tech-lead-stack 1.0.1

package/.ai/skills/qa-handover-generator.md

@@ -0,0 +1,182 @@
+---
+name: qa-handover-generator
+description: >
+  Produces a QA handover + universal smoke-test criteria document for a changed
+  feature and delivers it to ClickUp. Splits behaviour by architecture/state
+  pattern, states the single source of truth per pattern (from real code), and
+  emits smoke-test acceptance criteria that are both agent-ingestible (for
+  generating formal acceptance criteria) and directly followable by a human
+  tester. All ClickUp output is rendered through the shared clickup-format
+  module (single source of truth for ClickUp formatting).
+cost: ~2100 tokens
+modes: [read-only, write, mcp]
+surface: public
+category: Ship & Communicate
+how:
+  'Performs Phase 0 G-Stack discovery of state architecture, maps components to
+  server-driven vs client-side patterns, and renders ClickUp markup via the
+  clickup-format module.'
+useCase:
+  'Generating high-fidelity QA handovers and smoke test checklists for
+  developers and automated testing agents.'
+phase: deploy
+kind: skill
+domain: eng
+ownership:
+  drive: human-ai
+  approve: human
+targets: [local, api, subscription]
+minModelClass: small
+consumes: [review-report]
+emits: [release]
+policies:
+  - user-sovereignty
+  - diagnosis-first
+  - four-pillars
+---
+
+# QA Handover Generator
+
+## Runtime modes
+
+Produces a verifiable QA handover in read-only chat, and in an IDE/MCP agent
+renders it via the shared ClickUp formatter and creates it in ClickUp (with a
+file fallback).
+
+**Persistence & Quality Mindset**: There is no reward for completion. The reward
+comes from a handover accurate enough that a QA engineer — or an agent ingesting
+it — can derive correct acceptance criteria without re-reading the source.
+Persist until the architecture split and the single-source-of-truth per pattern
+are stated correctly and render correctly in ClickUp.
+
+> [!CAUTION] **ClickUp formatting is NOT hand-rolled.** All ClickUp output MUST
+> be produced via the shared module `scripts/clickup-format.ts` (headings, bold,
+> code, bullets, checklists, tables, document assembly). Never format ClickUp
+> markdown inline in this skill — the shared module is the single source of
+> truth so formatting stays consistent and testable across every ClickUp-
+> producing skill. If ClickUp rendering needs a fix, fix it in that module once.
+
+## 🎯 Handover Gates
+
+### Phase 0: Skill Acquisition & Architecture Discovery (MANDATORY)
+
+- **Skill Usage Enforcement (NON-NEGOTIABLE):**
+  - **FORBIDDEN:** Direct file access via `view_file` or `run_command` is
+    strictly prohibited for skill reading.
+  - **IDE / MCP-enabled Agent:** You MUST call the MCP `get_skills` tool (which
+    may be prefixed as `mcp_tech-lead-stack_get_skills` or
+    `tech-lead-stack_get_skills` depending on client prefixing).
+  - **Chat UI (/chat):** You MUST call the internal `get_skill` tool.
+
+- **Scope the change:** Identify exactly which modules/screens/components the
+  change touches. The handover covers the feature under test, nothing else.
+- **Discover the state architecture (the core of this skill):** From the real
+  code, determine how state is managed for the feature under test. Distinguish
+  the patterns that actually apply (name only what the code uses):
+  - **Server-driven** (e.g. URL query string is the source of truth; a hook
+    parses URL state and maps it to server query variables; sort/paginate/filter
+    round-trip to the server).
+  - **Client-side / in-memory / offline-first** (e.g. a static query fetches all
+    records once; sort/search/filter execute in-memory; URL may hold filter
+    state for deep-linking but no server round-trip on change).
+  - Any other real pattern (cursor-paginated, optimistic-update, event-driven…).
+- **Per pattern, extract the mechanics** a tester needs: single source of truth,
+  the owning hook/query/function BY REAL NAME, where filtering/sorting executes,
+  reset behaviours (e.g. offset reset on tab change), browser/history/offline
+  integration.
+- **Scoped discovery only:** exclude `node_modules`, `.next`, `.nx`, `dist`,
+  `build`. No unscoped recursive searches.
