tech-lead-stack 1.0.1

package/.ai/skills/dev-team-sub-max.md

@@ -0,0 +1,369 @@
+---
+name: dev-team-sub-max
+description: >
+  [DEV-TEAM · SUB-MAX · NO API KEYS · CROSS-MODEL VERIFY] Subscription-tier
+  ($100/mo) dev team orchestrator. Runs up to 2 parallel lanes with git
+  worktrees, enforces turn budgets and quota ledger checkpoints, hardens plans
+  via reflexion-loop-sub-max, manages multi-vendor model isolation (L0-L3) and
+  exhaustion limits without losing work, and keeps the full visual fidelity gate
+  intact without requiring API keys.
+cost: ~4300 tokens
+modes: [read-only, write, mcp]
+surface: public
+category: Orchestrators
+how:
+  'Multi-vendor model contract, Quota Ledger with active model tracking,
+  Findings Ledger, and context-firewalled reviewer isolation'
+useCase:
+  'Multi-lane parallel feature orchestration on a high-tier ($100/mo)
+  subscription without API keys'
+kind: orchestrator
+domain: eng
+spans: [intent, specify, plan, build, maintain, review, deploy]
+ownership:
+  drive: human-ai
+  approve: human
+targets: [api, subscription]
+minModelClass: large
+suggests:
+  [
+    dev-team-orchestrator,
+    dev-team-sub-pro,
+    mission-architect,
+    reflexion-loop-sub-max,
+    reflexion-loop,
+    visual-verifier,
+  ]
+policies:
+  - user-sovereignty
+  - diagnosis-first
+  - four-pillars
+---
+
+# Dev Team Orchestrator — Sub-Max Tier ($100/mo)
+
+> Tier siblings: dev-team-orchestrator (API keys, dual-model) · dev-team-sub-max
+> ($100 tier) · dev-team-sub-pro ($20 tier). See the tier table in the README.
+>
+> [!NOTE] **Tier profile ($100/mo subscription):** Designed for high-tier
+> subscription agents (e.g. Pro/Max or Team plans). Runs up to **2 parallel
+> lanes** with git worktree isolation, enforces turn budgets, uses
+> `reflexion-loop-sub-max` for plan hardening, handles model exhaustion limits
+> seamlessly without losing work, and requires **no external API keys**.
+
+## Runtime modes
+
+Produces a blueprint and hand-off in read-only chat; executes in an IDE/MCP
+agent.
+
+> [!IMPORTANT] **Anti-Micromanagement Litmus** Personas receive goals + gates,
+> never line-by-line instructions. The human appears only at gates. We advise;
+> the User Tech-Lead decides.
+
+<!-- -->
+
+> [!CAUTION] **RUNTIME MODE (DETERMINE FIRST — NON-NEGOTIABLE)**
+>
+> - **Read-only chat (`/chat`):** write/exec tools are forbidden; only
+>   `get_skill`, `list_skills`, `read_file` exist. Deliver a **verifiable
+>   blueprint + handoff**.
+> - **IDE / MCP-enabled Agent:** you have full write access and must execute and
+>   verify changes.
+
+<!-- -->
+
+> [!IMPORTANT] **EXECUTION DISCIPLINE (IDE/MCP MODE)**
+>
+> - **Produce, don't deliberate:** never call a sequential-thinking/planning
+>   tool more than twice consecutively without emitting a concrete output. If
+>   unsure, emit the current phase's artifact.
+> - **No stall commands:** run no further search/terminal command between
+>   finishing a phase's inputs and emitting its artifact.
+> - **DIRECT EDITS ONLY:** native file edit tools only (no regex/patch.js
+>   scripts).
+
+## Pre-Flight Model Contract (MANDATORY BEFORE PHASE 0)
+
+The orchestrator MUST read active models from the agent harness at runtime.
+Formulate and print the Pre-Flight Model Contract before Phase 0:
+
+| Role             | Model class assigned          | Isolation vs writer | Continuity Fallback        |
+| ---------------- | ----------------------------- | ------------------- | -------------------------- |
+| planner          | frontier                      | N/A (writer)        | Rung 1 -> Rung 2           |
+| reviewer / qa    | a different vendor's frontier | L0 (cross-vendor)   | Rung 1 -> Rung 2 -> Park   |
+| developer        | mid-tier                      | N/A (implementer)   | Rung 1 -> Rung 2 -> Rung 3 |
+| discovery / grep | small/fast                    | N/A (read-only)     | Rung 3                     |
+
+### Environment Check (`CLAUDE_CODE_SUBAGENT_MODEL`)
+
+Before claiming an isolation level, check the `CLAUDE_CODE_SUBAGENT_MODEL`
+environment variable.
