@event4u/agent-config 9.8.0 → 9.10.0

package/docs/contracts/adr-level-6-productization.md

@@ -1,145 +0,0 @@
----
-stability: stable
----
-
-# ADR — Level-6 Productization Closure
-
-> **Status:** Decided · 2026-05-14
-> **Context:** Closure record for `road-to-productization.md` and its
-> two sibling roadmaps (`road-to-proof-not-features.md`,
-> `road-to-better-skills-and-profiles.md` Block A). PR #43 lifted the
-> package from Level-4 (execution engine) to Level-5 (observable
-> decision system); this roadmap was the Level-5 → Level-6 jump:
-> **steerable + provable + onboardable**.
-> **Cross-links:** sibling roadmaps `road-to-productization`,
-> `road-to-proof-not-features`, and `road-to-better-skills-and-profiles`
-> — see `agents/roadmaps/` (or `agents/roadmaps/archive/` for the
-> closed ones) for current status.
-
-## What shipped
-
-### Decision-Engine steerability (Phase 2)
-
-- [`decision-engine-gates.md`](decision-engine-gates.md) — additive
-  `decision_engine:` block in `.agent-settings.yml` with
-  `min_confidence`, `block_on_risk`, `require_memory_hits`, `on_block`,
-  `ask_timeout_seconds`, `on_block_fallback`. Absent block = unchanged
-  observe-only behaviour.
-- Gate-conflict resolution matrix (P2.1a) + non-TTY timeout fallback
-  (P2.1b) shipped before the gates themselves; the engine refuses to
-  evaluate downstream gates after the first rejection and falls back
-  to `on_block_fallback` in non-interactive contexts.
-- Confidence-band gate (P2.2) and risk-class gate (P2.3) wired into
-  the scoring path. Memory-required policy (P2.4) unblocks on P6.2
-  shipping (`affected` keys in the decision trace).
-
-### UX simplification (Phase 3)
-
-- README "Quickstart" block — install → `/onboard` → `/work "first
-  real task"`, contributor detail moved below the `## For contributors`
-  fold.
-- Default `rule_loading_tier` flipped from `minimal` to `balanced`;
-  rationale in [`cost-profile-defaults.md`](cost-profile-defaults.md).
-- `/onboard` step 11 prints the Quickstart command list inline.
-- CI gate: `task smoke-quickstart` runs the installer into a tmpdir
-  and validates the documented default surface deterministically.
-
-### Multi-stack skill depth (Phase 4)
-
-- `symfony-workflow` skill (~8.6 KB) — DI, Doctrine, Messenger,
-  voters, Twig, console.
-- `nextjs-patterns` skill (~9.9 KB) — App Router, RSC boundaries,
-  Server Actions, caching, route handlers, 14.x↔15.x deltas.
-- README stack table now separates Symfony / Next.js / Zend-Laminas
-  rows; "Deepest reference stack" paragraph names the workflow-grade
-  second tier explicitly.
-
-### Architecture cleanup (Phase 5)
-
-- Auto-rules (`non-destructive-by-default`, `scope-control-policy`)
-  audited: already refactored to trigger + Iron Law + pointer shape;
-  bound by the kernel-budget linter at 4 000-char override ceiling
-  (P5.1).
-- Rule-Interaction matrix marked rule-only by design;
-  [`rule-interactions.md`](rule-interactions.md) § "Out of scope —
-  orchestration surfaces" points at `decision-engine-gates`,
-  `decision-trace-v1`, `memory-visibility-v1`,
-  and the `ai-council` skill for Council × Memory × Work-Engine
-  interactions (P5.2).
-- `type: orchestrator` frontmatter tag exempts cluster routers from
-  the `command_missing_skill_references` linter check; 15 commands
-  carry the tag (P5.3).
-- Beta-review marker protocol shipped in [`STABILITY.md`](STABILITY.md)
-  § Beta-review markers; `scripts/check_beta_review_markers.py` wired
-  into `task ci`; 39 beta contracts back-filled (P5.4).
-- Test-redundancy audit produced the `step-5-test-cleanup` sibling
-  roadmap — audit-only, no deletions (P5.5).
-
-### Release-trunk discipline (Phase 1)
-
-- [`release-trunk-sync.md`](release-trunk-sync.md) protocol; CI gate
-  fails the release-prep branch when `main` is more than one tagged
-  release behind (P1.3).
