mission

name: mission description: "Orchestrate multi-phase development missions. Use when the user asks to 'start a mission', 'run a mission', or uses /mission with a description. Supports standard (6-phase) and minimal (3-phase) workflows with planning, implementation, testing, audit, and verification." argument-hint: "[description | status | skip | pause | resume | done | reset]"

Mission — Multi-Phase Development Orchestrator

Orchestrate structured development workflows with distinct phases: plan, review, implement, test, audit, and verify. Each phase has a specialized protocol that controls behavior, ensuring disciplined execution.

Argument Routing

Parse $ARGUMENTS to determine the action:

$ARGUMENTS

Reset signal: If reset is run while a phase is active, log a user-signal with type:"reset_mid_flow", role:"<current active phase's dominant role>", delta:-20 before clearing state: node "$MISSION_SCRIPT" user-signal '{"role":"<role>","phase":"<name>","type":"reset_mid_flow"}'

End-of-mission rating: On the final phase transition, for Low/Medium autonomy: prompt the user for a 1–5 rating (use AskUserQuestion on Claude Code; plain-text STOP on other tools), then call node "$MISSION_SCRIPT" rate-mission '{"rating":<n>}'. For High autonomy: call node "$MISSION_SCRIPT" rate-mission '{"skipReason":"high_autonomy"}'. The phase-transition command then prints the full Mission Scorecard and merges the run into $XDG_CONFIG_HOME/mission/profile.json (default: ~/.config/mission/profile.json).

Starting a New Mission

1. Check for Existing Mission

Read .missions/active-mission.json. If an active (non-completed) mission exists, inform the user and ask if they want to overwrite it. Do not proceed without confirmation.

1.5. Detect Host Tool

Run node "$MISSION_SCRIPT" detect-tool to identify the current AI coding tool. The script checks env vars in this cascade:

1. $CLAUDECODE=1 → claude-code
2. any $CODEX_* env var → codex
3. $AMP_API_KEY or any $AMP_* env var → amp
4. any $OPENCODE_* env var → opencode
5. cached profile value (if <30 days old and matches) → use cached
6. nothing matched → unknown

If the result is unknown and AGENTS.md is present, use AskUserQuestion (or plain-text STOP) to ask: "Which tool are you running this mission from? codex / opencode / other". Once known, call node "$MISSION_SCRIPT" detect-tool --confirm <tool> to persist it in the profile.

If the result is already claude-code, codex, amp, or opencode (either from env or cache), no user prompt is needed — call detect-tool --confirm <tool> only if the env-detected tool differs from the cached value.

2. Interactive Setup

Use AskUserQuestion to gather configuration:

If your tool doesn't have AskUserQuestion (Codex, OpenCode, etc.): present the question and options as plain text and STOP. Wait for the user to reply with their choice before continuing. Do not pick a default and proceed.

Question 1 — Template (read references/templates.md first for descriptions):

• Feature — adding new functionality (Standard, Medium autonomy)
• Bug Fix — diagnosing and correcting a defect (Minimal, Low autonomy)
• Refactor — restructuring without behavioral change (Standard, Medium autonomy)
• Investigation — exploring an unknown area (Minimal, High autonomy)
• Custom — no template, configure manually

If a template is selected, apply its default mode and autonomy. Skip Questions 2 and 3 unless the user wants to override.

Question 2 — Mode (skip if template selected):

• Standard (6 phases): Architect, Review Plan, Implement, Test, Audit, Verify — full rigor for significant features
• Minimal (3 phases): Plan, Build, Verify — lightweight for smaller tasks

Question 3 — Autonomy (skip if template selected):

• Low: Pause after every phase, wait for explicit "continue"
• Medium: Pause at phase boundaries for status check (Recommended)
• High: Run to completion, only stop on critical failures

Question 4 — Constraints (optional): Ask: "Any constraints or out-of-scope boundaries? (e.g., 'don't touch auth module', 'no new dependencies'). Press enter to skip."

Question 5 — Model Assignment:

Run node "$MISSION_SCRIPT" load-model-defaults to load the saved map for the detected tool from the profile. Then present it:

Host tool: <tool> (auto-detected)
Loaded defaults from $XDG_CONFIG_HOME/mission/profile.json:

  Explorer           <model>
  Planner            <model>
  Worker             <model>
  Business Reviewer  <model>
  Security Reviewer  <model>
  Edge Case Reviewer <model>
  Reviewer           <model>
  Verifier           <model>

Override any role? (reply "defaults" to accept, or "role=model" lines)
If you override: save to profile for future missions? [yes / no / this-role-only]

Wait for the user's reply. Then:

• Merge any overrides into the map and store in active-mission.json as modelAssignment.
• If user replied yes: call node "$MISSION_SCRIPT" save-model-defaults '<full-map-json>' to persist the full updated map.
• If user replied this-role-only: call save-model-defaults with only the overridden role(s).
• If user replied no or defaults (no overrides): skip the profile write.

