mission-driver

mission-driver — Roadmap creation and mission execution

This skill operates the tools/mission-driver/ engine shipped with this repo. It is stack-agnostic: it works for Node, Python, Java, Go, or any project whose verification commands can be expressed as shell strings. Project-specific values (test command, module layout, commit format) are read from missions/base.json, docs/context/project-context.md, and AGENTS.md rather than hardcoded here.

Companion docs (read on demand, not all at once):

• tools/mission-driver/docs/user-manual.zh.md / user-manual.en.md — full training handbook
• tools/mission-driver/README.md — command cheat-sheet
• tools/mission-driver/EXECUTION-PRINCIPLE.md — internal execution deep dive
• tools/mission-driver/TROUBLESHOOTING.md — diagnostics when stuck
• docs/plans/00-plan-authoring-and-execution-guide.md — plan format and lifecycle
• references/mission-config-schema.md — full mission.json schema (companion to this file)
• references/roadmap-template.md — roadmap structure with annotated example

Mental model

A mission = a fixed config (missions/<name>.json) for one development goal. The engine reads it and enters a loop:

CHECK (health: tests/build pass)
  → REVIEW_PLANS (review draft plans → promote active; empty forEach = passthrough)
  → EXEC_PLANS (one plan-execution subflow per active plan:
                   EXECUTE → CLOSURE_SCRIPT_CHECK → CLOSURE_AUDIT → BUILD_VERIFY)
  → DRAFT_PLANS (draft 1-3 plans from roadmap → sub-agent review → promote active)
  → [loop back to REVIEW_PLANS]
  → nothing to draft → DEEP_AUDIT (multi-audit + open-audit → draft remediation plans)
                       → REVIEW_PLANS (execute audit-created plans)
  → ... until maxAuditRounds exhausted and nothing outstanding

Key facts:

• Roadmap precedes mission. The draft command will generate both roadmap and missions/<name>.json from a single description, but only after the brief stage confirms scope is clear.
• missions/<name>.json is pure static config — no runtime state. The engine's runtime state lives in _tmp/<runDir>/run-state.json.
• Every AI step is a child opencode run process; logs land in _tmp/<ts>-mission-driver/.
• Plan lifecycle: draft → (REVIEW_PLANS) → active → (EXEC_PLANS) → completed. Subflow plan-execution runs once per active plan.
• Ctrl-C / SIGTERM is caught by main.js for graceful process-tree cleanup.
• Plan format is a fixed contract enforced by tools/mission-driver/src/plan-check.mjs.