v0.9.2

Added

• OpenSpec is a first-class layout. The compass documented the spec-kit
  quartet (spec, plan, tasks, checklist) and said nothing about OpenSpec,
  so a host whose contract lives in docs/openspec/ got guidance for a layout it
  does not use — and spec-validation-map reported every one of its capabilities
  as missing a plan, which is the wrong question rather than a finding.
  docs/workflows/openspec.md is the reference: the
  twelve workflows, the status gate, the vocabulary mapped onto spec-kit's, and
  the nine rules. spec-driven-development.md now resolves the layout before the
  first write instead of assuming one.
• openspec-guard enforces the chain that openspec validate cannot see.
  Validate asks whether the files present are well-formed; the guard asks whether
  the files that should be present are. Eight checks — root, config, chain,
  deltas, stale, workflows, ready, orphans — plus a
  .openspec-guard.json ratchet for changes that predate a gate, where an entry
  that now passes is itself a failure. Three of them come from real faults found
  in a host on the first run: a change that had held a proposal and nothing else
  for four days while the gate reported 23 passed, an empty second root left by
  a move that made the CLI answer No changes exist on six active changes, and
  six of the twelve agent workflows never installed — verify, the pre-archive
  gate, among them.
• openspec-lifecycle skill and an openspec profile, detected from the
  resolved root, so a host with an OpenSpec root gets the procedure and the gate
  and a host without one gets neither.
• templates/specs/openspec-config.example.yaml. A rule in the root's
  config.yaml reaches the agent through openspec instructions at the moment it
  writes that artifact, which lands where the same sentence in AGENTS.md does
  not. The template shows the rules.<artifact> and operations.<op>.guidance
  shape — and the guard checks it, because an unquoted ": " in a list item makes
  the whole document fail to parse and the CLI then reports an empty project.
• Nine knowledge instincts, harvested from two host projects:
  openspec-artifact-chain,
  cap-as-ratchet-with-baseline,
  generated-file-whole-run,
  build-success-needs-an-artifact,
  evidence-needs-a-control,
  untrusted-text-inside-a-fence,
  no-undo-write-needs-a-receipt,
  test-proves-presence-not-truth,
  worktree-base-before-work,
  parallel-locale-artifacts,
  and changelog-entry-in-the-same-task.

Changed

• spec-validation-map detects its layout and says which one it read. On an
  OpenSpec root it reports one row per active change and a capability
  traceability table; on a spec-kit layout it reports the quartet as before.
  --spec-kit forces the old reading.
• One resolver for the OpenSpec root, in scripts/lib/openspec.mjs, read by the
  guard, the map and profile detection. The declaration in
  agent-compass.commands.json (paths.openspec) wins over convention; the
  commands template now carries a paths block for it.

v0.9.1

Fixed

• A stale install is no longer silent. An install is a snapshot of a pin, so
  it now records one: .agent/external-skills.json (or
  ~/.agent-compass/external-skills.json for --global) holds the source, the
  commit it came from, the skills, and the targets. external-skills --check
  compares that against the current pins offline, and --upgrade re-installs
  every record at the current pin. The check runs on three paths without being
  asked — the session-start hook, recommend, and install --doctor — because
  for the operational corpus a moved pin means the safety gate and the
  argv-secret narrowings were regenerated, which is not cosmetic.
• A tracked package version is now part of the verified contract. anydoc
  had moved 0.1.9 → 0.2.3 while its commit pin looked current, leaving the
  convert-documents-to-markdown skill telling agents to run a version four
  releases old. A reference source can now declare package and version:
  --verify fails when any local file (including a tool_version frontmatter
  field) pins a different version, and --update rewrites every occurrence when
  the pin moves.
• Refreshed the four stale source pins — taste-skill, caveman,
  i-have-adhd, and anydoc. All nine sources are now current, and the anydoc
  package pin advanced to 0.2.3 in the skill along with it.

v0.9.0

Fixed

• release is idempotent. A second run for the same version inserted a second,
  empty ## [<version>] heading, and the changelog extractor then matched the
  empty one — so every forge release got a Release v<version> placeholder body
  instead of the notes. It now leaves an existing section alone.

