@event4u/agent-config 9.8.0 → 9.10.0

package/docs/archive/CHANGELOG-6.0.0-overview.md

@@ -1,953 +0,0 @@
-# CHANGELOG — 6.0.0 release overview (archived)
-
-> Drained from `CHANGELOG.md` `## [Unreleased]` on 2026-07-21. This was the
-> hand-authored "6.0.0 at a glance" release overview, staged under
-> `[Unreleased]` during 6.0.0 readiness and never removed after the release
-> shipped. It is preserved here verbatim for history; `[Unreleased]` now
-> starts clean. The dated per-version 6.0.0-era entries live under the
-> `# Era: pre-7.0.0 — archived` pointer in `CHANGELOG.md`.
-
----
-
-### 6.0.0 at a glance — release overview
-
-> A reader who never saw the PRs should understand 6.0.0 in under two
-> minutes. Seven headlines here; migration steps in **Breaking changes**
-> below; granular per-roadmap detail in the entries further down.
-
-6.0.0 is the release that turns agent-config into a **workplace**, not just a
-rule pack:
-
-1. **Workspace.** A task-orchestration layer owns the host-session lifecycle,
-   continuation, and drive health — the engine behind `/work` and
-   `/implement-ticket`.
-2. **Multi-turn continuation.** A long task survives across turns: the
-   workspace resumes where it left off instead of restarting cold.
-3. **Secure local stores.** Session state and sensitive local data live in an
-   encrypted-at-rest store under `~/.event4u/agent-config/`, not loose files.
-4. **MCP server.** agent-config exposes its skills and commands over MCP, so
-   MCP-aware hosts can reach them directly.
-5. **AI-video Creative Pack (optional).** The AI-video surface graduates from
-   prototype to validated provider adapters, shipped as an **opt-in Creative
-   Pack** — not core identity, not a new platform.
-6. **Breaking — condensed output relocated.** The repo-root `.agent-src/` tree
-   is removed; condensed output now lives at `dist/agent-src/` ([ADR-058]).
-7. **Upgrade cleanup.** Upgrading reaps stale deployed files (renamed/removed
-   skills and commands) so old installs don't rot; `agent-config --dry-run`
-   previews exactly what gets removed before any deletion.
-
-Pack-scoped projection (opt-in) and the TypeScript CLI + local-server
-foundation also land — both non-breaking by default.
-
-### Breaking changes (6.0.0)
-
-**1. Condensed output moved: root `.agent-src/` → `dist/agent-src/`** ([ADR-058])
-
-- **Who is affected:** consumers whose scripts or tool configs hard-code
-  `.agent-src/…` paths (plugin-marketplace clones, custom symlinks, CI globs).
-  Default installs are **not** affected — regenerated projections pick up the
-  new location automatically.
-- **What breaks:** the repo-root `.agent-src/` tree is removed, so any
-  hard-coded `.agent-src/<x>` path no longer resolves.
-- **How to migrate:** replace `.agent-src/` with `dist/agent-src/` in any
-  path you hard-code.
-- **Which command:** `agent-config install` (consumers) — or `task sync` +
-  `task generate-tools` (this repo) — regenerates projections at the new
-  location; the upgrade reaper removes the stale root tree. No manual path
-  edits beyond your own hard-codes.
-- **Rationale:** the root tree conflated source and build output; moving
-  condensed output under `dist/` makes "generated, do not edit" unambiguous
-  and keeps the repo root clean.
-
-**2. Settings key renamed: `cost_profile` → `rule_loading_tier`**
-
-- **Who is affected:** anyone with a hand-edited `cost_profile` in
-  `.agent-settings.yml`.
-- **What breaks:** nothing immediately — `cost_profile` is read as a fallback
-  during the grace period. The colliding memory-cadence meaning moved to its
-  own `memory.cadence` key (`auto`/`always`/`never`).
-- **How to migrate:** automatic.
-- **Which command:** `agent-config install` / `setup` migrates existing
-  settings on the next run. No manual action required.
-- **Rationale:** one key carried two unrelated meanings (rule-loading footprint
-  vs. memory-line cadence); splitting them removes the collision.
-
-[ADR-058]: docs/decisions/ADR-058-condensed-output-relocation-to-dist.md
-
-### Feedback wanted — multi-agent MCP config (demand probe, `ADR-086`)
-
-We are deciding whether to build a cross-agent MCP discovery / install helper.
-One question — answer in [Discussions](https://github.com/event4u-app/agent-config/discussions):
-
-> **Do you run more than one AI coding agent in the same project, and is keeping
-> MCP-server config in sync across them painful?**
-> (a) very painful · (b) somewhat · (c) single-agent, not applicable · (d) I already
-> self-script it.
-
-This gates the read-only MCP discovery helper (`agent-config mcp:search`) and the
-deferred auto-install question (`ADR-086`).
-
-### Added — humanizer (AI-tell removal for deliverable prose) (`road-to-humanized-writing` + `road-to-humanizer-hardening`)
-
-- **`humanizer` skill + `/humanize` command** (gtm-marketing pack) — remove
-  AI-writing tells from deliverable prose (posts, articles, drafts) via a
-  draft→audit→final loop; scoped to deliverable text only, never chat replies
-  or repo docs.
-- **`detect_ai_tells.ts` detector** — deterministic hard-tell / weighted-cluster
-  / em-dash-density scan (EN + DE), with a hidden-unicode ingestion scan and a
-  ReDoS length bound on untrusted input.
-- **Behavior change — write-engine step 4b is default-on.** Ghostwriter /
-  `post-as` drafts now pass a humanize audit before delivery; opt out per run
-  with `--raw`. The mechanical detector pass runs when a Node runtime is
-  present and degrades to a prose-only audit otherwise — the new
-  `humanizer-runtime` doctor check reports which path an install gets.
-- **Disclosure footer stays inviolable** through the audit; quoted/secondhand
-  text is never rewritten.
-
-### Added — `agents/` directory contract + gitignore hygiene (`road-to-agents-dir-and-gitignore-hygiene`)
-
-- **`docs/contracts/agents-layout.md`** — authoritative contract for every top-level entry in `agents/` (purpose, git policy, retention, consumer-scope); includes the User Inbox Workflow (`agents/tmp/` → `agents/tmp.old/`).
-- **`src/config/agents-paths.yml`** — machine-readable classification manifest for all agent-managed paths; feeds the new CI freshness gate.
-- **`agents/tmp/` + `agents/tmp.old/` now covered in consumer `gitignore-block.txt`** — both dirs were previously unignored for consumers.
-- **New CI gates** (all wired into `task ci`):
-  - `check-gitignore-freshness` — verifies every manifest entry is in the consumer block.
-  - `check-generator-output-coverage` — every `generate_*` output root in `condense.ts` must be classified.
-  - `check-tracked-but-ignored` — fails if any git-tracked file is now covered by an ignore pattern (makes the bench-run class of drift impossible to reintroduce).
-  - `check-generated-artefact-headers` — validates freshness markers on regenerable analysis artefacts.
-- **`lint_agents_layout.ts` extended** — in source-repo mode, unknown **directories** at the `agents/` root are now CI errors (previously only flat files were checked); `CONSUMER_EXPECTED_ENTRIES` extended with the full consumer-scope allowlist.