-
-### Proof + cognition layers (Phases 6 + 7)
-
-- Memory-consequence in the trace: `affected` keys in
-  [`decision-trace-v1.md`](decision-trace-v1.md) (sibling P2.1a–c).
-- README three-audience split (sibling P2.2a–c).
-- Hook doctor (sibling P2.3).
-- Persona spine: Core-tier 5-section + Specialist-tier 7-section
-  spines locked in [`persona-schema.md`](persona-schema.md) (sibling
-  Block A).
-
-## What got cancelled
-
-- **P6.1 — Three real showcase sessions** (sibling P1.1–P1.4).
-  Cancelled upstream — capturing real host-agent sessions requires a
-  hosted-LLM runner that is out of scope for this roadmap. P1.0
-  pre-flight shipped; the capture surface is ready when a runner
-  exists. Reopen as `road-to-showcase-capture.md` once a runner is
-  on the table.
-- **P8.1 — End-to-end Level-6 smoke** — same gating as P6.1.
-  Structural coverage (`task smoke-quickstart` + decision-engine
-  schema validator + gate-evaluator unit tests) covers the
-  configuration surface deterministically; the live smoke remains
-  the manual pre-tag gate.
-
-## What stayed beta
-
-39 contracts carry `keep-beta-until: 2026-08-12` (next audit
-deadline). None met the 30-day promotion floor at audit time.
-First-commit age range: 0–12 days. Audit cap is 90 days from the
-audit date; CI rejects undated betas, multiple markers, and
-keep-beta-until dates beyond the window.
-
-## What got deferred to siblings
-
-- **Showcase capture** → future `road-to-showcase-capture.md` when a
-  hosted-LLM runner is on the table.
-- **Test-suite deletion** → `step-5-test-cleanup` sibling roadmap
-  (audit-only sibling spawned by P5.5; non-destructive by default).
-- **Persona Block B** (Architect / Risk-Officer extension) —
-  anti-recommended per the sibling closure decision; not deferred,
-  closed.
-- **Distribution / adoption** →
-  `road-to-distribution-and-adoption.md`, gated on this roadmap
-  closing (which this ADR records).
-- **MCP server work** — own strand, out of scope.
-
-## Consequences
-
-- **Steerable:** the Decision Engine now gates on configurable
-  thresholds; the configuration surface is documented and CI-tested.
-- **Provable:** memory hits/misses surface as `affected` keys in the
-  decision trace; the trace shape is contract-stable.
-- **Onboardable:** a fresh user can land at a working `/work`
-  invocation in three Quickstart steps without scrolling past the
-  fold.
-- **Multi-stack credible:** Laravel stays the deepest reference;
-  Symfony and Next.js shipped at workflow-grade depth; other stacks
-  remain project-analysis-only with the honest delta language in the
-  README.
-- **Architecturally tidy:** orchestrator commands no longer warn,
-  beta contracts cannot rot undated, and the contract surface itself
-  carries a periodic review obligation.

package/docs/contracts/adr-mcp-runtime.md

@@ -1,128 +0,0 @@
----
-stability: stable
----
-
-# ADR — MCP server runtime: Anthropic `mcp` Python SDK
-
-> **Status:** Decided · 2026-05-10 (recorded 2026-05-15).
-> **Context:**
-> [`mcp-phase-1-scope.md`](mcp-phase-1-scope.md),
-> [`mcp-cloud-scope.md`](mcp-cloud-scope.md).
-
-## Decision
-
-The MCP server at `scripts/mcp_server/` runs on **Python 3.11+** using the
-official Anthropic **`mcp` Python SDK** (PyPI; pinned to `mcp==1.27.1`
-per [`scripts/mcp_server/requirements.txt`](../../src/scripts/mcp_server/requirements.txt)).
-**FastMCP** (the higher-level decorator wrapper) and the **MCP TypeScript SDK**
-are explicitly rejected for this surface.
-
-The hosted Cloudflare Worker bridge (`internal/workers/mcp/`) is the only place a
-non-Python runtime is allowed, and it stays bound to the same wire contract
-(see [`mcp-cloud-scope.md`](mcp-cloud-scope.md)).
-
-## Why this was a real question
-
-The package already ships Python under `scripts/` (work engine, AI Council,
-skill linter, install driver helpers) and ships zero Node-runtime code paths
-outside the npx dispatcher. Picking a runtime for the MCP server had three
-candidates that all could have shipped:
-
-1. **MCP Python SDK** (low-level `Server` + `stdio_server` handlers).