Model IDs: Use full, dated model IDs (e.g. claude-haiku-4-5-20251001, claude-opus-4-8, claude-sonnet-4-6) — not short aliases (haiku/opus/sonnet), which can resolve to a stale snapshot (Claude Code issue #25588). The values shown above must match DEFAULT_MODEL_DEFAULTS["claude-code"] in scripts/mission-state.mjs; keep them in sync when either changes.

Note: Mechanical checks (tests, lint, TODO scan, secret detection) are now run by scripts/mission-checks.mjs rather than LLM agents. This saves ~25K tokens per standard mission — the verifier role focuses on reasoning over script output.

Optional — aggressive token modes (off by default): for tight-budget runs the user may opt into extra Audit-phase savings by setting optimizations: { "gateSecurityReviewer": true, "microMissionMode": true } in active-mission.json. These skip/merge reviewers based on mechanical scope detection and carry real quality risk — only enable when asked, and never gateSecurityReviewer on a repo handling auth/PII/payments. See references/protocol-audit-aggressive.md. Auto-gating of the Async and Performance reviewers by scope is always on and is safe (it dispatches when uncertain).

If your tool doesn't have AskUserQuestion (Codex, OpenCode, etc.): present the defaults as plain text and STOP. Wait for the user to reply ("defaults" or overrides) before continuing.

Question 6 — Test/Lint Commands (optional): Ask: "What commands run your tests and linter? (e.g., 'test: pytest, lint: ruff check'). Press enter to auto-detect."

If provided, store as checks in the state file: { "test": "<cmd>", "lint": "<cmd>" }. mission-checks.mjs reads this first before auto-detection.

Question 7 — Token modes (optional, default OFF): Ask: "Enable extra token-saving modes? Each trades some audit thoroughness for cost. Press enter to keep all off (recommended)."

• Verifier on Haiku — final Verify gate runs on Haiku 4.5 (it mostly reads scripted check output). Set modelAssignment.verifier to "claude-haiku-4-5-20251001".
• Planner on Sonnet — spec written by Sonnet instead of Opus for tight-budget/well-scoped missions. Set modelAssignment.planner to "claude-sonnet-4-6".
• Gate Security reviewer by scope — skip it when the diff touches no auth/input/IO/crypto. Set optimizations.gateSecurityReviewer: true.
• Micro-mission mode — merge Async+Perf reviewers on tiny diffs. Set optimizations.microMissionMode: true.
• JSON audit synthesis — reviewers emit JSON; a script merges/dedups findings. Set optimizations.jsonSynthesis: true.

Store enabled flags under optimizations and/or the model overrides in active-mission.json. Never enable gateSecurityReviewer on a repo handling auth/PII/payments. Full risk notes: references/protocol-audit-aggressive.md. (The Async/Perf scope-gating in the Audit phase is always on and is safe — it dispatches when uncertain.)

3. Read Project Instructions & Set Up Worktree

Read the project instructions file (if present): Check the project root and any parent directories for a tool-specific instructions file. Read whichever is present (in priority order):

• CLAUDE.md — Claude Code
• AGENTS.md — Codex, OpenCode, and other OpenAI-compatible tools
• amp.md — Amp
• .cursor/rules — Cursor
• Any other agent config file your tool uses

This file contains project-specific instructions, conventions, and constraints that override generic behavior. Apply everything in it throughout the mission. When protocol files say "project instructions file", they mean whichever file was found here.

Set up a git worktree (auto, skippable): After reading project instructions, create an isolated worktree for the mission:

node ${CLAUDE_PLUGIN_ROOT:-.}/skills/git-worktree/scripts/worktree-manager.mjs \
  create mission/<mission-slug>

Parse the trailing JSON line {"worktreePath": "...", "branch": "..."} and store worktreePath in the mission state (see §4). All subsequent phase commands run with cwd: worktreePath.

If the user has declined worktree isolation in their answers to Question 2 (autonomy/constraints), skip this step and operate on the current branch.

