Skip to content

feat: SDK optimization — lifecycle, stall detection, communication, performance#187

Open
michael-wojcik wants to merge 11 commits intomainfrom
feat/sdk-optimization
Open

feat: SDK optimization — lifecycle, stall detection, communication, performance#187
michael-wojcik wants to merge 11 commits intomainfrom
feat/sdk-optimization

Conversation

@michael-wojcik
Copy link
Collaborator

Summary

Full audit of PACT plugin's Task SDK and Agent Teams SDK usage, identifying and implementing optimizations across lifecycle management, stall detection, communication, cost, and performance.

  • Add TeamDelete to wrap-up and session_end cleanup (fixes resource leak)
  • Use last_assistant_message in validate_handoff (less fragile than transcript parsing)
  • Enrich worktree_guard error with corrected path suggestion (faster agent retry)
  • Add TeammateIdle hook for stall detection and idle cleanup (closes critical gap)
  • Make non-critical hooks async (reduced latency)
  • Use broadcast for HALT algedonic signals (team-wide emergency stop)
  • Add Terminate outcome to imPACT with TaskStop (handle unrecoverable agents)
  • Set model: haiku for memory-agent (cost optimization for mechanical tasks)
  • Clarify pact-memory vs SDK memory scope in agent documentation

Test plan

  • 1079 tests pass (75 new + 1004 existing), zero regressions
  • New test_hooks_json.py validates hooks.json structure and async flags
  • TeammateIdle hook: 54 tests (36 coder + 18 test engineer) covering stall detection, idle cleanup, concurrent tracking
  • validate_handoff: 31 tests covering last_assistant_message preference and fallback
  • worktree_guard: 20 tests including path suggestion edge cases
  • session_end cleanup: 16 tests with mocked filesystem
  • Verify TeammateIdle hook fires on agent idle events
  • Verify broadcast HALT stops all running teammates

Closes design items #1-8, #10, #11, #12 from docs/plans/2026-02-19-sdk-optimization-design.md
Deferred items tracked in #186

Add team cleanup section to wrap-up command (shutdown teammates,
then TeamDelete). Add best-effort cleanup_stale_teams() to
session_end.py as safety net for sessions ending without wrap-up.
Prefer last_assistant_message field over transcript parsing in
SubagentStop hook, with fallback for older SDK versions. Add 16
new tests covering both paths.
When blocking edits outside the worktree, include "Did you mean:
{corrected_path}" in the error message to help agents retry faster.
Add 3 new tests for path suggestion logic.
New hook with two responsibilities:
- Stall detection: alerts orchestrator when agent goes idle with
  in_progress task (enables automated imPACT triggering)
- Idle cleanup: tracks consecutive idle events, suggests shutdown
  at 3, instructs orchestrator to send shutdown_request at 5

Also marks session_end, file_size_check, and file_tracker hooks
as async for reduced latency on non-critical paths.

Includes 36 new tests.
- HALT algedonic signals now use SendMessage broadcast to stop all
  teammates simultaneously instead of manual relay
- Add Terminate as fifth imPACT outcome using TaskStop for
  unrecoverable agents (infinite loops, context exhaustion)
- Set model: haiku for pact-memory-agent (purely mechanical tasks)
- Clarify in all agent skill tables and pact-agent-teams SKILL.md
  that pact-memory is for project-wide institutional knowledge only,
  not agent-level domain expertise (handled by SDK memory frontmatter)
Edge cases and error paths across all modified modules:
- session_end cleanup_stale_teams: 16 tests (filesystem mocking)
- teammate_idle stall/cleanup interaction: 18 tests
- validate_handoff last_assistant_message: 15 tests
- worktree_guard path suggestion: 17 tests
- hooks.json structural validation: 13 tests (new file)

1079 total tests pass, zero regressions.
- Fix task ID string comparison in teammate_idle.py (lexicographic
  "3" > "20" bug) — use int() with fallback for non-numeric IDs
- Scope cleanup_stale_teams to current session's team only via
  CLAUDE_CODE_TEAM_NAME env var (prevents cross-session deletion)
- Fix TOCTOU in write_idle_counts — acquire lock before truncating

1078 tests pass.
- Reset idle count to 0 when teammate is assigned new work (detects
  task ID change via structured idle_counts.json format)
- Validate file belongs to same project root before suggesting
  worktree path (prevents misleading suggestions from nested repos)
- Require project marker at common ancestor in fallback path
  suggestion (prevents false matches across unrelated projects)
- Test summary truncation >80 chars in session_end snapshot
- Validate pipe-separated matcher syntax in hooks.json
- Validate SubagentStart matcher covers all 11 agent types
  (auto-detects new agents from agents/ directory)

1083 tests pass.
New features: TeammateIdle hook, TeamDelete cleanup, broadcast HALT,
TaskStop in imPACT, async hooks, model: haiku for memory-agent,
last_assistant_message in validate_handoff, enriched worktree_guard
error messages, memory scope documentation.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant

Comments