Add wingman automated code review system

[dipree]

Summary

Adds wingman, an automated code review system that reviews agent-produced code changes after each commit and delivers actionable suggestions back to the agent. The system prioritizes visible delivery — the user sees the agent reading and applying the review in their terminal.

How It Works

Review Generation

After each commit (or auto-commit checkpoint), wingman spawns a detached background process that:

1. Waits 10 seconds for the agent turn to settle
2. Computes a branch-level diff against main/master merge-base (holistic view, not just one commit)
3. Loads session context (prompts, commit message, checkpoint metadata)
4. Calls Claude (sonnet) with read-only tool access (Read, Glob, Grep) for an intent-aware review
5. Writes suggestions to .entire/REVIEW.md

Review Delivery (Two Paths)

Primary — Visible injection (when a live session exists):

Commit → wingman review → REVIEW.md written → live session detected → defer

User sends next prompt → UserPromptSubmit hook
  → REVIEW.md exists → inject as additionalContext
  → Agent reads REVIEW.md, applies suggestions, deletes file
  → User sees everything in their terminal
  → Agent then proceeds with user's actual request

The additionalContext hook response field adds the instruction directly into Claude's context as a mandatory pre-step (stronger than systemMessage which the agent can deprioritize).

Fallback — Background auto-apply (when no sessions are alive):

Review finishes → no live sessions → claude --continue --print
  → Applies review silently in background

This only fires when all sessions are ENDED (user closed everything). Three trigger points cover all scenarios:

• Review process itself (review finishes, no live sessions)
• SessionEnd hook (last session closes with pending review)
• Stop hook (turn ends with no live sessions — edge case)

Decision Flow

              REVIEW.md written
                    │
                    ▼
          ┌─────────────────┐
          │ Any live session │
          │   exists?        │
          └────────┬────────┘
              │          │
             Yes         No
              │          │
              ▼          ▼
        Defer to     Background
        next prompt  auto-apply
        (visible)    (invisible)

Key Behaviors

Stale Review Cleanup

• REVIEW.md from a different session → deleted (stale)
• REVIEW.md without state file → deleted (orphan)
• REVIEW.md older than 1 hour → deleted (TTL expired)
• REVIEW.md from current session + fresh → kept (skip new review)

Concurrency & Safety

• Atomic lock file (O_CREATE|O_EXCL) prevents concurrent review spawns; stale locks (>30m) auto-cleaned
• File hash dedup prevents re-reviewing identical change sets within the same session
• ENTIRE_WINGMAN_APPLY=1 env var prevents post-commit hook from triggering another review during auto-apply (recursion prevention)
• --setting-sources "" disables hooks on review/apply subprocesses
• GIT_ env stripping* prevents git index corruption in subprocesses
• ApplyAttemptedAt field prevents infinite auto-apply retry loops

Session Detection

• hasAnyLiveSession() scans .git/entire-sessions/ for any non-ENDED session
• isSessionIdle() reads session state files directly using repoRoot (not git rev-parse, which fails in detached processes running from cwd=/)
• Handles both normal repos and git worktrees (parses .git file for gitdir: pointer)

Strategy Integration

• Manual-commit: Review triggered from git post-commit hook via triggerWingmanFromCommit()
• Auto-commit: Review triggered from stop hook after SaveChanges() creates the commit

Files

File	Purpose
wingman.go	State management, trigger logic, dedup, lock files, stale cleanup, CLI commands
wingman_review.go	Detached review subprocess, diff computation, Claude API call, session detection, auto-apply
wingman_prompt.go	Review prompt construction from diff + context
wingman_instruction.md	Embedded instruction injected into agent context
wingman_spawn_unix.go	Detached process spawning (Unix)
wingman_spawn_other.go	No-op stubs (non-Unix)
wingman_test.go	25 tests covering state, dedup, stale cleanup, session detection
hooks_claudecode_handlers.go	Prompt-submit injection, stop hook trigger, session-end trigger
hooks_geminicli_handlers.go	Session-end trigger for Gemini CLI
hooks_git_cmd.go	Post-commit hook trigger for manual-commit strategy
hooks.go	additionalContext hook response field, outputHookResponseWithContext()
settings/settings.go	IsWingmanEnabled() setting
docs/architecture/wingman.md	Comprehensive architecture documentation

Configuration

entire wingman enable   # Enable
entire wingman disable  # Disable + cleanup
entire wingman status   # Show status

Test plan

• mise run fmt && mise run lint && mise run test:ci passes
• Enable wingman in a test repo, make agent produce changes, verify REVIEW.md is generated
• Verify prompt-submit injection fires on next user prompt (visible in terminal)
• Verify agent addresses REVIEW.md before user's request
• Verify background auto-apply only fires when no live sessions exist
• Start new session with stale REVIEW.md from old session — verify cleanup
• Verify ApplyAttemptedAt prevents infinite retry loops
• Verify concurrent review spawns are blocked by lock file

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR adds wingman, an automated code review system that reviews agent-produced code changes after each commit and delivers actionable suggestions back to the agent. The system prioritizes visible delivery where users see the agent reading and applying reviews in their terminal.

Changes:

• Adds wingman review system with detached background processes for reviewing code changes using Claude
• Implements dual-delivery mechanism: visible injection via prompt hooks (primary) and background auto-apply (fallback)
• Includes comprehensive state management with deduplication, stale review cleanup, and retry prevention
• Adds CLI commands for enabling/disabling/status checking of wingman functionality
• Integrates with existing hook system for Claude Code and Gemini CLI agents

Reviewed changes

Copilot reviewed 14 out of 14 changed files in this pull request and generated 13 comments.

Show a summary per file

File	Description
docs/architecture/wingman.md	Comprehensive architecture documentation covering components, lifecycle, and configuration
cmd/entire/cli/wingman.go	Core state management, trigger logic, deduplication, lock files, and CLI commands
cmd/entire/cli/wingman_review.go	Detached review subprocess implementation with diff computation, Claude API calls, and session detection
cmd/entire/cli/wingman_prompt.go	Review prompt construction from diff and session context with truncation for large diffs
cmd/entire/cli/wingman_instruction.md	Embedded instruction injected into agent context for applying reviews
cmd/entire/cli/wingman_spawn_unix.go	Unix-specific detached subprocess spawning with process group detachment
cmd/entire/cli/wingman_spawn_other.go	No-op stubs for non-Unix platforms
cmd/entire/cli/wingman_test.go	Unit tests covering state management, deduplication, session detection, and stale cleanup
cmd/entire/cli/hooks_claudecode_handlers.go	Prompt-submit injection, stop hook trigger, and session-end trigger integration
cmd/entire/cli/hooks_geminicli_handlers.go	Session-end trigger for Gemini CLI agent
cmd/entire/cli/hooks_git_cmd.go	Post-commit hook trigger for manual-commit strategy
cmd/entire/cli/hooks.go	Hook response structure with additionalContext field
cmd/entire/cli/settings/settings.go	Wingman enable/disable settings and Agent field addition
cmd/entire/cli/root.go	Wingman command registration