-2. **FastMCP** (higher-level Pythonic decorators built on the same SDK).
-3. **MCP TypeScript SDK** (Node runtime, separate package).
-
-Without an ADR, this choice would have stayed implicit in the code and
-re-litigated every time a contributor read `scripts/mcp_server/server.py`.
-
-## Why MCP Python SDK (low-level) wins
-
-| Criterion | MCP Python SDK | FastMCP | MCP TypeScript SDK |
-|---|---|---|---|
-| Runtime already in repo | ✅ Python is the `scripts/` runtime | ✅ Same | ❌ Adds Node-runtime path for one server |
-| A0 safety boundary fit (read-only `prompts/list`, `prompts/get`, narrow `tools/*` allowlist) | ✅ Direct handler control matches the [Phase 1 scope contract](mcp-phase-1-scope.md) | ⚠️ Decorator sugar can obscure the unimplemented-tool guard | ✅ Possible but duplicates Python helpers |
-| Import-surface guard (`tests/test_mcp_server.py` asserts no `subprocess`, `os.system`, `os.popen`, no HTTP client in `scripts.mcp_server.prompts/tools`) | ✅ Trivial to enforce — one module set to audit | ⚠️ FastMCP pulls in extra deps that widen the audit surface | ❌ Would need a TS-side equivalent |
-| Reuse of existing project helpers (`scripts/skill_linter.py`, `scripts/chat_history.py`) | ✅ Direct in-process call | ✅ Same | ❌ Cross-runtime IPC or duplicated logic |
-| Pin / supply-chain footprint | One pin (`mcp==1.27.1`) + `PyYAML` | Adds FastMCP version coupling on top | Node toolchain (`npm`, `tsc`, `dist/`) |
-| Smoke-test path | `task mcp:setup && task mcp:run` (already shipped) | Would re-wrap the same SDK | Separate test runner |
-
-Evidence the decision is already realised in code:
-
-- [`scripts/mcp_server/server.py`](../../src/scripts/mcp_server/server.py) — uses
-  `mcp.server.Server`, `mcp.server.stdio.stdio_server`, `InitializationOptions`
-  directly (no FastMCP decorators).
-- [`scripts/mcp_server/__init__.py`](../../src/scripts/mcp_server/__init__.py) —
-  pins `__version__` and declares stability/contract pointer.
-- [`scripts/mcp_server/requirements.txt`](../../src/scripts/mcp_server/requirements.txt)
-  — `mcp==1.27.1`, no FastMCP, no Node tooling.
-- [`scripts/mcp_setup.sh`](../../src/scripts/mcp_setup.sh) — onboarding writes
-  the Claude Desktop config snippet against `python -m scripts.mcp_server`.
-
-## Tool surface (Phase 1 scoping)
-
-Locked separately by [`mcp-phase-1-scope.md`](mcp-phase-1-scope.md) Phase 4
-amendment. The current ALLOWLIST is exactly two tools, registered as a
-hardcoded module-level tuple in
-[`scripts/mcp_server/tools.py`](../../src/scripts/mcp_server/tools.py):
-
-| Tool | Mode | Source |
-|---|---|---|
-| `lint_skills` | read-only | wraps `scripts.skill_linter.lint_file` |
-| `chat_history_append` | path-scoped write | wraps `scripts.chat_history.append`; writes restricted to `agents/runtime/.agent-chat-history` (current default), `agents/.agent-chat-history`, or `.agent-chat-history` under the consumer root |
-
-No `push`, `merge`, `commit`, or prod-write surface is exposed. The
-unimplemented-tool envelope from
-[`mcp-tool-stub-envelope.md`](mcp-tool-stub-envelope.md) governs the rest of
-the [`consumer_tool_catalog.json`](../../src/scripts/mcp_server/consumer_tool_catalog.json)
-entries.
-
-`agent-config init`, `agent-config skills list`, and
-`agent-config council estimate` (the speculative tool surface in the
-step-14 stub) are **not** exposed today. They stay terminal-gated because
-their natural shape is a stateful CLI. The AI Council (claude-sonnet-4-5 +
-gpt-4o, 2026-05-10, 2 rounds, $0.06) converged on a hardcoded module-level
-ALLOWLIST with mandatory path-scoping for any write tool, and locked the
-rule that engine-state-bearing operations stay off the MCP wire until a
-real consumer ask justifies amending the A0 boundary.
