@nklisch/pi-agile-workflow 0.15.3

package/.claude-plugin/plugin.json

@@ -0,0 +1,10 @@
+{
+  "name": "agile-workflow",
+  "description": "Markdown-based work-tracking substrate for AI-driven projects. Items as files in .work/, late-binding releases, gates that produce items, goal-backed autopilot queue runner. See docs/VISION.md.",
+  "version": "0.15.3",
+  "author": {
+    "name": "nklisch"
+  },
+  "repository": "https://github.com/nklisch/skills",
+  "license": "MIT"
+}

package/.codex-plugin/plugin.json

@@ -0,0 +1,28 @@
+{
+  "name": "agile-workflow",
+  "version": "0.15.3",
+  "description": "Markdown-based work-tracking substrate for AI-driven projects. Items as files in .work/, late-binding releases, gates that produce items, goal-backed autopilot queue runner.",
+  "author": {
+    "name": "nklisch"
+  },
+  "repository": "https://github.com/nklisch/skills",
+  "license": "MIT",
+  "skills": "./skills/",
+  "hooks": "./hooks/hooks.json",
+  "interface": {
+    "displayName": "Agile Workflow",
+    "shortDescription": "Substrate tracking with goal-backed autopilot",
+    "longDescription": "Skills for substrate-driven work tracking. Items as files in .work/, late-binding releases, gates that produce items (security, tests, cruft, docs, patterns), and a goal-backed autopilot queue runner.",
+    "developerName": "nklisch",
+    "category": "Productivity",
+    "capabilities": [
+      "Read",
+      "Write",
+      "Bash"
+    ],
+    "defaultPrompt": [
+      "Use agile-workflow to inspect the substrate queue and recommend the next ready item."
+    ],
+    "websiteURL": "https://github.com/nklisch/skills"
+  }
+}

package/CHANGELOG.md

@@ -0,0 +1,591 @@
+# Changelog
+
+## Unreleased
+
+### Dynamic subagent prompting
+
+- **No shipped custom agents** — removed agile-workflow's Claude/Pi agent
+  manifest entries, Pi `subagents` package metadata, Codex `aw-*` templates, and
+  the Pi extension's global agent sync. The plugin now ships skills and runtime
+  commands only.
+- **Structured generic delegation** — skills now route design, implementation,
+  review, scanner, and explorer work through the host's existing
+  generic/general-purpose subagent mechanism using the dynamic prompt postures in
+  `skills/principles/references/subagents.md`.
+
+## v0.14.0
+
+### Backlog grooming + staleness signal
+
+- **`groom` skill** — new user-invocable backlog-hygiene sweep
+  (`/agile-workflow:groom`). Classifies `.work/backlog/` items
+  DONE/SUPERSEDED/DUPLICATE/STALE/MERGEABLE/VALID via mechanical signals plus a
+  grounded semantic pass, and writes a triage report. Propose-not-prune:
+  dispositions are operator-confirmed and route through the terminal-tier
+  retention convention; it never auto-prunes and is not a release gate. Opt-in.
+- **`work-view --stale`** — lists backlog items whose last-touched date
+  (`updated` if present, else `created`) exceeds the new `backlog_staleness_days`
+  `.work/CONVENTIONS.md` key. Absent key ⇒ inert notice + exit 0, so a project
+  that doesn't opt in is unaffected. Backlog-tier only; std-only, local-time.
+- **`updated` on the backlog contract** — documented as an optional backlog
+  frontmatter field (`== created` at birth, written by `park`; bumped by the
+  PostToolUse hook on edit; treated as `created` when absent). `BACKLOG_REQUIRED`
+  unchanged — the field stays optional.
+
+## v0.12.1
+
+### Gate configurability
+
+- **Gate finding routing** — documented `gate_finding_routing` in
+  `.work/CONVENTIONS.md` so release gates normalize severity/priority/confidence
+  to one project-level item-placement policy.
+- **gate-refactor scan roots** — documented `gate_refactor_scan_library_roots`
+  so monorepos can share scan-rule libraries while preserving the default
+  `.agents/skills` then `.claude/skills` discovery order.
