CLI Forge Chef

Optimization: Heavy content lives in references/. Load on demand.

Language rule: Skill instructions are written in English. When generating user-facing files (prompts, shared-state, tmuxinator comments, reports), detect the project's primary language (from README, comments, docs, commit messages) and produce those files in that language. If the project is bilingual, ask the user which language to use before proceeding. The Brigade vocabulary (Menu, Commis, Sous-Chef, Chef, Plat, Mise en place, etc.) stays in French regardless — it's the pattern's canonical terminology, not prose.

Gotchas: Read ../gotchas.md AND references/gotchas-chef.md before producing output.

CLI Forge Boss — Brigade de Cuisine

"On ne cuisine pas seul. On cuisine en brigade." — Auguste Escoffier ("You don't cook alone. You cook in a brigade.")

The Brigade

┌─────────────────────────────────────────────────────┐
│                      CLIENT                          │
│  (user, ticket, roadmap)                             │
│                                                      │
│  Orders plats via the MENU (PERT)                    │
│  "I want: decision tokens + E2E + sokolsky"          │
└──────────────────────┬──────────────────────────────┘
                       │ order
┌──────────────────────▼──────────────────────────────┐
│                  CHEF DE CUISINE                     │
│  (plans, decides, orchestrates)                      │
│                                                      │
│  - Reads the menu (PERT / roadmap)                   │
│  - Assigns plats to the commis                       │
│  - Decides the order in which to send them out       │
│  - Asks the Sous-Chef to taste before sending        │
│  - Calls "Envoyez !" when it's ready                 │
│  - NEVER COOKS                                       │
└──────────────────────┬──────────────────────────────┘
                       │ SendMessage
┌──────────────────────▼──────────────────────────────┐
│              3 VOTING SOUS-CHEFS                     │
│  (taste, judge — quorum 2/3)                         │
│                                                      │
│  sous-chef-scope   : is the plat in your station?    │
│  sous-chef-secu    : any broken glass in the plat?   │
│  sous-chef-qualite : does it taste right?            │
│                                                      │
│  2/3 APPROVE → passes                                │
│  1 DENY → sent back to the kitchen                   │
│  1 ESCALATE → call the patron (human, ~5% of cases)  │
├──────────────────────────────────────────────────────┤
│              SOUS-CHEF MERGE                         │
│  (plates up, sends to the pass)                      │
│                                                      │
│  - Quality gates (/cli-audit-*)                      │
│  - Merge + CI (tmux send-keys gate)                  │
│  - Updates the carnet (shared-state.md)              │
│  - Runs in bypassPermissions                         │
│  - CWD = {project}-wt-gate                           │
│    (push + PR allowed, apply DENIED)                 │
├──────────────────────────────────────────────────────┤
│              APPLY PANE (only if infra present)      │
│  (executes tofu/helm/kubectl after 3/3 quorum)       │
│                                                      │
│  - Receives APPLY_REQUEST from any agent             │
│  - Runs the plan (tofu plan, helm diff, kubectl diff)│
│  - Submits plan to 3 voting sous-chefs — 3/3 needed  │
│  - Executes only after unanimous APPROVE             │
│  - Prod target → extra PATRON ACK                    │
│  - CWD = {project}-wt-apply                          │
│    (tofu/helm/kubectl allowed, git push DENIED)      │
│  - See references/apply-quorum.md                    │
└──────────────────────┬──────────────────────────────┘
                       │ SendMessage
┌──────────────────────▼──────────────────────────────┐
│              COMMIS (x N)                            │
│  (cook, prep, plate)                                 │
│                                                      │
│  - Each has their own station (worktree / project)   │
│  - Prepare their plat (code, test, commit)           │
│  - Announce "Pret !" to the sous-chef                │
│  - DO NOT DECIDE when to send out                    │
└──────────────────────┬──────────────────────────────┘
                       │ "Plat au passe"
┌──────────────────────▼──────────────────────────────┐
│              MAITRE D'HOTEL                          │
│  (pass → client, landing watchdog)                   │
│                                                      │
│  - Receives PRs once auto-merge is enabled           │
│  - Polls until MERGED (not just "auto-merge on")     │
│  - Rattrapage: rebase BEHIND branches when main moves│
│  - Relance: rerun transient CI failures (max 2x)     │
│  - Renvoi: real failures go back to the Sous-Chef    │
│  - Encaissement: tag verified + shared-state updated │
│  - Reports "Client content" only when all 6 done     │
│                                                      │
│  Read references/maitre-dhotel.md for the full role  │
└──────────────────────┬──────────────────────────────┘
                       │ "Client content + tag"
┌──────────────────────▼──────────────────────────────┐
│                   CLIENT (happy)                     │
│  Release tag cut, prod deploy green, sprint done     │
└─────────────────────────────────────────────────────┘