4. Create State File

Create the directory .missions/ if it doesn't exist, then write active-mission.json:

Standard mode phases:

[
  { "name": "Architect", "emoji": "\ud83d\udcd0", "status": "active", "startedAt": "<now>", "completedAt": null },
  { "name": "Review Plan", "emoji": "\ud83d\udc41\ufe0f", "status": "pending" },
  { "name": "Implement", "emoji": "\ud83d\udd28", "status": "pending" },
  { "name": "Test", "emoji": "\ud83e\uddea", "status": "pending" },
  { "name": "Audit", "emoji": "\ud83d\udd0d", "status": "pending" },
  { "name": "Verify", "emoji": "\u2705", "status": "pending" }
]

Minimal mode phases:

[
  { "name": "Plan", "emoji": "\ud83d\udccb", "status": "active", "startedAt": "<now>" },
  { "name": "Build", "emoji": "\ud83d\udd28", "status": "pending" },
  { "name": "Verify", "emoji": "\u2705", "status": "pending" }
]

Beyond description, mode, autonomy, template, constraints, checks, modelAssignment (above), and phases, the file also carries performanceLog, failureLog, gamification, userSignals, userRating, paused, pauseHistory, progressLog, startedAt, completedAt. You don't hand-author these — mission-state.mjs writes them. See references/state-schema.md for the full annotated schema.

If a template is set, include its constraints in every phase prompt alongside constraints. Read references/templates.md for the template's phase emphasis and inject the relevant section when entering each phase.

5. Create Phase Tasks

Use TaskCreate to create one task per phase. Use the phase emoji and name as the subject (e.g., "📐 Architect"). Set the first phase task to in_progress via TaskUpdate.

6. Begin First Phase

Read the appropriate protocol file from this skill's references/ directory:

Also read references/autonomy-levels.md and apply the selected autonomy level's behavior throughout the mission.

Follow the protocol's instructions. Each protocol file includes Phase Transition instructions at the end that tell you exactly how to advance to the next phase.

Phase Execution Loop

The /mission command is the orchestrator. It owns all decisions: which subagents to spawn, when to retry, when to escalate, and when to hand off. Subagents are workers — they report results back, they don't make orchestration decisions.

Compaction Resilience

Context compaction can happen at any time — after a subagent returns, between phases, or mid-phase. The orchestrator MUST be stateless between steps. Never rely on conversation memory for mission state.

Rule: re-read state before every decision. Specifically:

1. After every subagent returns — run node ~/.claude/skills/mission/scripts/mission-state.mjs get phases to know where you are. Don't assume you remember which work item was being processed.
2. Before every phase transition — re-read the full state file. Compaction may have removed the setup context (template, constraints, model assignment).
3. Before dispatching a subagent — re-read failureLog and performanceLog from state. Don't rely on conversation memory for attempt counts or scores.
4. If you feel disoriented — run node ~/.claude/skills/mission/scripts/mission-state.mjs status. The state file is the single source of truth.

The state file + scripts make the orchestrator resilient to compaction. Even if 100% of conversation context is lost, the mission can continue from the state file alone.

After Every Subagent Returns

1. Re-read state: node ~/.claude/skills/mission/scripts/mission-state.mjs get phases — confirm which phase is active, which work item is current
2. Check for completion marker — scan the subagent's output for <!-- SUBAGENT_DONE::
   • Marker present → subagent completed normally. Continue to step 3.
   • Marker absent → subagent hit context limit mid-task. Handle as context exhaustion (see below) — do NOT log as a failure attempt or burn retry budget.
3. Evaluate the subagent's output using references/protocol-scoring.md
4. Score quality (1-5), completeness (1-5), efficiency (1-5) — omit composite and verdict; the script derives both (composite = q×0.5 + c×0.3 + e×0.2, verdict from the band). To preview the numbers without touching state, run mission-state.mjs score-compute '<json>'.
5. Write specific, actionable feedback
6. Log the score: node ~/.claude/skills/mission/scripts/mission-state.mjs score '<json>' (or score-batch '[...]' for a whole phase in one call). Raw {quality,completeness,efficiency} is enough — no hand-computed composite required.
7. Clear the checkpoint: node ~/.claude/skills/mission/scripts/mission-state.mjs checkpoint-clear
8. Feed scores into the next subagent's prompt

Note: Each phase protocol also carries an inline scoring step (Step 1.5, Step 2.5, etc.) placed right next to the dispatch — this is immune to context compaction of SKILL.md. The score-batch command logs all scores for a phase in one Bash call.

Context Exhaustion (missing SUBAGENT_DONE marker)

node ~/.claude/skills/mission/scripts/mission-state.mjs checkpoint-read

If checkpoint exists → spawn a new subagent for the same work item, prepending:

RESUMING FROM CHECKPOINT — do not restart from scratch.
Already completed: [checkpoint.completedSteps]
Remaining steps:   [checkpoint.remainingSteps]
Last commit:       [checkpoint.lastCommit]
Files touched so far: [checkpoint.filesChanged]

Continue from the first remaining step only.

