anchored 0.5.1

Full Changelog: v0.4.0...v0.5.1

anchored 0.3.0

Added

• /setup skill — a conversational anchored.yml configuration
  assistant. Auto-triggers whenever the user wants to create, change, or
  extend their anchored.yml (add a step, wire an agent/skill into a
  stage, tune a gate, set up commit/PR/TDD automation) — even without
  naming the file. It translates the user's stated requirements into
  correct, schema-valid config, writes name + instructions on every
  custom step, and validates before finishing. It clarifies genuine
  ambiguities with the user and advises on request (backed by
  plugin/references/power-user-setups.md), but never pushes a setup the
  user didn't ask for — no power-user onboarding funnel.
• Custom steps gain instructions (any step) + type (use steps), and
  use: dispatch is now real. A lifecycle step
  (plan/refine/build/wrap.steps) can now declare:
  • instructions: — free-prose documentation of what the step does and
    why, allowed on any step. The recommended base for every custom
    step is name + instructions. On a run: step it self-documents
    the shell (rationale + audit trail); on a use: step it is
    additionally threaded into the invoked worker, the step-level analogue
    of the reserved slots' instructions (build.implement.instructions,
    …).
  • type: agent | skill — on a use: step, picks the invocation
    mechanism. agent (the default when omitted; back-compatible with
    existing use: anchored/implement steps) spawns an isolated
    subagent via the Agent tool; skill invokes the worker via the
    Skill tool in the orchestrator's own session. Rejected on a run:
    step (no worker to type).
  • A single canonical dispatch contract
    (plugin/references/step-dispatch.md) replaces the previously
    unspecified "invoke the named tool … depending on how the user has
    wired it" hand-wave; all four lifecycle skills now point at it.
  • See plugin/EXTENDING.md → "Add a custom step to any stage".

Changed

• Decluttered the shipped default-config.yml — dropped the decorative
  # ─── stage ─── divider headers and the redundant schema line; the
  inline per-slot explanations stay. Cosmetic; no shape change.

anchored 0.2.0

Added

• Dynamic Workflow executor — optional per-phase fan-out for
  /impl-build. A phase can opt into executor: workflow (default,
  field-absent, stays implement — byte-identical to before) to run as
  a Claude Code Dynamic Workflow: its acceptance criteria fan out across
  parallel unit-workers (≤16 concurrent / 1000 total) instead of one
  sequential implement agent.
  • New phase field executor: implement | workflow (schema +
    RESERVED_FIELD_NAMES), a phase.executor.set op, the
    task__set_phase_executor MCP tool, and the
    anchored phase executor set <slug> <phase-slug> <executor> CLI.
  • New workflow agent: a fan-out unit-worker with the inverse
    write-contract — it writes its own evidence/failures to the
    task-file via the anchored CLI (invoked as
    npx -y -p @chaafoo/anchored-mcp anchored …, the same
    package-resolution that loads the MCP server — no global install or
    PATH entry needed).
  • /impl-build feature-detects the Workflow runtime and falls back
    to the sequential implement path when it is unavailable — never
    a hard error. The task-validate + code-validate gates run once
    over the merged phase result, and the failures-driven retry loop /
    retry-limit are unchanged.
  • See plugin/EXTENDING.md → "Running phases as a Dynamic Workflow"
    (incl. the CLI allowlist note for background workflows).
• Behavioral eval suite (evals/). An LLM-in-the-loop harness for
  the lifecycle skills (plan / refine / build / wrap), with per-stage
  task-file fixtures reusable as seeds.

Changed

• Build autonomy reworked from persisted levels to ephemeral
  walk-style + global stop-conditions. The persisted task.autonomy
  field (ask_all / ask_high_only / decide_all) is gone. In its
  place:
  • /impl-refine Stage 0 now picks an ephemeral walk-style
    (AI-all / high-together / all-together) that governs only that
    one Q&A walk — it is never written to the task-file. /impl-build
    runs the same ephemeral walk to clear any still-open questions
    before its run (open-question detection is programmatic).
  • /impl-build runs maximally autonomous over emergent
    build-time decisions: retry → decide → document via
    question_resolve(source='ai', reasoning). It halts ONLY when an
    emergent decision matches a rule in anchored.yml.build.stop
    (shipped default: a single rule, 'a decision deviates from the plan'), judged by the new stop-check evaluator + the implement
    worker's own self-report (double safety net). The old three-way
    autonomy branch in the failures loop is removed.
  • /impl-wrap now surfaces every source='ai' resolution with its
    reasoning, grouped by phase, as a decisions review.
• MCP tool surface: 38 typed tools. Net of this release: the
  task.autonomy op was removed (build-autonomy rework) and
  task__set_phase_executor was added (workflow executor). Old on-disk
  task-files carrying a top-level autonomy: key still load — the
  parser silently drops the legacy field.

Fixed

• Rule-file discovery no longer silently skips underscore-prefixed
  directories. The rules + rules-check agents relied on a
  recursive glob (.claude/rules/** / **/*.md) that can return zero
  files when rules live under _concern/ / _pattern/, making an agent
  report "no rules exist" and drop all rule-coverage. Both now fall back
  to Grep / explicit subdir globbing before concluding the folder is
  empty.
• /impl-build flips refined → build at entry. The forward-only
  state machine made the terminal build → wrap transition illegal when
  the task was left at refined; the flip now happens up front.

anchored 0.1.3

Fixed

• MCP server reported itself as v0.2.0 regardless of the published
  package version — a stale hardcoded string in src/mcp/server.ts
  left over from the V0.2 era. Server now reads its version from
  package.json at startup via createRequire, so the
  serverInfo.version advertised to MCP clients and the
  anchored-mcp v<X> ready startup line stay in sync with the
  published version automatically. No more manual sync on release.

anchored 0.1.2

First successful npm publish via CI. (0.1.1 tagged but never reached
npm registry due to npm CLI bundled with Node 20 not handling
trusted-publisher OIDC reliably.)

Fixed

• mcp/package.json had duplicate keys (keywords, license,
  repository) from a botched merge. The duplicate repository
  carried the wrong URL (anchored/anchored instead of
  chafoo/anchored), which npm normalized into the published
  metadata for 0.1.0. Removed the duplicates; canonical
  repository: chafoo/anchored is the single source of truth now.

Changed

• CI publish workflow (.github/workflows/publish.yml) switched
  from NODE_AUTH_TOKEN to npm trusted publishing via OIDC. No
  long-lived tokens; each release is authenticated by GitHub's
  short-lived id-token signed against the workflow + repository
  registered with npm. Removes the supply-chain attack surface
  from stolen registry tokens.
• CI now pins Node 22 (npm 11.x) and upgrades npm to latest before
  publish — npm 10.x's trusted-publisher implementation had
  intermittent registry-auth bugs.