its-magic 0.1.2-48 → 0.1.2-55

package/README.md

@@ -2,38 +2,78 @@
 
 [GitHub Repository](https://github.com/fl0wm0ti0n/its-magic)
 
-Happy coding! Build something awesome.
-
-Drop-in template repo that implements a structured its-magic workflow in Cursor:
-intake -> discovery -> architecture -> sprint plan -> execute -> QA -> release,
-with pause/resume, decision gates, and persistent artifacts.
+You bring the idea; its-magic is your structured **AI dev team** in Cursor — PO, Tech
+Lead, Dev, QA, Release, and Curator — that turns ideas into shipped software through
+explicit phases and handoff artifacts.
+
+State lives in repo files (`docs/product`, `handoffs`, `sprints`, `decisions`) — not
+chat-only memory. Run `/intake` with your idea, then follow intake → discovery →
+architecture → sprint plan → execute → QA → release; pause/resume and decision gates
+keep you in control when you want to steer. Implementers: see `docs/developer/README.md`
+for the DEV shard.
+
+When you want hands-off delivery, enable **`AUTO_FLOW_MODE=full_autonomy`**
+(default-off), run **`/auto` once in Cursor**, and let the native in-chat auto-chain
+drain your backlog — self-verify UAT, bounded block retry, and advance to the next OPEN
+story or bug without re-invoking each phase manually. The outer driver is **optional**
+(**fallback** for headless/CI or when native chain is unavailable). Guided and
+decision-gated modes remain the default.
 
 ## Features (what its-magic can do)
 
-- Structured phase workflow with explicit artifacts.
-- Artifact-first memory (state in files, not chat only).
-- Decision gate + escalation (`decisions/DEC-xxxx.md`).
-- Pause/resume with checkpoints (`handoffs/resume_brief.md`).
-- Automated execute/QA loop with safety caps (optional).
-- 3-layer quality chain: AI loop → local validate-and-push → CI auto-fix.
-- User-visible metadata guard for operator-facing scripts/CLI (`US-0071` / `DEC-0053`):
-  `python scripts/check-user-visible-metadata.py` (see `docs/engineering/runbook.md`).
-- CI/CD templates driven by `docs/engineering/runbook.md`.
-- Team-friendly local overrides (`scratchpad.local.md`).
-- Optional remote/docker execution and autonomous installs.
-- Built-in benchmarks (live, prompted, headless).
-- Multiplatform distribution (npm, Chocolatey, Homebrew).
+### Autonomous AI workflow
+
+- Run `/intake` through `/release` with explicit phase handoffs and fresh subagent contexts.
+- Use `/pause` and `/resume` with checkpoints when you want to steer; escalate blocking
+  choices to `decisions/DEC-xxxx.md`.
+- Enable **`AUTO_FLOW_MODE=full_autonomy`** (default-off), run **`/auto` once in Cursor**,
+  and drain backlog in-chat; outer driver is **optional** / **fallback** for headless/CI.
+- Team mode routes work across PO, Tech Lead, Dev, QA, Release, and Curator roles.
+- Backlog and bug drain advance OPEN items without re-invoking each phase manually.
+- See the catalog in **Commands and workflow** for phase commands and orchestration details.
+
+### Quality & verification gates
+
+- 3-layer quality chain: AI execute/QA loop → local `validate-and-push` → CI auto-fix.
+- Phase gates include `/plan-verify`, `/qa`, `/verify-work`, and `/uat` with fail-closed stops.
+- `/acceptance` blocks README ↔ backlog drift; user-visible metadata guard on operator scripts.
+- Browser UAT probes with structured fallback when live browser checks are unavailable.
+- Release gates enforce coverage, parity, and evidence before publish.
+- See the catalog in **Features** (`/acceptance`) and **Commands and workflow** for gate commands.
+
+### Distribution & install
+
+- Global install via npm, npx, Chocolatey, or Homebrew; apply to any repo with
+  `its-magic --target`.
+- Modes: `missing` (safe merge), `overwrite` (+ `--backup`), `upgrade` (framework only),
+  and `--clean-repo`.
+- Lifecycle QA matrix validates fresh install, upgrade, backup, and clean-repo paths.
+- Multi-target release publish with confirmation gates for npm/choco/brew.
+- See the **Feature coverage catalog** below for distribution-tagged items.
+
+### Operator control & ergonomics
+
+- Scratchpad flags and `scratchpad.local.md` tune behavior without rewriting framework files.
+- Guided intake packs structure your first `/intake` conversation.
+- Caveman voice mode and optional input compression for terse operator UX.
+- `TOKEN_PROFILE` cost profiles slim context packs without changing workflow semantics.
+- Voice input shortcuts and permissions/runtime connectivity for remote execution.
+- See the catalog in **Other useful capabilities** for scratchpad and governance flags.
 