When the patron (human) steps in

~5% of cases. The human is called only when a Sous-Chef votes ESCALATE:

Everything else passes without human intervention.

Brigade vocabulary

Plat lifecycle (state machine)

A plat moves through a fixed sequence of states between Ready and Envoye. A DENY from a sous-chef loops back to Cuisson; a critical failure (same diff rejected twice, regression, 30-min idle) triggers Apoptosis and the plat re-enters the pool.

stateDiagram-v2
    [*] --> Ready
    Ready --> MiseEnPlace: dispatched to commis
    MiseEnPlace --> Cuisson: write-set locked in shared-state
    Cuisson --> Dressage: code + tests pass locally
    Dressage --> Gouter: commit + lint clean, sent to Sous-Chef
    Gouter --> Envoi: 2/3 APPROVE (normal) or 3/3 (sensitive)
    Gouter --> Renvoi: DENY or CONCERN with solution
    Renvoi --> Cuisson: commis applies the proposed fix
    Envoi --> PlatAuPasse: Sous-Chef creates PR with auto-merge + F8 conflict scan
    PlatAuPasse --> EnSalle: Maitre d'hotel receives the plat
    EnSalle --> Surveillance: polling loop every 45s
    Surveillance --> Surveillance: CLEAN / HAS_HOOKS / UNSTABLE (wait)
    Surveillance --> Rattrapage: BEHIND (main moved)
    Rattrapage --> Surveillance: rebase + force-push-with-lease + re-enable auto-merge
    Rattrapage --> Renvoi: rebase conflict
    Surveillance --> Relance: BLOCKED + transient CI
    Relance --> Surveillance: gh run rerun --failed
    Relance --> Escalade: 3rd transient in a row
    Surveillance --> Renvoi: BLOCKED + real failure OR DIRTY merge conflict
    Surveillance --> Encaissement: MERGED
    Encaissement --> ClientContent: tag verified + shared-state Valid merges updated + cascade Rattrapage on other in-flight
    ClientContent --> [*]: Chef notified, plat fully served
    Escalade --> [*]: > 2h timeout, Chef takes over
    Cuisson --> Apoptosis: same diff rejected 2x / idle 30min / regression
    Dressage --> Apoptosis: same diff rejected 2x / regression
    Apoptosis --> Ready: revert + locks released, plat re-enters the pool
    Gouter --> AppelAuPatron: 3 rounds without consensus
    AppelAuPatron --> Cuisson: human decision applied
    AppelAuPatron --> [*]: human aborts the plat

This is the authoritative plat lifecycle — if prose anywhere else disagrees, fix the prose.

Communication flow

Commis edits a file
  → Chef sends the diff to the 3 voting Sous-Chefs (parallel)
  → 2/3 APPROVE → passes automatically (zero human)
  → 1 DENY → sent back to the commis with a reason
  → 1 ESCALATE → Chef notifies the patron (human)

Commis finishes cooking
  → "Pret !" to the Sous-Chef (SendMessage)
  → Sous-Chef tastes (quality gates)
  → IF good:
      → Sous-Chef plates up and sends (merge + CI)
      → Sous-Chef to Chef: "Plat sent, table served"
      → Chef to dependent commis: "Envoyez the next one!"
  → IF not good:
      → Sous-Chef to Commis: "Renvoi! Too much salt in fn X"
      → Sous-Chef to Chef: "Renvoi on plat Y, commis is fixing it"

Input

$ARGUMENTS is the target project path or name.