-
-## Install surface
-
-The step-14 stub speculated about an `agent-config install --mcp` flag.
-That shape was rejected in favour of two existing entrypoints, both
-already shipped:
-
-- **One-liner onboarding:** `task mcp:setup` runs
-  [`scripts/mcp_setup.sh`](../../src/scripts/mcp_setup.sh) — creates
-  `.venv-mcp/`, installs `mcp`, and prints the Claude Desktop JSON snippet
-  the operator pastes into
-  `~/Library/Application Support/Claude/claude_desktop_config.json`
-  (with the per-OS variants documented in
-  [`docs/mcp-server.md`](../mcp-server.md)).
-- **Config writer:** `./agent-config mcp:render --claude-desktop` writes
-  the user-scope Claude Desktop config directly.
-
-Writing the global Claude Desktop config from the npx dispatcher without
-an operator pasting JSON is **not** part of this contract — Claude Desktop
-restarts and config-merge semantics make silent rewrites a footgun. The
-copy-paste path stays the canonical install shape until non-dev recruit
-evidence under `agents/evidence/eval-findings/` demonstrates the manual step is the
-actual adoption blocker.
-
-## Consequences
-
-- Adding a third tool to the MCP server is a code-review event against
-  `ALLOWLIST` in `scripts/mcp_server/tools.py`. No settings flag, no env
-  var, no dynamic registration — see
-  [`mcp-phase-1-scope.md`](mcp-phase-1-scope.md) Phase 4 amendment.
-- Picking up a future protocol break (MCP SDK 2.x) is one pin bump in
-  `scripts/mcp_server/requirements.txt`, gated on the 12 import-surface +
-  behaviour tests in `tests/test_mcp_server.py` staying green.
-- Re-opening FastMCP or the TypeScript SDK requires a new ADR that
-  supersedes this one with evidence (Python SDK shipping a deprecation
-  or FastMCP closing the safety-audit gap on `tools/*`).
-
-## See also
-
-- [`mcp-phase-1-scope.md`](mcp-phase-1-scope.md) — Phase 1–6 hard contract.
-- [`mcp-cloud-scope.md`](mcp-cloud-scope.md) — hosted Worker bridge scope.
-- [`mcp-tool-stub-envelope.md`](mcp-tool-stub-envelope.md) — unimplemented-tool wire shape.

package/docs/contracts/adr-product-ui-track.md

@@ -1,381 +0,0 @@
----
-stability: stable
----
-
-# ADR — Product UI Track: audit-as-hard-gate, design-review loop, stack dispatch
-
-> **Status:** Decided · R3 Phases 1–6 shipped · 2026-05-01
-> **Context:** [`ui-track-flow.md`](ui-track-flow.md)
-> **Builds on:** [`adr-prompt-driven-execution.md`](adr-prompt-driven-execution.md)
-> — R2 envelope routing and the band-action gate that R3 widens to UI.
-> **Defers to:** the visual-review-loop track for headless-browser
-> screenshot capture and visual-regression assertions.
-
-## Decision
-
-R3 ships four directive sets — `backend` (R1/R2), **`ui`**, **`ui-trivial`**,
-and **`mixed`** — dispatched at the engine boundary on
-`state.directive_set`. The slot wiring is fixed by
-[`directives/ui/__init__.py`](../../.agent-src.uncondensed/templates/scripts/work_engine/directives/ui/__init__.py),
-[`directives/ui_trivial/__init__.py`](../../.agent-src.uncondensed/templates/scripts/work_engine/directives/ui_trivial/__init__.py),
-and [`directives/mixed/__init__.py`](../../.agent-src.uncondensed/templates/scripts/work_engine/directives/mixed/__init__.py);
-the contract for each lives in [`ui-track-flow.md`](ui-track-flow.md).
-
-The UI set drives `audit → design → apply → review → polish → report`
-with three load-bearing properties:
-
-1. **Existing-UI-audit is a hard gate.** No `apply` runs without
-   `state.ui_audit` populated. The gate lives at directive level **and**
-   at always-on rule level ([`ui-audit-gate`](../../.agent-src.uncondensed/rules/ui-audit-gate.md))
-   so an agent acting outside the engine cannot bypass it.
-2. **Design brief is locked microcopy.** `apply` consumes the brief
-   verbatim — `PLACEHOLDER_PATTERNS` (`<placeholder>`, `lorem`, `todo:`,
-   `tbd`, `xxx`) reject at producer and consumer.
