for /session-end: active feature + counts, init.sh under a 60s watchdog, state-archive dry-run + ledger-validate relays, two-column (state/source) tree scan, TOC structural-change hint (A/D/R + untracked, capped at 20).
  • validate-anchor [10/10] — the four mechanism scripts must be executable and parse, and every {CLAUDE_PLUGIN_ROOT}/scripts/* reference in commands/ + agents/ must resolve (thin wrappers made script paths a single point of failure); plus [9b] — template existence is now cross-checked against scaffold.sh's map (the old [9] check went silently empty once the command mds stopped naming template paths).
  • Unit suites for all four scripts (tier precedence, three-state verdicts incl. timeout, rerun byte-inertness by checksum, --overwrite allowlist, --render fidelity, cpp dotfile drops, refusal exit codes 3/4, engine-degradation via PATH shims, duplicate-id relay).

Changed

  • /status, /anchor, /cpp-init, /session-end are now thin wrappers over the scripts above — the script is the single source of truth for the mechanical half; the markdown keeps judgment and interaction (AskUserQuestion conflict round-trips, Default-FAIL flips, consent-gated archival, flywheel). Only read-only /status retains a manual degraded path. Saves the template/gathering round-trips through context (~6-18 tool calls → 1-2 per command) and pins the outputs byte-level.
  • drift-analyst runs the golden-rules mechanical tier via golden-rules-check.sh and keeps judgment: adjudicating FINDINGS lines (expected vs violation), reviewing MANUAL rules, surfacing CHECK-ERROR as a broken Check rather than a pass.

0.10.0 - 2026-07-05

Added

  • cpp-gated, slimmed SessionStart injection. The meta-skill body is now injected as a pure filter of using-harness-anchor/SKILL.md: YAML frontmatter stripped, and <!-- cpp-only-start --> / <!-- cpp-only-end --> regions (the four cpp-* sibling skills, /cpp-init, /sanitize) dropped in non-C/C++ projects — catching invariant #5 up at the injection layer. The file itself is untouched for the Skill-tool path. New contract test pins both modes plus the skip-leak guard; validate-anchor checks the regions are flat — sequenced, non-nested, closed (a nested pair would leak past the filter's single skip boolean with start/end counts still equal) — and that every cpp-only line is exactly one of the two markers.
  • measure-context.sh second pass on a bare generic fixture, so the generic fixed-cost baseline (the common case) is measured alongside the C/C++ e2e fixture.

Fixed

  • SessionStart watchdog: kills are now SIGKILL and the watchdog's stdio is detached. Measured on macOS bash 3.2: a subshell blocked on sleep 5 defers SIGTERM until the sleep completes, so (a) the parent's wait $watchdog_pid burned the full 5-second window on every session start even though main() finished in ~0.4s — any consumer waiting on the hook saw ~5s of wall per invocation — and (b) a genuinely runaway main() was never actually killed on that bash (the deferred TERM aborts the subshell before its kill line runs), leaving invariant #7's enforcement partly fictional. Exposed on CI by the rescaled 600-dir deep-repo fixture: slow runners pushed main() past 5s and the still-armed watchdog hard-killed it ("no output emitted"). Deep-repo hook wall: 5047ms → 502ms; the timeout contract (genuine overrun → silent, exit 0) re-verified at 5054ms.
  • SessionStart JSON escaping is O(n) via python3 (pure-bash fallback retained). The ${var//…} escaper is quadratic in matches×length: a ~12KB, ~700-line payload (a deep-repo directory map at the raised cap) burned the remaining watchdog window in this one step on pessimal macOS CI runners (~0.15s on a fast machine — which masked it). json.dumps is linear and also escapes control characters the bash path misses; environments without python3 keep the old escaper. Deep-repo hook wall: 502ms → 165ms.

Changed

  • Injection budget raised: ≤ 2000 → ≤ 3000 tokens (8000 → 12000 chars) — invariant #2. The old cap was 94% consumed and the squeeze fell entirely on the project-specific <project-toc> block (the banner never truncates). With the slimmed body a generic project's fixed cost drops to ~4.6KB (≈1160 tokens measured) and the TOC budget grows ~5×, so repos up to ~150 files get the full ## Files view instead of the degraded directory map. All cap reference points (hook, measure script, three test files, CLAUDE.md, README, both context-budget references) moved in lockstep; the deep-repo contract fixture rescaled 200→600 dirs so the degradation path stays exercised.
  • Meta-skill body compressed ~5.9KB → ~5.0KB — packaging only: rules, trigger keywords, read order, and command timing preserved item-for-item (contract-suite verified; a live three-scenario spot-check — scope-jump, TOC-before-Glob, no-done-without-evidence — gates the merge).