-- **Janitor** (`task janitor` / `task janitor-apply`) — TTL sweep for `agents/tmp.old/` (30 days) and `agents/runtime/tmp/` (7 days); reports user inbox state and `.harvest-local/` size; never auto-sweeps the user inbox.
-- **`/sync-gitignore:fix` extended** (Steps 5–6) — now also reports ignored-but-tracked files with the exact `git rm --cached` commands, and runs the manifest coverage check.
-- **`/roadmap:create`** — moves consumed `agents/tmp/` files to `agents/tmp.old/` in the same reply (inbox workflow).
-- **`agent-docs-writing` skill** — documents the user inbox contract; redirected `prediction-pool` agent scratch from `agents/tmp/` to `agents/runtime/tmp/`.
-
-#### Migration
-
-If you have a consumer project:
-
-1. Run `agent-config sync-gitignore --cleanup-legacy` to add `/agents/tmp/` and `/agents/tmp.old/` to your `.gitignore`.
-2. Run `git ls-files -ci --exclude-standard` — if any files appear, run the printed `git rm --cached` commands and commit.
-
-### Added — pack-scoped projection (opt-in, `road-to-6.0.0-b`)
-
-The projector can now write only the **active profile + packs'** artefacts into
-the host-tool trees instead of all 150 commands / 223 skills (ADR-040 —
-projection-time filtering, not a runtime resolver).
-
-- **Non-breaking by default.** A new `projection.mode` setting defaults to
-  `legacy-all` — `npm update` to 6.0.0 changes nothing. Scoped projection is
-  **opt-in** and never inferred from `profile.id`.
-- **Opt in** with `agent-config use --profile=<id>` (sets `projection.mode:
-  scoped`), then `agent-config refresh` to re-project. Selected packs are the
-  profile's base `packs` unioned with the `runtime.active_packs` session
-  overlay, expanded over the `requires` graph. `agent-config use
-  --profile=legacy-all` restores the full surface.
-- **Profiles are self-sufficient** — every `skills_hint` skill resolves from
-  the profile's base packs (the new `profile.packs` field). Resolved in the AI
-  council, 2026-06-03.
-- **Atomic.** A failed scoped projection restores the full (legacy-all) tree.
-  The committed plugin marketplace (`.claude-plugin/`) and the Augment tree
-  (`.augment/`) always project the full set in 6.0.0.
-
-**Staged rollout (this release ships 6.0.0 only):**
-
-1. **6.0.0** — default `legacy-all` (opt-in profiles). *This release.*
-2. **6.1.0** — default flips to profile/scoped mode with a `--legacy` escape.
-3. **7.0.0** — removes `legacy-all` **only if** telemetry shows <10% usage.
-
-The default flip and removal are explicitly **out of scope** for 6.0.0 and
-gated on the telemetry from 6.0.0-C. Deferred to 6.0.0-C: a
-`skills_discoverable` field, reactive just-in-time pack activation, and the
-per-pack budget lint.
-
-### Added — team mode: governed cross-model review (default-off, `road-to-team-mode`)
-
-A new default-off `/team` command family (`review`, `adversarial`, `delegate`,
-`status`) that governs the official `openai/codex-plugin-cc` as the Claude-Code
-transport — a second strong model reviewing the real working-tree diff (depth),
-the complement to the council's neutral-artefact breadth.
-
-- **Default-off + fail-closed.** With `ai_team.enabled: false` (default) no
-  command is suggested and invocation prints an enable pointer — byte-identical
-  to pre-roadmap behaviour. On Claude-Code hosts the wrappers are thin
-  delegations to `/codex:*`; absent the plugin they fail closed with the
-  `agent-config doctor --check team` remediation, never a silent no-op.
-- **Multi-host fallback (review lens only).** On non-Claude-Code hosts
-  `/team:review` sends a size-capped repo-diff bundle through the existing
-  subscription `OpenAICliClient`, with an honest capability-delta header.
-- **Governed, not rebuilt.** Nothing upstream-maintained is reimplemented; the
-  Review-Gate stays upstream's (we add an opt-in loop-bound + ledger line), and
-  `/team:delegate` (write access) sits behind a second opt-in
-  `ai_team.allow_delegate: false`. Config: `docs/contracts/ai-team-config.md`.
-- **No quality claim yet.** Cross-model defect-finding lift is unproven until
-  the Phase 5 benchmark runs (spend-gated); the docs describe workflow value
-  only, never a lift claim.
-
-### Breaking — v4.0.0 unified setup (`road-to-unified-setup`)
-
-`v4.0.0` is a **hard-cut** release. The legacy Python installer at
-`scripts/install.py` and the standalone TypeScript installer workspace
-at `packages/core/installer/{src,tests}` are no longer the canonical
-install path; the wizard now boots a single Fastify process at
-`agent-config install` / `agent-config setup` and drives the whole
-plan + apply through the TypeScript engine under `src/install/`.
-
-**Migrating from v3.x**
-
-1. The first `agent-config install` run detects an existing v3 tree at
-   `~/.event4u/agent-config/` and renders a **backup screen** before
-   any write. Pick **"Backup v3 and proceed"** — the wizard copies the
-   v3 layout to `~/.event4u/agent-config.v3.bak/` atomically and runs
-   the v4 install against a clean slate.
-2. No auto-migration runs. Rolling back is a single rename:
-   `mv ~/.event4u/agent-config.v3.bak ~/.event4u/agent-config`.
-3. Re-running `npx @event4u/agent-config install` against an active v3
-   layout (without backup) is refused — silent overwrite is gone.
-
-**Removed / retired**
-
-- `packages/core/installer/{src,tests,package.json,tsconfig.json,
-  vitest.config.ts}` (the standalone `@event4u/installer` workspace)
-  and its workspaces declaration in the root `package.json`.
-- `/install-via-agent` command — its JSON agent-mode protocol shipped
-  with the removed installer.
-- `task installer-typecheck` / `task installer-test` gates in
-  `taskfiles/ci-fast.yml`.
-
-**Kept for one release**
-
-- `scripts/install.py` remains so the `curl | bash setup.sh` CURL
-  fallback + `smoke-public-install.yml` CI matrix keep working. It is
-  **not** invoked from the modern npx flow. Full removal is scheduled
-  for v5.
-
-**New features (Phase B / C / E)**
-
-- Hard-stop **continue-screen** after Step 3 (modules) when entering
-  via `agent-config install`; `setup` lands directly at Step 4
-  (identity) via `?mode=setup`. (B5)
-- `WIZARD_READY <url>` line on stdout once Fastify is bound; the new
-  `scripts/bootstrap.sh` watches for it so a Node bootstrap does not
-  poll the port. (B4 + C1)
-- **Recovery-pre-step** — `GET /api/v1/install/recovery` surfaces an
-  abort marker (Resume / Rollback / Ignore) when the previous run
-  closed mid-apply. (B4 + Council Finding #24)
-- **v3 backup screen** — `GET /api/v1/install/legacy-v3` +
-  `POST /api/v1/install/backup-v3` drive the migration. (E2 +
-  Council Finding #21)
-- **Batch-conflict resolution** at ≥ 5 conflicts (B3 + Council
-  Finding #19).