-3. **Polish has a hard 2-round ceiling.** Schema-level validation
-   (`work_engine.state._validate_ui_polish`) rejects `rounds > 2` on
-   disk; after round 2 the engine halts with ship-as-is / abort / hand-off.
-
-Halt budget on the happy path: **2 user halts** — audit pick + design
-sign-off. Apply / review / polish run silently when their producers
-write clean envelopes.
-
-## Why this was a real question
-
-R1 + R2 produced a competent backend executor. UI work was the gap. The
-package already shipped UI skills (`fe-design`, `blade-ui`, `livewire`,
-`flux`) but they were flat tools: an agent could call `flux` directly,
-write a component that duplicated three existing primitives, ignore the
-project's design tokens, and ship microcopy with `<placeholder>` strings
-intact. Tests passed; the result felt undesigned.
-
-Three options were on the table:
-
-1. **More UI skills, no directive set** — add `react-shadcn-ui`,
-   `existing-ui-audit`, `ui-design-brief` as flat skills and let
-   `/work` route to them via the existing dispatcher. Rejected:
-   the audit step is the load-bearing piece, and a flat skill cannot
-   enforce "no apply before audit". Rules can advise; only the
-   dispatcher can refuse.
-2. **One mega-skill (`product-ui`) that orchestrates internally** —
-   single SKILL.md that runs audit → design → apply → review →
-   polish in one shot. Rejected: the engine's halt-budget,
-   sentinel-based replay, and Golden Transcript suite are designed
-   for slot-level granularity. Folding five steps into one skill
-   blinds the freeze-guard to mid-flow regressions.
-3. **A new directive set per intent (chosen)** — `ui` for build /
-   improve, `ui-trivial` for provably bounded edits, `mixed` for
-   prompts that touch both layers. Adopted because each slot keeps
-   its own sentinel, halt surface, and replay coverage; the audit
-   gate is enforced at the dispatcher boundary and at the agent
-   boundary; and stack-specific implementation is dispatched without
-   widening the directive set.
-
-## Audit as a hard gate (the load-bearing piece)
-
-The Lovable-grade differentiator is **"audit existing UI first, design
-before code, polish before ship"**. Everything in R3 follows from that.
-
-The audit gate refuses `apply` until `state.ui_audit` is well-formed:
-either `≥1 components_found` entry, or `greenfield=True` with a user-chosen
-`greenfield_decision` ∈ `{scaffold, bare, external_reference}`. An empty
-dict, `None`, or a populated dict without those keys is **not** findings;
-the gate emits `@agent-directive: existing-ui-audit` and refuses to advance.
-
-Two enforcement layers, deliberately redundant:
-
-- **Dispatcher layer** —
-  [`directives/ui/audit.py`](../../.agent-src.uncondensed/templates/scripts/work_engine/directives/ui/audit.py)
-  refuses to write `outcomes["refine"] = "success"` without a populated
-  audit. Purely structural; no LLM, no heuristic.
-- **Agent layer** — [`ui-audit-gate`](../../.agent-src.uncondensed/rules/ui-audit-gate.md)
-  is an always-on rule that fires when the agent is about to write a
-  component file outside the engine (free-form edit, side conversation,
-  cloud surface). The rule encodes the same Iron Law in prose so cloud
-  agents that don't ship the engine still honour it.
-
-**Why two layers.** A single layer would leak: cloud surfaces and
-free-form edits bypass the dispatcher entirely. A rule alone would not
-hold under engine-driven runs because rules don't refuse exit codes.
-Belt-and-suspenders is cheap (no shared state, no double-write) and
-the failure modes are different enough to justify it.
-
-### Confidence-path resolution
-
-Audit findings carry a confidence label and per-candidate similarity.
-[`directives/ui/audit.py::_decide_path`](../../.agent-src.uncondensed/templates/scripts/work_engine/directives/ui/audit.py)
-resolves to one of:
-
-- `high_confidence` — confidence `high` + ≥1 match with similarity
-  ≥ `STRONG_SIMILARITY = 0.7` and no runner-up within
-  `TIE_GAP = 0.05`. Audit folds findings into the design brief — no
-  separate halt.
-- `ambiguous` — populated but below the high-confidence threshold or
-  with a tie. Numbered-options halt records `audit_path = "ambiguous"`
-  + `candidate_pick`.