• If a path: analyze the project, generate config for it
• If a name: create a new project directory and generate config
• If empty: use the current working directory

Phase 0 — Mise en place

First, check whether a sprint is currently running or paused:

Read references/sprint-persistence.md. Check .claude/sprint-history/current/:

• If a sprint is PAUSED → offer: Resume / Rewind / Fresh / Abandon
• If a sprint is DONE → read the gotchas-learned and sensitive zones for the new sprint
• If no history → fresh sprint

Before generating anything, understand the project:

0.1 — Detect project type

If no manifest found → ask the user: "What type of project is this? (Rust, Go, JS/TS, Python, Infra, Other)"

Use the detected type to:

• Filter build tool permissions in settings.local.json (only include relevant tools)
• Set correct build/test/lint commands in commis prompts
• Choose appropriate quality gates

0.2 — Read the project

1. Read the project — README, CONTRIBUTING.md, docs/explanation/architecture.md, manifest file, src/ structure
2. Read git state — branches, worktrees, CI config
3. Identify the menu — what needs to be done? Check issues, roadmap, design docs

0.3 — Check prerequisites

Tools

# Check all prerequisites
for tool in tmux tmuxinator claude gh; do
  which "$tool" >/dev/null 2>&1 || echo "MISSING: $tool"
done
grep -q "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS" ~/.claude/settings.json 2>/dev/null || echo "AGENT_TEAMS_NOT_ENABLED"

If a tool is missing → show the install command for the detected OS and stop.

VCS — git or jj

The project MUST be a version-controlled repository. Check:

# Detect VCS type
if [ -d .jj ]; then
  VCS="jj"
elif git rev-parse --git-dir >/dev/null 2>&1; then
  VCS="git"
else
  echo "FATAL: not a git or jj repository. Initialize with 'git init' or 'jj git init' first."
  exit 1
fi

If not a repo → stop and ask the user to init. The Chef never creates a repo — that's a user decision.

Branching model detection

Auto-detect the project's branching model. Never assume Git Flow.

# Detect GitHub repo
REPO=$(gh repo view --json nameWithOwner,defaultBranchRef -q '.nameWithOwner' 2>/dev/null)
DEFAULT_BRANCH=$(gh repo view --json defaultBranchRef -q '.defaultBranchRef.name' 2>/dev/null)

# Detect branching model by examining existing branches
HAS_DEVELOP=$(git branch -r 2>/dev/null | grep -c 'origin/develop')
HAS_RELEASE_BRANCHES=$(git branch -r 2>/dev/null | grep -c 'origin/release/')
HAS_RELEASE_PLZ=$(test -f release-plz.toml && echo 1 || echo 0)

if [ "$HAS_DEVELOP" -gt 0 ] && [ "$HAS_RELEASE_BRANCHES" -gt 0 ]; then
  BRANCH_MODEL="gitflow"
  BASE_BRANCH="develop"
  RELEASE_BRANCH="main"
elif [ "$HAS_DEVELOP" -gt 0 ]; then
  BRANCH_MODEL="github-flow-develop"
  BASE_BRANCH="develop"
  RELEASE_BRANCH="main"
elif [ "$DEFAULT_BRANCH" = "main" ] || [ "$DEFAULT_BRANCH" = "master" ]; then
  BRANCH_MODEL="github-flow"
  BASE_BRANCH="${DEFAULT_BRANCH}"
  RELEASE_BRANCH="${DEFAULT_BRANCH}"
else
  BRANCH_MODEL="trunk"
  BASE_BRANCH="${DEFAULT_BRANCH:-main}"
  RELEASE_BRANCH="${BASE_BRANCH}"
fi

Impact on the brigade:

• github-flow / trunk: commis branch from main, PRs target main, no sync-main needed, release-plz (if used) targets main directly
• github-flow-develop: commis branch from develop, PRs target develop, sync-main needed after each release tag
• gitflow: same as above, plus release branches for stabilization

GitHub repo & branch protection (G26)

if [ -z "$REPO" ]; then
  echo "WARNING: not a GitHub repo or gh not authenticated. PR mode unavailable."
  MERGE_MODE="direct"
else
  # Detect if base branch requires PRs (G26)
  PROTECTED=$(gh api repos/${REPO}/rulesets 2>/dev/null | grep -c 'pull_request')
  if [ "$PROTECTED" -gt 0 ]; then
    MERGE_MODE="pr"
    echo "DETECTED: branch protection requires PRs → Sous-Chef will use PR mode"
    