+
+## v0.11.3
+
+### Research-substrate linkage fields + agentic-research adoption
+
+- **work-view research linkage** — the Item model gains `research_refs` (Arrow 1: a work item
+  tracks/consumes research) and `research_origin` (Arrow 2: research that spawned the item), parsed
+  from frontmatter and queryable via `--research-refs` / `--research-origin`. Mirrors the
+  `gate_origin` precedent; both fields parse harmlessly when unset, so a plain `.work/` substrate is
+  unaffected.
+- **convert** — research-substrate fields are conditional on the `agentic-research` plugin. The
+  canonical AGENTS block ships the base field list and only advertises `research_refs` /
+  `research_origin` (plus the `.work/` ↔ `.research/` handoff pointer) when that plugin is installed;
+  convert never scaffolds `.research/` on its own.
+
+## v0.11.2
+
+### Clarify `--stage review` semantics in seeded docs
+
+- **convert / docs** — reworded the canonical AGENTS section template and architecture docs so
+  `work-view --stage review` reads as "items awaiting an agent review pass (`/agile-workflow:review`)"
+  rather than "items waiting on user." The review stage is the agent's review queue (advance or
+  bounce), not a hold for human sign-off.
+
+## v0.11.1
+
+### Terminal-tier archival: merged `archived_atop` late-binding into delete-refs
+
+- **`archived_atop` late-binding** — archived items are bodyless stubs stamping an immutable
+  `archived_atop` baseline + `git_ref`; `release-deploy` late-binds all unbound archived stubs into
+  one release summary.
+- **Archive-stub gate hydration** — gate skills include `.work/archive/` stubs in the release
+  bundle and hydrate historical bodies from `git_ref` when a body is needed.
+- **convert** — one merged terminal-retention convention; sync detects/offers it and offers to prune
+  existing retained terminal bodies to stubs (stamping `archived_atop` + `git_ref`).
+
+## v0.11.0
+
+### Delete-after-release terminal tiers
+
+- Archived items become bodyless ref stubs; a release collapses bound items into one
+  `releases/<version>/release-<version>.md` summary; full bodies live in git history. `convert` seeds
+  the `terminal-tier retention` convention (default `delete-refs`).
+
+_(0.10.x release notes: see git history.)_
+
+## v0.9.5
+
+### Three-channel distribution
+
+- **Pi package channel** - Added Pi package metadata for supported plugins, kept
+  package versions in lockstep with Claude/Codex metadata, and documented the
+  three-channel distribution model across root and plugin guides.
+- **Pi extension surface** - Added the agile-workflow Pi extension shell with
+  `/aw` queue snapshots, parent/blocking lookups, and handoffs to shared board,
+  autopilot, and scope skills.
+- **Delegation policy** - Updated agile-workflow skills so worker/scout/reviewer
+  routing names Pi-native subagents where available while preserving the shared
+  skill contracts.
+
+### Board and dependency view
+
+- **Expanded browsing** - Added expandable epic and tag groups to the board
+  sidebar without changing filter semantics.
+- **Dependency canvas polish** - Added zoom, hand-pan, layout controls, compact
+  web layout nodes, ephemeral dragging, clearer impact counts, and fixes for
+  resize edge drift, node click jitter, Hand-mode detail opening, clipping, and
+  the null-sentinel module import failure.
+
+### Tests and release gates
+
+- **Board JS behavior harness** - Added a no-build ES-module test harness plus
+  coverage for markdown safety, filter composition, dependency/table behavior,
+  kanban/detail behavior, and expanded browsing.
+- **Hook concurrency and PostCompact audit** - Added deterministic coverage for
+  hook state-file interleaving and verified Codex PostCompact output behavior.
+- **Release-gate follow-ups** - Added Pi package metadata tests, Pi extension
+  command tests, filtered dependency stub coverage, keyboard reachability
+  coverage for tag expansion, and removed dead board-test harness cleanup code.
+
+## v0.9.0
+
+### Cross-vendor `.agents/rules/` auto-load
+
+- **Generic rules loader** - A hook now injects `.agents/rules/*.md` into agent
+  context in both Claude Code and Codex, so dense project rules and the pattern
+  digest load automatically without per-tool wiring. Hook state-file writes are
+  concurrency-hardened (unique temp names, fail-open on error, `fcntl` locking
+  when available).