 <!-- readme-feature-coverage-catalog -->
 
 ### Feature coverage catalog (US-0091)
 
+- `/acceptance` — README ↔ backlog/acceptance feature coverage backfill + blocking drift gate (`US-0091`).
+- `README.md` — Visionary intro + tiered feature hierarchy (autonomous AI dev team positioning, root/template parity) (`US-0094`).
+- `/auto` — Native in-chat auto-chain + full-autonomy mode (`US-0095`, `US-0092`).
 - `/bin` — POSIX npm installer + Linux remote test targets (WSL / SSH / Docker) (`US-0084`).
 - `/choco` — Configurable Multi-Target Release Publish with Confirmation Gate (`US-0054`).
 - `/devops` — First-Class Bug Issue Workflow (Open/Closed) (`US-0079`).
 - `/engineering` — Agent-Driven Codebase Map Bootstrap (`US-0082`).
 - `/engineering` — Remote Runtime Connectivity Contract for QA/Release/Publish (`US-0064`).
 - `/install` — Template/install payload omits intake gate scripts (`BUG-0001`).
+- `/installed` — its-magic ships its OWN packaging CI into generated repos, breaking CI in every created project (`BUG-0009`).
 - `/intake` — Optional Caveman-style input compression (safe file scope) (`US-0090`).
 - `/lint` — CI/CD Workflows (`US-0007`).
 - `/or` — Cursor Caveman mode (scratchpad-configurable terse responses) (`US-0089`).
@@ -1140,6 +1180,7 @@ workflow stops and points you to `scripts/validate-and-push` for local fixing.
 - `/acceptance` — Mandatory Intake Question Packs for First and Small Intakes (`US-0068`).
 - `/ask` — /ask Command: Context-Aware Questions Without Workflow (`US-0020`).
 - `/ask` — Context Compaction and Tiered Token-Cost Optimization Mode (`US-0053`).
+- `/auto` — Architecture triad archiver ignores `## US-xxxx` headings, blocking `/auto` with `STATE_ARCHIVE_BOUNDARY_AMBIGUOUS` (`BUG-0010`).
 - `/auto` — Configurable Auto Phase Selection Policy (`US-0070`).
 - `/auto` — Continuous `/auto` Backlog-Drain Mode with Fine-Tune Switches (`US-0044`).
 - `/auto` — Fresh Subagent Context Per Phase and /auto Orchestration (`US-0023`).
@@ -1184,6 +1225,7 @@ workflow stops and points you to `scripts/validate-and-push` for local fixing.
 - `/repetitive` — Delegable Intake Clarification Without Hard Blocks (`US-0083`).
 - `/research` — Knowledge Curation & Early Research (`US-0029`).
 - `/risk` — Intelligent Intake Decomposition and Risk-Aware PO Questioning (`US-0051`).
+- `/scratchpad` — Caveman mode missing voice compression rules (US-0089 incomplete delivery) (`BUG-0011`).
 - `/scratchpad` — Executable Scratchpad-Driven Sync and Auto-Push Wiring (`US-0076`).
 - `/scratchpad` — Scratchpad Delivery Simplification (Example-Only Install Policy) (`US-0073`).
 - `/scratchpad` — Upgrade Scratchpad Example–First Refresh (Fix Example Drift vs Materialized Baseline) (`US-0075`).
@@ -1191,6 +1233,7 @@ workflow stops and points you to `scripts/validate-and-push` for local fixing.
 - `/sprint-plan` — Explicit `/sprint-plan --bulk` Mode (`US-0046`).
 - `/sprint-plan` — Sprint Sizing Rules and Configurable Sprint Planning (`US-0022`).
 - `/story` — Optional Feature User Guide Generation (`US-0032`).
+- `/uat` — Cursor browser-integrated UAT self-test (browser_smoke + automatable manual UI) (`US-0093`).
 - `/uat` — UAT Artifact Lifecycle and Ownership (`US-0027`).
 - `SKILL` scratchpad flag — Skill and Templates (`US-0004`).
 - `US-0001` scratchpad flag — Core Workflow Commands (`US-0001`).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "its-magic",
-  "version": "0.1.2-48",
+  "version": "0.1.2-55",
   "description": "its-magic - AI dev team workflow for Cursor.",
   "license": "MIT",
   "bin": {

package/scripts/check_intake_template_parity.py

@@ -7,6 +7,8 @@ Scoped modes (DEC-0073 §10 / US-0090):
   --scope=readme-feature-coverage DEC-0074 README feature-coverage pair table.
   --scope=downstream-ci-guard   DEC-0075 downstream CI guard script pair table.
   --scope=us-0092               DEC-0078 full-autonomy outer driver + probe surfaces.
+  --scope=us-0093               DEC-0079 browser UAT probe surfaces.
+  --scope=us-0095               DEC-0080 native in-chat auto-chain surfaces.
   --scope=all              union of all tables.
 """
 