-
-**TypeScript CLI shell + local server foundation (`typescript-cli-and-local-gui-foundation`)** —
-Phase-5 cutover of the `agent-config` binary from a 955-line Bash
-dispatcher to a compiled TypeScript entry point. The shell handles
-native commands (`versions`, `doctor-shell`, `ui:serve`) and forwards
-every other subcommand to the legacy logic, now living at
-`scripts/_dispatch.bash`. New surfaces:
-
-- **TS bin entry** — `package.json#bin` flipped to
-  `dist/cli/agent-config.js`. `npm run build:cli` compiles and
-  `chmod +x`-es the entry; `scripts/prepack-check.mjs` runs on
-  `npm pack` / `npm publish` and asserts the compiled file exists,
-  is executable, and carries the Node shebang.
-- **Bash shim** — the old path `scripts/agent-config` is now a
-  ~30-line deprecation shim that forwards to the TS binary. The real
-  dispatch logic moved to `scripts/_dispatch.bash` to prevent the
-  forward-and-back loop. Doctor probes both paths.
-- **Embedded Fastify server** — `agent-config ui:serve` boots a
-  local server bound to `127.0.0.1`, on the first free port in
-  `41000–41999`. Wire shape locked in
-  [`docs/contracts/local-server-api.md`](docs/contracts/local-server-api.md)
-  (beta): per-process bearer token, Host-header guard, Origin
-  allow-list, `GET /api/v1/ping` liveness probe.
-- **CI gates** — `task lint-ts`, `task typecheck-ts`, `task test-ts`,
-  `task build-ts` wired into `task ci` and `task ci-strict`. New
-  GitHub Actions job `TypeScript Tests` runs the same harness across
-  `ubuntu-latest` and `macos-latest` on Node 20, including the
-  `prepack-check.mjs` gate.
-- **Test coverage** — 47 Vitest tests across
-  `tests/cli/{paths,registry,cli-e2e}.test.ts`,
-  `tests/server/{app,port,token}.test.ts`, and
-  `tests/ui/build.test.ts`. CLI-e2e exercises the shim → dispatcher
-  → TS round trip; server tests cover the auth gate and 421/403/401
-  failure modes.
-
-Decision record: [`docs/decisions/ADR-012-typescript-cli-shell.md`](docs/decisions/ADR-012-typescript-cli-shell.md).
-
-**Discovery polish — `--root` override, doctor diagnostics, install-mode marker (`step-8-discovery-polish`)** —
-follow-up to step-7 driven by AI Council review of PR #157. Adds an
-explicit escape hatch for monorepos, surfaces the resolution decision
-path, and replaces install-mode heuristics with an authoritative
-marker file. New surfaces:
-
-- **Global `--root <dir>` flag** — parsed by the bash dispatcher
-  before any subcommand and wins over every other channel (subcommand
-  `--project`, `AGENT_CONFIG_PROJECT_ROOT` env-pin, anchor walk).
-  Implemented as `AGENT_CONFIG_ROOT_OVERRIDE=1` +
-  `AGENT_CONFIG_PROJECT_ROOT` plus a new `ORIGIN_ROOT_FLAG` so doctor
-  can attribute the resolution. Wrapper-coupling warning emitted on
-  stderr when `--root` overrides a wrapper-pinned env-pin.
-- **Fail-loud root validation** — `--root`, `--project`, and
-  `AGENT_CONFIG_PROJECT_ROOT` all reject non-existent paths and
-  non-directories via `ProjectRootError` mapped to exit code `2`. No
-  silent CWD fallback when the operator explicitly named a path.
-- **`agent-config doctor --trace-root`** — read-only diagnostic that
-  prints every ancestor probed during root discovery, the winning
-  anchor, and the originating channel. `--json` for machine-readable
-  output. Short-circuits the drift report.
-- **`agent-config doctor --context`** — prints effective project
-  root, origin, install mode + source, settings-layer chain, wrapper
-  state, and env-pin/override flags. `--json` available.
-- **Install-mode marker file** — `agents/.agent-state/install-mode.txt`
-  written by `install.py` with `minimal\n` or `full\n`. Doctor uses
-  the marker authoritatively for new installs; back-compat heuristic
-  (`AGENTS.md` + copilot bridges) only fires when the marker is absent.
-- **Minimal-install upgrade hint** — `install.py` emits a stderr
-  hint after a `--minimal` install describing how to upgrade to a
-  full install when the user is ready.
-- **Test coverage** — `tests/test_root_override.py` (precedence,
-  fail-loud for each channel, end-to-end CLI exit-2 check),
-  `tests/test_doctor_trace.py` (text + JSON shape for both flags,
-  marker-vs-heuristic install-mode detection, origin propagation).
-  Twenty new tests, full suite green.
-
-Council artefacts:
-`agents/council-sessions/step-8-discovery-polish-decision.md` — 2
-rounds + peer-review, $0.12. Decision cut `template-gen` (A4) and
-`anchor-freeze AC` (A6), kept `--trace-root` and `--context` as two
-distinct flags (host override of Claude's consolidation proposal),
-and added the install-mode marker file as the hybrid resolution to
-Q5 (marker for new installs, filesystem heuristic for back-compat).
-
-**Agent-folder discovery + `--minimal` init (`step-7-agent-folder-discovery-and-minimal-init`)** —
-project-root discovery widens beyond `.git` so non-git checkouts,
-monorepo sub-trees, and `agent-config`-only worktrees resolve
-correctly, and a new minimal-init payload lets teams keep the
-runtime global while committing only the per-project shell. New
-surfaces:
-
-- **Boundary anchors (D1+D3)** — `.git` plus `agents/` containing
-  **any** of `roadmaps/`, `.ai-council.yml`, `roadmaps-progress.md`.
-  Bare `agents/` is **not** an anchor (false-positive guard). When
-  no boundary anchor exists in any ancestor, the **outermost**
-  `.agent-settings.yml` becomes the root so the layered-settings
-  cascade survives. Same-level tiebreaker:
-  `.agent-settings.yml` > `agents/` > `.git` (diagnostic anchor
-  name only — resolved path is identical).
-- **`resolve_project_root(arg)` helper** (`scripts/_lib/agent_settings.py`)
-  — single entry point used by every `cmd_*.py`. Precedence:
-  explicit `--project`/`--target` → `AGENT_CONFIG_PROJECT_ROOT` env
-  → anchor walk → CWD fallback. Returns `(root, origin)` so doctor
-  can surface which step resolved.
-- **`AGENT_CONFIG_PROJECT_ROOT` env pin** — the `./agent-config`
-  wrapper exports its own directory before forwarding to the master
-  CLI, so subcommands invoked from a subdirectory (or after
-  `os.chdir`) stay pinned to the right root without re-walking.
-- **`agent-config init --minimal`** — writes exactly three files:
-  `.agent-settings.yml`, `agents/.gitkeep`, and the `./agent-config`
-  wrapper. No `.augment/` / `.claude/` / `.cursor/` / `AGENTS.md`.
-  Refuses to run when an ancestor already carries an anchor
-  (nested-install guard); override with explicit `--target`.
-  Does **not** pin `agent_config_version` (D4) — consumer follows
-  the global CLI.
-- **`AGENT_CONFIG_LEGACY_ANCHOR=1` kill-switch (D5)** — reverts
-  discovery to the pre-step-7 `.git`-only walk for one minor-version
-  soak. Scheduled for removal after the soak window unless
-  telemetry surfaces a missing case.