+- **`convert` extracts rules** - `convert` lifts the dense agent-instruction
+  block into `.agents/rules/agile-workflow.md`, guarded by a content-integrity
+  gate that verifies content exists at its canonical replacement before any
+  destructive op.
+- **`gate-patterns` writes the digest** - The patterns gate emits a generated,
+  hook-loaded `.agents/rules/patterns.md` digest (slug + one-liner index with a
+  drift-detecting `src-sha256`) alongside the canonical pattern skills.
+- **Skills ground on `.agents/rules/`** - The design/implement/review-family
+  skills read `.agents/rules/*.md` during grounding.
+
+### work-view freshness & anti-drift lifecycle
+
+- **`work-view --version`** - Both the Rust binary and the bash fallback report
+  the plugin version (`work-view <semver>`), kept in lockstep with `plugin.json`
+  by `bump-version.sh`, so a single string compare answers "is this installed
+  copy current?".
+- **Self-heal install lifecycle** - The SessionStart hook and `convert` detect a
+  missing or stale `.work/bin/work-view` and reinstall the version-stamped
+  entrypoint; the probe fails open (never blocks the session) on timeout or
+  error.
+
+### Documentation
+
+- Rolled the SPEC forward: the kanban surface is documented as parent swimlanes
+  with verbatim per-stage columns (not a fixed 5-column collapse), and the plugin
+  source layout now includes the `work-view/` Rust workspace.
+- Refreshed pattern-skill example `file:line` references after board modules
+  shifted positions.
+
+### Internal
+
+- New test coverage: `bump-version.sh` lockstep projection, `work-board.sh` shim
+  routing, `convert` content-integrity structural guard, self-heal version-probe
+  timeout fail-open, and a `$TMPDIR`-invariant substrate-root test.
+- Codified 4 new code patterns (board view-module contract, local DOM
+  text-element builder, fail-open subprocess probe, hand-written error `Display`).
+- Removed dead board-view scaffolding (`placeholderView`, `registeredViews`, and
+  orphaned CSS).
+
+## v0.8.9
+
+### Board UI polish
+
+- **Mock-aligned board filters and metadata** - Removed the interactive board's
+  Release filter section and release-binding metadata from cards, details, and
+  the table surface. Missing values now render with the mock's dash treatment
+  instead of a visible `(none)` sentinel, and null parent/release values are not
+  exposed as filter chips.
+
+## v0.8.8
+
+### Interactive Substrate Board
+
+- **Live `work-view board` surface** - Added the compiled localhost board host
+  with `/healthz`, `/api/substrate`, embedded assets, loopback Host header
+  hardening, busy-port scanning, and browser-open handling. `work-view serve`
+  aliases the same read-only board server.
+- **Board shell and views** - Added the no-build browser shell with shared
+  filters, auto-hide, safe markdown rendering, shared item cards, diagnostics,
+  theme/accent controls, and detail surface. Delivered kanban stage/swimlane,
+  dependency graph, and sortable/filterable table views over the same substrate
+  feed.
+- **Board skill** - Added the user-invocable `agile-workflow:board` skill for
+  launching `.work/bin/work-view board`, with clear failure handling when a
+  project still has a non-board-capable bash fallback.
+- **Legacy `/board` command removed** - Deleted the Claude-only `/board`
+  command surface. `scripts/work-board.sh` remains only as a compatibility shim
+  that delegates to `.work/bin/work-view board` when the installed binary
+  supports it.
+
+### Documentation
+
+- Rolled forward the agile-workflow SPEC and ARCHITECTURE docs, plus the root
+  substrate-access architecture note, from static generated board language to
+  the live `work-view board` model.
+
+## v0.8.6
+
+### Substrate CLI
+
+- **Compiled `work-view` CLI** - The `.work/` query tool is now a compiled Rust
+  binary (`work-view-core` library + `work-view` CLI) with full flag/output
+  parity over the bash script, on a shared query core (frontmatter parsing, the
+  item model, dependency-graph evaluation, composable filters) that the
+  forthcoming web board reuses. Ships per-platform with a pure-bash fallback.
+- **Stage-aware `--ready` / `--blocked`** - `--ready` no longer hides
+  design-ready (`drafting`) and review-ready items whose dependencies are
+  satisfied; it surfaces any active-tier item at `drafting`, `implementing`, or
+  `review` with all `depends_on` terminal (and `--blocked` the inverse). The
+  prompt-context hook no longer double-lists a `review` item under both Ready
+  and Review.