+
+### Gate 1: Architecture Overview (verified against code)
+
+- **Positive (Pass):** The handover opens with an Architecture Overview split by
+  the state patterns actually found. For EACH pattern: target modules, single
+  source of truth, mechanism (named hooks/queries/functions), and
+  pattern-specific behaviours (resets, history, offline). Every claim traces to
+  real code.
+- **Negative (Fail):** Generic overview, a pattern the code does not use, or
+  named symbols that do not exist. Rendered with `clickup-format` headings +
+  tables/lists.
+
+### Gate 2: Universal Smoke-Test Acceptance Criteria
+
+- **Positive (Pass):** For EACH pattern, smoke-test criteria covering general
+  usage and core user flows (NOT edge cases): sort, paginate, filter, search,
+  tab switch, navigate — each with its expected observable result and any
+  pattern-specific gotcha (e.g. "server-side pagination offset is zero-indexed:
+  offset=0 is page 1"; "client-side table issues no server request on filter
+  change — manipulation is immediate/in-memory").
+- **Dual-audience rule:** Each criterion MUST be (a) concrete enough for an
+  agent to convert into formal acceptance criteria, AND (b) followable
+  step-by-step by a human doing it manually. Render criteria as ClickUp
+  checklist items via `clickup-format.checklist(...)` so QA can tick them off.
+
+### Gate 3: Testability & Environment Notes
+
+- **Positive (Pass):** States what the tester needs to run the smoke tests
+  locally: which modules/URLs to visit, auth/role requirements, offline/PWA
+  considerations, how to observe state (e.g. URL query string for server-driven
+  tables), and where behaviour differs by environment (live vs seeded/offline)
+  so QA does not report false failures.
+
+### Gate 4: ClickUp Delivery
+
+- **Render:** Build the entire document through `scripts/clickup-format.ts`
+  (`h2`/`h3`, `bold`, `code`, `bullets`, `checklist`, `renderTable`, `section`,
+  `assembleDocument`). Do not concatenate raw markdown by hand.
+  - **Tables:** call `renderTable(table, mode)`. Default `mode` is `'list'`
+    (guaranteed to render correctly in ClickUp). Only pass `'pipe'` if
+    pipe-tables have been confirmed to render in the target ClickUp context. The
+    mode is the ONLY table decision — never hand-write table syntax.
+- **Create in ClickUp (primary path):** When the ClickUp MCP is connected AND a
+  destination is provided (space/folder/list/doc id + title), create the
+  handover via the ClickUp MCP tools — prefer `clickup_create_document` /
+  `clickup_create_document_page` for a handover Doc, passing the rendered
+  content. Confirm the created doc's headings, table, checkboxes and code render
+  correctly.
+- **File fallback (no destination / no MCP):** write the same rendered content
+  to `.ai/output/qa-handovers/<feature>-handover.md` for manual paste into
+  ClickUp.
+- **Opt-in:** never create in ClickUp without an explicit destination.
+
+## Handover Structure (rendered via clickup-format)
+
+```md
+# QA Handover & Universal Smoke Test Criteria: <Feature>
+
+## 1. <Feature> Architecture Overview
+
+<framing paragraph>
+### A. <Pattern name>   (e.g. Server-Side / URL-Driven)
+- **Target Modules:** <real names>
+- **Single Source of Truth:** <what owns state>
+- **Mechanism:** `<hook/query/fn>` — <how controls map to state/server>
+- <pattern-specific behaviours>
+### B. <Pattern name>   (e.g. Client-Side / In-Memory / Offline-First)
+- **Target Module:** <real names>
+- **Mechanism:** `<query/wrapper>` — <fetch/hold data>
+- **Filtering & Search:** <where filtering executes>
+- **Performance / Offline:** <what to expect>
+
+## 2. Universal Smoke Test Acceptance Criteria
+
+### <Pattern A> Smoke Tests (Verify on <modules>. Note: <gotcha>.)
+
+- [ ] <interaction> → <expected observable result>
+
+### <Pattern B> Smoke Tests (Verify on <module>. Important: <gotcha>.)