    # Check auto-merge is enabled in repo settings
    AUTO_MERGE=$(gh api repos/${REPO} --jq '.allow_auto_merge' 2>/dev/null)
    if [ "$AUTO_MERGE" != "true" ]; then
      echo "FIX: enabling auto-merge on the repo (required for PR mode)"
      gh api repos/${REPO} --method PATCH -f allow_auto_merge=true 2>/dev/null
    fi
  else
    MERGE_MODE="direct"
  fi
fi

Record ALL detected config in shared-state.md so the Sous-Chef and commis know at runtime:

## Sprint config
| Key | Value |
|-----|-------|
| VCS | {git or jj} |
| Branch model | {github-flow / github-flow-develop / gitflow / trunk} |
| Base branch | {main / develop / master} |
| Release branch | {main / master} |
| Merge mode | {direct or pr} |
| Repo | {owner/repo} |
| Default branch | {from GitHub API} |
| Sync-main needed | {yes or no} |

gh auth scopes

# Check gh has the required scopes
gh auth status 2>&1 | grep -q "repo" || echo "MISSING: gh scope 'repo'"

If the project has CI workflow files that may need editing, warn:

NOTE: pushing .github/workflows/ changes requires the 'workflow' scope.
If needed during the sprint: gh auth refresh -s workflow
After the push: gh auth refresh (to remove it)

Agent isolation model (CRITICAL — G26 + industry best practice 2026)

Every commis MUST work in an isolated worktree with its own branch. Never on the base branch directly.

RULE: 1 commis = 1 worktree = 1 branch = 1 PR

Base branch (develop or main) is NEVER modified directly by a commis.
All work flows through: worktree → branch → push → PR → auto-merge.

This is the pattern used in production by teams running 4-7 parallel agents (incident.io, 2026). It prevents:

• Commis stepping on each other's files
• Push rejected by branch protection (G26)
• Merge conflicts discovered too late
• Work stranded locally when push fails

Worktree setup per commis:

# Chef creates worktrees in Phase 0 (in on_project_start or manually)
git worktree add ../{project}-wt-{commis_name} -b {branch_name} 2>/dev/null || true

Conflict pre-detection (optional but recommended for 3+ commis):

If clash is installed (which clash), add to the ccheck or Sous-Chef loop:

clash status  # shows conflict matrix across all worktrees
clash check {file}  # check before writing a specific file

If not installed, the coupling matrix from Phase 0.5 (/cli-audit-tangle) serves as the static conflict prevention.

Merge sequencing:

When multiple commis are done:

1. The Sous-Chef merges PR 1 first (oldest or least coupled)
2. Waits for auto-merge to complete
3. Remaining commis rebase their branches: git rebase origin/{base_branch}
4. Sous-Chef merges PR 2, and so on

Never merge all PRs simultaneously — sequential merge with rebase between each avoids the quadratic conflict explosion.

Exception: missing tmuxinator is NOT a blocker. If Ruby/gem is not available (air-gapped env, immutable OS like Bluefin/Silverblue), generate the raw-tmux fallback instead:

• Read references/raw-tmux-fallback.md
• Generate {project}/.claude/scripts/boss-{session}.sh (chmod +x)
• Document in the Phase 4 report: "tmuxinator missing → raw-tmux fallback generated"
• Launch with bash {project}/.claude/scripts/boss-{session}.sh instead of tmuxinator start {session}

Auto-fix configuration (G19, G20, G22) — do ALL of these automatically, don't ask:

1. ~/.claude/settings.json: ensure CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 + skipDangerousModePermissionPrompt=true
2. ~/.claude.json: ensure "teammateMode": "tmux"
3. {project}/.claude/settings.local.json: ensure all permissions (build tools, shared-state, external repos, Obsidian vault, Team tools)
4. Tmuxinator on_project_start: ensure auto-kick message with sleep 12 && tmux send-keys ... & (G20)
5. If project has an Obsidian vault: add Edit/Read permissions for the vault path (G22)
6. Warn user: do NOT run brew upgrade claude-code while the session is running (G18)

0.4 — Choose model and brigade size (Mitosis)

Read references/simplified-model.md. Two models exist — pick based on tier. The S/M/L/XL semantics are canonical in ../shared/tiering.md — this skill's tier table specializes them to brigade composition (1 commis → full XL brigade w/ sous-chef clusters).