+- **Platform install path** - `install-work-view.sh` selects the matching
+  prebuilt binary (uname → target triple, smoke-test, atomic install), wired
+  into `/agile-workflow:convert`, with the bash fallback when no binary matches.
+  Added a CI cross-compile workflow and the `dist/` layout. Binaries are
+  CI-produced; until the manual refresh job populates `dist/`, installs use the
+  bash fallback (no regression).
+
+### Also in this release (previously unreleased)
+
+- **Review skill progressive disclosure** - Split the review skill into a
+  substrate-first core plus references for target resolution, review lenses,
+  deep-review mechanics, and substrate side effects. Supports standalone
+  branch/PR/commit reviews without `.work` mutations and keeps deep-review mode
+  adaptable across contracts, data, concurrency, release, and product/UX
+  dimensions.
+- **work-view single-pass parsing** - Rewrote `work-view.sh` frontmatter
+  parsing to a single `awk` pass over the whole tree instead of spawning an
+  `awk` (plus `tr`) per field per file. Measured on a 617-item repo: default
+  table view drops from ~7.9s to ~0.12s (66x); output is now deterministically
+  sorted by path, otherwise byte-for-byte identical to before.
+- **Dual Codex/Claude hook set** - Replaced the eager SessionStart queue dump
+  with prompt-gated context injection. Actionable workflow prompts get a compact
+  queue snapshot plus once-per-session principles capsules; idle chat and
+  explainer prompts stay silent. Capsule state lives in the host plugin data
+  directory when available, with non-worktree fallbacks.
+- **Substrate maintainer hook** - Expanded the PostToolUse hook from `updated:`
+  bumping to deterministic touched-item validation: required frontmatter, valid
+  kind/stage, filename/id match, parent/dependency existence, and dependency
+  cycles reachable from the touched item.
+
+### Documentation
+
+- Foundation docs (VISION, SPEC, ARCHITECTURE) rolled forward: work-view bash
+  script → compiled binary + fallback, and the stage-aware ready/blocked
+  semantic.
+
+### Internal
+
+- First project pattern catalog under `.agents/skills/patterns/` (5 patterns
+  extracted from the CLI work).
+- Expanded test coverage: binary-level proof of the `--ready` drafting fix,
+  exit-3 fatal-IO, the `--blocked` review case, the hook review-dedup, a
+  `convert` install-routing guard, and bash-parity empty-result tests.
+- Added an `actionlint` CI workflow for GitHub Actions files; removed dead code
+  (unused `From<io::Error>` impl, an unused parameter, a vestigial no-op); added
+  a root `.gitignore`.
+
+## v0.7.7 - Skill metadata hotfix
+
+- **Codex skill loading** - Shortened long agile-workflow skill frontmatter
+  descriptions so every `SKILL.md` stays under Codex's 1024-character
+  description limit. This fixes `perf-design` being skipped by Codex after the
+  v0.7.6 release.
+- **Goal-backed autopilot** - Reshaped `autopilot` as a model- and
+  user-invocable queue policy that can be selected from a harness goal
+  statement such as "Use agile-workflow autopilot to drain --all." Removed
+  current docs' dependence on `/loop`, watchdog ticks, and `--resume`; the
+  harness goal/continuation feature now owns long-running persistence.
+- **Caller-awareness alignment** - Updated delegated skills and docs so
+  "autopilot mode" means an active autopilot run or goal, not only a literal
+  slash invocation in the transcript.
+
+## v0.7.6 - Perf and refactor design alignment
+
+Tightens performance-design guidance for low-level systems and makes refactor
+conventions a cohesive extension of agile-workflow rather than a standalone
+planning path.
+
+- **Perf-design profiling depth** - The optimization hierarchy now calls out
+  data locality and microarchitecture work explicitly, including cache locality,
+  pointer chasing, allocation pressure, branch predictability, false sharing,
+  NUMA effects, and benchmark/counter evidence for CPU-bound systems.
+- **Profiling taxonomy** - Discovery mode now classifies workload baseline,
+  on-CPU, memory/GC, I/O/serialization, off-CPU/synchronization, and
+  cache-line/NUMA bottlenecks before emitting perf items.