+
+- [ ] <interaction> → <expected observable result>
+
+## 3. Testability & Environment Notes
+
+- **Local run:** <URLs/modules>
+- **Auth/role:** <requirement>
+- **Data/offline:** <seeded data / PWA / env differences>
+```
+
+## Telemetry
+
+When invoked via MCP skill tools, pass telemetry overrides
+`{ teamRole: "qa", actorType: "AGENT", loopRunId: "<MISSION_ID>" }` so the
+handover generation is attributed on the Agentic Health dashboard.

package/.ai/skills/reflexion-loop-local.md

@@ -0,0 +1,39 @@
+---
+name: reflexion-loop-local
+description:
+  '[LOOP · LOCAL · SAME-MODEL] Fully offline model loop with same-model
+  sequential self-critique, governed by a token and wall-clock budget.'
+phase: plan
+kind: skill
+domain: eng
+ownership:
+  drive: ai
+  approve: human
+targets:
+  - local
+minModelClass: small
+cost: ~250 tokens
+modes: [read-only, write, mcp]
+surface: public
+category: Plan & Harden
+policies:
+  - four-pillars
+---
+
+# Reflexion Loop (Local Tier)
+
+This workflow runs the exact same sequential self-correcting plan loop as the
+standard reflexion-loop, but constrained to a single model instance. Because
+local models are often smaller or slower, it uses the identical model for both
+the generator and the critic steps, governed by an absolute wall-clock limit
+(via `REFLEXION_MAX_WALLCLOCK_MS`) and a token limit.
+
+- **Constraints**: No USD budget is applied (since it runs locally for free).
+- **Enforcement**: Model isolation is downgraded to `same-model` via the
+  `TIER_POLICY`.
+
+Usage:
+
+```bash
+npm run reflexion -- --tier local "Your feature brief"
+```

package/.ai/skills/reflexion-loop-sub-max.md

@@ -0,0 +1,214 @@
+---
+name: reflexion-loop-sub-max
+description: >
+  [LOOP · SUB-MAX · NO API KEYS · CROSS-MODEL VERIFY] $100/mo tier
+  context-isolated plan hardening loop. Manages multi-vendor model isolation
+  (L0-L3) and exhaustion limits without losing work, delivering cross-model
+  verified plans without requiring API keys. (Note: The stated token cost is per
+  loop/run).
+cost: ~2400 tokens
+modes: [read-only, write, mcp]
+surface: public
+category: Plan & Harden
+how:
+  'Multi-vendor model contract, Findings Ledger, and context-firewalled critic
+  isolation'
+useCase:
+  'Plan hardening on a $100/mo subscription without requiring external API keys'
+phase: plan
+kind: skill
+domain: eng
+ownership:
+  drive: human-ai
+  approve: human
+targets: [local, api, subscription]
+minModelClass: small
+consumes: [spec]
+emits: [plan]
+suggests:
+  [clean-code, regression-bug-fix, reflexion-loop-sub-pro, reflexion-loop]
+policies:
+  - user-sovereignty
+  - diagnosis-first
+  - four-pillars
+---
+
+# Reflexion Loop ($100/mo Tier - No API Keys)
+
+> Tier siblings: reflexion-loop (API keys, dual-model) · reflexion-loop-sub-max
+> ($100 tier) · reflexion-loop-sub-pro ($20 tier). See the tier table in the
+> README.
+>
+> [!NOTE] **Tier profile ($100/mo subscription):** Hardens implementation plans
+> using multi-vendor model isolation (L0–L3) without requiring API keys. Handles
+> model exhaustion limits seamlessly while preserving verification integrity.
+
+## Pre-Flight Model Contract (MANDATORY BEFORE PHASE 0)
+
+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      |
+| ------------------ | ----------------------------- | ------------------- | ------------------------ |
+| Generator (Writer) | frontier                      | N/A (author)        | Rung 1 -> Rung 2         |
+| Critic (Auditor)   | a different vendor's frontier | L0 (cross-vendor)   | Rung 1 -> Rung 2 -> Park |
+
+### Environment Check (`CLAUDE_CODE_SUBAGENT_MODEL`)
+
+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.
+
+### Four-Level Isolation Ladder
+
+- **L0 (Cross-Vendor)**: Generator and critic run on models from different
+  vendors. Default target.
+- **L1 (Cross-Family, Same Vendor)**: Generator and critic run on different
+  model families from same vendor (shared lineage limitation apply).