0.9.1 - 2026-07-05

Fixed

  • index-builder.mjs: --target/--output require a real value. Both flags feed the PROJECT-TOC.md write path; the lax argv[++i] parse let a missing or empty value (e.g. an unset shell variable in --target "$DIR") silently fall back and rewrite the current directory's index, and a following flag was eaten as the value (probe-confirmed: a junk ./--output/.harness-anchor/ dir and a file literally named --target). A missing/empty/flag-like value is now a one-line usage error — exit 1, distinct from the runtime-fatal exit 2, raised before any target-derived path (including the error-log dir) is computed — the same argument contract state-archive.mjs adopted in 0.9.0. New unit test pins all six refusals plus valid-value consumption.

0.9.0 - 2026-07-04

Added

  • State-file entropy governance. The SessionStart injection was already hard-capped, but the state files themselves grew without bound on long-lived projects — and the startup ritual reads them every session (a year-scale project puts feature_list.json at ~100KB / 20k+ tokens per session start). Now:
    • scripts/state-archive.mjs — deterministic, idempotent checkpoint archival: moves progress.md sections beyond the newest 20 to progress-archive.md, and pass features beyond the 10 most recently completed — evidence intact — to feature_archive.json (same schema shape). Archive-first write order + verbatim- duplicate convergence make it crash-safe; malformed JSON — or a ledger with duplicate feature ids — aborts with no writes (it never "repairs" the ledger, and never operates on a corrupt one). History is moved, never deleted; archives are git-tracked and grep-only.
    • /session-end budget step — after the ledger update, a --dry-run backlog check offers archival (explicit confirmation; the archives ride the same state-file commit); flags .harness-anchor/ > ~5MB as deletable runtime evidence (informational).
    • SessionStart state-budget sentinel (warn-only) — one banner line when a budgeted file exceeds its cap (progress 64KB · feature_list 32KB · golden-rules 8KB · AGENTS 8KB · handoff 4KB), pointing at /session-end. Observation point + residual blind spots documented per hook rule 5 (byte size is a proxy: a quiet sentinel is not proof of context health). New contract test.
    • feature-list-validate.mjs is archive-aware — feature_archive.json shares the id namespace: --check treats archived ids as taken (suggestion clears both files), whole-file mode reports hot∩archive collisions, and a corrupt archive is a hard error rather than a silently disabled guard.

Changed

  • Read discipline for long-lived state (zero structural change): AGENTS.md.tpl startup rules now say read the head of feature_list.json (actionable-first keeps live entries on top) and Grep, don't full-read a large PROJECT-TOC.md; project-indexing adds the ~400-line hard read rule; feature-state-keeper documents the hot windows and grep-only archives; context-budget-discipline carries the budget table.
  • Templates document their budgets (progress hot window; handoff ≤ 300 words / ~4KB; golden-rules ~30 rules / 8KB with prune-not-archive). Existing projects adopt the updated template wording by re-running /anchor (Overwrite/Skip/Diff prompt); the archival step, sentinel, and archive-aware validation are plugin-side and need no migration.
  • /status merges archived pass counts (pass: N (+M archived)) and adds a state-budget line to Harness health; /anchor points over-budget legacy projects at /session-end instead of archiving during scaffolding.
  • scripts/validate-anchor.sh [2/9] now also node --checks every scripts/*.mjs (glob, not an enumerated list).
  • templates/context-budget.md.tpl: Tier-1 table re-measured — its estimates predated several releases (harness-anchor's injection was listed at ~700 tokens vs ~1900 actual), and the row structure misrepresented the banner / TOC-head / handoff-head lines as blocks separate from that injection. Rows now mirror the real blocks (state banner + adaptive TOC view + meta-skill body) under the stated 8000-char cap, with a re-measure pointer (${CLAUDE_PLUGIN_ROOT}/scripts/measure-context.sh); the Tier-2 note carries the measured largest skill. context-budget-discipline's Tier-1 row now states cap vs measured (treat as full), and its sibling reference context-budget-template.md — whose same-era table even summed superpowers into the "≤2000 Tier-1 total" (that cap is harness-anchor's own) and described the pre-adaptive TOC truncation — is rebuilt on the same corrected structure, plus a watch-point for the new State budget: sentinel line.

0.8.0 - 2026-07-03

Fixed

  • templates/cpp/sanitizer-build.sh.tpl: per-OS detect_leaks. LeakSanitizer is unavailable on macOS/Apple toolchains; the previous unconditional detect_leaks=1 made every ASan run abort at startup there ("detect_leaks is not supported on this platform") — the old comment even called it a no-op. Now Darwin → 0, else → 1.
  • cpp-sanitizers skill: corrected the LSan platform claim. macOS is not "supported but off by default" — forcing it aborts ASan; standalone -fsanitize=leak is likewise Apple-unavailable. Use leaks(1)/Instruments on macOS, or run LSan on Linux CI.
  • PostToolUse clang-tidy signal fidelity. On macOS the hook now injects the SDK sysroot via xcrun; when a TU still fails to parse (any clang-diagnostic-error) it suppresses the unreliable diagnostics and emits one honest notice instead of garbage warnings (previously the false positives from the half-parsed TU were injected as if real — field use had to adopt a "don't trust the hook" project rule, the opposite of a guardrail). New contract test pins both behaviours. Warn-only contract unchanged.