Tiers S/M/L: the Chef generates files in Phase 0 and then disappears. Commis self-organize via shared-state.md (Boids + stigmergy). The single Sous-Chef reviews with 3 lenses. Commis that fail self-terminate (apoptosis).

Tier XL: the full brigade is kept for regulatory compliance (vote audit trail, decision traceability).

For tier S: skip tmuxinator entirely. Use a single claude invocation with --append-system-prompt and one worktree.

Tier XL — Sous-Chef clusters (Brigades de parties)

For large projects (monorepos, multi-domain, regulated), each domain has its own cluster of 3 sous-chefs. Each cluster votes internally (2/3), then the clusters vote among themselves (2/3 normal, 3/3 sensitive).

Chef de Cuisine
  │
  ├── SECURITY cluster (3 sous-chefs) — internal vote 2/3
  │   ├── sous-chef-secrets     (DLP, credentials, env vars)
  │   ├── sous-chef-deps        (supply chain, advisories, licenses)
  │   └── sous-chef-perimeter   (scope, access, file permissions)
  │   → Produces 1 consolidated verdict: APPROVE / DENY+solution
  │
  ├── QUALITY cluster (3 sous-chefs) — internal vote 2/3
  │   ├── sous-chef-coherence   (mission, design, architecture)
  │   ├── sous-chef-tests       (coverage, pyramid, regression)
  │   └── sous-chef-proprete    (code smells, dead code, naming)
  │   → Produces 1 consolidated verdict
  │
  └── OPS cluster (3 sous-chefs) — internal vote 2/3
      ├── sous-chef-ci          (pipelines, workflows, matrix)
      ├── sous-chef-deploy      (infra, docker, helm)
      └── sous-chef-perf        (benchmarks, timeouts, resources)
      → Produces 1 consolidated verdict

Inter-cluster vote:
  Normal zone: 2/3 clusters APPROVE → passes
  Sensitive zone: 3/3 clusters APPROVE → passes
  DENY: the cluster that DENYs proposes a solution
  Resolution rounds: identical to tier M

Tier XL means 9 sous-chefs + 1 merge = 10 validation agents. Use it only when the cost of errors justifies the cost of the agents (regulated, defense, finance).

Selection logic (the Chef decides in Phase 0):

The Chef detects the tier at startup:
  - Number of tasks → S/M/L
  - Multi-repo → L minimum
  - CONTRACTS.md file or compliance/ → XL
  - Mention of "regulated", "defense", "finance" in README/CONTRIBUTING.md → XL
  - User can force: /cli-forge-chef --tier XL

Cost confirmation for tier XL (MANDATORY — auto-detect or flag):

Tier XL costs roughly 5-8x the tier L token spend (10 validation agents instead of ~2, each diff voted on by 9 sous-chefs instead of 3, audit trail generated every round).

Before spawning the XL brigade, always display and ask for confirmation:

=== TIER XL DETECTED ===
Reason: {detection_reason}  # e.g., "CONTRACTS.md present" or "--tier XL flag"
Brigade: Chef + 9 Sous-Chefs (3 clusters) + 1 Sous-Chef + N Commis
Estimated cost: ~5-8x tier L (each diff voted on by 9 agents instead of 3)
Estimated duration: ~2-3x tier L (inter-cluster resolution rounds)

Recommended justification:
  - Regulated project (defense, finance, healthcare)
  - Legal audit trail required
  - Cost of errors > cost of agents

Continue with XL? [y/N/downgrade-to-L]

If the user answers downgrade-to-L, switch to tier L and note in the Phase 4 report: "Tier XL detected but downgraded to L at user request".

This confirmation is as mandatory as the Phase 0.6 confirmation for EXPLORATION.

0.5 — Build coupling matrix

Read references/conflict-resolution.md. Before assigning tasks:

1. Run /cli-audit-tangle on the project
2. Build the coupling matrix (file × task: R/W)
3. If W/W conflict on the same file → assign to the same commis OR sequence in the PERT
4. Mark "hot files" (Cargo.toml, ci.yml, mod.rs) → managed by the Sous-Chef only

0.6 — Identify exploration candidates

Read references/parallel-exploration.md. For each task, the Chef asks:

• Is there a non-trivial architecture choice? → mark as EXPLORATION
• Is there a complex bug with unknown root cause? → mark as HYPOTHESES
• Did the user ask to "compare approaches"? → mark as EXPLORATION

EXPLORATION tasks use 2-3 commis on the same problem with different approaches. The Sous-Chef runs the comparison grid and the Chef picks the winner. Confirm with user before launching (2-3x token cost).