+- **L2 (Fresh Sub-Agent)**: Same model, fresh sub-agent context receiving ONLY
+  plan text + rubric + diagnosis.
+- **L3 (Fresh Session)**: Same model, plan pasted into a new session cold.
+
+**Rule:** Critic is FORBIDDEN from seeing writer's drafting conversation or
+rationale.
+
+### Critic Resolution Ladder (MANDATORY, BEFORE PHASE 1)
+
+Resolve the critic with the probe, never by guessing a CLI:
+
+```bash
+./.ai/rtk-run run resolve-critic --writer <anthropic|google|openai> --writer-model <generator-model>
+# in the stack repo itself: node scripts/resolve-critic.mjs --writer <vendor> --writer-model <id>
+```
+
+`--writer-model` is required for a Claude writer so the Claude rung picks a
+different model. The probe is the single source of truth: it smoke-tests each
+CLI rung in order and prints JSON for the first that answers. Every rung works
+on a subscription login or a pay-as-you-go API key:
+
+1. **Rung G — Gemini** (L0; skipped for a Google writer): `rung: "gemini-cli"`
+   (standalone `gemini` with `GEMINI_API_KEY`, Vertex AI or enterprise Code
+   Assist), else `rung: "agy"` (Antigravity CLI, consumer Google plans).
+2. **Rung X — Codex** (`rung: "codex"`, L0; skipped for an OpenAI writer):
+   `codex exec` on a ChatGPT plan or an OpenAI API key.
+3. **Rung C — Claude** (L0, or L1 for a Claude writer):
+   `rung: "claude-subagent"` inside Claude Code — spawn a fresh sub-agent on
+   exactly the probe's `model`, never the writer's; otherwise
+   `rung: "claude-cli"` (`claude -p` on a Claude plan or `ANTHROPIC_API_KEY`).
+4. **Rung H — Another harness model** (probe returned `rung: "harness"`, or the
+   probe could not run): pick a different vendor (L0) or family (L1) in the
+   harness / sub-agent model setting, preferring Claude.
+5. **Rung S — Same model** (nothing else available): fresh sub-agent (L2) or
+   fresh session (L3). Verdict is forced to `PROVISIONAL` and `state.json` MUST
+   carry `criticAdvisory` (below). Use STATE 2, with its second line naming the
+   unavailable rungs instead of a usage limit.
+
+For a CLI rung, run the probe's `command` array with the critic prompt (plan +
+rubric + diagnosis only) appended as its **final argument**, e.g.
+`<command...> "$(cat .loop-out/<runId>/critic-prompt-r1.txt)"`. Never pipe the
+prompt on stdin — `agy` ignores stdin.
+
+> [!WARNING] The standalone `gemini` CLI stopped serving personal Google logins
+> (AI Pro/Ultra, free Code Assist) on 18 June 2026; a pay-as-you-go
+> `GEMINI_API_KEY` still works. A `GOOGLE_CLOUD_PROJECT` error from it means
+> **unsupported account type**, not a missing setting — the probe moves on to
+> `agy`. Never ask the user to set `GOOGLE_CLOUD_PROJECT` unless they confirm an
+> enterprise Code Assist licence.
+
+Record the probe output plus the final rung as `criticResolution` in
+`state.json` (on probe failure, set `probeError` and start at Rung H). On Rung S
+also write:
+
+```json
+"criticAdvisory": {
+  "level": "STRONG",
+  "code": "SAME_MODEL_CRITIC",
+  "message": "The critic ran on the writer's model. This audit is NOT independent; treat the score as self-assessment.",
+  "rungsTried": ["<criticResolution.skipped[].rung: reason>", "harness: <why no other model>"],
+  "remediation": "Set GEMINI_API_KEY or sign in to agy, sign in to codex, or install the claude CLI, then re-run the critic and deep-review this output before use."
+}
+```
+
+## Quota Discipline & Findings Ledger
+
+- **Turn Budget:** 12 agent turns per run. Print
+  `[turn N/12 | model: <active-model>]`.
+- **Findings Ledger (`.loop-out/<runId>/findings.md`):** Write stack facts, file
+  paths, domain boundaries, decisions taken, and **OPTIONS REJECTED WITH
+  REASONS** (mandatory) to survive model swaps without re-discovery.
+- **State File (`.loop-out/<runId>/state.json`):** Checkpoint after EVERY phase.