@@ -78,6 +80,40 @@ US0092_PAIRS: tuple[tuple[str, str], ...] = (
     ("docs/engineering/runbook.md", "template/docs/engineering/runbook.md"),
 )
 
+US0093_PAIRS: tuple[tuple[str, str], ...] = (
+    ("scripts/uat_probe_lib.py", "template/scripts/uat_probe_lib.py"),
+    (".cursor/commands/verify-work.md", "template/.cursor/commands/verify-work.md"),
+    (".cursor/commands/qa.md", "template/.cursor/commands/qa.md"),
+    (".cursor/commands/execute.md", "template/.cursor/commands/execute.md"),
+    (
+        ".cursor/scratchpad.local.example.md",
+        "template/.cursor/scratchpad.local.example.md",
+    ),
+    ("docs/engineering/runbook.md", "template/docs/engineering/runbook.md"),
+    (
+        "docs/engineering/auto-orchestration-reference.md",
+        "template/docs/engineering/auto-orchestration-reference.md",
+    ),
+    (
+        "docs/engineering/context/installer-owned-paths.manifest",
+        "template/docs/engineering/context/installer-owned-paths.manifest",
+    ),
+)
+
+US0095_PAIRS: tuple[tuple[str, str], ...] = (
+    (".cursor/commands/auto.md", "template/.cursor/commands/auto.md"),
+    (
+        "docs/engineering/auto-orchestration-reference.md",
+        "template/docs/engineering/auto-orchestration-reference.md",
+    ),
+    ("docs/engineering/runbook.md", "template/docs/engineering/runbook.md"),
+    ("README.md", "template/README.md"),
+    (
+        "scripts/check_intake_template_parity.py",
+        "template/scripts/check_intake_template_parity.py",
+    ),
+)
+
 DOWNSTREAM_CI_GUARD_PAIRS: tuple[tuple[str, str], ...] = (
     (
         "scripts/check_downstream_ci_guard.py",
@@ -95,12 +131,16 @@ SCOPES: dict[str, tuple[tuple[str, str], ...]] = {
     "readme-feature-coverage": README_FEATURE_COVERAGE_PAIRS,
     "downstream-ci-guard": DOWNSTREAM_CI_GUARD_PAIRS,
     "us-0092": US0092_PAIRS,
+    "us-0093": US0093_PAIRS,
+    "us-0095": US0095_PAIRS,
     "all": (
         INTAKE_TEMPLATE_PAIRS
         + CAVEMAN_COMPRESS_PAIRS
         + README_FEATURE_COVERAGE_PAIRS
         + DOWNSTREAM_CI_GUARD_PAIRS
         + US0092_PAIRS
+        + US0093_PAIRS
+        + US0095_PAIRS
     ),
 }
 

package/template/.cursor/commands/auto.md

@@ -71,19 +71,95 @@ Operators must follow the runbook recipe
 `stop_reason` vocabulary: `completed`, `decision_gate`, `missing_input`,
 `pause_request`, `loop_max`, `error`, `blocked`.
 