-- **Performance budget (D6)** — anchor walk is `O(depth)` with at
-  most three `Path.exists()` calls per level. Test asserts
-  `< 5 ms` at depth 20 (`tests/test_anchor_perf.py`).
-- **Test coverage** — `tests/test_project_root_anchors.py`
-  (precedence, mixed-anchor edge case, bare `agents/` rejection),
-  `tests/test_kill_switch.py`, `tests/test_minimal_init.py`
-  (payload shape + nested-install refusal), `tests/test_subdir_invocation.py`
-  (env-var precedence, wrapper-pin surviving `chdir`),
-  `tests/test_anchor_perf.py`.
-
-Council artefacts: `agents/council-sessions/step-7-d3-cascade-conflict-{question,decision}.md`
-— D3 strict closest-leaf-wins was relaxed to boundary-vs-layer
-split after the analysis lens flagged four pre-existing cascade
-tests as the contract, not edge cases.
-
-**PR #150 follow-up hardening (`step-9-pr150-feedback-hardening`)** —
-twelve-phase response to the Claude+GPT review findings on PR #150 plus
-three user-requested settings-shape changes (CLI-default, preferred-single
-member, low-impact dispatch) under one roadmap. New surfaces:
-
-- **`defaults.member_mode: cli | api`** (default `cli`) — flips the
-  global transport default to CLI-first onboarding (subscription auth,
-  no API key). Per-member explicit `mode:` keeps overriding.
-- **`routing.solo_member_fallback_chain: [provider, ...]`** — ordered
-  preference list for single-member dispatch. Disabled members are
-  skipped; duplicates rejected at config-load time; all-invalid
-  escalates to full council (never fails the decision). 15-minute
-  auth cache, 3-second timeout per probe.
-- **`low_impact.dispatch: full | single`** (default `full`) — opt-in
-  cost-efficient routing for low-impact decisions through
-  `routing.solo_member_fallback_chain`. **Iron Law:** `high_impact`
-  and `user_required` dispatch are NOT configurable — config validator
-  rejects both top-level and `decision_resolution.classes.*` forms,
-  including YAML `<<:` anchor smuggling attempts. Sentinel in
-  `tests/test_iron_law_config.py`.
-- **Shadow-mode safety net** — when `low_impact.dispatch: single` is
-  active, sample (default `0.1`) decisions go to both solo + full
-  council. Disagreements append to `agents/council-shadow-log.jsonl`
-  (privacy-redactor enforced). `agent-config council shadow-report`
-  computes the 7-day rolling SLO (`OK <5%` / `WARN 5–8%` / `BREACH
-  >8%`); pre-flight cost disclosure surfaces the verdict. No auto-
-  revert — humans decide.
-- **Confidence gate — auto-escalation on uncertain solo runs**
-  (`scripts/ai_council/confidence_gate.py`) — every solo response is
-  scored before the verdict is returned. Four escalation reasons in
-  priority order: `refusal` (empty / refusal markers), `split`
-  (two-verdict / option-A-vs-B / Variante 1/2), `short_response`
-  (< 20 chars), `low_confidence` (explicit `Confidence: 0.X` marker
-  or hedge-word density below floor). Zero extra LLM calls —
-  heuristics run in-process. Escalations append `escalated: true` +
-  `escalation_reason` to `agents/council-shadow-log.jsonl`; the SLO
-  banner appends the rolling auto-escalation rate so quiet uncertainty
-  cannot hide behind a flat disagreement rate.
-- **`low_impact.solo_confidence_floor: float`** (default `0.7`,
-  range `[0.0, 1.0]`) — threshold for the confidence gate.
-  **Iron Law:** `solo_confidence_floor` is also rejected on
-  top-level `high_impact:` and `user_required:` — those classes
-  never run solo, so a solo-specific knob there is incoherent.
-- **`AGENT_CONFIG_FORCE_FULL_COUNCIL=1`** — per-invocation kill-switch
-  that forces full council regardless of `low_impact.dispatch`.
-- **Airgap detection** (`scripts/ai_council/airgap.py`) — installer
-  first-run probes DNS for `api.anthropic.com`, `api.openai.com`,
-  `generativelanguage.googleapis.com` with 1s timeout per host. All
-  fail → banner `airgapped environment detected — defaulting to
-  mode: api` and `defaults.member_mode: api` recommendation.
-- **Fast-path marker visibility Iron-Law rule**
-  (`.agent-src.uncondensed/rules/fast-path-marker-visibility.md`,
-  kernel-tier, always-active) — host agents MUST surface
-  `Resolved via low-impact council fast-path: …` markers verbatim;
-  swallowing them is a rule violation.
-- **`/memory learn-low-impact --preview`** (default) / `--apply` —
-  preview shows promoted entries, refused entries with redaction
-  reasons, source-project-stripped diff, and the upstream PR body
-  draft before anything is written.
-- **Corpus parser hardening** — typed `CorpusParseError(reason=...)`
-  replaces silent skip; seven failure-mode fixtures in
-  `tests/fixtures/corpus-robust/` (missing anchor · renamed heading ·
-  bullet-without-quotes · duplicate entry · anti-example wrongly
-  under Validated · malformed ISO timestamp · redactor-bypass).
-  Contract in `docs/contracts/low-impact-corpus-format.md`.
-- **Fuzzy corpus matching (opt-in)** —
-  `low_impact.fuzzy_match.{enabled,threshold}` (defaults `false`,
-  `0.92`) using `difflib.SequenceMatcher`. High-impact trigger veto
-  + anti-example veto preserve the Iron Law.
-- **CLI-binary UX** — pre-flight cost disclosure now prints
-  `member X skipped: binary not found, install via <hint>` per
-  skipped CLI member; hints sourced from
-  `scripts/ai_council/cli_hints.py`.
-- **Doctor CLI checks** — `agent-config doctor --check council-cli`
-  reports per CLI member: binary present · auth probe · parse
-  fixture · quota remaining · billable flag.
-- **Iron-Law bypass linter** — `audit_cloud_compatibility.py
-  --iron-law` scans `scripts/` for raw YAML loads of `ai-council.yml`
-  that skip the config validator; baseline is 0 findings; suppress
-  legitimate cases with `# iron-law-ok: <reason>` on the load line.
-
-Surface delta: **4 new config knobs** (`defaults.member_mode`,
-`routing.solo_member_fallback_chain`, `low_impact.dispatch`,
-`low_impact.solo_confidence_floor`),
-**1 new env-var kill-switch** (`AGENT_CONFIG_FORCE_FULL_COUNCIL`),
-**1 new CLI subcommand** (`council shadow-report`), **1 new CLI flag**
-(`council run --single`), **1 new always-active rule**
-(`fast-path-marker-visibility`), **2 new local-only files**
-(`agents/council-shadow-log.jsonl`, gitignored;
-`scripts/ai_council/airgap.py`). Backward-compat: defaults map the
-existing behaviour (`dispatch=full`, empty chain, `member_mode=cli`
-in fresh installs; existing configs read unchanged). No migration
-script — defaults are silent-safe.
-
-**Low-impact corpus YAML lockfile (`step-10-corpus-yaml-lockfile`)** —
-follow-up to the PR #150 review (Punkt 4): the Markdown parser used to
-be the runtime source-of-truth for the low-impact decisions corpus,
-making every human edit-variation (curly quotes, bullet style,
-whitespace, heading rename) a potential parser bomb on the hot path.