+
+- If `CLAUDE_CODE_SUBAGENT_MODEL` is set to anything other than `inherit`, all
+  sub-agents collapse onto a single model.
+- **Action:** Issue a warning plainly in chat and cap the claimed isolation
+  level at **L2**. Never claim L0 or L1 when environment variables override
+  sub-agent model selection.
+
+### Four-Level Isolation Ladder
+
+Claim the level ACHIEVED based on runtime model selection:
+
+- **L0 (Cross-Vendor)**: Writer and reviewer run on models from different
+  vendors. Default target for reviewer.
+- **L1 (Cross-Family, Same Vendor)**: Writer and reviewer run on different model
+  families from the same vendor (shared training lineage limitation apply).
+- **L2 (Fresh Sub-Agent)**: Same model, fresh sub-agent context receiving only
+  diff + criteria + commands.
+- **L3 (Fresh Session)**: Same model, work pasted into a new session cold.
+
+### Reviewer / QA Resolution Ladder
+
+Resolve reviewer and QA models with the Critic Resolution Ladder defined in
+`reflexion-loop-sub-max`
+(`./.ai/rtk-run run resolve-critic --writer <vendor> --writer-model <planner-model>`):
+Gemini (`gemini-cli`, then `agy`) -> `codex` -> Claude (`claude-subagent` inside
+Claude Code, else `claude-cli`) -> another harness model -> same model. Each
+rung works on a subscription or an API key. Never guess a CLI: the standalone
+`gemini` CLI no longer serves personal Google logins (since 18 June 2026), and
+its `GOOGLE_CLOUD_PROJECT` error means unsupported account type. On the
+same-model rung the slice is `PROVISIONAL` and the lane state file MUST carry
+the STRONG `Critic Advisory` line.
+
+## Phase 0A — Cold Resume Protocol (MANDATORY FIRST STEP)
+
+On invocation, before performing any codebase search or stack discovery:
+
+1. Check for existing lane state files in `.dev-team/lanes/*.md`.
+2. If any lane file exists and status is NOT `Complete`:
+   - Read `.dev-team/quota.md` and print
+     `[turns: used/budget | model: <active-model>]`.
+   - Read Findings Ledger `.dev-team/analysis/<lane-id>.md` and incomplete lane
+     state file(s).
+   - **DO NOT re-run Phase 0 stack discovery** for a lane that already recorded
+     its stack — that wastes turn budget in a fresh window.
+   - Immediately resume execution from recorded `Phase` following
+     `Resume Instruction`.
+3. If no incomplete lane files exist, proceed to Phase 0B.
+
+## Phase 0B — Stack Discovery & Mission Frame
+
+- **Skill acquisition (NON-NEGOTIABLE):** IDE/MCP agent MUST call `get_skills`
+  tool; Chat UI MUST call `get_skill`. Never raw-read `.ai/skills/`.
+- **Discovery Budget:** Scoped searches only (exclude `node_modules`, `.next`,
+  `.nx`, `dist`, `build`). Print Phase 1 sizing scores immediately after
+  discovery.
+- **Stack ID & Mission Frame:** Formulate stack summary, one-sentence mission,
+  and success metric.
+- **Findings Ledger Creation:** Immediately write stack facts, file paths, and
+  domain boundaries to `.dev-team/analysis/<lane-id>.md`.
+
+## Phase 1 — Crew Sizing Gate
+
+Evaluate task based on five-signal 0–2 rubric (Surface area, Novelty, Risk,
+Ambiguity, Parallelism). **Scores MUST be printed first before any work
+begins.**
+
+_(Note: The ceiling limits below are generated/derived — see
+`TIER_POLICY['sub-max']` in `src/lib/ai/tier-policy.ts` for the authoritative
+code policy.)_
+
+| Size | Score | Crew                                          | Max Parallel Lanes | Loop Hardening                                                   |
+| ---- | ----- | --------------------------------------------- | ------------------ | ---------------------------------------------------------------- |
+| XS   | 0–1   | Developer only                                | 1                  | self-check + autoeval                                            |
+| S    | 2–3   | Developer + Reviewer                          | 1                  | reviewer gate                                                    |
+| M    | 4–5   | Planner + Developer + Reviewer                | 1–2                | plan gate + review gate                                          |
+| L    | 6–8   | PM-analyst + Planner + Dev ×N + Reviewer + QA | 2                  | `reflexion-loop-sub-max` recommended                             |
+| XL   | 9–10  | mission-architect strategy + L crew           | 2                  | `reflexion-loop-sub-max` mandatory + Tech-Lead confirmation gate |
+
+> [!CAUTION] **Sub-Max Sizing Constraints:** Max 2 parallel lanes (enforced by
+> `tier-policy.ts`). XL requires a Tech-Lead confirmation gate in
+> `.dev-team/inbox.md` stating expected turn cost (~60 turns across lanes)
+> before worktree creation.