Changed

• No external skill is stored in this repository any more. All nine sources
  moved to "strategy": "reference", which removed 166 vendored skill folders
  and skills/upstream-lock.json. The registry keeps the pin, the licence, the
  full upstream inventory, and a new recommended field holding Agent Compass's
  own curation — 146 of 163 operational skills, 3 of 20 caveman skills, 10 of 13
  design skills. Nothing about the update lifecycle changed: one cached
  --check-updates covers all nine, and --verify still runs offline.
• The operational safety adapter moved from vendoring time to install time.
  The corpus was vendored so Agent Compass could rewrite it — a safety gate on
  every skill and eight passages with the secret taken out of argv. Tracking it
  would have lost that, so scripts/lib/upstream-skills.mjs now runs inside the
  installer instead. No uncorrected copy exists anywhere, an override whose
  upstream target was reworded still fails rather than being dropped, and the
  guarantee is now asserted against install output for all 146 skills.
• One install path for local and tracked skills. skills-sync --only routes
  each requested name to the local copy or the tracked source automatically, and
  stack profiles name external skills in an external field that selectAssets
  merges into skills. recommend, adopt, and setup-wizard therefore behave
  exactly as they did when the skills were vendored — a caller passes one list and
  never needs to know which kind a name is. global-setup --style does the same
  user-wide.
• skills answers for a tracked skill. agent-compass skills <name> reports
  the source, pin, licence, pack membership, and install command for a skill that
  is not on disk; --grep matches a hyphenated slug from a spaced query
  ("github actions" finds github-actions); --pack <id> reports the pack and
  how to install it.
• convert-documents-to-markdown is now compass-authored guidance for the pinned
  @firecrawl/anydoc CLI rather than a copy of the upstream skill. The anydoc
  source stays tracked so the version pin remains visible.
• Retired check-skill-quality and the lint:skill-quality gate. What they
  guarded is now covered by upstream-skills --verify plus the install-output
  tests.

Added

• external-skills — one installer for Claude Code, Codex, and Copilot.
  Fetches a tracked source at its pin and writes .claude/skills/ and
  .agents/skills/, plus a .github/instructions/ file because Copilot has no
  skills directory. --global installs user-wide, --recommended takes the
  compass curation, --skill <a,b> or --all take a selection, --dry shows the
  plan. Executable payloads are refused unless --allow-scripts is passed, and
  each source's licence notice — including PolyForm's verbatim Required Notice
  line — is written beside the install.

• Compass-authored routers replacing the deleted copies:
  operational-skills,
  design-taste-skills, and
  working-style-skills — each carrying
  the curation reasoning that used to be implicit in which folders existed.

• Pointer documents: operational-skills.md
  (the eight narrowings in full) and
  style-and-design-skills.md.

• compass-external-source mission skill. The guided path for adding,
  listing, curating, refreshing, or removing a tracked source: the licence gate
  runs first (no licence means refuse; a use restriction is escalated, not
  absorbed), then the registry entry, the computed inventory, the pointer
  document, the narrowings, the fit wiring, and the tests. Routed from
  MISSIONS.md.

• Native mobile coverage, tracked instead of copied. Two published skill
  corpora are now registered in skills/upstream-sources.json:
  android/skills (21 skills, Apache-2.0,
  Google LLC) and
  dpearson2699/swift-ios-skills
  (86 skills, PolyForm Perimeter 1.0.0). Neither is vendored. The Apple corpus
  carries a noncompete term, and Agent Compass redistributes skills to host
  projects, so copying it would engage that term; the Android corpus ships its
  own installer and keeps its mirrored documentation fresher than a mirror would.
  See ADR 002.

• A third source strategy, reference. A reference source is pinned,
  update-checked, and documented, but owns no local file. It records an install
  command, an inventoryRoot, an inventoryDoc, pointers, and the
  upstreamSkills inventory at the pinned commit. upstream-skills --verify
  fails offline when a pointer disappears, stops naming its repository, or when
  the generated inventory block drifts. upstream-skills --update <id> re-reads
  the inventory from the new tree with git ls-tree and git show, moves the
  pin, rewrites the inventory block, and prints added and removed upstream
  skills — copying nothing and executing nothing. Both new sources join the
  existing cached --check-updates path with no separate command.

• native-mobile-skills skill. Routes a native Android or Apple-platform
  task to the vendor skill that covers it, installs it with the vendor's own
  installer, and keeps the compass gates in force: validation on the real
  toolchain, an emulator or simulator screenshot for a screen change, and no
  third-party skill relaxing §4. Self-contained, so it still routes after
  skills-sync copies it into a host without the compass tree.