+## Native in-chat auto-chain (US-0095 / DEC-0080)
+
+When **`AUTO_FLOW_MODE=full_autonomy`** runs in **Cursor IDE**, the orchestrator
+**self-chains in-chat** across intersected lifecycle phases and backlog-drain
+segment boundaries via a **foreground sequential** Task/subagent loop in the
+**same /auto orchestrator session** — without mandatory outer driver or manual
+re-invocation between segments.
+
+### Activation gate
+
+| # | Condition |
+|---|-----------|
+| 1 | Merged scratchpad **`AUTO_FLOW_MODE=full_autonomy`** (exact literal) |
+| 2 | Invocation context = **Cursor IDE** (default Agent panel `/auto` without `--invoke-cmd`) |
+| 3 | Task tool available for foreground subagent spawn |
+
+Set **`native_chain_active=true`** in `state.md` phase boundary when all hold.
+
+### Continuation loop (reference Step 5 — IDE primary)
+
+1. Resolve next `phase_id` from intersected schedule or drain-advance target.
+2. **US-0069** preflight (role matrix + capability gate).
+3. **Spawn fresh subagent** (Task tool, foreground — blocks until done).
+4. Verify isolation evidence + **DEC-0038** strict-proof tuple in `state.md`.
+5. Increment **`outer_cycle_index`**; check **`AUTO_LOOP_MAX_CYCLES`**.
+6. Branch stop matrix → continue spawn, drain-advance, block-retry, or hard stop.
+
+**Loop invariants** (spawn-only — **BUG-0006** unchanged):
+
+1. Orchestrator **must not** stop after one phase or one story segment solely due to Cursor turn boundaries when continuation is schedulable.
+2. Each phase completes only via **fresh subagent spawn** + artifacts — orchestrator **must not** execute phase-role work in-band (**`AUTO_ORCHESTRATOR_PHASE_EXECUTION`** forbidden).
+3. **`stop_reason=completed (segment exhausted)`** is **invalid** when next phase, drain target, or relaxable retry is schedulable.
+
+Preflight/post checks per **US-0069** / **DEC-0051** at every boundary.
+
+### Fail-closed: `NATIVE_CHAIN_UNAVAILABLE`
+
+Emit when Task tool denied, spawn depth limit hit, or IDE context cannot schedule foreground subagent. Hard stop for native path. Optional fallback hint only: `python scripts/auto_outer_driver.py --repo .` (**optional** / **fallback** for headless/CI). **Non-suppressible** under **`AUTO_QUIET=1`**.
+
+### IDE drain-advance-without-pause
+
+Deterministic **7-step** algorithm when **`full_autonomy`** + drain policy active (**`AUTO_BACKLOG_DRAIN=1`** or bug-queue per **US-0087** mutex).
+
+**Trigger** (all required): `stop_phase=refresh-context`; `stop_reason=completed`; drain enabled; budget remaining.
+
+| Step | Action |
+|------|--------|
+| **1** | **READ** latest phase-boundary block in `docs/engineering/state.md` |
+| **2** | **ASSERT** **DEC-0069** pairing — completed phase refreshed **`resume_brief`** + **`state.md`**; stale → **`RESUME_BRIEF_STALE`** (fail-closed, no advance) |
+| **3** | **SELECT** next work item (story or bug per drain mutex) |
+| **4** | **RELOAD** scratchpad; **MATERIALIZE** `resolved_phase_plan` (**US-0070**) |
+| **5** | **PREPEND** `handoffs/resume_brief.md` with segment pointers |
+| **6** | **APPEND** `state.md` materialization breadcrumb for new segment |
+| **7** | **IMMEDIATELY** spawn first phase subagent — **without operator re-`/auto`**, **no** mandatory outer-driver instruction |
+
+**DEC-0069 pairing mandate**: every phase boundary and drain advance **must** refresh **`resume_brief`** + **`state.md`** before scheduling in-chat continuation. Stale brief → **`RESUME_BRIEF_STALE`** fail-closed (no advance).
+
+### Native-chain stop matrix (US-0095)
+
+Native chain **does not weaken** **DEC-0078** hard gates. Hard stops (no relaxation): **`decision_gate`**, isolation/strict-proof violations, security deny, **`BACKLOG_MAX_STORIES_REACHED`**, **`AUTO_LOOP_MAX_CYCLES`**, unrecoverable **`error`**, **`pause_request`**. Relaxable transient stops per **DEC-0078** when configured.
+
+### `AUTO_QUIET` under native chain (US-0095)
+
+| Event | `AUTO_QUIET=0` | `AUTO_QUIET=1` |
+|-------|----------------|----------------|
+| Routine phase PASS | May notify | **Suppress** |
+| In-chat phase continuation | Compact breadcrumb OK | **Suppress** |
+| Drain advance | Segment notify OK | **Suppress** routine prose; **no** outer-driver wait |
+| Gates, caps, errors, **`NATIVE_CHAIN_UNAVAILABLE`** | **Always** | **Always** |
+
+**Forbidden** in IDE-primary `full_autonomy` prose: mandatory `run the outer driver`; `re-run /auto` between drain segments; `segment exhausted` as terminal when continuation pending; unqualified `python scripts/auto_outer_driver.py`.
+
+Full detail: **`docs/engineering/auto-orchestration-reference.md`**.
+
 ## Full-autonomy mode + outer driver (US-0092 / DEC-0078)
 
-**`AUTO_FLOW_MODE=full_autonomy`** (exact literal, default-off) enables the shipped
-stdlib outer driver **`scripts/auto_outer_driver.py`**. The driver **loops hook
-invocations** — spawn-only preserved (**BUG-0006**); it never performs phase-role work.
+**`AUTO_FLOW_MODE=full_autonomy`** (exact literal, default-off) enables hands-off
+orchestration. **IDE primary path** (US-0095): run **`/auto` once in Cursor** —
+native in-chat auto-chain above. **Optional fallback**: stdlib outer driver
+**`scripts/auto_outer_driver.py`** for headless/CI or when **`NATIVE_CHAIN_UNAVAILABLE`**.
 