+
+## Phase 2 — Lane & Quota Ledgers
+
+### Quota Ledger (`.dev-team/quota.md`)
+
+- Mission turn budget: **60 turns** | Per-lane turn budget: **25 turns**
+- Print `[turns: used/budget | model: <active-model>]` at every gate boundary.
+
+| Lane ID | Active Model | Turns Used | Turn Budget | Headroom Last Poll | Swap Count | Last Checkpoint | Status |
+| ------- | ------------ | ---------- | ----------- | ------------------ | ---------- | --------------- | ------ |
+| ...     | ...          | ...        | 25          | ...                | 0          | ...             | ...    |
+
+### Lane Ledger & CHECKPOINT-BEFORE-GATE
+
+| lane-id | task | size | crew | branch+worktree | state-file | status | next-gate |
+| ------- | ---- | ---- | ---- | --------------- | ---------- | ------ | --------- |
+| ...     | ...  | ...  | ...  | ...             | ...        | ...    | ...       |
+
+- **Isolation:** One git worktree per lane (`rtk git worktree add ...`), single
+  writer per lane. Mandatory worktree bootstrap (deps, `.env`, clients, builds)
+  before testing.
+- **CHECKPOINT-BEFORE-GATE (MANDATORY):** Write `.dev-team/lanes/<lane-id>.md`
+  BEFORE entering any gate.
+
+**State File Template (`.dev-team/lanes/<lane-id>.md`):**
+
+```md
+# Lane: <lane-id>
+
+- Task: <description>
+- Status: <status> (Active | Complete | PROVISIONAL | UNREVIEWED)
+- Next Gate: <gate>
+- Phase: <current-phase-number>
+- Active Model: <model-name>
+- Isolation Level: <L0-L3>
+- Critic Rung: <gemini-cli | agy | codex | claude-subagent | claude-cli |
+  harness-model | same-model> (<skipped rungs: reasons>)
+- Critic Advisory: <none | STRONG — SAME_MODEL_CRITIC: review ran on the
+  writer's model and is NOT independent; deep-review before merge, then set up
+  gemini/agy, codex or claude and re-review>
+- Turns Used: <turns-count> / 25
+- Last Checkpoint: <timestamp>
+- Resume Instruction:
+  <one imperative sentence telling a fresh-context agent exactly what to do next>
+- Current Artifacts: <links/paths>
+```
+
+## Phase 3 — Persona Execution, Model Continuity, UNREVIEWED vs PROVISIONAL Slices
+
+### Model Exhaustion Classification
+
+- **Mode A (MODEL-SCOPED):** Quota spent on one model. Run fallback ladder and
+  continue.
+- **Mode B (ACCOUNT-WIDE):** Limit reached across models. Consolidate into
+  Findings Ledger, PARK, report reset window. State 3 disclosure applies.
+- **Mode C (SILENT DOWNGRADE):** Harness swapped models mid-session without
+  error. Poll active model at every phase boundary. Any change is recorded as a
+  swap event.
+
+### Sideways-Before-Downward Fallback Ladder
+
+Reason about VENDOR and CLASS separately:
+
+1. **Rung 1 (Same Class, Different Vendor):** Move to an equivalent frontier
+   class model on another vendor. Isolation level preserved (L0).
+2. **Rung 2 (Same Vendor, Lower Class):** Move to lower model class on same
+   vendor. Recompute isolation level (if lands on writer's model, claim drops to
+   L2).
+3. **Rung 3 (Small/Fast Class):** THROUGHPUT ROLES ONLY (developer, discovery,
+   QA capture). Assurance roles NEVER fall to Rung 3; PARK instead.
+4. **Rung 4 (No Capacity Anywhere):** Consolidate into Findings Ledger, PARK,
+   report reset window in State 3.