Added

  • templates/cpp/lint.sh.tpl — sysroot-correct clang-tidy wrapper dropped by /cpp-init as scripts/lint.sh: the stable lint entry point for agents, docs, and done_criteria (field use kept having to reinvent exactly this script on macOS). Locates compile_commands.json across root / .build/ / build/ / builddir/ (the PostToolUse hook's search order) with a clear generate-hint when absent, and soft-falls-back to no sysroot args when xcrun cannot report an SDK path.
  • /sanitize INFRA-FAIL verdict — a sanitizer-infrastructure abort (e.g. an unsupported ASAN_OPTIONS flag) is now reported as INFRA-FAIL: not a code finding, and never CLEAN. Previously the report shape had no honest slot for "the tooling itself failed before exercising anything".

Changed

  • cpp-static-analysis skill documents the macOS '<header>' file not found failure mode (Homebrew clang-tidy without the SDK sysroot) and the rule that diagnostics from a failed parse are garbage — do not act on them.
  • Friction-point skill wiring: hook clang-tidy warnings now carry a self-correction-loop / cpp-static-analysis pointer tail; /sanitize and the cpp template scripts point at cpp-sanitizers / docs-lookup on infra failures — the knowledge existed in the skills but nothing delivered it at the moment of friction.
  • scripts/validate-anchor.sh [9/9] now also cross-checks templates referenced from commands/cpp-init.md (previously only /anchor's references were guarded).
  • CI hardening: the hook-contract and script-unit steps now glob their test directories instead of enumerating files (the hardcoded list had already silently omitted this release's new fidelity test); ShellCheck now also lints the shipped shell templates (templates/**/*.sh.tpl); new tests/unit/lint-template.sh pins the lint wrapper's DB-search / explicit-args / no-DB / sysroot behaviours; the e2e cpp fixture now models /cpp-init's scripts/lint.sh + scripts/sanitizer-build.sh outputs and CI asserts them.

0.7.2 - 2026-07-01

Fixed

  • verification-runner wrote evidence logs without creating .harness-anchor/ first. It captured build/test/lint output to .harness-anchor/verify-<step>-<ts>.log without ensuring the dir exists; .harness-anchor/ is gitignored and no hook creates it, so when /verify was the first gate run in a fresh clone or worktree the shell redirect failed with "No such file or directory" and evidence capture silently broke. Added the mkdir -p .harness-anchor guard its sibling evidence-writers (coverage-analyst, drift-analyst) already use, and enforced it: validate-anchor.sh [5/9] now asserts a fresh-dir contract — any agent that writes to .harness-anchor/ must mkdir -p it first (read-only agents stay exempt).

Changed

  • Extended superpowers complementarity — closed 3 more seams the post-v0.3.3 surface opened. Additive only — no skill description changed, so triggering is unaffected (mirrors the v0.3.3 audit).
    • skills/feature-state-keeper: the Altitude sync-contract now names parallel / subagent dispatch (superpowers:dispatching-parallel-agents / subagent-driven-development) as a second shared-state writer — dispatched workers don't each write the state trio or run the subagent-backed gates (single-level); the coordinating parent reconciles feature_list.json once after integration.
    • skills/init-verification: init.sh is documented as the superpowers:using-git-worktrees baseline (Step 2 Project Setup / Step 3 Verify Baseline), and a fresh worktree's absent, gitignored .harness-anchor/ is expected (recreated on demand), not un-anchored.
    • skills/self-correction-loop: a fresh-context sensor's findings (/verify · /test-plan · /gc) are triaged with superpowers:receiving-code-review rigor — verify each, push back with reasoning, don't blind-apply — most importantly inside /verify --fix.
    • skills/using-harness-anchor: Hard Rule #5 now spells out that dispatched workers must not run the subagent-backed gates (the parent does), plus a one-line interop pointer to the three skills above.