• android-compose and swift-ios stack presets, with detection.
  detectStacks now recognises an Android Gradle plugin, AndroidManifest.xml,
  an Xcode project or workspace, Package.swift, and Podfile — and holds back
  when Expo, React Native, or Flutter is present, because those toolkits generate
  the android/ and ios/ trees rather than owning them. A Gradle build only
  counts as Android when a build file applies an Android plugin, so a Kotlin
  service does not match.

• platform-skill-before-memory instinct. Why recall fails on a
  yearly-release platform, and which of the two failure modes reaches review.

v0.8.1

Fixed

• Every externally sourced skill now carries its MIT notice. The LICENSE
  file reached only the first skill of each multi-skill source, and none of the
  146 skills of the operational corpus. A host that installs one skill folder
  therefore received MIT content without the notice the terms require. 152 skill
  folders gained the file, and the safety gate accepts LICENSE beside
  SKILL.md while it still refuses every executable payload.
• The session-start update hook finds a vendored compass again. It searched
  docs/agent-compass/scripts/ alone, so a host that vendors anywhere else got
  no notice. It now searches the common locations and honours
  AGENT_COMPASS_HOME. A host that installs skill folders alone, and carries no
  compass tree, still leaves quietly.
• The external source cache is ignored wherever it is written. The
  .gitignore entry held a slash, which anchors a pattern to the directory of
  the .gitignore. It therefore never matched
  docs/agent-compass/.agent/.upstream-source-check.json. The pattern is now
  unanchored, and migration 0.8.1 widens it for a host that already took the
  anchored form.

v0.8.0

Added

• A pinned external-source registry, cached update notices, reviewed refresh
  commands, and provider startup hooks now keep externally sourced skills
  traceable and easy to update without automatic merges.
• convert-documents-to-markdown adds a pinned, local document-ingestion
  fallback for office files and text PDFs.
• codebase-to-specs drafts explicitly inferred specifications, decision
  records, and architecture sketches from an existing codebase for human review.

Changed

• Four stale external skill sources were refreshed to their reviewed remote
  heads while local safety metadata and guidance were preserved.
• diagram-to-adr and diagram-to-likec4 supersede the Excalidraw-only skills.
  They add draw.io, embedded draw.io SVG, Mermaid, and appropriate BPMN support.

v0.7.9

Fixed

• An evidence bundle no longer asks to be committed. .agent/evidence/ and
  .agent/changes/ join the generated artifacts the gitignore template already
  lists. Both hold copied screenshots, so a host that adopted that template would
  otherwise have carried megabytes of regenerable images into its history on the
  first run. Publish a bundle as a CI artifact instead.

v0.7.8

Added

• agent-compass evidence builds the artifact a completion claim owes. It
  collects every JUnit report and screenshot into a self-contained bundle at
  .agent/evidence/ — index.html for a human, summary.md for the pull
  request — and the status is binary: complete only when nothing failed and the
  promised screenshots exist. Configure it with the evidence key of
  agent-compass.commands.json. Without that key it discovers the files and
  says so in the report, because a discovered count is a weaker claim than a
  declared one.
• A before/after report for a spec'd change. evidence --change <slug> --phase start records the proof and a SHA-256 snapshot of the workspace;
  --phase finish re-runs, diffs the snapshot, lists the changed and tested
  files, reads the acceptance criteria out of the spec, and renders the before
  and after screenshots side by side under .agent/changes/<slug>/. It is a
  gate: it exits non-zero unless the after-proof is complete and something
  actually changed, so a green suite over an empty diff cannot pass.
• docs/guidelines/definition-of-done.md. The formalism behind the
  Completion Gate: the AUTO / PR / REVIEW / N/A proof types, the rule
  that no level compensates another, the blocking conditions, the actions an
  agent may never take against a test, and an exception procedure that needs an
  owner and a deadline and then expires.
• docs/guidelines/accessibility.md, the first accessibility rules in the
  compass, plus the keyboard-path-before-done instinct. An automated audit
  reads the rendered tree; it cannot tell you that Tab skips your new button.
• specs/change-spec-template.md, whose ## Acceptance criteria and
  ## Expected proof scenarios headings are read literally by the change
  report, and the evidence-outlives-the-claim instinct.

Changed

• docs/guidelines/testing-tdd.md gained the proof half of testing: a fixed
  screen-profile matrix, the difference between regenerable execution evidence
  and committed visual baselines that only a human may update, the evidence
  commands, and the forbidden anti-patterns.