-Switched the runtime to a generated YAML lockfile while keeping
-Markdown as the human-editable source. New surfaces:
-
-- **`scripts/ai_council/compile_corpus.py`** — build tool that runs
-  `parse_corpus_strict()` over `agents/low-impact-decisions.md` and
-  emits `agents/low-impact-decisions.lock.yaml` (schema-v1, sorted
-  keys, deterministic). `--check` mode exits `1` on drift, `2` on
-  parse error.
-- **`agents/low-impact-decisions.lock.yaml`** — generated, **committed**
-  lockfile. Schema `{schema_version: 1, provenance: {source_path,
-  source_sha256, last_upstreamed}, validated, probation, anti_examples}`.
-  Diffed in PR reviews alongside the Markdown source.
-- **`low_impact_corpus.load_corpus_lock()`** — new primitive that
-  loads the YAML lockfile and re-materialises a `CorpusParseResult`.
-  Schema-version mismatch raises `CorpusParseError` with reason
-  `schema_version_mismatch`.
-- **Lenient loaders prefer YAML** — `load_validated_phrases` and
-  `load_anti_example_phrases` (consumed by `necessity.py` and
-  `low_impact.py` on the hot path) check for the sibling `.lock.yaml`
-  first. Missing / malformed lockfile → silent fallback to lenient
-  Markdown parsing, so a fresh clone before `task sync` still works.
-- **CI gate** — `task compile-corpus` runs as part of `task sync` and
-  `task consistency` (before `git diff --quiet`). A stale lockfile
-  fails CI the same way `.agent-src/` drift does.
-- **Preview tool unchanged** — `learn_low_impact_preview.py` keeps
-  reading the Markdown directly via `parse_corpus_strict`; it runs
-  *before* `task sync` rebuilds the lockfile, so it must read whatever
-  the user just edited.
-
-The Markdown parser stays in the repo as a build-time dependency, not
-a runtime dependency — a parser regression breaks `task consistency`,
-never the live council. Privacy floor (`low-impact-corpus-privacy-floor`)
-unchanged: the redactor scans entry text, not file format.
-
-**Council quota & necessity transparency (`step-8-quota-necessity-transparency`)** —
-the council's two pre-flight gates (`cli_call_budget` and
-`necessity_classifier`) become observable and aligned with the
-"Council always active when enabled" mental model. New surfaces:
-
-- `council run` / `council debate` print a one-line
-  `council:quota · <provider> used/limit · …` summary before the
-  first member fires. Only providers with a configured per-day cap
-  appear; uncapped providers are omitted (no false metering).
-- `cli_call_budget.warn_at` (float, default `0.8`) — once
-  `used / max_calls_per_day >= warn_at`, the summary line is
-  prefixed `⚠️` and a `council:quota · WARN` line names the
-  providers near the wall.
-- New `agent-config council quota` subcommand dumps today's
-  `~/.event4u/agent-config/cli-calls.json` state plus the configured
-  caps. `--reset <provider> --confirm` clears today's counter for one
-  provider for manual rollover.
-- **Necessity tier split (D2):** `necessity_classifier.user_explicit_mode`
-  (default `warn-only`) separates the agent and user_explicit tiers.
-  User-typed `/council` calls proceed by default with an annotated
-  stdout line; agent-initiated dispatches keep `educate` behaviour.
-  New `warn-only` mode joins the existing `off | educate | block`
-  enum — annotates the verdict but never skips.
-- **Persistent events log (D3):** every necessity-gate decision
-  (`proceed` / `skip_necessity`) and every quota block
-  (`block_quota`) appends one JSON line to `agents/council-events.log`
-  (gitignored; never committed). Schema v1 with `schema_version`,
-  `ts_utc`, `lens`, `invocation`, `action`, `verdict`,
-  `provider_caps`, `original_ask_hash`. `original_ask` is hashed
-  `sha256[:12]` before write — the raw prompt is never persisted
-  (privacy floor per `agents/low-impact-decisions.md`).
-- **Kill-switch (D5):** `AGENT_CONFIG_NO_EVENTS_LOG=1` disables
-  every events-log write in-process. Mirrors the
-  `AGENT_CONFIG_LEGACY_ANCHOR=1` pattern from step-7.
-
-Surface delta: **1 new subcommand** (`council quota`), **2 new
-config knobs** (`cli_call_budget.warn_at`,
-`necessity_classifier.user_explicit_mode`), **1 new local-only file**
-(`agents/council-events.log`, gitignored). Backward-compat:
-existing `.agent-settings.yml` and `agents/.ai-council.yml` work
-unchanged; new knobs are optional. Docs:
-[`docs/contracts/ai-council-config.md`](docs/contracts/ai-council-config.md)
-+ [`docs/installation.md`](docs/installation.md) § AI Council local
-state.
-
-Four roadmaps land in this release.
-
-**Tier-0 trim (`road-to-surface-discipline` Phase 1)** — six CLI
-commands moved from Tier-0 to Tier-1 in `./agent-config --help` to
-collapse the daily-driver surface down to the seven commands a new
-contributor actually needs in their first session. Commands stay
-fully invokable by full name — only the default `--help` view
-changed. Source of truth: `docs/contracts/command-surface-tiers.md`.
-
-Pre/post diff of `./agent-config --help` Tier-0 block:
-
-| Command | Pre | Post | Rationale |
-|---|:-:|:-:|---|
-| `init` | Tier-0 | Tier-0 | daily-driver entrypoint |
-| `sync` | Tier-0 | Tier-0 | daily-driver entrypoint |
-| `validate` | Tier-0 | Tier-0 | daily-driver entrypoint |
-| `work` | Tier-0 | Tier-0 | daily-driver entrypoint |
-| `implement-ticket` | Tier-0 | Tier-0 | daily-driver entrypoint |
-| `help` | Tier-0 | Tier-0 | help meta-command |
-| `--version` | Tier-0 | Tier-0 | help meta-command |
-| `first-run` | Tier-0 | **Tier-1** | one-time setup; not in daily loop |
-| `keys:install-anthropic` | Tier-0 | **Tier-1** | one-time credential setup |
-| `keys:install-openai` | Tier-0 | **Tier-1** | one-time credential setup |
-| `council:estimate` | Tier-0 | **Tier-1** | on-demand review tool |
-| `council:run` | Tier-0 | **Tier-1** | on-demand review tool |
-| `council:render` | Tier-0 | **Tier-1** | on-demand review tool |
-
-Net surface delta: **0 new commands, 0 removed commands.** Only the
-`--help` surfacing changed. Run `./agent-config --help --tier=1` for
-the full power-user view (15 commands) or `--tier=all` for
-maintenance / hooks / MCP / telemetry (26 additional commands).
-
-**Diagnostic Hub (`road-to-surface-discipline` Phase 2)** — the
-existing `./agent-config doctor` is repositioned as the single
-entrypoint for health checks. New `CHECK_IDS` registry plus
-`doctor --check <id>` filter; `--list-checks` enumerates every
-runner; failing checks print the literal command line that
-reproduces the failure. Surface delta: **0 new commands** — only
-flags and registry plumbing.