0.7.1 - 2026-06-22

Changed

  • Evidence contract now covers deliverable state, not just the working tree. anti-hallucination-gates gains a "Deliverable committed & reproducible" criterion (review the full git status, confirm the committed HEAD builds — not only the dirty working tree) plus an anti-pattern against dismissing uncommitted changes as "old / unrelated" without a git diff; verification-runner now reports working-tree clean/dirty and flags that green local evidence does not prove a buildable HEAD; /session-end surfaces uncommitted source (the whole tree, not just state files) with a HEAD-buildability caveat before offering its state-file commit (it still never auto-commits source). Closes the failure where a feature marked pass could leave its own source uncommitted and the committed HEAD unbuildable.
  • Coverage obligations extended to behavioural-contract regressions and liveness. test-coverage-design's risk checklist + coverage-analyst now derive two classes the sensors previously missed: (1) behavioural-contract substitution — swapping a container / algorithm / impl behind a stable API can silently regress a guaranteed observable property (ordering / stability / idempotency / documented no-op) or a public signature, caught by a characterization / metamorphic test; (2) liveness / termination under adversarial structure — cycles, self-edges, or already-satisfied preconditions in graphs / dependencies must terminate or diagnose, not hang. The shared-mutable row also now names check-then-act (TOCTOU) on a composite predicate.
  • Drift detection now flags dead stores. drift-analyst gains a computed-but-never-used heuristic (a buffer / accumulator / timestamp built then never read) — wasted work that looks like real logic and previously slipped the scan.
  • Synced plugin.json / marketplace.json descriptions with the GitHub repo About — they now reflect coverage gates (v0.5.0), entropy governance (v0.6.0), and the warn-only / zero-dependency identity, not just the pre-v0.5.0 blurb (metadata-only change carried over from the prior unreleased state).
  • Bumped plugin.json / marketplace.json to 0.7.1 (synced); refreshed docs/commands.md (/session-end, /verify) and its doc-align marker.

0.7.0 - 2026-06-16

Added

  • Action-side scope-creep detector (hooks/post-tool-use): a warn-only check that fires when a new code module is created via Write while a feature is in-progress, surfacing agent-initiated scope expansion the prompt-side UserPromptSubmit guardrail cannot observe (the "observation-point mismatch" of #6). Edits, overwrites, test/doc files, git-ignored files, and non-git projects stay silent by construction. Resolves #6.
  • Guardrail authoring rule (CLAUDE.md): new hooks must state the failure's manifestation surface (Y) and the observed signal (X), assert X ⊇ Y, or document the residual blind spot in the hook header.

Changed

  • feature-state-keeper / using-harness-anchor: scope discipline is now enforced on both the prompt side and the action side (the new post-tool-use check gives "record new scope as planned first" action-layer teeth).

0.6.1 - 2026-06-16

Fixed

  • golden-rules count off-by-one in the SessionStart banner (Golden rules: N) and /status harness-health. The count pattern ^### GR- also matched the commented-out ### GR-1 example inside the freshly-scaffolded golden-rules.md, so a project with zero real rules reported 1. Fixed by counting only numbered real rules (^### GR-[0-9]) and changing the template's example heading to the non-counted placeholder ### GR-N. New tests/hook-contracts/session-start-banner.sh assertions guard it: real rules counted, commented example excluded, the as-shipped empty template → 0, absent file → no banner line (the missing contract assertion that let this ship — mirrors the v0.3.1 cpp-init-hint pattern).

Changed

  • scripts/validate-anchor.sh [5/9] now also asserts every agents/*.md ends with the single-level constraint line ("Do not invoke other subagents from this one.") — mechanizing invariant #3, which was previously maintained by hand.

0.6.0 - 2026-06-16

Added

  • Entropy governance / feedback flywheel (Concept ⑥ — the one canonical harness-engineering concept harness-anchor had no mechanism for, built in the lightweight, report-only form a solo / small-team harness needs: "grow from failure, 3 rules not 30", and Böckeler's "continuous drift sensors"). Warn-only, zero-dependency, never auto-refactors:

    • templates/golden-rules.md.tpl (NEW) — a project state file for accumulated taste / anti-pattern rules, each GR-<n> tied to a concrete past failure with a Check that escalates manual → grep → lint by frequency × impact. Ships empty (seed on real recurrence); scaffolded by /anchor and Skip-by-default on re-anchor so accumulated rules are never wiped. Kept separate from AGENTS.md so the map stays a map.
    • skills/capturing-golden-rules/ (NEW, generic) — the ratchet: turn a recurring failure into a durable rule ("blame the process, not the agent"). Routes the four Feedback-Flywheel signal types to their homes (failure → golden-rules; context → AGENTS.md; instruction / workflow → skill / AGENTS.md) so the file doesn't become a dump.
    • agents/drift-analyst.md (NEW, read-only, fresh-context) — scans changed code against golden-rules.md + generic drift heuristics (duplicated helpers, inconsistent error handling, copy-paste, oversized files, TODO pileup, doc-drift), grades findings must / should / nice, persists .harness-anchor/drift-<ts>.md. Explicitly non-overlapping with verification-runner (build/test/lint) and coverage-analyst (coverage / run-scope).
    • commands/gc.md (NEW) — dispatches drift-analyst; mirrors /verify & /test-plan's fresh-context, read-only shape; report-only (offers scoped fixes / rule capture, never bulk-refactors). Not git gc.
    • Flywheel wiring: /session-end gains a one-question reflection ("did anything recur worth capturing?") — anchored to the existing checkpoint, not a new ceremony (Garg's Feedback Flywheel). /status gains a lightweight ### Harness health section (rule count, last drift scan, staleness) — a few signals, deliberately not a dashboard. The SessionStart banner surfaces the golden-rules count.