-- `greenfield` — no `components_found`, `greenfield = True`, user picks
-  `scaffold` / `bare` / `external_reference`.
-
-Constants are named, exported, and imported by tests so a re-tune
-re-captures Goldens explicitly rather than drifting silently.
-
-## Design-review polish loop
-
-`apply` writes the rendered envelope. `review` (stack-dispatched) emits
-`findings` + `review_clean`. `polish` runs a bounded fix loop with a
-hard ceiling of `POLISH_CEILING = 2` rounds, validated at three layers
-(in-memory state, on-disk schema, dispatcher).
-
-| `review_clean` | `rounds` | Behaviour |
-|---|---|---|
-| `True` | any | `SUCCESS` — advance to report |
-| `False` | `< 2` | `BLOCKED` + `@agent-directive: ui-polish-<stack>`; agent applies fixes, re-runs review, increments `rounds` |
-| `False` | `== 2` | `BLOCKED` numbered options: ship-as-is / abort / hand off |
-
-**Why ceiling at 2.** Three rounds is one round too many: by round 3
-the agent is either making cosmetic tweaks the user can't notice, or
-spinning on a design problem the engine cannot resolve. Halting at 2
-hands the decision to the user with the cheapest possible context
-(two completed rounds of evidence). Rate of false-ceiling-hits will
-be visible in delivery reports; a tune to 3 stays a one-line constant
-change with a Golden re-capture.
-
-**Token-violation extraction.** Findings with `kind == "token_violation"`
-carry `category` and `value`. Polish classifies against
-`state.ui_audit.design_tokens`:
-
-- Matched value → fix uses the named token.
-- Unmatched value repeated > `TOKEN_REPEAT_THRESHOLD = 2` times →
-  emits `polish_token_extraction_pending` to extract a new token
-  before the next round runs. One-off unmatched values stay inline.
-
-This is the only place R3 mutates the design system. It runs only
-after a finding identifies a violation, and only when the same value
-appears three or more times — single-use values are not worth the
-ceremony.
-
-## Halt budget — why max 2 on the happy path
-
-The Lovable feel comes from "user decides once, agent runs". Three
-halts (audit + design + polish-final) is the obvious shape but
-empirically wrong: by the time polish needs sign-off, the user is
-context-switched away. Two halts that pin the **decisive** choices
-(which existing UI to extend, what microcopy to ship) buy the agent
-the runway to finish silently.
-
-Additional halts surface only on real ambiguity:
-greenfield-undecided, shadcn-version-mismatch, audit-ambiguous,
-placeholder rejection, polish round (per dirty review, capped at 2),
-polish ceiling. GT-U11 (high-confidence) and GT-U12 (ambiguous) pin
-the budget at 1 and 2 halts respectively; a regression that adds a
-third halt fails replay.
-
-## Trivial path and reclassification
-
-For provably bounded edits (single class swap, copy tweak, one-prop
-adjustment), the Phase-1 intent classifier writes
-`directive_set = "ui-trivial"`. The slot wiring collapses to
-`refine → ⊘ → ⊘ → ⊘ → apply → test → ⊘ → report` with
-`MAX_FILES = 1` and `MAX_LINES_CHANGED = 5` enforced inside
-[`directives/ui_trivial/apply.py`](../../.agent-src.uncondensed/templates/scripts/work_engine/directives/ui_trivial/apply.py).
-
-**Mandatory reclassification at apply time.** When a trivial edit
-exceeds the preconditions, apply flips
-`state.directive_set = "ui"` and the dispatcher restarts at audit.
-The reclassification is loud (delivery report records it) and
-counted (delivery report tracks the rate). Silent skips are the
-failure mode to watch for.
-
-**Why the bypass exists.** Without it, "fix the typo on the login
-button" runs the full audit + design loop. That is six halts the
-user does not want to take. The bypass restores common sense without
-weakening the audit gate — the gate stays in force whenever the
-preconditions don't hold.
-
-## Stack detection and dispatch
-
-[`scripts/work_engine/stack/detect.py`](../../.agent-src.uncondensed/templates/scripts/work_engine/stack/detect.py)
-reads `composer.json` and `package.json` once, applies a four-rule
-priority table, and writes `state.stack.frontend`:
-
-1. `livewire/livewire` + `livewire/flux` → `blade-livewire-flux`
-2. `react` + (`@radix-ui/*` OR `shadcn-ui` OR `components.json`) → `react-shadcn`
-3. `vue` (any major) → `vue`
-4. otherwise → `plain`
-
-Cached on `state.stack` against the manifest `mtime`. A `composer.json`
-or `package.json` change invalidates the cache; same conversation
-re-detects on the next dispatch.