-Operator recipe: set scratchpad keys → run
+The driver **loops hook invocations** — spawn-only preserved (**BUG-0006**); it
+never performs phase-role work. Headless/CI recipe: set scratchpad keys → run
 `python scripts/auto_outer_driver.py --repo .` once → interpret exit table in
-**`docs/engineering/runbook.md`** § **Full-autonomy outer driver (US-0092)**.
+**`docs/engineering/runbook.md`** § **Full-autonomy outer driver (US-0092)** (**fallback**).
 
 **Drain-advance-without-pause**: with **`full_autonomy`** + **`AUTO_BACKLOG_DRAIN=1`**
 (or bug-queue policy), segment completion schedules the next OPEN story/bug
-**immediately** without operator re-`/auto`; **`resume_brief`** +
+**immediately** **without operator re-`/auto`**; **`resume_brief`** +
 **`state.md`** refresh per **DEC-0069** at every boundary.
 
 ### Full-autonomy stop matrix (US-0092)

package/template/.cursor/commands/execute.md

@@ -48,6 +48,32 @@ as stale isolation evidence).
 
 Release gate semantics (US-0039): mandatory gates (check-in test, QA, UAT) and no-bypass/override contract are enforced at `/release`; see `.cursor/commands/release.md` and `.cursor/commands/qa.md`.
 
+### Browser UAT self-test (US-0093)
+
+When **`UAT_BROWSER_PROBE_MODE=cursor`** (default) and acceptance steps classify as **`browser_smoke`**
+(or automatable **`manual_operator`** re-routed per **DEC-0079** §4), execute the **Cursor browser MCP**
+sequence during `/execute` when web UI verification applies (**BUG-0006**: **`scripts/uat_probe_lib.py`**
+never invokes MCP — subagent only):
+
+1. **Resolve target URL** — `docs/engineering/runtime-connectivity.md` first `http(s)://`; else
+   `package.json` dev/start + **`DEV_SERVER_PORT`** scratchpad override.
+2. **`browser_navigate`** — load URL; respect enterprise origin allowlist.
+3. **Step plan** — map automatable verbs to **`browser_click`** / **`browser_type`** / **`browser_scroll`**;
+   **never** fill password/credential fields; **never** read **`.env`** paths.
+4. **`browser_screenshot`** — write to **`sprints/Sxxxx/evidence/browser/<probe_id>-<seq>.png`** (max **5**).
+5. **Console + network evidence** — counts + summary path refs only (no inline secrets).
+6. **Verdict** — set **`passed`**, **`reason_code`**, **`browser_evidence_refs`** on the matching
+   **`probe_results[]`** row in **`uat.json`**. **`passed=true`** in **`cursor`** mode requires
+   non-empty **`navigation_url`** + at least one screenshot or console/network summary path — else
+   downgrade to **`UAT_BROWSER_PROBE_FAILED`**.
+7. **Fallback trigger** — MCP unavailable → record **`UAT_BROWSER_UNAVAILABLE`**; when
+   **`UAT_BROWSER_FALLBACK_CHAIN=1`**, stdlib HTTP/Playwright fallback runs via **`uat_probe_lib.py`**.
+   Both fail → **`UAT_BROWSER_PROBE_FAILED`** (no silent PASS).
+
+Optional validation: **`python scripts/uat_probe_lib.py --merge-result <fragment.json>`** —
+evidence-required-on-PASS in **`cursor`** mode. No auto-read **`.env`**, no credential auto-fill,
+no intake evidence mutation.
+
 ## Intake evidence tooling reference (US-0078 / DEC-0060)
 
 Stories that harden intake persistence ship **`scripts/intake_evidence_lib.py`**,
@@ -77,6 +103,29 @@ parity for listed paths: **`python scripts/check_token_cost_parity.py --repo .`*
 - `/execute` must not start/continue implementation solely based on
   non-canonical status evidence.
 