+- **Refactor conventions integration** - `refactor-conventions-creator` now
+  writes short style summaries into AGENTS plus a canonical
+  `.agents/skills/refactor-conventions/` catalog. `refactor-design` keeps its
+  built-in defaults and treats the catalog only as an extension lens.
+- **Convert one-pass alignment** - `convert --update` now covers AGENTS/CLAUDE
+  compatibility, work-view, pattern-skill mirrors, and refactor-conventions
+  catalog placement in one pass. Cleanup is opt-in: convert asks whether
+  destructive cleanup is in scope and defaults to preserve-only.
+- **Tool metadata fixes** - Skills that ask strategic questions now declare
+  `AskUserQuestion` in their allowed tools.
+
+## v0.7.5 - Tighten [refactor] tag routing
+
+Stops mistagged "major rework" items from leaking into `refactor-design` when
+they actually change observable behavior.
+
+- **Sharper tag semantics** - The AGENTS.md section now ships a `### Tag
+  semantics` subsection with the black-box test ("would any observable
+  behavior change for a caller of the public surface?") plus contrasting
+  examples on each side. Definition is re-read by agents mid-work, not just
+  buried in CONVENTIONS.md. SPEC.md's CONVENTIONS template entries for
+  `[refactor]` and `[perf]` are tightened to match.
+- **Misroute self-heal in refactor-design** - Per-feature mode Phase 1 now
+  applies the black-box test to the brief. If the item describes new
+  capability, an API change, swapped semantics, or "rework rather than
+  restructuring," it strips `[refactor]`, logs the misroute, commits, and
+  returns without advancing - autopilot reroutes to `feature-design` on the
+  next pass. Mirrors `feature-design`'s existing misroute guard.
+- **Tag application gate in scope** - Phase 5 (single-idea) and Phase B4
+  (batch clustering) now require the black-box test before applying
+  `[refactor]` or `[perf]`. Stops the bad tag at the source.
+- **Convert sync handles existing projects** - `convert` and
+  `convert --update` now detect verbatim-stale tag entries in
+  `.work/CONVENTIONS.md` (only the known prior plugin defaults — user-edited
+  lines are treated as intentional) and offer to refresh them via
+  AskUserQuestion. Non-destructive: only flagged lines are touched, never the
+  rest of CONVENTIONS.
+
+## v0.7.4 - AGENTS-backed legacy pattern rules migration
+
+- **Gate-patterns rules source** - `gate-patterns` now reads project rules from
+  AGENTS targets instead of treating `.claude/rules/patterns.md` as an input.
+  `.agents/skills/patterns/` remains the canonical structural-pattern catalog,
+  with `.claude/skills/patterns/` only as a legacy mirror.
+- **Convert sync for older repos** - `convert` and `convert --update` now detect
+  legacy `.claude/rules/patterns.md`, import non-duplicate content into the
+  selected AGENTS target, and replace the legacy file with a shim so AGENTS is
+  the single project-rules source.
+- **Docs alignment** - SPEC, ARCHITECTURE, and MIGRATION now document AGENTS as
+  the canonical rules home and `.agents/skills/patterns/` as the structural
+  pattern-skill home.
+
+## v0.7.2 - Codex sub-agent and AGENTS.md compatibility
+
+Improves the plugin's Codex path while preserving Claude compatibility.
+
+- **AGENTS.md canonicalization** - `convert` now writes the agile-workflow
+  section to AGENTS.md and keeps CLAUDE.md as a symlink or compatibility shim.
+  Root, `.agents/`, and `.claude/` instruction-file locations are detected.
+- **Sub-agent routing** - implementation and gate skills now carry explicit
+  inline guidance for Claude model choices and Codex reasoning effort. The
+  implement orchestrator owns bundle/wave planning and may run large
+  non-overlapping write paths in parallel.
+- **Codex plugin metadata** - the Codex manifest now declares capabilities and
+  a default prompt, and explicit-only skills ship Codex `agents/openai.yaml`
+  policy files.
+- **Hooks** - plugin hooks resolve both Codex and Claude plugin roots, and the
+  PostToolUse hook recognizes Codex `apply_patch` payloads.
+- **Pattern skills** - reusable pattern references are written under
+  `.agents/skills/patterns` with optional Claude mirrors.