+
+### Slice Status Definitions: UNREVIEWED vs PROVISIONAL
+
+- **PROVISIONAL Slices:** A slice whose review was completed under a degraded
+  assurance role (same model or L2/L3 isolation). Marked `PROVISIONAL` in Lane
+  Ledger and state file. PROVISIONAL slices are re-reviewed when capacity
+  returns, or require explicit Tech-Lead waiver.
+- **UNREVIEWED Slices:** A slice where the run stopped before the audit ran or
+  completed (Mode B park). UNREVIEWED work has received NO verification pass.
+- **Risk-2 Work Rule:** Auth, payments, customer data, and infrastructure
+  changes MAY NOT close on a PROVISIONAL or UNREVIEWED approval — PARK instead.
+
+### Findings Ledger Protocol (`.dev-team/analysis/<lane-id>.md`)
+
+Write stack facts, file paths, domain boundaries, Figma measurements, and
+**OPTIONS REJECTED WITH REASONS** (mandatory) the moment established.
+Consolidate at 20% headroom / 80% turn budget spent. Schedule swap for NEXT
+phase boundary.
+
+### Reviewer Isolation & Visual Gate
+
+Reviewer runs in fresh sub-agent (Context Firewall: diff + criteria + commands).
+Must **ACT, not read**. Figma spec in plan required. Gate 4 Layout Deviation
+Report & `visual-verifier` capture required for UI slices.
+
+## Phase 4 — Tech-Lead Interview at Gates
+
+Batch questions in `.dev-team/inbox.md`. PARK is a hard stop. Inbox is read-only
+after human responds. Never bypass gate failures silently.
+
+## Phase 5 — Friction Defect Protocol & Model Swap Triggers
+
+**Triggers:**
+
+- ≥2 rework loops on one gate.
+- Skill misbehaviour or missing tool/permission.
+- **Model Swap Trigger:** A lane swaps models more than 2 times, or any
+  assurance role falls to Rung 3.
+
+**Action:** Write `.dev-team/friction/<date>-<slug>.md`. Append draft command to
+inbox. Draft-only unless `DEV_TEAM_AUTOFILE_ISSUES=1`.
+
+> [!CAUTION] **ABSOLUTE RULE** `git push`, `git add`, and `merge` remain
+> STRICTLY FORBIDDEN regardless of mode or env var.
+
+## Telemetry
+
+Pass `{ teamRole: "<ROLE>", loopRunId: "<MISSION_ID>", actorType: "AGENT" }` on
+every skill call.
+
+## Three Mandatory End-State Disclosures (VERY FIRST LINE OF FINAL OUTPUT)
+
+The FIRST line of final output MUST emit exactly one of these three end states:
+
+- **STATE 1 — Separation Held (Auditor finished on a different model):**
+
+  ```text
+  Model separation held: written by <model-a>, audited by <model-b>.
+  ```
+
+- **STATE 2 — Separation Lost (Auditor FINISHED, but on the writer's model):**
+
+  ```text
+  MODEL SEPARATION LOST: <writer-model> wrote this work and also audited it.
+  <exhausted-model> hit its usage limit at <phase/step>, so the audit fell back to the same model that produced the work. This audit was not independent.
+  ```
+
+- **STATE 3 — Audit Incomplete (Run stopped before auditor finished):**
+
+  ```text
+  AUDIT NOT COMPLETED: the run stopped at <phase/step> before the audit finished.
+  <exhausted-model> hit an account-wide usage limit, so no model was available to continue. The work below is UNREVIEWED, not approved. Quota resets <window>.
+  ```
+
+### Selection Rule
+
+State 2 REQUIRES that an audit RAN TO COMPLETION on the writer's model. If the
+audit did not complete, State 3 applies — NEVER State 2. An unfinished audit is
+not a weak audit, it is an absent one.
+
+### Full Provenance Table (Emitted Beneath Disclosure Line)
+
+| Phase | Role | Model | Isolation | Reason for swap | Effect on the claim |
+| ----- | ---- | ----- | --------- | --------------- | ------------------- |
+| ...   | ...  | ...   | ...       | ...             | ...                 |
+
+## Four Pillars Compliance & Anti-Rationalization
+
+### Anti-Rationalization Protocol
+
+| Rationalization                                                               | Rebuttal / Required Behavior                                                                            |
+| ----------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------- |
+| "Tests pass so the slice is done."                                            | Test-pass is not design-pass. Run visual fidelity gate and paste Layout Deviation Report.               |
+| "I'll bootstrap the worktree later."                                          | Without bootstrap (deps, `.env`, clients, builds), `check-types` is meaningless. Bootstrap immediately. |
+| "The human is probably fine with this choice."                                | PARK and write to `.dev-team/inbox.md`. Guessing an answer is a defect.                                 |
+| "The audit passed anyway, so the notice would just worry them."               | A pass from the author is not a pass; the notice IS the finding. Emit disclosure line as line 1.        |
+| "The model swap was handled automatically, so it's an implementation detail." | Handling it seamlessly is why developer cannot see it, which is exactly why it must be stated.          |
+| "It is already recorded in the provenance table below."                       | A table row is not a disclosure; the first line is.                                                     |
+
+### Hooks (Ownership Gates)
+
+Before advancing to the next phase or gate, you MUST consult `.ai/hooks/`. If a
+guard is triggered and requires human approval (`require-human-approve`), you
+MUST append the question to the human inbox (`.dev-team/inbox.md`) rather than
+proceeding.