-
-Errors (missing file, malformed JSON) downgrade to `plain` rather than
-raising — a wrong stack label is recoverable (audit catches it, user
-can override), a crash mid-dispatch is not.
-
-`apply`, `review`, `polish` each route on `state.stack.frontend` to a
-stack-specific directive (`ui-apply-blade-livewire-flux`,
-`ui-apply-react-shadcn`, `ui-apply-vue`, `ui-apply-plain`). The
-dispatch table shape is identical across all three slots; adding a
-stack adds three rows, one per slot.
-
-## fe-design migration — reference, not executor
-
-`fe-design` previously functioned as both heuristics catalogue and
-direct executor. R3 splits the responsibilities:
-
-- **Reference (kept):** layout patterns, form / table design,
-  responsive strategy, a11y heuristics. Cited by
-  [`directives/ui/design.py`](../../.agent-src.uncondensed/templates/scripts/work_engine/directives/ui/design.py).
-- **Executor (removed):** code-writing responsibilities migrated to
-  the stack-specific apply / review / polish skills (`flux`,
-  `livewire`, `blade-ui`, `react-shadcn-ui`, `ui-apply-vue`).
-
-The split lets the design step stay framework-agnostic while
-implementation skills stay focused on a single stack's idioms.
-
-## Mixed orchestration — contract first, then UI, then stitch
-
-When a single input touches both layers, `mixed` runs
-`refine → memory → analyze → contract → ui → stitch → verify → report`.
-Sentinels:
-
-- `state.contract.contract_confirmed` — UI sub-flow refuses to start
-  without it (defense-in-depth even if `outcomes["plan"] == "success"`).
-- `state.ui_review.review_clean` — mixed `ui` step's success condition.
-- `state.stitch.verdict = "success"` — stitch's success condition;
-  `blocked` / `partial` halts unless `state.stitch.integration_confirmed`
-  flips.
-
-**Why contract first.** UI shape and microcopy depend on the API
-surface (field names, error codes, paginated vs streamed). Building
-UI before locking the contract leads to two rounds of churn — once
-when the contract surfaces a constraint the design ignored, once
-when the integration test reveals a field-name mismatch. Locking
-the contract first costs one halt and saves an entire polish round.
-
-## Tradeoffs accepted
-
-- **Three new directive sets** — bigger surface to maintain, more
-  Goldens (12 GT-U entries vs. R2's 4 GT-P). Mitigated by shared
-  refine/report handlers and the strict-verb replay harness.
-- **Hard 2-round polish ceiling** — some legitimate edge cases
-  (a11y compliance edits across multiple components) need 3 rounds.
-  Mitigated by ship-as-is / hand-off as explicit user choices at
-  the ceiling halt.
-- **Stack detection is heuristic, not full-AST** — a project with
-  `react` + `flux` (e.g. test scaffold) misclassifies as
-  `react-shadcn`. Mitigated by user-override path: explicit `intent`
-  in prompt overrides detection.
-- **Two-layer audit enforcement** — duplicates the gate logic in
-  rule prose and dispatcher Python. Accepted: cloud surfaces don't
-  ship the engine, and engine-driven flows can't rely on rule
-  enforcement. Both layers are needed; the cost is documentation.
-
-## Non-goals
-
-- Does **not** replace `/implement-ticket` or `/work` — same engine,
-  new directive sets only.
-- Does **not** introduce a `/build` or `/build-screen` entrypoint.
-  Intent classification at refine is enough; new commands wait on
-  evidence of mis-routing.
-- Does **not** pin a removal date for `fe-design`; the reference
-  positioning is stable.
-- Does **not** ship visual review (headless browser, screenshot
-  capture, a11y tooling) — Roadmap 4 stub.
-- Does **not** ship project-local design-memory beyond what the
-  audit step already extracts; the audit is the memory.
-
-## Consequences — unblocks
-
-- **Roadmap 4** (visual review loop) can register a post-polish
-  visual-assertion step against the existing `state.ui_apply`
-  envelope.
-- **New stacks** (Svelte, SolidJS, Astro) plug in via the
-  extension recipe ([`ui-stack-extension.md`](ui-stack-extension.md)) —
-  detector heuristic + apply/review/polish skills + Golden fixture.