-
-**MCP beta gating (`road-to-surface-discipline` Phase 3)** — MCP
-promotion criteria pinned in
-[`docs/contracts/mcp-beta-criteria.md`](docs/contracts/mcp-beta-criteria.md)
-with a four-gate contract (offline-readiness, scope, drift, runtime).
-`doctor --check mcp-beta-readiness` enforces the gates; failing tests
-under `tests/test_mcp_beta_gates.py` codify the promotion bar.
-Cloud-scope behaviour cross-referenced from
-[`docs/contracts/mcp-cloud-scope.md`](docs/contracts/mcp-cloud-scope.md).
-
-**Architecture refresh (`road-to-surface-discipline` Phase 4)** —
-[`docs/architecture.md`](docs/architecture.md) and
-[`docs/mcp-server.md`](docs/mcp-server.md) re-anchor the 6-layer
-system model (consumer → tools → installer → package → kernel →
-runtime). `AGENTS.md` re-trimmed to honour the Thin-Root contract
-(< 3,000 chars). No code churn — docs only.
-
-**Tier-usage telemetry (`road-to-surface-discipline` Phase 5)** —
-empirical retiering signal added behind a default-off opt-in. New
-`telemetry.tier_usage` namespace in `.agent-settings.yml`; signal
-contract pinned in
-[`docs/contracts/command-clusters.md`](docs/contracts/command-clusters.md)
-§ tier-usage (whitelist: `ts_bucket`, `command`, `tier`, `outcome`,
-`user_hash`; hourly time buckets; 16-char salted user hash; no paths,
-no argv). Retiering thresholds (≥ 20 invocations and ≥ 3 distinct
-users over 30 days) live in the same contract. A new
-`tier_usage_report.py` template script aggregates the local log into
-a frequency table and refuses to render rows that violate the
-privacy floor. The existing `doctor` gains a `tier-usage-readiness`
-check (warn when disabled or empty; fail when every record is
-poisoned; ok when ≥ 1 record passes the floor). Surface delta: **0
-new commands, 0 new skills, 0 new personas** — telemetry rides on
-the existing dispatcher and the existing `doctor` entrypoint.
-
-
-**Package consolidation** — the standalone wrapper package
-`@event4u/create-agent-config` is retired. `npx @event4u/agent-config init`
-is now the canonical one-shot entrypoint; the bundle package gains an
-`init` subcommand that delegates to `scripts/install`. `packages/create-agent-config/`
-is removed from the repo, and every doc / template / help string is
-updated to the new command. Existing users running the old wrapper
-should switch — the old package name will stop receiving releases.
-
-**Global-First Install (R4)** — `npx @event4u/agent-config init`
-now defaults to a **global** install (`~/.claude/`, `~/.cursor/`, …)
-when run outside a project, and a **project** install when run inside
-one. A new global lockfile at `~/.config/agent-config/installed.lock`
-records `agent_config_version` + `tools[]` and refuses mismatched
-re-installs (exit 1) until `update` or `--force` realigns it. New
-`agent-config export --tool=<id> --output=<path>` subcommand ejects a
-named tool's resolved content into the project tree (idempotent;
-`--force` overrides drift). 12 supported tool ids (claude-code,
-cursor, windsurf, cline, aider, codex, gemini, copilot,
-claude-desktop, continue, kilocode, zed, jetbrains, kiro). Windows
-CI matrix added for the lockfile + export paths. See
-[`docs/decisions/ADR-007-agent-discovery-scopes.md`](docs/decisions/ADR-007-agent-discovery-scopes.md)
-and [`docs/decisions/ADR-008-installed-tools-manifest.md`](docs/decisions/ADR-008-installed-tools-manifest.md).
-
-Three roadmaps land in this release.
-
-**Universal Execution Engine (R1)** — the `/implement-ticket` runtime is
-renamed and re-shaped into a universal dispatcher. **No user-visible
-behavior change** — the `/implement-ticket` slash command and the
-`./agent-config implement-ticket` CLI are byte-stable, gated by the
-new Golden-Transcript replay harness.
-
-**Prompt-Driven Execution (R2)** — a new `/work` command drives free-form
-prompts through the same `work_engine` dispatcher with a
-confidence-band gate at `refine`. R1 goldens remain byte-equal across
-the R2 changes.
-
-**Product UI Track (R3)** — three new directive sets (`ui`,
-`ui-trivial`, `mixed`) turn UI work from "writes the code" into
-"produces UI that feels designed". An existing-UI audit is enforced
-as a hard gate before any `apply`; design briefs lock microcopy
-verbatim; polish has a 2-round ceiling. Stack-aware dispatch routes
-implementation to `blade-livewire-flux` / `react-shadcn` / `vue` /
-`plain` skill bundles. Happy-path halt budget is 2 (audit pick +
-design sign-off). R1 + R2 goldens stay byte-equal; 12 new GT-U
-baselines pin the R3 contract.
-
-### Features
-
-* **engine:** universal `work_engine` Python module with explicit
-  `version: 1` state schema, `directive_set` envelope (backend / ui-stub
-  / mixed-stub), and `input.kind`-based dispatch ready for prompt-driven
-  and UI directives in R2 / R3.
-* **migration:** `work_engine.migration.v0_to_v1` auto-migrates
-  `.implement-ticket-state.json` → `.work-state.json` on first run;
-  v0 file preserved as `.implement-ticket-state.json.bak`. Idempotent
-  and refuses to overwrite an existing v1 destination.
-* **tests:** Golden-Transcript replay harness with 5 live-captured
-  baseline transcripts (`tests/golden/baseline/GT-{1..5}/`),
-  `CHECKSUMS.txt` SHA-256 manifest, four strict comparators
-  (exit codes, state-snapshot structure, halt-marker shape, delivery
-  report headings) with allow-listed free-text drift on `questions`
-  and `report` bodies. Both pytest (`tests/golden/test_replay.py`) and
-  CLI (`python3 -m tests.golden.harness`) entry points.
-* **command:** `/work` — sibling entrypoint to `/implement-ticket` that
-  drives free-form prompts through the same `work_engine` dispatcher
-  with `input.kind="prompt"`. Backed by the `command-routing` and
-  `refine-prompt` skills.
-* **engine (R2):** prompt resolver (`work_engine.resolvers.prompt`)
-  builds the prompt envelope; `directives/backend/refine.py::_run_prompt`
-  reconstructs acceptance criteria, scores the prompt on five
-  dimensions, and dispatches on the resulting band.
-* **scoring:** deterministic, heuristic-only confidence scorer at
-  `work_engine.scoring.confidence` — single source of truth for the
-  rubric (`goal_clarity`, `scope_boundary`, `ac_evidence`, `stack_data`,
-  `reversibility`) and band thresholds (`high ≥ 0.8`, `medium ≥ 0.5`,
-  `low < 0.5`).
-* **band-action gate:** silent proceed on `high`, assumptions-report
-  halt on `medium`, one-question halt on `low` (per the
-  `ask-when-uncertain` Iron Law).
-* **tests (R2):** four new Golden Transcripts (`GT-P1` high-band happy,
-  `GT-P2` medium-band release, `GT-P3` low-band one-question halt,
-  `GT-P4` UI-intent rejection — superseded by the R3 UI track but kept
-  to pin pre-R3 behaviour) pinned alongside the R1 goldens.