+## Browser UAT self-test (US-0093)
+
+When acceptance steps classify as **`browser_smoke`** (or automatable **`manual_operator`**
+reclassified per **DEC-0079** §4) and **`UAT_BROWSER_PROBE_MODE=cursor`**, execute the **Cursor
+browser MCP** sequence during execute-phase verification — **lib never calls MCP directly**
+(**BUG-0006**):
+
+1. **Resolve target URL** — `docs/engineering/runtime-connectivity.md` first `http(s)://`; else
+   dev-server port from `package.json` + scratchpad **`DEV_SERVER_PORT`**.
+2. **`browser_navigate`** — respect origin allowlist.
+3. **Step plan** — **`browser_click`** / **`browser_type`** / **`browser_scroll`** when
+   automatable; **never** fill password/credential fields or read **`.env`**.
+4. **`browser_screenshot`** → **`sprints/Sxxxx/evidence/browser/<probe_id>-<seq>.png`** (max **5**).
+5. Console + network summary path refs only (no inline secrets).
+6. Write **`browser_evidence_refs`** to **`uat.json`** `probe_results[]`; mirror
+   **Runtime browser evidence** in **`qa-findings.md`**.
+7. MCP unavailable → **`UAT_BROWSER_UNAVAILABLE`** + stdlib fallback when
+   **`UAT_BROWSER_FALLBACK_CHAIN=1`**; validate with
+   **`python scripts/uat_probe_lib.py --merge-result <fragment.json>`**.
+
+Reason codes: **`UAT_BROWSER_UNAVAILABLE`**, **`UAT_BROWSER_PROBE_FAILED`**,
+**`UAT_BROWSER_PROBE_TIMEOUT`**. **No silent PASS** without agent evidence in **`cursor`** mode.
+
 ## Steps
 1. Implement one task at a time.
 2. Update summary and engineering state.

package/template/.cursor/commands/qa.md

@@ -56,6 +56,35 @@ acceptance steps, execute probes where stack profile resolves, record evidence i
 Fail closed with **`UAT_PROBE_UNRESOLVED`** (not PASS) when no probe maps.
 Forbidden: auto-read **`.env`**, mutate intake evidence — **`UAT_PROBE_FORBIDDEN`**.
 
+Reason codes extend with **`UAT_BROWSER_UNAVAILABLE`**, **`UAT_BROWSER_PROBE_FAILED`**,
+**`UAT_BROWSER_PROBE_TIMEOUT`** (**US-0093** / **DEC-0079**).
+
+### Browser UAT self-test (US-0093)
+
+When **`UAT_BROWSER_PROBE_MODE=cursor`** (default) and step classifies as **`browser_smoke`**
+(or automatable **`manual_operator`** reclassified per **DEC-0079** §4), execute the **Cursor
+browser MCP** sequence — **lib never calls MCP directly** (**BUG-0006**):
+
+1. **Resolve target URL** — `docs/engineering/runtime-connectivity.md` first `http(s)://`; else
+   dev-server port from `package.json` + scratchpad **`DEV_SERVER_PORT`**.
+2. **`browser_navigate`** — load health/app URL; respect enterprise origin allowlist.
+3. **Step plan** — map acceptance verbs to **`browser_click`** / **`browser_type`** /
+   **`browser_scroll`** when automatable; **never** fill password/credential fields or read **`.env`**.
+4. **`browser_screenshot`** — write to **`sprints/Sxxxx/evidence/browser/<probe_id>-<seq>.png`**
+   (max **5** per probe).
+5. **Console + network evidence** — counts + summary path refs only (no inline secrets).
+6. **Verdict** — set **`passed`**, **`reason_code`**, **`browser_evidence_refs`** in
+   **`uat.json`** `probe_results[]`; **`passed=true`** in **`cursor`** mode **requires** non-empty
+   **`navigation_url`** + at least one screenshot or console/network summary path — else
+   **`UAT_BROWSER_PROBE_FAILED`**. Mirror under **`qa-findings.md`** **Runtime browser evidence**.
+7. **Fallback** — MCP unavailable → record **`UAT_BROWSER_UNAVAILABLE`**, run stdlib fallback when
+   **`UAT_BROWSER_FALLBACK_CHAIN=1`**; validate via
+   **`python scripts/uat_probe_lib.py --merge-result <fragment.json>`**.
+
+**No silent PASS** — stdlib alone does **not** PASS **`browser_smoke`** in **`cursor`** mode
+without agent **`browser_evidence_refs`**. Security: no auto-read **`.env`**, no credential
+auto-fill, no intake evidence mutation — **`UAT_PROBE_FORBIDDEN`** unchanged.
+
 ## Steps
 0. If `SECURITY_REVIEW=1`, verify `docs/engineering/security-review.md` exists
    and has no unresolved `critical` findings before proceeding. If unresolved

package/template/.cursor/commands/verify-work.md

@@ -103,12 +103,39 @@ Record evidence in **`sprints/Sxxxx/uat.json`** `probe_results[]` and
 **No silent PASS** — unresolvable steps must record fail-closed reason codes:
 
 `UAT_PROBE_UNRESOLVED`, `UAT_STACK_PROFILE_UNKNOWN`, `UAT_PROBE_TIMEOUT`,
-`UAT_PROBE_FAILED`, `UAT_PROBE_FORBIDDEN`, `UAT_PROBE_PASS`.
+`UAT_PROBE_FAILED`, `UAT_PROBE_FORBIDDEN`, `UAT_PROBE_PASS`,
+`UAT_BROWSER_UNAVAILABLE`, `UAT_BROWSER_PROBE_FAILED`, `UAT_BROWSER_PROBE_TIMEOUT`.
 