If no checkpoint → the subagent made no progress before running out of context. Treat as attempt 1 failure and follow the normal failure escalation path.

See references/protocol-scoring.md for the full rubric.

Phase Loop

For each phase:

1. Read the current phase's protocol from references/
2. Follow the protocol instructions exactly
3. When the protocol's completion criteria are met, follow the Phase Transition steps in that protocol file:
   • Update the state file (mark current phase done, next phase active)
   • Update Tasks (current to completed, next to in_progress)
   • Log the transition in progressLog
   • Read the next phase's protocol
4. Apply autonomy gates:
   • Low: After every phase transition, STOP and summarize. Wait for user to say "continue"
   • Medium: After every phase transition, STOP. Generate the 3-line progress summary with node ~/.claude/skills/mission/scripts/mission-state.mjs progress-summary (deterministic — no hand-written prose), present it, then wait. Do NOT proceed until the user replies "continue" (or "pause"). (This matches autonomy-levels.md — Medium is a hard pause-gate, not a soft one.)
   • High: Continue automatically through all phases

When the final phase completes, set completedAt on the mission state and present a final summary.

Changing Models Mid-Mission

You can edit .missions/active-mission.json between phases to adjust any role in modelAssignment. The orchestrator re-reads state at every phase boundary, so the next spawned subagent will use the new value. An in-flight phase keeps the value it read at start — wait for it to finish (or pause) before editing if that matters.

Orchestrator Failure & Handoff Loop

The orchestrator manages all retries, escalations, and handoffs. Subagents never decide these — they just return success or failure.

Subagent returns failure
        ↓
Orchestrator logs attempt to failureLog in state file
        ↓
Total attempts ≥ 6? ──yes──→  HARD STOP: mark work item as blocker, ask user
        ↓ no
Attempt < 3 (this session)?  ──yes──→  Spawn new subagent (different approach)
        ↓ no
Escalate to a **powerful debug agent** (use `modelAssignment.planner` — your most capable model) with full failure log
        ↓
Debug agent succeeds? ──yes──→  Continue to next work item
        ↓ no
Auto-generate handoff.md → Pause mission → Inform user

Guards Against Infinite Loops

• Per-work-item ceiling: Max 3 attempts per session + 1 powerful-model escalation. This is enforced by counting attempts in the failureLog entry for this work item.
• Cross-session ceiling: Max 6 total attempts across ALL sessions for any single work item. If a handoff resumes and the same item has already been tried 6 times total, STOP — do not retry. Mark the work item as a blocker and ask the user: "This item has failed 6 times across sessions. Skip it (/mission skip), fix it manually, or abort?"
• Autonomy override: The failure/handoff loop always pauses after the debug agent fails, regardless of autonomy level. Even on High autonomy, exhausted retries force a pause. This prevents runaway loops.
• Only the orchestrator writes state: Subagents never write to active-mission.json. They return results to the orchestrator, which is the single writer. This prevents race conditions when parallel subagents run.

How it works step by step

1. Before dispatching, run node ~/.claude/skills/mission/scripts/mission-state.mjs failure-check "<work item>" --session <n> — it returns the escalation decision so you never count attempts from memory:
   • shouldHardStop: true (≥6 total attempts across sessions) → HARD STOP, ask user
   • shouldEscalate: true (≥3 attempts this session) → skip to powerful-model escalation
   • otherwise → spawn a new subagent, feeding the returned priorFailures as "do NOT repeat these approaches"
2. Orchestrator dispatches a subagent for the work item
3. Subagent returns success or failure
4. Orchestrator receives the result and decides:
   • Success → log it, move to next work item
   • Failure, attempt 1-2 → log the error and approach to failureLog, spawn a new subagent with instructions: "Previous attempts failed: [details]. Do NOT repeat these approaches."
   • Failure, attempt 3 → log it, spawn a powerful debug agent (modelAssignment.planner) with the full failure history
   • Debug agent succeeds → mark resolved: true, continue
   • Debug agent fails → auto-write handoff.md, pause mission, inform user
5. New session runs /mission → orchestrator reads state + handoff.md → checks total attempt count before retrying

What subagents know

Subagents receive:

• The work item goal, files, approach
• Previous failure details (so they don't repeat)
• Project instructions file summary and constraints

Subagents do NOT:

• Track their own attempt count
• Decide to escalate or hand off
• Write to failureLog or active-mission.json (the orchestrator does this)
• Know about other subagents or the mission state

Lifecycle Commands

Each /mission <verb> (from the Argument Routing table) is handled below. Most are script-backed; read references/lifecycle-commands.md for the full step-by-step before executing the manual path.