-
-### Changed
-
-* **engine (refactor):** internal Python module renamed
-  `implement_ticket` → `work_engine`. Public CLI surface, slash command,
-  user-facing prompts, halts, and delivery report are unchanged. State
-  filename moves from `.implement-ticket-state.json` to `.work-state.json`
-  with auto-migration.
-* **engine (R2):** `refine` SUCCESS paths now mirror
-  `data["reconstructed_ac"]` into `state.ticket["acceptance_criteria"]`
-  as an independent list copy, so downstream gates (`analyze`, `plan`)
-  read the same AC slot regardless of envelope kind.
-
-### Deprecated
-
-* **`implement_ticket` Python module** — re-export shim retained for
-  backwards compatibility; emits `DeprecationWarning` on import.
-  Internal Python consumers should migrate to `from work_engine import …`.
-  Removal is a separate user-driven decision, not pinned to a release.
-
-### CI
-
-* **freeze-guard:** named `task golden-replay` step in `Taskfile.yml`
-  and a dedicated `Golden Replay` step in `.github/workflows/tests.yml`,
-  both invoked before the main pytest sweep so structural regressions
-  surface first. Freeze-guard workflow rejects baseline edits outside
-  `R1-P1`-tagged commits.
-* **roadmap-progress-check** wired into `task ci` (1.13.0 carry-over).
-
-### Documentation
-
-* **ADR (R1):** [`agents/contexts/adr-work-engine-rename.md`](agents/contexts/adr-work-engine-rename.md)
-  — rationale, scope of the rename, compatibility shim policy, state
-  migration, golden-test contract, tradeoffs, non-goals.
-* **ADR (R2):** [`docs/contracts/adr-prompt-driven-execution.md`](docs/contracts/adr-prompt-driven-execution.md)
-  — naming decision (`/work` over `/do`), confidence-band gate,
-  AC-projection fix, R3 deferral boundary, and golden contract.
-* **flow:** `docs/contracts/implement-ticket-flow.md` gains a "Replay
-  protocol — Strict-Verb comparison" section (R1) and a
-  "Prompt envelopes and confidence bands (R2)" section pinning the
-  band-action mapping to the scorer module.
-* **README + AGENTS.md template:** document `/work` as the sibling
-  entrypoint, the confidence-band rubric, and the
-  `/work` vs. `/implement-ticket` selection rule.
-* **rules:** `scope-control` forbids release / version / deprecation-date
-  language in roadmaps, plans, and ADRs; introduces a `Decline = silence`
-  policy preventing branch-switch and PR proposals from being re-asked
-  on the same task.
-
-### Features (R3 — Product UI Track)
-
-* **directive sets:** `directives/ui/` (`audit → design → apply →
-  review → polish → report`), `directives/ui_trivial/` (provably
-  bounded edits, `MAX_FILES = 1` and `MAX_LINES_CHANGED = 5`), and
-  `directives/mixed/` (`contract → ui → stitch`). Dispatched at the
-  engine boundary on `state.directive_set`; the audit gate is enforced
-  at the dispatcher AND at always-on rule level
-  ([`ui-audit-gate`](.agent-src/rules/ui-audit-gate.md))
-  so cloud surfaces and free-form edits cannot bypass it.
-* **stack detection:** `scripts/work_engine/stack/detect.py` reads
-  `composer.json` + `package.json` and labels the frontend as
-  `blade-livewire-flux` / `react-shadcn` / `vue` / `plain`. Cached on
-  `state.stack` against manifest `mtime`; recoverable on parse error
-  (downgrades to `plain`). `apply` / `review` / `polish` route on the
-  label.
-* **audit gate:** `state.ui_audit` must carry `≥ 1 components_found`
-  OR `greenfield = True` with a user-chosen `greenfield_decision` ∈
-  `{scaffold, bare, external_reference}`. Empty dict, `None`, or
-  populated-without-keys is rejected; the dispatcher emits
-  `@agent-directive: existing-ui-audit` and refuses to advance.
-* **design brief lock:** `apply` rejects components carrying
-  `PLACEHOLDER_PATTERNS` (`<placeholder>`, `lorem`, `todo:`, `tbd`,
-  `xxx`); both producer and consumer enforce.
-* **polish ceiling:** hard `POLISH_CEILING = 2` rounds, validated at
-  in-memory state, on-disk schema, AND dispatcher layers. After round
-  2 the engine halts with ship-as-is / abort / hand-off.
-* **token-violation extraction:** polish classifies
-  `kind == "token_violation"` findings against
-  `state.ui_audit.design_tokens`; matched values use the named token,
-  unmatched values repeated > `TOKEN_REPEAT_THRESHOLD = 2` times
-  trigger token extraction before the next round.
-* **trivial-path reclassification:** `directives/ui_trivial/apply.py`
-  flips `state.directive_set = "ui"` when the bounded preconditions
-  fail, restarting the dispatcher at audit. Reclassification is loud
-  and counted in the delivery report.
-* **mixed orchestration:** sentinels `state.contract.contract_confirmed`
-  (UI sub-flow refuses to start without it) and
-  `state.stitch.verdict = "success"` (stitch's success condition;
-  `blocked` / `partial` halts unless `integration_confirmed` flips).
-* **skills:** new `existing-ui-audit`, `react-shadcn-ui`, plus
-  stack-specific apply / review / polish bundles. `fe-design`
-  repositioned as a framework-agnostic reference cited by
-  `directives/ui/design.py`.
-* **tests (R3):** twelve new Golden Transcripts (`GT-U1`..`GT-U4`,
-  `GT-U9`..`GT-U12`) covering high-confidence happy path
-  (1 halt budget), ambiguous (2 halts max), greenfield decisions,
-  trivial-path apply, polish ceiling hit, and mixed orchestration.
-  All R1 + R2 + R3 baselines auto-discovered by `tests/golden/test_replay.py`.
-
-### Documentation (R3)
-
-* **ADR (R3):** [`docs/contracts/adr-product-ui-track.md`](docs/contracts/adr-product-ui-track.md)
-  — audit-as-hard-gate rationale, design-review loop, halt-budget
-  reasoning, trivial-path-and-reclassification, stack-detection
-  strategy, fe-design migration, tradeoffs, non-goals.
-* **flow contract:** [`docs/contracts/ui-track-flow.md`](docs/contracts/ui-track-flow.md)
-  — slot-by-slot wiring for `ui` / `ui-trivial` / `mixed`, the
-  audit-path table (`STRONG_SIMILARITY = 0.7`, `TIE_GAP = 0.05`),
-  design-brief lock + placeholder patterns, stack-dispatch tables,
-  polish ceiling, trivial preconditions, mixed sentinels,
-  idempotency table, declared ambiguities across all eight directives.
-* **extension recipe:** [`docs/contracts/ui-stack-extension.md`](docs/contracts/ui-stack-extension.md)
-  — how to add a new stack (Svelte, SolidJS, Astro, …): label
-  conventions, detector heuristic, three required skills, dispatch
-  wiring, version anchor, Golden fixture, end-to-end verification.
-* **README + AGENTS.md template:** UI track flow table (`ui` /
-  `ui-trivial` / `mixed`), audit-gate property, design-brief lock,
-  polish ceiling, stack-dispatch summary; the pre-R3 "rejected,
-  backend-only" wording is removed.