 Probe catalog: **`build`**, **`test`**, **`api_health`**, **`process_health`**,
 **`browser_smoke`**, **`cli_smoke`**, **`manual_operator`**. CLI:
 `python scripts/uat_probe_lib.py --repo . --step "<acceptance step>" --report`.
 
+### Browser UAT self-test (US-0093)
+
+When **`UAT_BROWSER_PROBE_MODE=cursor`** (default) and step classifies as **`browser_smoke`**
+(or automatable **`manual_operator`** reclassified per **DEC-0079** §4), execute the **Cursor
+browser MCP** sequence — **lib never calls MCP directly** (**BUG-0006**):
+
+1. **Resolve target URL** — `docs/engineering/runtime-connectivity.md` first `http(s)://`; else
+   dev-server port from `package.json` + scratchpad **`DEV_SERVER_PORT`**.
+2. **`browser_navigate`** — load health/app URL; respect enterprise origin allowlist.
+3. **Step plan** — map acceptance verbs to **`browser_click`** / **`browser_type`** /
+   **`browser_scroll`** when automatable; **never** fill password/credential fields or read **`.env`**.
+4. **`browser_screenshot`** — write to **`sprints/Sxxxx/evidence/browser/<probe_id>-<seq>.png`**
+   (max **5** per probe).
+5. **Console + network evidence** — counts + summary path refs only (no inline secrets).
+6. **Verdict** — set **`passed`**, **`reason_code`**, **`browser_evidence_refs`** in
+   **`uat.json`** `probe_results[]`; **`passed=true`** in **`cursor`** mode **requires** non-empty
+   **`navigation_url`** + at least one screenshot or console/network summary path — else
+   **`UAT_BROWSER_PROBE_FAILED`**. Mirror under **`qa-findings.md`** **Runtime browser evidence**.
+7. **Fallback** — MCP unavailable → record **`UAT_BROWSER_UNAVAILABLE`**, run stdlib fallback when
+   **`UAT_BROWSER_FALLBACK_CHAIN=1`**; validate via
+   **`python scripts/uat_probe_lib.py --merge-result <fragment.json>`**.
+
+**No silent PASS** — stdlib alone does **not** PASS **`browser_smoke`** in **`cursor`** mode
+without agent **`browser_evidence_refs`**. Security: no auto-read **`.env`**, no credential
+auto-fill, no intake evidence mutation — **`UAT_PROBE_FORBIDDEN`** unchanged.
+
 ## Steps
 1. Convert acceptance criteria into testable UAT steps. Derive steps directly from the story's acceptance criteria in `docs/product/acceptance.md`. Each AC should map to at least one UAT step. Run **`scripts/uat_probe_lib.py`** for each step where automation applies; record `probe_results[]` in **`uat.json`**.
 2. Populate UAT artifacts: write derived steps into `uat.json` (with description and result per step, accurate pass/fail counts) and `uat.md` (step list with results, summary section). Ensure UAT artifacts are in **populated** state per DEC-0009 — not placeholder.

package/template/.cursor/scratchpad.local.example.md

@@ -238,6 +238,27 @@ DOC_DETAIL_LEVEL=balanced
 # - README_FEATURE_COVERAGE_ENFORCE: 0|1 (default 0 until backfill + --report green)
 README_FEATURE_COVERAGE_ENFORCE=1
 