  • AGENTS.md template upgraded to the validated content formula (templates/AGENTS.md.tpl): commands-first (they were an empty block at the bottom), a ## Git workflow section (the missing sixth domain, kept value-neutral — no hardcoded GitFlow / commit convention), a code-example stub in Conventions (examples > prose), and a pointer to golden-rules.md. Still a map (≤ ~80 lines). Existing projects adopt it by re-running /anchor (Overwrite/Skip/Diff prompt).

Changed

  • skills/self-correction-loop: added the edit budget (doom-loop) alongside the existing re-run budget — ~3–5 edits to one file chasing the same goal without the signal clearing means STOP, re-read the spec (not your own diff), switch approach or escalate; and don't "fix" a failing test by rewriting it to match the code (LoopDetection-middleware discipline).

0.5.0 - 2026-06-15

Added

  • Test-coverage design capability (targets a confident-wrong failure mode: a fixed-width-accumulator overflow that only fires on large inputs and ships because its only triggering binary is built but never add_test-registered, so the sanitizer never runs it and a false "clean" is claimed). Post-implementation, and complements superpowers' (deliberately code-blind) TDD — TDD owns the pre-impl, spec-driven test-first pass; this owns the code-aware post-impl pass.

    • agents/coverage-analyst.md (NEW, read-only, fresh-context) — derives test obligations from code + spec (gray-box), diffs them against the suite and the verified run scope (flags binaries the runner never executes), and recommends a minimal oracle-independent-first test set (metamorphic / differential / property: correctness comes from a relation, not the model's judgement — so it survives the correlated blind spot when code and tests are both LLM-generated); persists evidence to .harness-anchor/coverage-<ts>.md.
    • skills/test-coverage-design/ (NEW, generic) + coverage-reference.md — the post-impl discipline + a design-technique catalog (EP / BVA / pairwise / metamorphic …) and a living risk-construct checklist (C/C++ arm cross-links cpp-sanitizers/ub-failure-patterns.md).
    • commands/test-plan.md (NEW) — dispatches coverage-analyst; mirrors /verify's fresh-context, read-only shape.
    • Warn-only cross-links: anti-hallucination-gates gains a "Coverage obligations" criterion (a green suite that skips the risk path is a false pass); cpp-sanitizers documents the run-scope caveat; verification-runner flags orphan binaries; using-harness-anchor lists both.

  • Feature id uniqueness enforcement (closes a gap the JSON Schema can't: draft-07 has no per-field uniqueness, so feature_list.schema.json was structurally blind to id collisions — and id is the lookup key for status/evidence updates, so a duplicate silently corrupts the source of truth). Warn-only, defense-in-depth:

    • scripts/feature-list-validate.mjs (NEW, node, read-only) — default mode flags duplicate ids (exit 3); --check <id> is a pre-write candidate test that suggests the first free -N suffix. Single responsibility = uniqueness (id format stays the schema's job).
    • Pre-write (primary): feature-state-keeper gains a "Feature id uniqueness" section — check the candidate id against existing ones (the file was just read to edit it; or --check) and qualify a colliding id before writing.
    • At-write safety net: hooks/post-tool-use warns (warn-only, fail-silent) the moment a feature_list.json write introduces a duplicate id, naming it + suggesting a free one.
    • Pre-commit backstop: /session-end runs the validator before offering a commit.
    • New tests/unit/feature-list-validate.sh + a duplicate-id hook-contract case + an e2e-fixture uniqueness assertion; schema carries an inert $comment documenting why uniqueness is enforced imperatively, not in the schema.

Fixed

  • scripts/toc-freshness.sh: exclude PROJECT-TOC.md from its own staleness counts (pathspec :(exclude) on both the committed-diff and working-tree checks). Previously the TOC counted itself — committing a regenerated TOC always advanced HEAD past its own anchor, so "fresh" was unreachable in the canonical tracked-TOC workflow (the very one /index-project recommends: git add PROJECT-TOC.md && git commit) and the "stale … run /index-project" nudge fired permanently, prescribing a cure that couldn't work. Now the regenerate→commit loop converges to fresh, and stale means real drift. Status words/format unchanged (consumers — SessionStart banner, /status, index-curator — unaffected). Rewrote the unit-test stale fixture that had pinned the old behavior, added a canonical-workflow regression case, and corrected three docs/comments that had rationalized the false-stale as "an inherent limitation" (tests/e2e-cpp-fixture/bootstrap.sh, docs/troubleshooting.md §3, skills/project-indexing algorithm note) plus the inaccurate "always stale" wording in /anchor docs (a no-git project actually reports not-git).