Fixed

• code-intel status reports an indexed repository as indexed. The project
  listing is read for a root path under five key names, and
  codebase-memory-mcp 0.10.5 returns the one that was missing, root_path.
  Every status read therefore ended on "this repository indexed: no", whatever
  the graph held. doctor never read that field and was correct throughout.

v0.7.7

Fixed

• The Windows install path works. agent-compass code-intel install now
  downloads and runs install.ps1 through pwsh or powershell instead of
  refusing with "POSIX-only". Upstream parses $args, not a param() block, so
  the flag is --skip-config on both platforms — the manual command previously
  printed -SkipConfig, which PowerShell would have ignored while the installer
  rewrote every agent's configuration.
• Windows executable discovery looks in %LOCALAPPDATA%\Programs\codebase-memory-mcp
  and its bin\ subdirectory, the real install.ps1 default, instead of the
  POSIX ~/.local/bin. PATH splits on ; on Windows and : elsewhere. Every
  platform difference is now a pure function of (platform, env) and the whole
  matrix is tested from any host.
• POSIX install falls back to wget when curl is absent, and a host with
  neither fails with the missing tool named plus the exact manual command. The
  downloaded installer is removed afterwards instead of being left in the
  temporary directory.
• agent-compass mcp-probe no longer reports the code-intelligence MCP example
  as a readiness gap on hosts that never opted in. The example ships to every
  host as a catalogue entry; it is probed once the host selects the layer.

Added

• docs/tooling/codebase-memory.md gains a platform matrix, Windows
  troubleshooting rows, and a worked false-positive example: a real
  detect_changes run on this repository returned an unrelated shell function
  among eight impacted symbols, with nothing marking the row as wrong.

v0.7.6

Added

• Structural code intelligence via codebase-memory-mcp (optional). A new
  agent-compass code-intel command (status, install, configure, setup,
  doctor) installs the CBM binary with --skip-config and lets Agent Compass
  keep ownership of provider and MCP configuration. scripts/lib/codebase-memory.mjs
  is the single place that finds, reads, installs, and configures the binary.
  Setup enables auto_index and auto_watch, creates
  .mcp/codebase-memory.example.json from the new template, ignores the
  generated .codebase-memory/ graph, and records
  "codeIntelligence": "codebase-memory" in the answers file.

• AGENTS.md §1b Codebase navigation: query the code graph before broad
  grep, read the exact files it returns, and never turn one query into an
  exhaustive or negative claim without checking index coverage and corroborating
  in source. Adds the four-source table (code, docs/ADRs, graph, projectmem),
  the query hierarchy, token-discipline rules, and the ADR-ownership rule that
  keeps docs/decisions/ canonical against CBM's own manage_adr feature.

• docs/tooling/codebase-memory.md — purpose, when not to use it, install, CLI,
  MCP, auto-index/watch, cache and ignore policy, the opt-in shared graph
  artifact, the projectmem split, ADR ownership, depgraph comparison,
  troubleshooting, update/uninstall, and privacy.

• Doctor, recommend, mcp-probe, and the setup wizard now know the layer.
  Repo-level facts (ignore rule, MCP example) are required once a host selects
  it; machine-level facts stay advisory so a plain clone and CI never fail.
  Five eval scenarios cover navigation, impact, fallback, durable history, and
  the negative-claim rule.

• A release request now ends at a published forge release, not at a pushed
  tag. AGENTS.md §10 names the three artifacts a release owes: the version in
  the project files, the tag on every remote, and the published release on every
  remote forge. scripts/release.mjs gains --release, which implies --push,
  uses the changelog section of the version as the release body, publishes once
  per repository behind the remotes, and treats an existing release as success.

v0.7.5

Added

• A release request means publish. AGENTS.md §10 now states that "push a
  release", "release a patch", "release a new version", or "cut a release" is
  the explicit ask that the safety rule requires, and that it authorizes the
  whole chain: bump, validate, commit, tag, then push the commit and the tag to
  every remote that git remote lists. A local tag alone is not a release.
  scripts/release.mjs gains --push to run that push in one command, and
  docs/workflows/releasing.md replaces the origin-only push with a loop over
  every remote.
• Three behaviour rules in AGENTS.md §2 that no rule covered before: answer a
  question instead of implementing it; act on cheap reversible in-scope work and
  fix a defect you find rather than reporting it back; run independent work in
  parallel with one writer per file. AGENTS.md §4 adds the all-items rule —
  deliver every item, or name the exact blocker in one sentence.

Fixed

• scripts/release.mjs now bumps the two README.md version markers with
  package.json and CHANGELOG.md. scripts/check-release.mjs requires those
  markers, so npm run lint:release used to fail on every release until a
  person edited the file by hand.