AUTO/YOLO agent pipeline: fix CI + code review via claude -p

[bearmug]

Summary

Implements the full agent pipeline for PR autopilot modes (AUTO/YOLO):

• Fix CI — when checks fail, spawn claude -p to analyze failures and push fixes
  • AUTO: --permission-mode acceptEdits --allowedTools "Bash Edit Write Read Glob Grep"
  • YOLO: --permission-mode bypassPermissions
• Code review — when checks pass, spawn claude -p in read-only mode to review the diff
  • Outputs structured JSON findings (critical/important/minor)
• Fix review — when Critical/Important issues found, spawn claude -p to fix them
  • Max 2 review+fix cycles, then stops

Pipeline state machine

checks_failing → spawnFixCI → pushes fix → checks re-run
checks_passing → spawnCodeReview → parse findings
has_issues → spawnFixReview → pushes fixes → re-review
clean → AUTO: wait for approval | YOLO: auto-merge

Safety

• AgentRunning field prevents double-spawn (only one agent per PR at a time)
• 15-minute timeout per agent, 20-minute stale state auto-clear
• $5/$3/$5 budget caps per agent invocation
• Review state resets on check regression (full pipeline re-runs after CI fixes)
• Clone to /tmp for isolation — never touches local repos

TUI

• Agent status with elapsed time in PR zoom panel
• Review findings by severity with file:line locations
• Accumulated cost display in PR header

Test plan

• All daemon tests pass (14 new model tests + 11 agent tests)
• All TUI tests pass
• Track a PR with failing CI in AUTO → verify agent spawns
• After green checks → verify code review runs
• YOLO mode → verify full pipeline + auto-merge