0.4.0 - 2026-06-09

Added

  • Test & CI hardening (no plugin-behavior change; closes audit gaps G1–G6):
    • tests/unit/ — black-box unit tests for the two most complex scripts: index-builder.mjs summary extraction (every comment marker incl. the -->/--!> js/bad-tag-filter fix, 80-char truncation, ## Decisions preservation, binary/lockfile skipping) and toc-freshness.sh (all status branches incl. the missing-anchor git cat-file guard).
    • tests/unit/feature-list-schema.sh — exercises feature_list.schema.json itself (anti-drift) and proves the Default-FAIL rule (status=pass ⇒ evidence) rejects a violating doc. Python stdlib only.
    • tests/cpp-detection/non-cpp-fixture/ — negative fixture asserting cpp-detect.sh → is_cpp_project:false (guards invariant #5).
    • tests/skill-triggering/ — 7 new adversarial prompts (11/11 sibling skills now covered) + check-coverage.sh, a no-LLM CI guard enforcing the "one prompt per skill" authoring rule.
    • CI (validate.yml): runs the new unit / coverage / negative-fixture steps on the ubuntu+macOS matrix, plus a shellcheck job. All hooks/scripts/tests were brought shellcheck-clean at warning — fixed SC2064 (trap expanded at set-time, not signal-time) and SC2164 (unguarded cd) across the test scripts, and for-over-find → while read in validate-anchor.sh — so the gate runs at --severity=warning (notes/style surfaced informationally).
  • C/C++ analysis-tool coverage made honest + selective expansion (skills/cpp-static-analysis + skills/cpp-sanitizers; no description changed, so triggering is unaffected):
    • cpp-sanitizers: documented LeakSanitizer as ASan's already-running leak component — on by default on Linux, ASAN_OPTIONS=detect_leaks=1 on macOS (supported, off by default), plus the lighter standalone -fsanitize=leak mode and LSAN_OPTIONS suppression.
    • cpp-static-analysis: clarified that clang-tidy's clang-analyzer-* is the Clang Static Analyzer (per-TU), so scan-build is only for whole-build / HTML / cross-TU passes — preventing redundant "new analyzer" runs.
    • cpp-static-analysis: added GCC -fanalyzer (GCC 10+) as a different-engine second opinion, explicitly flagged C-only (the manual: "only suitable for use on C code") so it is never misapplied to C++. Mirrored in tool-comparison.md.
    • Deliberately rejected to stay lean: RealtimeSanitizer (niche), an MSan build recipe (high-friction), valgrind helgrind/DRD, and CodeChecker / Infer / PVS-Studio (heavyweight / commercial).
  • docs-lookup: first-party ecosystem docs MCPs as a preferred Step-1 source. Generalized Step 1 from "Context7" to "structured docs MCP — Context7, or a first-party MCP for its ecosystem," naming Microsoft Learn (microsoft_docs_search → microsoft_docs_fetch) as the prime example for .NET/Azure/Windows/MSVC topics — relevant to the plugin's own C/C++-on-Windows surface (MSVC errors, Win32/SDK headers). Optional + prefer-when-present: same graceful fall-through to Context7 → WebSearch, so the zero-dependency ethos holds (Context7 is already treated this way). Stated as a general principle, not a vendor list, to avoid chain bloat. description unchanged; the README companion-plugins table is synced to match.
  • State files scale to heavyweight, long-running projects (new deterministic scripts/*.mjs tools + adaptive SessionStart injection; non-breaking — hook JSON shape + warn-only contract unchanged):
    • scripts/feature-list-sort.mjs — reorders feature_list.json actionable-first (in-progress → blocked → planned → pass) at /session-end; deterministic, idempotent, and lossless (preserves unknown top-level keys, evidence, 2-space formatting). Lets the agent read the head of a long ledger and stop.
    • scripts/progress-prepend.mjs — inserts a new progress.md entry after the header without loading the whole file into context (newest-first; safe on a malformed/headerless file).
    • scripts/index-builder.mjs now emits a ## Directory map (one line per directory — direct-file + subdir counts) above ## Files.
    • hooks/session-start adaptively injects the TOC: full ## Files on a small repo, else the directory map, else (huge repo) the shallowest directories first — always within the ≤8000-char Tier-1 budget (now tested at scale). A 3-tier read path (top dirs → map → files) without a multi-file tree.
    • Wired through /session-end, /index-project, project-indexing, feature-state-keeper, the AGENTS.md / PROJECT-TOC.md templates, and docs/architecture.md. Existing projects self-heal (no migration). New unit tests + a budget-at-scale hook-contract case.