+
+## v0.6.1 — UI/UX alignment promoted to scope/epic tier
+
+Shifts where `ux-ui-design` mockups get produced. Epic-design Phase 4.6 is
+now the PRIMARY mockup tier and errs on the side of mocking; feature-design
+is the FALLBACK, used only for minor surfaces not covered upstream. The
+`--only-questions` modes always run the mockup pass — they're the visual
+alignment gate, not just the textual one.
+
+- **`epic-design`** Phase 4.6 — primary mockup tier: palette + screens
+  (for every net-new surface) + flows (for every journey). Under
+  autopilot, defers with a `## UI alignment deferred` note. Child feature
+  template now carries `## Mockups` for inheritance.
+- **`epic-design --only-questions`** — mockup pass added to the workflow.
+- **`feature-design`** Phase 4.6 — fallback tier. Inherits parent-epic
+  mocks by reference; only invokes ux-ui-design for surfaces upstream
+  missed.
+- **`feature-design --only-questions`** — Phase 4.6 added with the same
+  fallback rules.
+- **`scope`** Phase 1.8 — large-scope + UI actively invokes `:palette`
+  during the Phase 4 roll-forward and `:flows` for cross-feature
+  journeys clear at scope time.
+- **`ux-ui-design:ux-ui-principles`** — new "Where in the workflow to mock"
+  section encoding the tier ordering; description rewritten in ALWAYS
+  style to match agile-workflow conventions.
+
+## v0.4.3 — orchestrator becomes default, scope-driven, cross-feature
+
+Shifts the routing for `stage: implementing` work so the orchestrator is the
+default path and is no longer scoped to a single parent feature.
+
+- **`implement-orchestrator`** — now the default for everything at
+  `stage: implementing` (features with or without children, lone stories,
+  parentless items). Accepts a scope arg in any of four forms: a feature id
+  (current behavior), an epic id, `--all`, or an explicit list of item ids.
+  Phase 1 grounding reads every parent feature in the work set, not just
+  one. Phase 2 builds a unified `depends_on` graph across the whole scope —
+  cross-feature dependencies are first-class, not partitioned by parent.
+  Phase 3 introduces a file-overlap conflict check before each wave; either
+  serialize conflicting items into a later sub-wave or spawn the wave with
+  `isolation: "worktree"`. Phase 9 advances **every** parent feature whose
+  children are all at `review` (multi-parent), and leaves parents with
+  out-of-scope implementing children at `implementing` until a later run
+  finishes them.
+- **`implement`** — repositioned as the inline alternative. Same single-stride
+  workflow as before; the trigger now points at the orchestrator as the
+  default and lists the cases where inline is the right call (small / focused
+  work, user explicitly asks to land inline, mid-flow continuation). No
+  numeric thresholds — the choice stays a judgment call.
+- **`autopilot`** Phase 5 — when the picked anchor is at `stage: implementing`,
+  hands the **full autopilot scope** (epic id, `--all`, or the reduced list
+  from a free-text directive) to the orchestrator in one call. The
+  orchestrator drains the entire in-scope implementing band, advancing
+  parents along the way. Phase 4 documents that the picked item is an
+  "anchor" only on the implementing branch — drafting and review continue
+  to be processed one item at a time.
+- **`docs/ARCHITECTURE.md`** — pipeline diagram, queue-selection algorithm,
+  and skill table updated to reflect orchestrator-as-default and the
+  scope-driven shape. Queue filter expanded to include `review` (autopilot
+  has handled review autonomously since v0.3.x; the doc was stale).
+
+The throughput motivation: with stories defaulting to the orchestrator, a
+ready story from feature A and a ready story from feature B can pack into the
+same 3-agent wave instead of running sequentially. Single-feature scopes
+behave exactly like before — only multi-feature or `--all` scopes see the new
+fan-out.
+
+## v0.4.2 — human-facing work-board view
+
+Adds a slash command and renderer that produce a self-contained HTML
+kanban board from the project's `.work/` substrate.
+
+- **`/agile-workflow:board`** — new slash command. Walks `.work/`, parses
+  every item's frontmatter, and writes a single HTML file (no external
+  assets) with five columns mapped from stage: Backlog → Drafting → In
+  Progress → Review → Done. Auto-opens via `xdg-open` / `open` /
+  `wslview` / `start` when a launcher is available.