+#
+# ## Browser UAT self-test (US-0093 / DEC-0079)
+# Two-tier browser probe: stdlib lib classifies + agent owns Cursor browser MCP (BUG-0006).
+# - UAT_BROWSER_PROBE_MODE: cursor|http_fallback|playwright_fallback (default cursor)
+#   - cursor: agent executes MCP sequence; lib emits plan + UAT_PROBE_UNRESOLVED until evidence
+#   - http_fallback: stdlib HTTP GET (CI recipe — set this in CI)
+#   - playwright_fallback: subprocess Playwright primary; HTTP fallback when missing
+# - UAT_BROWSER_FALLBACK_CHAIN: 0|1 (default 1; enable HTTP → Playwright after MCP unavailable)
+# - UAT_PROCESS_HEALTH_POLL_SECONDS: positive int (default 60; process_health readiness cap)
+# - UAT_PROCESS_HEALTH_POLL_INTERVAL_SECONDS: positive int (default 2; poll interval)
+# - DEV_SERVER_PORT: int (optional; URL/port inference override)
+# - DEV_SERVER_COMMAND: shell command (optional; process_health startup override)
+# Interaction: orthogonal to PERMISSION_MODE and Cursor browser approval modes (manual / allow-list /
+#   auto-run per vendor docs). Health URLs from docs/engineering/runtime-connectivity.md first.
+UAT_BROWSER_PROBE_MODE=cursor
+UAT_BROWSER_FALLBACK_CHAIN=1
+UAT_PROCESS_HEALTH_POLL_SECONDS=60
+UAT_PROCESS_HEALTH_POLL_INTERVAL_SECONDS=2
+DEV_SERVER_PORT=
+DEV_SERVER_COMMAND=
+
 #
 # ## Caveman mode (US-0089)
 # Response-side voice toggle. Default off. Composition is orthogonal to
@@ -247,10 +268,18 @@ README_FEATURE_COVERAGE_ENFORCE=1
 # - CAVEMAN_MODE: 0|1 (default 0; absence = 0)
 # - CAVEMAN_LEVEL: lite|full|ultra (empty; with MODE=1 empty -> treat as full;
 #   unknown value -> CAVEMAN_LEVEL_UNKNOWN and fall back to pre-US-0089 voice)
-# - CAVEMAN_COMPRESS_INPUT: 0|1 -- reserved for US-0090; inert in US-0089;
-#   no behavior until compression story ships
-# - CAVEMAN_FILE_SCOPE: string -- reserved for US-0090; inert in US-0089;
-#   no behavior until compression story ships
+#
+# ## Caveman input compression (US-0090 / DEC-0073)
+# Input-side prose minification via scripts/caveman_compress_input.py. Default off.
+# Orthogonal to CAVEMAN_MODE (reply voice) and TOKEN_PROFILE (context breadth).
+# - CAVEMAN_COMPRESS_INPUT: 0|1 (default 0) -- activation gate; must be 1 for --write
+# - CAVEMAN_FILE_SCOPE: string (empty default) -- allow-list of files eligible for compression:
+#     * empty: no files in scope (fail-closed on --write with CAVEMAN_COMPRESS_SCOPE_EMPTY)
+#     * named profile: e.g. docs-prose-only (user-guides, runbook, state-archive, handoffs/archive)
+#     * raw globs: e.g. docs/user-guides/**/*.md,handoffs/archive/*.md (forward slashes only)
+#     * hybrid: profile:docs-prose-only;globs:handoffs/archive/*.md
+#   Mutation requires COMPRESS_INPUT=1 + non-empty scope + CLI --write; use --dry-run first.
+#   Originals land in docs/.caveman-originals/<path>; deny-list always wins over allow.
 CAVEMAN_MODE=0
 CAVEMAN_LEVEL=
 CAVEMAN_COMPRESS_INPUT=0

package/template/.cursor/scratchpad.md

@@ -240,3 +240,23 @@ USER_GUIDE_MODE=0
 DOC_AUDIENCE_PROFILE=both
 DOC_DETAIL_LEVEL=balanced
 
+#
+# ## Browser UAT self-test (US-0093 / DEC-0079)
+# Two-tier browser probe: stdlib lib classifies + agent owns Cursor browser MCP (BUG-0006).
+# - UAT_BROWSER_PROBE_MODE: cursor|http_fallback|playwright_fallback (default cursor)
+#   - cursor: agent executes MCP sequence; lib emits plan + UAT_PROBE_UNRESOLVED until evidence
+#   - http_fallback: stdlib HTTP GET (CI recipe — set this in CI)
+#   - playwright_fallback: subprocess Playwright primary; HTTP fallback when missing
+# - UAT_BROWSER_FALLBACK_CHAIN: 0|1 (default 1; enable HTTP → Playwright after MCP unavailable)
+# - UAT_PROCESS_HEALTH_POLL_SECONDS: positive int (default 60; process_health readiness cap)
+# - UAT_PROCESS_HEALTH_POLL_INTERVAL_SECONDS: positive int (default 2; poll interval)
+# - DEV_SERVER_PORT: int (optional; URL/port inference override)
+# - DEV_SERVER_COMMAND: shell command (optional; process_health startup override)
+# Interaction: orthogonal to PERMISSION_MODE and Cursor browser approval modes (manual / allow-list /
+#   auto-run per vendor docs). Health URLs from docs/engineering/runtime-connectivity.md first.
+UAT_BROWSER_PROBE_MODE=cursor
+UAT_BROWSER_FALLBACK_CHAIN=1
+UAT_PROCESS_HEALTH_POLL_SECONDS=60
+UAT_PROCESS_HEALTH_POLL_INTERVAL_SECONDS=2
+DEV_SERVER_PORT=
+DEV_SERVER_COMMAND=