+
+## Code Modification Convention
+
+**REQUIREMENT:** When modifying files, you MUST use the `apply_patch` tool with
+minimal SEARCH/REPLACE blocks instead of rewriting whole files.
+
+- Never emit a full-file rewrite.
+- Never restate unchanged code.
+- **Rule:** Include only the lines that change plus minimal surrounding anchor
+  context.

package/.ai/skills/dev-team-sub-pro.md

@@ -0,0 +1,288 @@
+---
+name: dev-team-sub-pro
+description: >
+  [DEV-TEAM · SUB-PRO · NO API KEYS · CROSS-MODEL VERIFY] Subscription-tier
+  ($20/mo) dev pair orchestrator. Single-lane, branch-based execution without
+  worktrees, enforcing turn budgets, builder/checker roles, cross-vendor model
+  isolation, Mode B quota handling, and tier-ceiling enforcement without
+  requiring API keys.
+cost: ~3500 tokens
+modes: [read-only, write, mcp]
+surface: public
+category: Orchestrators
+how:
+  'Single-lane Builder/Checker model contract, compressed Findings Ledger, Mode
+  B consolidate-and-park, and mandatory three-state end-state disclosure'
+useCase:
+  'Frugal single-lane feature orchestration on a standard ($20/mo) subscription
+  without API keys'
+kind: orchestrator
+domain: eng
+spans: [intent, specify, plan, build, maintain, review, deploy]
+ownership:
+  drive: human-ai
+  approve: human
+targets: [api, subscription]
+minModelClass: large
+suggests:
+  [
+    dev-team-orchestrator,
+    dev-team-sub-max,
+    reflexion-loop-sub-pro,
+    reflexion-loop,
+    vertical-slice-decomposer,
+  ]
+policies:
+  - user-sovereignty
+  - diagnosis-first
+  - four-pillars
+---
+
+# Dev Team Orchestrator — Sub-Pro Tier ($20/mo)
+
+> Tier siblings: dev-team-orchestrator (API keys, dual-model) · dev-team-sub-max
+> ($100 tier) · dev-team-sub-pro ($20 tier). See the tier table in the README.
+>
+> [!NOTE] **Tier profile ($20/mo subscription):** Honest promise: **A
+> disciplined pair, not a team.** Designed for standard subscription tiers (e.g.
+> Pro/Plus plans). Operates as a single-lane **Builder + Checker pair** directly
+> on a feature branch (no git worktrees), capped at size M / Risk 1 tasks, with
+> a strict 20-turn slice budget and **no external API keys**.
+
+<!-- -->
+
+> [!NOTE] **Frugal Compression Notice:** To preserve token frugality on a $20
+> plan, multi-lane rebalancing and rung-probing sequences are **OMITTED**. The
+> L0–L3 ladder, fallback rungs, and Findings Ledger structure are
+> **COMPRESSED**. For detailed protocol mechanics, refer to `dev-team-sub-max`.
+> The Pre-Flight Model Contract, `CLAUDE_CODE_SUBAGENT_MODEL` check, Mode B
+> consolidate-and-park protocol, and Mandatory Three-State Disclosure rules are
+> included **FULL and uncompressed**.
+
+## Runtime modes
+
+Produces a blueprint in read-only chat; executes in IDE/MCP mode.
+
+> [!IMPORTANT] **Anti-Micromanagement Litmus** Personas receive goals + gates,
+> never line-by-line instructions. The human appears only at gates. We advise;
+> the User Tech-Lead decides.
+
+<!-- -->
+
+> [!CAUTION] **RUNTIME MODE (DETERMINE FIRST — NON-NEGOTIABLE)** Read-only chat
+> (`/chat`): write/exec forbidden; deliver verifiable blueprint + handoff.