+- **`scripts/work-board.sh`** — pure-bash renderer. Reuses the
+  `work-view` frontmatter parsing patterns. Item title and excerpt are
+  base64-encoded into the embedded JSON so quotes, backticks, and
+  `</script>` sequences round-trip safely. Computes ready/blocked
+  signals and surfaces unmet dependencies on each card.
+- **`scripts/work-board.template.html`** — embedded CSS/JS, light + dark
+  modes via `prefers-color-scheme`, kind-colored card borders, kind/tag
+  filter pills, fuzzy id/title search ('/' to focus), and a collapsed
+  Releases section grouping items by `release_binding` /
+  `.work/releases/<version>/` directory.
+
+Flags: `--print` (skip auto-open), `--out <path>` (custom output path),
+`--serve [port]` (tiny `python3 -m http.server` for remote access).
+
+The renderer is read-only. Re-run the command to refresh — there's no
+watcher, no daemon, no background process. Stays consistent with the
+substrate philosophy: items are the source of truth.
+
+## v0.4.1 — drop model: frontmatter, skills inherit session model
+
+Removed the `model:` frontmatter field from every skill in the plugin
+(previously a mix of `opus` and `sonnet`). Skills now run at whatever model
+the session is on, rather than overriding for the current turn.
+
+Rationale: the `model:` override changed the model only for the turn the
+skill loaded into and reverted on the next prompt — useful only when a
+specific tier was genuinely required. For a workflow plugin where the user
+chooses the session model deliberately, forcing a different model per skill
+fragmented behavior and silently consumed extra tier capacity. Inheriting
+the session model is the right default. Sub-agent dispatches inside the
+skill bodies still set model explicitly via the Agent tool's `model:`
+parameter (e.g. the gates spawn opus sub-agents) — that's unchanged.
+
+Also brings the frontmatter into alignment with the skilltap official
+schema, which doesn't include `model:` at all.
+
+## v0.4.0 — gates run in opus sub-agents, autopilot drives continuously
+
+Two related shifts in how heavy thinking gets dispatched and how the queue
+runner paces itself.
+
+- **Gates are sub-agent driven.** All five quality gates (`gate-security`,
+  `gate-cruft`, `gate-docs`, `gate-tests`, `gate-patterns`) now delegate
+  their full analysis to a single opus sub-agent. The orchestrator skill
+  becomes a thin shell: prepare bundle context, dispatch the sub-agent
+  with a methodology brief, convert the structured findings it returns
+  into items (or pattern files, for `gate-patterns`), commit. The heavy
+  reasoning runs in an isolated context — domain audits, drift detection,
+  contract extraction, pattern discovery — leaving the orchestrator's
+  context lean and the analysis quality consistent. Orchestrator model
+  dropped from `opus` to `sonnet` for the four item-producing gates;
+  `gate-patterns` was already sonnet.
+- **Autopilot drives continuously; watchdog ticks are a hidden
+  safeguard.** Earlier wording around "session survives compaction" and a
+  "context approaching ~600k tokens" stop condition made the agent
+  self-throttle — pacing work against the watchdog ticks instead of
+  draining the queue. The skill now states explicitly: drive the queue
+  without pausing, the ticks are a harness-level safeguard you do not
+  reason about once scheduled, and there are exactly three stop
+  conditions (queue empty / truly stuck / user stop signal). Compaction
+  prediction is removed entirely — the harness handles it.
+
+## v0.3.2 — autopilot delegates to loop skill instead of prescribing args
+
+v0.3.1 prescribed `Skill(skill="loop", args="30m /agile-workflow:autopilot
+--resume")` directly in autopilot's Phase 1, which still over-specified the
+invocation — agents either guessed at the args format or skipped the loop
+skill entirely and went straight to CronCreate. Phase 1 and Phase 8 now just
+say "load the loop skill via the Skill tool and follow its instructions" and
+describe the desired schedule (30m + 3h, --resume), letting the loop skill
+handle the actual scheduling mechanics.
+
+## v0.3.1 — autopilot /loop invocation fix
+
+The `autopilot` skill instructed the agent to run `/loop ...` in bare code
+blocks, which agents interpreted as bash commands and ran via the Bash tool —
+silently dropping the watchdog scaffolding. Phase 1 (schedule) and Phase 8
+(cancel) now spell out the `Skill` tool invocation explicitly:
+`Skill(skill="loop", args="...")`.