Phase 1 — The menu (auto-detection, DO NOT ASK)

LANGUAGE RULE: run git log --oneline -10 + head -20 README.md. If the project is in French → ALL generated files (prompts, shared-state, tmuxinator comments, reports) are in French. If English → English. The Brigade vocabulary (Menu, Commis, Sous-Chef) stays in French regardless of the project language.

COMMITS RULE: commis and the Sous-Chef use cli-git-conventional for ALL commits. Ghostwriter style, zero AI marker, project language.

RULE: do NOT ask any question if the answer is in the project. Read EVERYTHING before asking anything:

1. Which plats? → Read in this order until found:

   • {project}/.claude/shared-state.md "Backlog" section (work from the previous sprint)
   • Obsidian roadmap (if vault is specified in shared-state)
   • gh issue list (open issues)
   • git log --oneline -20 (recent work, what's missing)
   • README.md TODO/roadmap section
   • CONTRIBUTING.md roadmap/next steps section Only if NOTHING found → ask

2. How many commis? → Count the tasks found in step 1, apply the tier (S/M/L/XL). Don't ask.

3. Sending order? → Compute the PERT. Read references/pert-computation.md end-to-end. Summary:

   • First, check for .claude/tangle-partition.json (exported by cli-audit-tangle). If present, each plat is matched to its Fiedler cluster and one commis owns one cluster for the sprint — file-exclusion becomes structural, not a runtime filter. Boundary functions become mandatory fan-in barriers in the PERT. If absent, run /cli-audit-tangle first on any sprint with > 5 plats; otherwise fall back to file-level coupling detection.
   • Edges: Fiedler boundaries (if artefact present) + cli-audit-tangle couplings + explicit semantic dependencies. W on a file created by A → depends on A. Otherwise → parallel.
   • For each plat, estimate O/M/P from the tier table (XS/S/M/L/XL based on file count + LOC + sensitive zone). Compute E=(O+4M+P)/6 and σ=(P-O)/6.
   • Forward/backward pass → ES, LS, slack per plat. Critical path = slack 0.
   • Dispatch priority = longest path from the plat to done (not readiness alone). Critical-path plats go to the first free commis.
   • File-exclusion filter: never dispatch two plats whose write-sets intersect. With a Fiedler partition, this check is already satisfied by the cluster-based assignment and becomes a sanity check only.
   • Write the PERT (Mermaid flowchart LR with :::critical, see pert-computation.md §5) + the companion text table to shared-state.md under ## PERT. This is the scheduling contract for the whole sprint. Don't ask.

4. Plats off-menu? → Scan:

   • ls {workspace}/ for neighbouring repos
   • shared-state.md "Shared context" section (links to other repos)
   • README.md / CONTRIBUTING.md (references to other projects) Don't ask.

5. What tasting level? → Derive from the tier:

   • S → /cli-audit-code only
   • M → minimum viable (code + drift)
   • L → standard (code + drift + test + sync)
   • XL → complete Don't ask.

6. Service branch? → git branch -a | head + CONTRIBUTING.md. It's always written down. Don't ask.

Only ask the user if:

• No source contains the information (no roadmap, no issues, no backlog)
• There is a critical ambiguity (2 contradictory roadmaps)
• Tier XL is detected (token cost confirmation)

Present the deduced menu and ask for confirmation: "Here's what I'm going to launch. OK?"

Phase 2 — The brigade sets up

Read references/templates.md for templates. Generate:

2.1 — The carnet de cuisine ({project}/.claude/shared-state.md)

Read references/shared-state-template.md and customize.

2.2 — The Chef's instructions ({project}/.claude/prompts/chef-{session}.md)

ABSOLUTE REQUIREMENT: read references/chef-prompt-template.md BEFORE generating the prompt. The generated prompt MUST contain the 3 voting Sous-Chefs + the Sous-Chef. If the prompt does not contain "sous-chef-scope", "sous-chef-secu", "sous-chef-qualite", "sous-chef" → the prompt is INVALID.