+> IDE/MCP mode: full write access via native edits.
+
+## Pre-Flight Model Contract (FULL)
+
+Read active models from agent harness at runtime. Formulate and print:
+
+| Role                      | Model class assigned          | Isolation vs writer | Continuity Fallback |
+| ------------------------- | ----------------------------- | ------------------- | ------------------- |
+| Builder (Plan + Dev)      | frontier / mid                | N/A (writer)        | Consolidate & Park  |
+| Checker (Review + Verify) | a different vendor's frontier | L0 (cross-vendor)   | L1 -> L2 -> Park    |
+
+> **Sub-Pro Note:** Sub-pro is **throughput-limited** (quota, single lane, crew
+> ceiling M, tighter turn budget), not model-limited. Cross-vendor L0 isolation
+> is reachable on the entry tier wherever the platform offers one lineup across
+> paid tiers. Model availability is platform-dependent; read the harness's
+> actual model list at runtime instead of assuming a tier ceiling.
+
+### Environment Check (`CLAUDE_CODE_SUBAGENT_MODEL` — FULL)
+
+If `CLAUDE_CODE_SUBAGENT_MODEL` is set to anything other than `inherit`, warn
+plainly in chat and cap claimed isolation level at **L2** (same-model
+sub-agent). Never claim L0 or L1 when overridden by environment variables.
+
+### Isolation Ladder & Fallback Rungs (COMPRESSED)
+
+- **L0 (Cross-Vendor)**: Different vendor models. Default target for Checker.
+- **L1 (Cross-Family)**: Same vendor, different model family.
+- **L2 (Fresh Sub-Agent)**: Same model, fresh sub-agent context.
+- **L3 (Fresh Session)**: Same model, cold paste into fresh session.
+- **Fallback Rungs:** Rung 1 (same class, different vendor) -> Rung 2 (same
+  vendor lower class, claim drops to L2 if lands on writer) -> Rung 3
+  (small/fast, throughput roles only; assurance roles PARK).
+- **Checker Resolution (COMPRESSED — full ladder in `reflexion-loop-sub-max`):**
+  run
+  `./.ai/rtk-run run resolve-critic --writer <vendor> --writer-model <builder-model>`;
+  order is Gemini (`gemini-cli`, then `agy`) -> `codex` -> Claude
+  (`claude-subagent` inside Claude Code, else `claude-cli`) -> another harness
+  model -> same model, each on a subscription or an API key. The standalone
+  `gemini` CLI no longer serves personal Google logins (since 18 June 2026) — a
+  `GOOGLE_CLOUD_PROJECT` error means the probe moves on to `agy`. Record
+  `Critic Rung:` in `.dev-team/lanes/lane-1.md`; on the same-model rung the
+  slice is `PROVISIONAL` and the file MUST carry a STRONG `Critic Advisory:`
+  line (SAME_MODEL_CRITIC — review not independent; deep-review before merge,
+  and re-review once a Gemini, `codex` or Claude critic works).
+
+## Phase 0A — Cold Resume Protocol (MANDATORY FIRST STEP)
+
+1. Check for existing state file `.dev-team/lanes/lane-1.md`.
+2. If incomplete: print `[turns: used/20 | model: <active-model>]`, read
+   `.dev-team/analysis/lane-1.md` and state file. **DO NOT re-run Phase 0
+   discovery**. Resume immediately from recorded `Phase`.
+3. If no incomplete lane file exists, proceed to Phase 0B.
+
+## Phase 0B — Stack Discovery & Mission Frame
+
+- **Skill acquisition:** Call `get_skills` for `dev-team-sub-pro`. Never
+  raw-read `.ai/skills/`.
+- **Discovery Budget:** Scoped searches (exclude build dirs). Print sizing
+  scores immediately after discovery.
+- **Findings Ledger:** Write stack facts, file paths, and domain boundaries to
+  `.dev-team/analysis/lane-1.md`.