Fixed

  • hooks/post-tool-use: quote the prefix in rel_path="${file_path#"$project_root"/}" (SC2295) so the relative-path computation does literal prefix removal instead of glob-pattern matching — robust to project paths containing [ ] * ?. Behavior-identical for normal paths (post-tool-use contract test 6/0); the shell suite is now shellcheck-clean at all severities.
  • templates/cpp/sanitizer-build.sh.tpl: corrected a factually-wrong comment claiming "macOS doesn't support leak detection" — macOS does support ASan leak detection (off by default; detect_leaks=1 enables it, which the script already sets). Comment-only; build behavior unchanged.

0.3.3 - 2026-06-04

Changed

  • Made superpowers complementarity explicit. An audit of harness-anchor against superpowers found a few overlapping process/record seams; this closes the one with real drift risk and cross-references the rest. Additive only — no skill description changed, so triggering is unaffected.
    • skills/feature-state-keeper: added an "altitude" reconciliation contract so harness-anchor's feature_list.json (durable project feature ledger / source of truth) and superpowers' plan-docs + TodoWrite (ephemeral step-level execution) don't drift or double-book — on disagreement, feature_list.json (with evidence) wins.
    • skills/anti-hallucination-gates: cross-references superpowers:verification-before-completion — same Iron Law; one verification run satisfies both gates (don't re-verify).
    • skills/self-correction-loop: names the switch threshold to superpowers:systematic-debugging (1–2 minimal fixes → switch; the cap-at-3 budgets are intentionally aligned).
    • commands/session-end.md: clarifies it is a session-pause checkpoint; branch/PR/merge belongs to superpowers:finishing-a-development-branch.

0.3.2 - 2026-06-02

Fixed

  • Stop hook emitted JSON invalid for the Stop event. hooks/stop wrapped its wrap-up reminder in hookSpecificOutput / additionalContext — valid for PostToolUse / UserPromptSubmit / SessionStart, but the Stop event has no additionalContext channel, so Claude Code rejected the output (Hook JSON output validation failed — (root): Invalid input) and the reminder never surfaced. The reminder now uses a top-level systemMessage — non-blocking and schema-valid (still never decision:"block" / stopReason, per invariant #1). The tests/hook-contracts/stop-wrap-up.sh contract test gained a real Stop-schema assertion (rejects hookSpecificOutput and any blocking field); it previously checked only JSON syntax + substrings, which is why the bad shape shipped.

0.3.1 - 2026-06-01

Added

  • SessionStart /cpp-init recommendation: the state banner now recommends /cpp-init when a project is detected as C/C++ and is already anchored (feature_list.json present) but lacks .clang-format/.clang-tidy — the same "missing artifact → run the command" pattern already used for /anchor and /index-project. Manual commands never auto-trigger, so the hook surfaces this one at the moment it is needed. Warn-only; fires only on that precise state (no nagging when the config exists or the project is un-anchored). A new C/C++ fixture in tests/hook-contracts/session-start-banner.sh asserts the hint fires when clang config is absent and is suppressed once it exists.

Changed

  • using-harness-anchor command catalog → when-to-recommend guidance: the meta-skill that SessionStart injects every session now frames each command by when to recommend it (parallel to its ## Sibling Skills section) instead of a flat what-it-does list, and back-fills the missing /sanitize entry (shipped in v0.3.0 but never cataloged).

0.3.0 - 2026-05-30

Added

  • /sanitize command (Tier 2): commands/sanitize.md builds a C/C++ project under ASan+UBSan (TSan separately) and runs its tests, reporting in verification-runner-style fixed sections with a mandatory .harness-anchor/sanitize-*.log evidence path. Reuses templates/cpp/sanitizer-build.sh.tpl; refuses on non-C/C++ projects via cpp-detect.sh. skills/cpp-sanitizers references it.
  • PostToolUse /sanitize nudge: the hook appends a one-line "consider /sanitize" suggestion to an existing warning on a C/C++ source edit — bounded, never standalone, never runs sanitizers inline (invariants #4/#7).
  • /verify --fix auto-fix loop (Tier 2): opt-in only; on a NOT READY verdict it applies the verification-runner's ### Recommendation-scoped fixes, then re-verifies with a fresh verification-runner, max 2 cycles, stopping on pass or budget exhaustion. Every change is surfaced; pass is never asserted without a fresh PASS (invariant #8). commands/verify.md gains Edit, Write for this path.

