implement（实现）

User request: $ARGUMENTS

Autonomously execute implementation in-place. Supports plan files, inline tasks, or interactive mode.

Fully autonomous: No pauses except as specified in Phase 3's "Pause ONLY when" list and Edge Cases.

Workflow

Phase 1: Resolve Input & Setup

Review flag: Review workflow runs by default after implementation. If arguments contain --no-review (case-insensitive), disable it. Remove flag from arguments before processing below.

Priority order:

1. File path (ends in .md or starts with /) → use plan file, optionally with --spec <path>
2. Inline task (any other text) → create ad-hoc single chunk: ## 1. [Task summary - first 50 chars (if longer, truncate at last space before char 50 if that space is at position 10+; otherwise truncate at char 47 and append "..."; if 50 chars or fewer, use as-is)] - Depends on: - - Tasks: [user's description] - Files: (list as created/modified during execution) - Acceptance criteria: task completed AND gates pass
3. Empty → search /tmp/plan-*.md (most recent by file modification time; if tied, use alphabetically last filename); if none found, ask user what they want to implement

For plan files, parse each chunk (## N. [Name] headers):

• Dependencies (Depends on: field, - = none)
• Files to modify/create with descriptions
• Context files (paths, optional line ranges)
• Implementation tasks (bullet list)
• Key functions/types

Invalid plan: Empty file, missing ## N. [Name] chunk headers, or malformed Depends on/Tasks fields (Depends on must be - or comma-separated chunk numbers; Tasks must be a bullet list with at least one item) → Error with path + expected structure.

Build dependency graph: No-dependency chunks first, then topological order.

Create flat todo list (Memento pattern—granular progress, resumable):

[ ] Read context for [Chunk]
[ ] [Task 1]...[ ] [Task N]
[ ] Run gates for [Chunk]
[ ] Commit chunk: [Chunk]
...
# Unless --no-review, append:
[ ] Run review on implemented changes
[ ] (Fix review issues - expand as findings emerge)

All todos created at once via update_plan (status pending). If update_plan unavailable, use /tmp/implement-progress.md with markdown checkboxes: - [] pending, - [~] in progress, - [x] completed, with timestamp prefix [HH:MM:SS].

Spec file (--spec <path>): Read before implementation for requirements/acceptance criteria. If path doesn't exist, add to Notes: "Warning: Spec not found: [path]" and continue. Spec is only used when explicitly provided via --spec.

Phase 2: Execute Chunks

CRITICAL: Execute continuously without pauses.

Per chunk:

1. Read context files from plan + files-to-modify, respect line ranges
2. Implement each task, marking in_progress→completed immediately
3. Run gates (Phase 3)
4. Commit chunk: git add [files created/modified] && git commit -m "feat(plan): implement chunk N - [Name]" (do NOT push)
5. Track created/modified files for summary

Phase 3: Auto-Fix Gates

Run gates in order: typecheck, then tests, then lint. Stop at first failure and iterate on that gate until it passes before proceeding to the next gate.

Gate command detection:

1. Check AGENTS.md for explicit commands (look for sections labeled "Development Commands", "Scripts", or "Gates"; identify typecheck/test/lint by command names like tsc, jest, eslint, mypy, pytest, ruff)
2. If AGENTS.md commands fail with "command not found", "not recognized", or exit code 127 → fall back to config detection
3. Use fallback detection (see Gate Detection section)

On failure—iterate:

1. Analyze: parse errors, identify files/lines, understand root cause
2. Fix by addressing root cause (not by suppressing errors, skipping tests, or adding // @ts-ignore)
3. Re-run the failing gate
4. Track attempts per issue by error message and file:line; if same error persists after 3 distinct fix strategies, escalate per "Pause ONLY when" rules

Distinct fix strategy: A strategy is distinct if it modifies different lines OR uses a categorically different technique: (1) adding/changing type annotations, (2) type assertions/casts, (3) refactoring logic/control flow, (4) adding null/undefined checks, (5) changing function signatures, (6) adding/modifying imports.

Pause ONLY when:

• Same error message and file:line persists after 3 distinct fix strategies
• Need info not available in codebase or context (API keys, credentials, external service configs)
• Fix requires modifying files not listed in the chunk's "Files:" field (creating new files is allowed only if they are: helper/utility modules, type definition files, or test files directly testing the chunk's code)
• Plan requirements contradict each other (both can't be satisfied simultaneously)

Report: what tried, why failed, what's needed.

Phase 4: Completion

## Implementation Complete
Chunks: N | Todos: M | Created: [list] | Modified: [list]
### Notes: [warnings/assumptions/follow-ups]
Run `$review` for quality verification.

Unless --no-review → proceed to Phase 5.

Phase 5: Review Workflow (default, skip with --no-review)

Skip if --no-review was set.

1. Mark "Run review" in_progress → invoke $review --autonomous → mark completed
2. If no issues → mark fix placeholder completed, done
3. Expand fix placeholder: [x] (Fix review issues - expand as findings emerge) [] Fix critical/high severity issues [] Re-run review to verify fixes [] (Additional fix iterations - expand if needed)
4. Mark "Fix critical/high" in_progress → invoke $fix-review-issues --severity critical,high --autonomous → mark completed
5. Mark "Re-run review" in_progress → invoke $review --autonomous → mark completed
6. Repeat fix/review cycle until clean or max 3 cycles

Edge Cases

Principles

• Autonomous: No prompts/pauses/approval needed except blocking issues listed in Phase 3 and Edge Cases
• Memento todos: One todo per action, granular visibility, resumable
• Persistent auto-fix: Iterate until gates pass (up to 3 distinct strategies per issue), escalate only when stuck
• Dependency order: Execute in order, skip failed chunk's dependents
• Gates non-negotiable: Fix root cause (no @ts-ignore, test skips, or suppressions); skip chunk only after 3 failed strategies
• Commit per chunk: Each successful chunk gets its own commit (no push until end); provides rollback points for recovery
• Simplicity: Prefer readable code over micro-optimizations; don't add complexity for marginal performance gains

Gate Detection

Priority: AGENTS.md → package.json scripts → Makefile → config detection

Skip any source that doesn't define relevant commands (test/lint/typecheck).

Fallback (if AGENTS.md doesn't specify or commands fail with exit code 127):

• TS/JS: tsconfig.json→tsc --noEmit, eslint.config.*→eslint., jest/vitest.config.*→npm test
• Python: pyproject.toml→mypy/ruff check, pytest config→pytest
• Go: go.mod→go build./..., golangci.yml→golangci-lint run
• Rust: Cargo.toml→cargo check, cargo test
• Other languages: Skip gates with warning "No gate commands detected for [language]; specify in AGENTS.md"