The Chef prompt MUST include, in this order:

1. TeamCreate
2. Spawn the 3 voting Sous-Chefs (scope, secu, qualite) — copy VERBATIM from the template
3. Spawn the Sous-Chef — copy VERBATIM from the template
4. Voting protocol with resolution rounds — copy VERBATIM from the template
5. Spawn the N Commis
6. PERT
7. Shutdown protocol (with sensitive-zone review)

The Chef:

• Creates the team (TeamCreate)
• Spawns 3 voting Sous-Chefs + 1 Sous-Chef + N Commis
• Commis permissions flow through the 3 Sous-Chefs (quorum 2/3 normal, 3/3 sensitive)
• The Sous-Chef handles quality gates, merges, CI
• Produces the service report

2.3 — The 3 voting Sous-Chefs (MANDATORY — embedded in the Chef prompt)

COPY the prompts of the 3 sous-chefs from references/chef-prompt-template.md section "Spawn the 3 Sous-Chefs".

Each votes independently: APPROVE / DENY+solution / CONCERN+solution. Resolution rounds if no consensus. Human escalation < 2%.

Prompts are CLI surfaces — ask once, default sensibly, recap before action, ship as committable files. The canonical 4 Laws + surface mapping is ../shared/cli-ergonomics.md.

2.4 — The Sous-Chef (MANDATORY — embedded in the Chef prompt)

COPY the prompt from references/chef-prompt-template.md section "Spawn the Sous-Chef".

Handles quality gates, merges, CI. Does not vote on permissions. On PR creation, runs the F8 parallel conflict scan (cli-forge-github), enables auto-merge, and hands off to the Maître d'hôtel — the Sous-Chef does NOT wait for the PR to land.

2.4b — The Maître d'hôtel (MANDATORY when tier ≥ M, release-plz present, or branch protection active)

COPY the prompt from references/chef-prompt-template.md section "Spawn the Maître d'hôtel". Read references/maitre-dhotel.md for the full role specification.

The Maître d'hôtel is the brigade's post-merge landing watchdog. He receives every PR the Sous-Chef hands off (with auto-merge already enabled), polls the GitHub merge queue every 45 s, and stays on the plat until it is MERGED, the release tag is cut (if applicable), cascading BEHIND rebases are done on every other in-flight PR, shared-state.md "Valid merges" is updated, and the Chef has received the "Client content" signal.

Why mandatory at tier ≥ M: without a Maître d'hôtel, "auto-merge enabled" is mistaken for "merged". The operator ends up rebasing PRs by hand every time release-plz bumps a version, exactly what this brigade pattern is designed to avoid. One Maître d'hôtel per brigade, centralized polling, cheap (Sonnet, no judgement).

Scale rule (from maitre-dhotel.md §7):

2.5 — Commis station sheets (embedded)

Each commis prompt includes:

• Shared state path (absolute)
• Mission (the plat to prepare)
• Steps (the recipe)
• SendMessage to the Sous-Chef when ready (NOT to the Chef)
• Permissions go through the quorum of the 3 Sous-Chefs (the commis does not know this)

2.5b — Setup script (MANDATORY — G35, G36)

Generate one script into {project}/scripts/:

• scripts/brigade-setup-worktrees.sh — copy verbatim from references/brigade-setup-worktrees.sh, chmod +x. Idempotent worktree + permission setup, recovers from orphan state, avoids SIGPIPE on git worktree add (G35), one branch per commis (G36).

The script creates every role's worktree, installs the per-worktree .claude/settings.local.json files, and creates the integration branch. It is called by on_project_start in the tmuxinator YAML.

Note on the previous launcher wrapper: earlier iterations of this skill also generated scripts/brigade-launch-agent.sh to work around $(cat prompt.md) shell-mangling (G34). That wrapper is no longer generated — use the native --append-system-prompt-file flag directly in the tmuxinator YAML (see §2.5c). The wrapper was overkill.

2.5c — Tmuxinator (~/.config/tmuxinator/{session}.yml)

Read references/tmuxinator-template.md.

