forge-auto（锻造汽车）

/forge-auto — FORGE Autopilot Mode

FORGE takes full control of the development pipeline. It analyzes, decides, executes, verifies, and iterates automatically until the objective is complete.

Principle

The user provides an objective → FORGE handles EVERYTHING else.
Planning → Architecture → Stories → Code → Tests → Verification → Deployment

Workflow

1. Load memory:

- Read .forge/memory/MEMORY.md for project context - Read the latest session from .forge/memory/sessions/ for continuity - Read .forge/sprint-status.yaml for the current state - Read .forge/config.yml for configuration - forge-memory search "<current objective>" --limit 3 → Load relevant past decisions and context

1. Analyze state and determine the phase: Inspect the project artifacts and sprint status, then execute the appropriate phase:

- No artifacts exist → Start with /forge-plan (generates the PRD) - PRD exists, no architecture → Launch /forge-architect - Architecture exists, no UX design → Launch /forge-ux - UX exists, no stories → Launch /forge-stories - Stories exist with "pending" status: - Count unblocked pending stories - 2+ unblocked stories AND Agent Teams available → Delegate to /forge-team build [STORY-IDs] (parallel execution), then continue with QA - Otherwise → Pick the next unblocked story, launch /forge-build STORY-XXX (sequential) - "in_progress" story exists → Resume /forge-build STORY-XXX - Story implemented (Dev tests pass) → Launch /forge-verify STORY-XXX - QA verdict = FAIL → Increment failure counter. Under 3 failures: fix and re-verify. At 3+ failures: escalate to /forge-loop "Fix STORY-XXX: [summary]" --mode hitl which iterates with sandbox guardrails until tests pass - QA verdict = PASS or CONCERNS → Launch /forge-review on the story's source code (adversarial analysis) - Review raises CRITICAL issues → Fix, re-run /forge-verify, re-run /forge-review - Review is clean → Move to the next story - All stories completed → Propose /forge-deploy or new stories

1. Execute with the appropriate agents:

- Each phase invokes the corresponding agent (PM, Architect, UX, SM, Dev, QA) - The agent loads its persona from ~/.claude/skills/forge/references/agents/ - The agent produces its artifacts in docs/ or src/ - Agent Teams acceleration: when entering the build phase with 2+ unblocked stories, and CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 is set, autopilot delegates to /forge-team build for parallel story implementation (up to 4 Dev + 1 QA). If Agent Teams is not available, stories are built sequentially as before.

1. Automatic quality gates:

- After each story: lint + typecheck + tests > 80% coverage - After each /forge-verify: mandatory QA verdict - After QA PASS/CONCERNS: mandatory /forge-review (adversarial review) - If /forge-review raises critical issues: fix → re-verify → re-review - Loop escalation: if 3 consecutive failures on the same story, autopilot delegates to /forge-loop in HITL mode with the QA failure summary as task. forge-loop iterates with sandbox guardrails (cost cap, circuit breaker, rollback) until tests pass, then returns control to autopilot for re-verification. - Ultimate circuit breaker: if forge-loop also fails (hits its own circuit breaker) → pause + report to user

1. Save memory:

- forge-memory log for each story completed during this session: forge-memory log "{STORY_ID} terminée: {N} tests, couverture {X}%" --agent dev --story {STORY_ID} - forge-memory log for the session summary: forge-memory log "Session autopilot terminée: {completed}/{total} stories, phase {PHASE}" - Consolidate session logs into MEMORY.md: forge-memory consolidate --verbose - Sync the full memory index: forge-memory sync - Update .forge/sprint-status.yaml

1. Human checkpoints (configurable):

- Default: checkpoint after each major phase (plan, architecture, stories) - --no-pause mode: no checkpoints (full autopilot) - --pause-stories mode: pause after story decomposition - --pause-each mode: pause after each story

Options

# Full autopilot — FORGE decides everything
/forge-auto

# Autopilot with a specific objective
/forge-auto "Implement the authentication system"

# Autopilot without pauses (warning: fully autonomous)
/forge-auto --no-pause

# Autopilot with pause after stories
/forge-auto --pause-stories

# Autopilot with pause after each story
/forge-auto --pause-each

# Resume autopilot after a pause
/forge-auto --resume

Progress Report

At each step, FORGE displays:

FORGE AUTOPILOT — Progress
──────────────────────────────
Phase     : Development (Story 3/8)
Last      : STORY-002 ✓ (QA: PASS)
Current   : STORY-003 — Implementation
Next      : STORY-004 (pending)

Metrics:
  Stories   : 2 completed / 1 in_progress / 5 pending
  Tests     : 47 pass / 0 fail
  Coverage  : 87%

Memory    : .forge/memory/MEMORY.md (up to date)
Session   : .forge/memory/sessions/YYYY-MM-DD.md

How /forge-auto Uses Other FORGE Tools

Difference with /forge-loop

Coexistence with Manual Mode

Autopilot and manual commands are 100% compatible:

• You can start with /forge-auto, pause, then continue manually
• You can work manually then launch /forge-auto --resume to continue
• Memory is shared: both modes read/write the same files
• /forge-status works in both modes

Notes

• Autopilot ALWAYS respects quality gates (no shortcuts)
• The circuit breaker protects against infinite loops
• Persistent memory ensures continuity between sessions
• Compatible with projects initialized via /forge-init
• Also works for resuming existing projects (analyzes the state)