+
+## Phase 1 — Crew Sizing & Tier Ceiling Gate
+
+Evaluate task using 0–2 rubric (Surface area, Novelty, Risk, Ambiguity,
+Parallelism). **Scores MUST be printed first.**
+
+_(Note: The ceiling limits below are generated/derived — see
+`TIER_POLICY['sub-pro']` in `src/lib/ai/tier-policy.ts` for the authoritative
+code policy.)_
+
+| Size   | Score | Execution Model              | Hardening                                     |
+| ------ | ----- | ---------------------------- | --------------------------------------------- |
+| XS     | 0–1   | Single Builder/Checker slice | Self-check                                    |
+| S      | 2–3   | Builder + Checker            | Optional `reflexion-loop-sub-pro` if Risk = 1 |
+| M      | 4–5   | Builder + Checker            | Optional `reflexion-loop-sub-pro` if Risk = 1 |
+| L / XL | 6–10  | **REFUSED (Exceeds Budget)** | Escalate to `dev-team-sub-max`                |
+
+> [!CAUTION] **Sub-Pro Tier Refusal Rules:** Score ≥ 6 (L/XL) or Risk signal = 2
+> (auth/payments/data/infra) -> Print scores and REFUSE: _"Exceeds this tier's
+> budget. Decompose with `vertical-slice-decomposer` or escalate to
+> dev-team-sub-max."_ STOP. (Enforced by `tier-policy.ts`)
+
+## Phase 2 — Single-Lane & Quota Ledger (COMPRESSED)
+
+- Slice turn budget: **20 turns max**. Branch-based (no git worktrees).
+- Quota Ledger (`.dev-team/quota.md`):
+  `[turns: used/20 | active_model | headroom | swap_count]`.
+- Lane Ledger: `lane-1` | task | size | Builder+Checker | current branch |
+  `.dev-team/lanes/lane-1.md` | Active | gate.
+- **CHECKPOINT-BEFORE-GATE:** Write `.dev-team/lanes/lane-1.md` BEFORE entering
+  any gate.
+
+## Phase 3 — Two-Hat Execution, Mode B Protocol & Slices (UNREVIEWED vs PROVISIONAL)
+
+### Mode B Consolidate-and-Park Protocol (FULL)
+
+On a single-lane $20/mo subscription, there is no headroom to spend probing
+fallback rungs.
+
+- On encountering Mode B account-wide limit, or reaching 20 turns:
+  1. Compact all stack facts, file paths, decisions, and **OPTIONS REJECTED WITH
+     REASONS** into `.dev-team/analysis/lane-1.md`.
+  2. Write state file `.dev-team/lanes/lane-1.md` with status `UNREVIEWED`.
+  3. PARK, report progress, and report reset window in State 3 disclosure.
+
+### Slice Status Definitions: UNREVIEWED vs PROVISIONAL (FULL)
+
+- **PROVISIONAL Slices:** Reviewed by a degraded Checker (same model or L2/L3
+  isolation). Marked `PROVISIONAL` in Lane Ledger and state file. Re-reviewed
+  when capacity returns or requires explicit Tech-Lead waiver.
+- **UNREVIEWED Slices:** Work where the run stopped before Checker ran or
+  completed (Mode B park). UNREVIEWED work has received NO verification pass.
+- **Mid-Flight Risk-2 Escalation:** Intake screening is imperfect. If work is
+  DISCOVERED to touch Risk-2 areas (auth, payments, data, infra) after
+  acceptance, the lane PARKS immediately. Record the discovery and evidence in
+  the lane state file and inbox, and escalate to a higher tier
+  (`dev-team-sub-max` or `dev-team-orchestrator`). Do not close the slice on a
+  PROVISIONAL or UNREVIEWED approval, and do not continue assuming the original
+  sizing was right. Mid-flight discovery is the expected failure mode this rule
+  catches.
+
+### Findings Ledger Structure (COMPRESSED)
+
+File: `.dev-team/analysis/lane-1.md`. Contains stack facts, file paths, domain
+boundaries, Figma measurements, decisions, and **OPTIONS REJECTED WITH REASONS**
+(mandatory).
+
+### Checker Isolation & Visual Gate
+
+Checker runs in fresh sub-agent context. Must **ACT, not read**. Figma spec in
+plan STILL required. Visual gate captures 2 viewports (Desktop/Mobile).
+
+## Phase 4 & 5 — Gates & Friction Defect Protocol
+
+Batch questions in `.dev-team/inbox.md`. PARK is a hard stop. Inbox is read-only
+after human responds.
+
+- **Friction Defect Trigger:** Write `.dev-team/friction/<date>-<slug>.md` if
+  rework ≥2 on a gate, skill misbehaviour, missing tool, **or if lane swaps
+  models >2 times or Checker falls to Rung 3**. Draft-only unless
+  `DEV_TEAM_AUTOFILE_ISSUES=1`.
+
+> [!CAUTION] **ABSOLUTE RULE** `git push`, `git add`, and `merge` are STRICTLY
+> FORBIDDEN.