# Chef — use --append-system-prompt-file (native, G34-safe)
- claude
    --dangerously-skip-permissions
    --permission-mode bypassPermissions
    --teammate-mode tmux
    --append-system-prompt-file {project}/.claude/prompts/chef-{session}.md
    "Demarre IMMEDIATEMENT Phase 0 selon le system prompt."

No wrapper script, no $(cat ...), no .initial.txt files.

2.6 — The contre-chef / ccheck (MANDATORY — G24)

Generate {project}/.claude/prompts/ccheck-{session}.md.

Read references/ccheck-prompt-template.md and customize with the project paths.

The ccheck is a dedicated Claude instance in its own tmux window that watches the Chef pane and auto-approves UI keyboard permissions in normal zones. Without it, the Chef blocks on every worker edit and the brigade stalls.

The ccheck:

• APPROVES edits in normal zones (src/, tests/, shared-state.md, docs/)
• SKIPS edits in sensitive zones (CI, deps, auth, secrets) — the Chef stays blocked, the user decides later
• LOGS every decision to /tmp/{session}-ccheck.log
• Re-reads the sensitive zone list every iteration (so the Chef can add zones mid-sprint)

This replaces the Phase 5 /loop approach which was fragile (depended on the user's session staying open) and optional (it should never have been optional).

2.6b — The contre-chef-inter (MANDATORY at tier ≥ M — G29)

Why this exists : ccheck handles only UI keyboard prompts (Do you want to make this edit?). In Agent Teams mode (tier ≥ M), commis request permissions through the team protocol — these appear as Permission request sent to team "<team>" leader cards that cannot be approved via keyboard. The Chef (team leader) must approve via SendMessage tool call. If the Chef is in a thinking loop or busy, these requests queue up for 30+ min and block commis. Observed on grob-s5 (2026-04-17) : commis-2 waited 45 min for a shared-state.md edit on normal zone.

Generate {project}/.claude/prompts/contre-chef-inter-{session}.md.

Read references/contre-chef-inter-prompt-template.md and customize with the project paths.

The contre-chef-inter is a Haiku-powered team member in its own tmux window that:

• Scans the Chef pane every 20s for Agent Teams permission cards
• Extracts (commis, tool, file) tuple from each pending card
• Classifies zone (normal / sensitive) via shared-state BOSS_SENSITIVE_PATHS
• Sends a SendMessage to the Chef with pre-chewed APPROVE or HOLD recommendation
• Escalates to user after 100s+ of persistent stall despite nudges
• Logs every decision to /tmp/{session}-inter.log

Scale rule:

Cost: ~$0.05 per 6h sprint (Haiku 4.5, ~50 iterations × ~2k input tokens).

Why different from ccheck: ccheck sends tmux send-keys Enter (keyboard); contre-chef-inter sends SendMessage (team protocol). They handle orthogonal permission pathways. Both are needed at tier ≥ M.

Note on tmux send-keys as rung 5: keyboard injection into a UI is rung 5 of ../shared/escalation-ladder.md — used here only because the underlying Claude Code UI exposes no programmatic permission-grant API (rungs 1-4 don't exist for this specific flow). It is gated on the ccheck's read-the-diff decision and logged for audit. Never extend this pattern to operations that have a real API.

2.7 — Permissions (one settings.local.json per worktree)

Read references/permissions-template.md.

Every role gets its own .claude/settings.local.json inside its own worktree — BUT the Agent Teams reality is more nuanced than « one file per role, filesystem-isolated » (see G32). In practice:

• Standalone claude panes (chef, ccheck, contre-chef-inter, apply pane): DO read their own cwd's settings.local.json. These are separate claude processes, each with its own permission file. Filesystem isolation works for them.
• Agent Teams teammates (sous-chefs, commis, maître-d'hôtel spawned via TeamCreate/Agent): DO NOT support cwd: on spawn. They inherit the team lead's (Chef's) settings.local.json. The per-worktree files for these roles exist on disk (useful for documentation and for potential future migration to standalone panes) but they are ignored at runtime.

Practical implication: the Chef's root settings.local.json is the only one that actually gates commis / sous-chef Bash commands. Put the strictest deny list there — teammates inherit it. Allow dangerous primitives only in standalone pane files (gate, apply, maitre).

Generate the following files (see the template for exact contents):

…(truncated)