+  Include `generatorModel`, `criticModel`, `activeIsolationLevel`,
+  `criticResolution`, `criticAdvisory` (null unless Rung S), and `status`.
+- **Cold Resume Protocol:** If `.loop-out/<runId>/state.json` exists, read it
+  and Findings Ledger, then resume from recorded phase. **Never re-run Phase 0
+  discovery on resume.**
+
+## Model Continuity, UNREVIEWED vs PROVISIONAL Verdicts
+
+### Exhaustion Modes A/B/C & Fallback Ladder
+
+- **Mode A (MODEL-SCOPED):** Quota spent on one model. Try Rung 1 (cross-vendor
+  frontier) -> Rung 2 (same vendor lower class). Recompute isolation level. If
+  the critic is exhausted, continue the Critic Resolution Ladder from the NEXT
+  rung and update `criticResolution`.
+- **Mode B (ACCOUNT-WIDE):** Consolidate into Findings Ledger, checkpoint state
+  as `PARKED`, report reset window. State 3 disclosure applies.
+- **Mode C (SILENT DOWNGRADE):** Poll model identity at every phase boundary.
+  Any change is treated as a swap event.
+- **PROVISIONAL Verdict:** If plan passed critique under a degraded critic (same
+  model or L2/L3 isolation), mark verdict as `PROVISIONAL`. PROVISIONAL plans
+  require explicit Tech-Lead sign-off.
+- **UNREVIEWED Verdict:** If the run stopped before the critic ran or completed
+  (Mode B park), mark status as `UNREVIEWED`.
+
+## Three Mandatory End-State Disclosures (VERY FIRST LINE OF VERDICT OUTPUT)
+
+The FIRST line of `.loop-out/<runId>/verdict.md` and final chat 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 <generator-model>, audited by <critic-model>.
+  ```
+
+- **STATE 2 — Separation Lost (Auditor FINISHED, but on the writer's model):**
+
+  ```text
+  MODEL SEPARATION LOST: <generator-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 Compliance & Anti-Rationalization
+
+### Anti-Rationalization Rebuttals
+
+| Rationalization                                                               | Rebuttal                                                                                         |
+| ----------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------ |
+| "The plan looks fine, let me skip the rubric."                                | Unscored is unhardened. Produce the rubric.                                                      |
+| "I'll harden it after I start coding."                                        | Rework after code is 10x more expensive.                                                         |
+| "I have no second model so the loop is pointless."                            | Context isolation is degraded, not absent — declare level and proceed.                           |
+| "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.                                              |
+
+## Exit States
+
+Concludes in one of: `PASSED`, `PROVISIONAL`, `PARKED`, `CAPPED`, `ABORTED`.

package/.ai/skills/reflexion-loop-sub-pro.md

@@ -0,0 +1,164 @@
+---
+name: reflexion-loop-sub-pro
+description: >
+  [LOOP · SUB-PRO · NO API KEYS · CROSS-MODEL VERIFY] $20/mo tier
+  context-isolated loop. Single-pass cross-model plan check enforcing Mode B
+  quota handling and mandatory disclosure without requiring API keys.
+cost: ~1850 tokens
+modes: [read-only, write, mcp]
+surface: public
+category: Plan & Harden
+how:
+  'Single-pass Generator/Critic model contract, Mode B consolidate-and-park, and
+  mandatory three-state end-state disclosure'
+useCase:
+  'Frugal single-pass plan verification on a standard ($20/mo) subscription
+  without API keys'
+phase: plan
+kind: skill
+domain: eng
+ownership:
+  drive: human-ai
+  approve: human
+targets: [local, api, subscription]
+minModelClass: small
+consumes: [spec]
+emits: [plan]
+suggests:
+  [clean-code, regression-bug-fix, reflexion-loop-sub-max, reflexion-loop]
+policies:
+  - user-sovereignty
+  - diagnosis-first
+  - four-pillars
+---
+
+# Reflexion Loop ($20/mo Tier - No API Keys)
+
+> Tier siblings: reflexion-loop (API keys, dual-model) · reflexion-loop-sub-max
+> ($100 tier) · reflexion-loop-sub-pro ($20 tier). See the tier table in the
+> README.