-  No engine change.
-- **Mixed entrypoints** (`/refactor`, `/spike`, future) compose by
-  building a new envelope kind plus a new directive set; the audit
-  gate is reusable when the work touches UI.
-
-## Follow-ups (not part of this ADR)
-
-- Track reclassification rate (`ui-trivial` → `ui` flips) in
-  delivery reports; if > 30 % of trivial runs reclassify, the
-  classifier needs tuning.
-- Track polish-ceiling-hit rate; if > 10 % of UI runs hit the
-  ceiling, raise to 3 with a Golden re-capture.
-- Decide whether `state.ui_audit` should expire on long-running
-  state files (currently sticky for the lifetime of the file).
-- Evaluate whether `react-shadcn` should split into
-  `react-shadcn-v2` / `react-shadcn-v3` once shadcn ships its next
-  major; the audit version-mismatch warning surfaces this signal
-  without forcing a split today.
-
-## R4 amendment — Visual Review Loop (2026-05-01)
-
-> **Status:** Decided · R4 Phases 0–4 shipped · Phase 5 in progress
-
-R4 narrows the polish-termination contract from "subjective ceiling
-only" to "subjective ceiling **plus** objective a11y block", adds a
-preview envelope on `state.ui_review.preview` (engine reads, skill
-writes — the engine never spawns a browser), and threads an a11y
-baseline through the existing audit step so pre-existing violations
-stay informational while new ones block.
-
-### Acceptance criteria — locked verbatim from the roadmap
-
-- **AC #1 — Objective polish anchoring:** Polish loop terminates when
-  (a) `findings` is empty OR (b) `rounds == POLISH_CEILING` AND no
-  `a11y_violation` entries remain at severity ≥ floor. Round 2 with
-  remaining a11y findings halts via `polish_a11y_blocking`, not via
-  `polish_ceiling_reached`.
-- **AC #2 — Preview envelope:** `state.ui_review.preview` shape
-  validated by engine; `render_ok: False` halts via
-  `preview_render_failed`; trivial path bypasses the envelope
-  entirely.
-- **AC #3 — A11y baseline:** `state.ui_audit.a11y_baseline` is read
-  by the review gate; only NEW/CHANGED violations are actionable.
-  Pre-existing violations stay in findings as informational, never
-  block polish.
-- **AC #4 — Goldens pinned:** GT-U13, GT-U14, GT-U15 captured and
-  replay-byte-equal. GT-U4 still passes (regression guard for the
-  narrowed `polish_ceiling_reached`).
-- **AC #5 — `task ci` green** with all three new baselines wired
-  into the harness and CHECKSUMS regenerated.
-- **AC #6 — Contracts updated:** `ui-track-flow.md` and
-  `adr-product-ui-track.md` reflect the new gates and the
-  polish-termination rewrite.
-
-### Termination rewrite — load-bearing change
-
-The pre-R4 polish step had one ceiling halt
-(`polish_ceiling_reached`) with three options: ship as-is / abort /
-hand off. R4 keeps that branch for **subjective** findings (visual
-polish, design tweaks) and adds an explicit **objective** branch
-when remaining findings include `a11y_violation` entries:
-
-- `polish_a11y_blocking` halt with options:
-  1. **Extend** — sets `state.ui_polish.extension_used = True`; the
-     schema validator widens `rounds` from `[0, 2]` to `[0, 3]` only
-     while the flag is set. Once spent, the Extend option disappears
-     on re-entry.
-  2. **Accept** — appends rule ids to
-     `state.ui_review.a11y.accepted_violations`; on replay the review
-     gate's `_apply_a11y_gate` filters those rule ids before
-     synthesising findings, so the run round-trips through `SUCCESS`
-     idempotently.
-  3. **Abort** — drops the UI request.
-
-The two halts share the polish step but never fire together: the
-explicit branch in `polish.run()` checks for `a11y_violation` entries
-before falling through to the subjective ceiling.
-
-### Iron law
-
-The engine **never** renders. Playwright integration lives in the
-stack-specific review skills. The engine reads
-`state.ui_review.preview`; the skill writes it. `render_ok: False`
-halts the user; `skipped: True` bypasses the gate idempotently;
-`render_ok: True` with `screenshot_path` set threads the path into
-the delivery report's `artifacts` list. This boundary keeps the
-package Python + Bash and pushes the browser dependency into
-consumer-project setup.