-* **golden test capture:** [`tests/golden/CAPTURING.md`](tests/golden/CAPTURING.md)
-  — central regeneration recipe, lock policy, when to relock; replaces
-  17 per-baseline `reproduction-notes.md` files (the loadable artefacts
-  — `transcript.json`, `state-snapshots/`, `halt-markers.json`,
-  `exit-codes.json`, `delivery-report.md` — stay).
-
-### Archived (R3)
-
-* **roadmap:** `agents/roadmaps/intent-based-orchestration.md` moved to
-  `agents/roadmaps/archive/`; superseded by R1 + R2 + R3.
-
-**Artefact-Engagement Telemetry** — opt-in, default-off measurement layer
-that records which skills, rules, commands, and guidelines the agent
-consults and applies during a `/implement-ticket` or `/work` run.
-Maintainer-targeted; consumers see no prompts.
-
-### Features
-
-* **telemetry:** `./agent-config telemetry:record` and
-  `./agent-config telemetry:report` CLI scripts + `telemetry/` Python
-  package (boundary, aggregator, renderer, schema). Default-off via
-  `telemetry.artifact_engagement.enabled` in `.agent-settings.yml`.
-* **schema:** `schema_version: 1` JSONL events with `task_id`,
-  `boundary_kind`, `consulted` / `applied` (kind→ids), `ts`,
-  `tokens_estimate`. Aggregator silently skips lines with unknown
-  schema versions.
-* **report:** quartile bucketing on applied/consulted ratio — Essential
-  (top 20 %), Useful (mid 60 %), Retirement Candidates (bottom 20 %) —
-  emitted as Markdown or JSON.
-* **rule:** `artifact-engagement-recording` — fires per phase-step
-  inside `/implement-ticket` and `/work`; no-op under the default-off
-  gate so the recording path costs nothing when disabled.
-* **privacy:** `check_id_redaction` validator enforces a redaction
-  floor (no path separators, no file extensions, no control characters,
-  no whitespace, non-empty) on **both** write and export. Tampered or
-  legacy logs cannot leak through reports — renderer re-validates.
-* **tests:** 36 redaction test cases plus end-to-end CLI coverage
-  in `tests/telemetry/`.
-
-### Documentation
-
-* **ADR:** [`agents/contexts/adr-artifact-engagement.md`](agents/contexts/adr-artifact-engagement.md)
-  — rationale, default-off doctrine, privacy contract, schema
-  versioning, deprecation horizon.
-* **flow:** `docs/contracts/artifact-engagement-flow.md` is the
-  cross-cutting reference for what gets recorded, when, and under
-  which constraints; includes the maintainer hand-audit recipe.
-* **AGENTS.md + README.md:** short *Maintainer telemetry (opt-in)*
-  pointer; consumers see nothing.
-* **`/onboard`:** Step 9 emits a one-screen maintainer-only hint
-  describing the feature; no question, no prompt.
-
-**Context-Aware Command Suggestion** — deterministic, read-only layer
-that surfaces eligible slash commands as numbered options when a
-non-`/`-prefixed user prompt matches their `suggestion.trigger_*`
-frontmatter. **Nothing auto-executes** — the user picks every time;
-the as-is option is always present and always last.
-
-### Features
-
-* **engine:** `scripts/command_suggester/` Python package
-  (`match` / `rank` / `cooldown` / `sanitize` / `render` / `loader`).
-  Heuristic-only scoring (substring + Jaccard token overlap +
-  structural-bonus boosts), per-conversation cooldown keyed on
-  `(command, evidence)`, and a sanitiser that strips fenced + inline
-  code blocks plus the suggester's own previous block shape before
-  scoring.
-* **rule:** `command-suggestion` (always-on) — emits one
-  numbered-options block per turn under the `user-interaction` Iron
-  Law. Subordinate to `scope-control`, `ask-when-uncertain`,
-  `verify-before-complete`, and any active role-mode contract or
-  engine halt; on conflict → silent.
-* **frontmatter:** every command carries
-  `suggestion.eligible` (default `true`) plus a flat
-  `trigger_description` + `trigger_context` pair (linter-validated,
-  ≥ 10 chars each). Locked eligibility table at
-  [`agents/contexts/command-suggestion-eligibility.md`](agents/contexts/command-suggestion-eligibility.md).
-* **settings:** `commands.suggestion` block in
-  `.agent-settings.yml` — `enabled` (global), `blocklist`
-  (per-command), `confidence_floor` (default `0.6`), `cooldown`
-  (default `10m`), `max_options` (default `4`, plus the always-extra
-  as-is option). Per-command overrides via frontmatter.
-* **opt-out paths (3):** global (`enabled: false`), per-command
-  (`blocklist`), per-conversation (`/command-suggestion-off` /
-  `/command-suggestion-on` directives detected by `cooldown.py`).
-* **anti-noise heuristics:** sub-floor suppression, lonely-match
-  guard within `floor + 0.1`, vague-prompt guard (< 6 words +
-  > 2 matches without structural bonus), continuation-phrase guard
-  (`ok`, `weiter`, `continue`, …). Structural bonuses (ticket key,
-  file path) override every suppressor.
-* **tests:** 84 unit cases in `tests/test_command_suggester.py`
-  (matcher, rank, cooldown, sanitiser, render, settings, directive)
-  plus 9 GT-CS goldens in `tests/test_command_suggester_goldens.py`
-  (single-match, tie-break, sub-floor, slash-bypass, as-is pick,
-  cooldown, settings off, clarification-wins, adversarial echo).
-* **no behavioural change to slash invocation:** explicit `/command`
-  bypasses the suggester entirely; per-command halts intact.
-
-### Documentation
-
-* **ADR:** [`docs/contracts/adr-command-suggestion.md`](docs/contracts/adr-command-suggestion.md)
-  — "suggest, never invoke" anchor, eligibility rubric, anti-noise
-  heuristics, hardening list, three opt-out paths.
-* **flow:** [`docs/contracts/command-suggestion-flow.md`](docs/contracts/command-suggestion-flow.md)
-  — scoring breakdown, evidence semantics, subordination order, and
-  hardening tests.
-* **README + AGENTS.md:** short *Context-aware command suggestion*
-  section pointing to the ADR and the flow doc.
-
-**Install-path pruning (relabel only)** — `docs/installation.md`
-reorders install paths by recommendation prominence. Composer + npm
-remain the default; manual / submodule / VS Code Git URL are tagged
-`advanced`; Claude.ai Web Skills UI is tagged `experimental`; Linear
-AI workspace guidance is tagged `staged`. **No installer or shipped
-artefact is removed in 1.15.0** — every path on the page still works
-and is still tested. The labels describe how prominent the path is in
-our recommendation order, not its support status.
-
-### Documentation
-
-* **`docs/installation.md`:** new label table at the top
-  (`(no label)` / `advanced` / `experimental` / `staged`) plus a
-  preamble that calls out the no-removal contract explicitly. Section
-  headers carry their label inline for skim-readers.
-* **rationale:** see R9 in
-  [`agents/roadmaps/archive/road-to-post-pr29-optimize.md`](agents/roadmaps/archive/road-to-post-pr29-optimize.md)
-  — relabelling addresses the "four install paths at zero external
-  users" tension without removing any path that an existing user
-  might rely on.