+>
+> [!NOTE] **Tier profile ($20/mo subscription):** Honest promise: **"One good
+> adversarial pass"**, not a hardened plan. Operates as a single-pass
+> cross-model check on a 5-turn budget without API keys.
+
+<!-- -->
+
+> [!NOTE] **Frugal Compression Notice:** Multi-lane rebalancing is **OMITTED**.
+> The L0–L3 ladder, fallback rungs, and Findings Ledger structure are
+> **COMPRESSED**. Detailed mechanics reference `reflexion-loop-sub-max` by name.
+> 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**.
+
+## 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 |
+| ------------------ | ----------------------------- | ------------------- | ------------------- |
+| Generator (Writer) | frontier / mid                | N/A (author)        | Consolidate & Park  |
+| Critic (Auditor)   | a different vendor's frontier | L0 (cross-vendor)   | L1 -> L2 -> Park    |
+
+> **Sub-Pro Note:** Sub-pro is **throughput-limited** (one critique pass,
+> 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 (COMPRESSED)
+
+- **L0 (Cross-Vendor)**: Different vendor models. **L1**: Cross-family, same
+  vendor. **L2**: Fresh sub-agent. **L3**: Cold paste.
+
+### Critic Resolution Ladder (COMPRESSED — full mechanics in `reflexion-loop-sub-max`)
+
+- Resolve the critic with
+  `./.ai/rtk-run run resolve-critic --writer <vendor> --writer-model <generator-model>`,
+  never by guessing a CLI. Order: 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.
+- Run a CLI rung's `command` with the critic prompt as its final argument; on
+  `claude-subagent`, spawn a fresh sub-agent on exactly the probe's `model`.
+- The standalone `gemini` CLI no longer serves personal Google logins (since 18
+  June 2026); its `GOOGLE_CLOUD_PROJECT` error means unsupported account type —
+  the probe moves on to `agy`.
+- Record `criticResolution` in `.loop-out/<runId>/state.json`. On the same-model
+  rung, verdict is `PROVISIONAL` and `state.json` MUST carry the STRONG
+  `criticAdvisory` defined in `reflexion-loop-sub-max`.
+
+## Single-Pass Mechanics & Quota Discipline
+
+_(Note: The limits below are generated/derived — see `TIER_POLICY['sub-pro']` in
+`src/lib/ai/tier-policy.ts` for the authoritative code policy.)_
+
+- **Deltas:** 1 critique pass, max 1 revision, pass threshold 7/10, budget 5
+  turns (`[turn N/5 | model: <active-model>]`), plan <= 400 words / <= 8 tasks.
+- **Risk-2 Refusal:** Risk signal = 2 (auth/payments/data/infra) -> MUST refuse
+  single-pass hardening and escalate to `reflexion-loop-sub-max` (Enforced by
+  `tier-policy.ts`).
+
+## Mode B Consolidate-and-Park Protocol (FULL)
+
+On encountering Mode B account-wide limit, or reaching 5 turns:
+
+1. Write brief, diagnosis, score, decisions, and **OPTIONS REJECTED WITH
+   REASONS** (mandatory) to `.loop-out/<runId>/loop.md`.
+2. Mark status as `UNREVIEWED` (or `PROVISIONAL` if degraded critic was used).
+3. Report progress and reset window in State 3 disclosure.
+
+## Three Mandatory End-State Disclosures (FULL — VERY FIRST LINE OF OUTPUT)
+
+The FIRST line of `.loop-out/<runId>/loop.md` and final chat 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 <generator-model>, audited by <critic-model>.
+  ```
+
+- **STATE 2 — Separation Lost (Auditor FINISHED, but on the writer's model):**
+
+  ```text
+  MODEL SEPARATION LOST: <generator-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
+
+| Rationalization                                                               | Rebuttal                                                                                         |
+| ----------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------ |
+| "The plan looks fine, let me skip the rubric."                                | Unscored is unhardened. Produce the rubric.                                                      |
+| "Risk-2 task can be done in Sub-Pro."                                         | Risk-2 MUST refuse single-pass hardening. Escalate.                                              |
+| "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.                                              |
+
+## Exit States
+
+Concludes in one of: `PASSED`, `PROVISIONAL`, `UNREVIEWED`, `PARKED`, `CAPPED`,
+`ABORTED`.