+
+## Telemetry
+
+Pass `{ teamRole: "<ROLE>", loopRunId: "<MISSION_ID>", actorType: "AGENT" }` on
+every skill call.
+
+## Three Mandatory End-State Disclosures (FULL — VERY FIRST LINE OF OUTPUT)
+
+The FIRST line of final output MUST emit exactly one of these three end states:
+
+- **STATE 1 — Separation Held (Auditor finished on a different model):**
+
+  ```text
+  Model separation held: written by <model-a>, audited by <model-b>.
+  ```
+
+- **STATE 2 — Separation Lost (Auditor FINISHED, but on the writer's model):**
+
+  ```text
+  MODEL SEPARATION LOST: <writer-model> wrote this work and also audited it.
+  <exhausted-model> hit its usage limit at <phase/step>, so the audit fell back to the same model that produced the work. This audit was not independent.
+  ```
+
+- **STATE 3 — Audit Incomplete (Run stopped before auditor finished):**
+
+  ```text
+  AUDIT NOT COMPLETED: the run stopped at <phase/step> before the audit finished.
+  <exhausted-model> hit an account-wide usage limit, so no model was available to continue. The work below is UNREVIEWED, not approved. Quota resets <window>.
+  ```
+
+### Selection Rule
+
+State 2 REQUIRES that an audit RAN TO COMPLETION on the writer's model. If the
+audit did not complete, State 3 applies — NEVER State 2. An unfinished audit is
+not a weak audit, it is an absent one.
+
+Emit Provenance Table
+(`| Phase | Role | Model | Isolation | Reason for swap | Effect on the claim |`)
+beneath disclosure line.
+
+## Four Pillars & Anti-Rationalization
+
+### Anti-Rationalization Protocol
+
+| Rationalization                                                               | Rebuttal / Required Behavior                                                                                     |
+| ----------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------- |
+| "It's an L task, but I can handle it in Sub-Pro."                             | Score L/XL or Risk 2 MUST be refused immediately. Escalate to Sub-Max or Dev-Team-Orchestrator.                  |
+| "The sizing said Risk-1, so this is fine."                                    | Sizing scored the brief, not the code; when discovery contradicts the brief, discovery wins — PARK and escalate. |
+| "Since we don't use worktrees, I'll commit directly."                         | `git add` and `git push` are strictly forbidden. Edits remain uncommitted.                                       |
+| "I'm the Builder so I can self-certify the review."                           | Checker must run in fresh sub-agent context and paste hard evidence.                                             |
+| "The audit passed anyway, so the notice would just worry them."               | A pass from the author is not a pass; the notice IS the finding. Emit disclosure line as line 1.                 |
+| "The model swap was handled automatically, so it's an implementation detail." | Handling it seamlessly is why developer cannot see it, which is exactly why it must be stated.                   |
+| "It is already recorded in the provenance table below."                       | A table row is not a disclosure; the first line is.                                                              |
+
+### Hooks (Ownership Gates)
+
+Before advancing to the next phase or gate, you MUST consult `.ai/hooks/`. If a
+guard is triggered and requires human approval (`require-human-approve`), you
+MUST append the question to the human inbox (`.dev-team/inbox.md`) rather than
+proceeding.
+
+## Code Modification Convention
+
+**REQUIREMENT:** When modifying files, you MUST use the `apply_patch` tool with
+minimal SEARCH/REPLACE blocks instead of rewriting whole files.
+
+- Never emit a full-file rewrite.
+- Never restate unchanged code.
+- **Rule:** Include only the lines that change plus minimal surrounding anchor
+  context.

package/.ai/skills/dummy-skill.md

@@ -0,0 +1,28 @@
+---
+name: Dummy Skill
+description: A dummy skill for testing purposes.
+phase: build
+cost: ~200 tokens
+modes: [read-only, mcp]
+surface: internal
+kind: skill
+domain: shared
+ownership:
+  drive: human-ai
+  approve: human
+targets: [local, api, subscription]
+minModelClass: small
+policies:
+  - user-sovereignty
+  - diagnosis-first
+---
+
+# Dummy Skill
+
+You MUST keep this line verbatim.
+
+This text is extra fluff that should be distilled away by the SLM to make the
+output shorter. We add a lot of extra words here to ensure that the compression
+algorithm has something to remove. The quick brown fox jumps over the lazy dog.
+The quick brown fox jumps over the lazy dog. The quick brown fox jumps over the
+lazy dog. The quick brown fox jumps over the lazy dog.