+
+## v0.3.0 — autonomous review, caller awareness, orchestrator default
+
+Autopilot now drains `stage: review` items end-to-end. Skills called by
+autopilot use judgment instead of halting on ambiguity.
+
+- **Autonomous review.** Autopilot routes `stage: review` items to the
+  existing `review` skill. No new `auto-review` skill — same skill, both
+  callers. When the last child reaches `done`, the review skill auto-advances
+  the parent (feature or epic) to `review`, so epic-rooted graphs drain.
+- **Circuit-breaker.** After an item bounces `implementing → review →
+  implementing` twice, autopilot halts on it (leaves stage at review,
+  escalates in summary, skips).
+- **Caller awareness (principles Part III).** Rule: if
+  `/agile-workflow:autopilot` was invoked this session, no AskUserQuestion,
+  no halts on ambiguity — judgment + log. Otherwise, asking is fine. Hard
+  halts only for substrate-missing, foundation-missing, depends_on cycle,
+  contradictory state. Applied to `feature-design`, `perf-design`,
+  `implement`, `implement-orchestrator`, `review`.
+- **Orchestrator is the default for features.** Single-story features run
+  as a one-agent wave; multi-story as multiple waves capped at 3 parallel
+  per wave (no upper limit on waves). Per-story prompts mirror `implement`'s
+  workflow (land mode, dep check, design-flaw rollback). `implement`
+  standalone is now for stories and feature-without-children.
+- **Free-text autopilot scope.** `autopilot finish the dangling work` etc.
+  is interpreted as a scope directive, logged in the run summary.
+- **Convert: per-design classification.** Each `docs/designs/` entry classified
+  into one of five buckets (drafting / implementing / review / done-shipped /
+  done-archived) with sensible defaults and user confirmation.
+- **Convert: capture in-flight work.** Phase 8.5 clusters uncommitted
+  working-tree files via Explore and scopes each cluster as a
+  `stage: implementing` feature. `implement`'s land mode lands them.
+- **Story heuristics.** `feature-design` Phase 7 now uses an explicit
+  four-condition test (parallelizable / non-trivial deps / multi-session /
+  heterogeneous acceptance) instead of vague "skip if small."
+- **Hook timezone.** `updated:` bump uses local time, not UTC.
+
+## v0.2.0 — design family covers epics
+
+Autopilot can now drain epic-rooted graphs end-to-end. Previously, an epic at
+`stage: drafting` was a dead end for autopilot (the `design` skill only
+handled features, so autopilot's queue would empty out as soon as the only
+in-scope item was an undecomposed epic).
+
+- **New skill: `epic-design`** — decomposes an epic at `stage: drafting` into
+  child features at `stage: drafting` with declared `depends_on` chains,
+  written into the epic body. Mirrors `epicize` one level down.
+- **Renamed: `design` → `feature-design`** — for symmetry with the new
+  `epic-design`. Breaking change for any external invokers using
+  `/agile-workflow:design`; update to `/agile-workflow:feature-design`.
+- **Autopilot routing updated** — `kind: epic` at `drafting` now routes to
+  `epic-design` instead of being skipped. Only `.work/backlog/` is out of
+  scope for autopilot.
+- Cross-references in `scope`, `epicize`, `refactor-design`, `perf-design`,
+  `implement-orchestrator`, ARCHITECTURE.md, MIGRATION.md, README, and the
+  agile-workflow guide updated to the new names.
+
+## v0.1.0 — initial release
+
+Substrate-based work-tracking plugin. Sibling to `workflow`.
+
+- 25 skills covering ideation, conversion, scoping, design (greenfield + refactor + perf), implementation, review, gates, release, and autopilot
+- `work-view` bash script for fast queries by stage, tag, kind, parent, release binding, and dependency state
+- Two hooks (SessionStart queue snapshot, PostToolUse `updated:` auto-bump), both flag-gated by `.work/CONVENTIONS.md` presence
+- Foundation docs: VISION, SPEC, ARCHITECTURE, PRINCIPLES, MIGRATION
+- Pre-1.0 signals that frontmatter and stage shapes may shift during the first real-project shakedown