0.2.1 - 2026-05-29

Added

  • R4 — command allowed-tools shape validation: scripts/check-allowed-tools.sh is the single source of truth for the rule that every commands/*.md must declare an allowed-tools: line shaped as a comma-separated list of tool-name tokens (identifier + optional (scope), e.g. Bash(git diff:*); MCP names allowed). It validates shape, not membership — no hard-coded tool registry to go stale. Wired into validate-anchor.sh [6/9] for real commands; negative fixtures under tests/command-fixtures/ exercised by a new CI step.
  • P1 — hook wall-clock benchmark: tests/bench/hook-timing.sh bootstraps the e2e fixture and times all four hooks against a time budget (warn ≥2s, fail ≥5s), guarding invariant #7 the way measure-context.sh guards the byte budget (invariant #2). Runs on both CI arms.

0.2.0 - 2026-05-29

Added

  • Layer A — Robustness:

    • scripts/index-builder.mjs writes .harness-anchor/last-error.log on failure; clears stale log on success (c9e1a37)
    • scripts/validate-anchor.sh extended with agent frontmatter (name + description) and command frontmatter (description only) checks (c9e1a37)
    • tests/hook-contracts/ — 5 contract tests for all 4 hooks (session-start banner + timeout, post-tool-use warn, stop wrap-up, user-prompt-submit scope-jump) (c9e1a37)
    • R1 total watchdog for hooks/session-start and hooks/post-tool-use: 5s tempfile-guarded cap, emits nothing on timeout (avoids partial-JSON pitfall) (c9e1a37)

  • Layer B — Anti-drift:

    • hooks/session-start reads version from plugin.json at runtime (no hardcoded v0.1.0) (5296b74)
    • README check count de-hardcoded ("count printed on run") (5296b74)
    • Removed dead scripts/escape-json.sh (zero references; all hooks inline escape_for_json) (5296b74)
    • Extracted scripts/toc-freshness.sh from session-start (shared by /status) (5296b74)

  • Layer C — Manifest validation + CI:

    • scripts/validate-manifests.sh: Python-stdlib-only validator for plugin.json and marketplace.json (name, description, version semver, version sync) (6162175)
    • scripts/schemas/{plugin,marketplace}.schema.json — reference docs (not executed) (6162175)
    • tests/manifest-fixtures/ — negative JSON fixtures (bad-keywords, bad-name, bad-version, version-mismatch) (6162175)
    • CI matrix: ubuntu-latest + macos-latest (6162175)
    • tests/posix-compat.sh — scans for GNU-only flags (6162175)

  • Layer D — Usability:

    • /status command — read-only project overview (6 Markdown sections) (56fa3c9)
    • Commit-hygiene section folded into feature-state-keeper/SKILL.md (56fa3c9)

  • Layer E — E2E fixture + observability:

    • Expanded tests/e2e-cpp-fixture/: 3 features (planned/in-progress/pass), C++ sources, gtest, full state files, .clang-format/.clang-tidy (d64eab5)
    • tests/e2e-cpp-fixture/bootstrap.sh — creates isolated git repo from fixture (d64eab5)
    • scripts/measure-context.sh — measures SessionStart output against the 8000-char cap (exit 1 if exceeded, warn at 90%) (d64eab5)

  • Layer F — Docs:

    • CHANGELOG.md (this file) — Keep-a-Changelog format
    • docs/troubleshooting.md — 5+ failure modes with diagnosis + fix

Fixed

  • TOC freshness algorithm: validate anchor commit exists (git cat-file) before diffing; sanitize arithmetic with tr -cd 0-9; extend grep pattern to include underscores (c9e1a37)
  • hooks/post-tool-use: stdin read before backgrounding (bash redirects bg stdin to /dev/null) (c9e1a37)
  • scripts/index-builder.mjs: exit(2) inside try block → throw Error() so catch block actually runs and writes error log (c9e1a37)

Changed

  • tests/self-correction/ → tests/hook-contracts/ (rename ripple: README, tests/README.md, CLAUDE.md, validate.yml) (c9e1a37)
  • validate-anchor.sh sections renumbered 7→9 (c9e1a37)
  • validate-anchor.sh excludes tests/manifest-fixtures/ from JSON parse (6162175)

0.1.0 - 2026-05-28

Added

  • Initial harness-anchor plugin: 12 skills, 3 subagents, 5 commands, 4 warn-only hooks (677b76c)
  • SessionStart hook with state banner, TOC freshness, project type detection (677b76c)
  • PostToolUse hook with regression-warn + clang-tidy (677b76c)
  • Stop hook with wrap-up reminders (677b76c)
  • UserPromptSubmit hook with scope-jump detection (677b76c)
  • C/C++ engineering suite: build systems, static analysis, formatting, sanitizers (677b76c)
  • docs-lookup skill with Context7 → WebSearch fallback chain (ba4e132)

Fixed

  • 5 plan-vs-implementation gaps closed across 2 rounds (f099f6c, 81c332f)

Docs

  • README rewrite, agent compression, docs-lookup test case (bdb0f99)