All confirmation gates and the exact state mutations are in references/lifecycle-commands.md.

Session Continuity

When /mission is invoked with no arguments in a new session:

1. Read .missions/active-mission.json
2. Health-check the state: run node ~/.claude/skills/mission/scripts/mission-state.mjs doctor. If verdict is not ok, surface the listed issues; on corrupt offer /mission reset and do not proceed; on issues fix what you can (e.g. fill a missing modelAssignment role) before resuming. Don't continue a structurally-broken mission silently.
3. If an active mission exists, display its status
4. Check for .missions/handoff.md:
   • If present, read it — this contains full context from the previous session including what was tried, what failed, and what's next
   • Delete the handoff file after reading (it's been consumed)
   • Resume the mission (set paused: false)
5. If the mission is not paused and not complete:
   • Read the current phase's protocol
   • If there are entries in failureLog for the current work, review them before retrying — do NOT repeat the same approaches that already failed
   • Continue execution from where the previous session left off

The state file + handoff document contain all information needed to fully reconstruct context.

State Validation

Every time the state file is read, validate before proceeding:

1. File exists? If not → "No active mission" (or offer to reconstruct from handoff.md if that exists)
2. Valid JSON? If parse fails → report the error, suggest /mission reset
3. Required fields present? Check: description, mode, phases (array with ≥1 entry), autonomy, startedAt. If any missing → report and suggest reset
4. modelAssignment complete? For standard mode, all 8 roles must be present. For minimal, at least explorer, planner, worker, verifier. If a role is missing → fill it with the user's balanced/worker model (e.g., "claude-sonnet-4-6" for Claude, "gpt-4o" for OpenAI) as fallback and warn the user
5. Exactly one active phase? If zero → mission may be complete (check completedAt) or stuck (suggest reset). If more than one → set the first active one as the real active phase, mark others as pending
6. Phase order valid? Done phases must come before active, active before pending. If out of order → warn and suggest reset

Important Notes

• The state file is the single source of truth. Always read before modifying, write after every change.
• Only the orchestrator writes to the state file. Subagents return results; the orchestrator updates state. This prevents race conditions.
• Use scripts for deterministic operations to save tokens. Run these via Bash instead of doing the work yourself:
  • ~/.claude/skills/mission/scripts/mission-state.mjs status — formatted mission status
  • ~/.claude/skills/mission/scripts/mission-state.mjs phase-transition — advance to next phase atomically
  • ~/.claude/skills/mission/scripts/mission-state.mjs pause / resume — toggle pause
  • ~/.claude/skills/mission/scripts/mission-state.mjs log — full progress timeline
  • ~/.claude/skills/mission/scripts/mission-state.mjs progress-summary [phase] — 3-line phase-gate summary (Medium autonomy)
  • ~/.claude/skills/mission/scripts/mission-state.mjs score '<json>' / score-batch '[...]' — append performance score(s); composite + verdict auto-derived from raw dimensions
  • ~/.claude/skills/mission/scripts/mission-state.mjs score-compute '<json>' — preview composite/verdict/XP without writing state
  • ~/.claude/skills/mission/scripts/mission-state.mjs failure '<json>' — append failure log entry
  • ~/.claude/skills/mission/scripts/mission-state.mjs failure-check "<workItem>" [--session <n>] — escalation decision (hard-stop / escalate / retry) + prior failures
  • ~/.claude/skills/mission/scripts/mission-state.mjs parse-usage '<usage block>' — extract {totalTokens,toolUses,durationMs} from a subagent usage block
  • ~/.claude/skills/mission/scripts/mission-state.mjs tokens — token usage report by phase and role
  • ~/.claude/skills/mission/scripts/mission-state.mjs doctor — validate state (required fields, one active phase, modelAssignment, phase order)
  • ~/.claude/skills/mission/scripts/mission-state.mjs get <field> — read a field from state
  • ~/.claude/skills/mission/scripts/mission-state.mjs checkpoint-write '<json>' — save subagent progress checkpoint
  • ~/.claude/skills/mission/scripts/mission-state.mjs checkpoint-read — read checkpoint (returns null if none)
  • ~/.claude/skills/mission/scripts/mission-state.mjs checkpoint-clear — delete checkpoint after successful completion
• Each protocol file in references/ is self-contained with its own completion criteria and phase transition instructions.
• Never skip the Review/approval gate — it exists to prevent wasted implementation effort.
• If handoff.md exists but state file is missing, offer to reconstruct state from the handoff document or reset.
• Template constraints and user constraints are additive. If they conflict, template constraints take priority. Warn the user at setup if a conflict is detected.