Improve token-safe worker orchestration and observability

[SpadesZ]

Context from rickie-second-brain orchestration: Codex should be brain/orchestrator, Jules handles large implementation, Antigravity handles independent verification, and local work is only small blockers/validation. Current MCP flow made this harder and more token-expensive than needed.

Needed improvements:

1. Filtered low-token job status API: repo/session/status filters; expose job_list consistently.
2. Compact worker summaries: job id, session id, status, changed files, validation results, patch artifact id, blockers; no huge patch dumps unless requested.
3. Completion-trigger semantics: monitor helpers that stay quiet until the whole project plan is verifiably complete, not each subtask.
4. Normalized Jules + Antigravity lifecycle: queued/in_progress/completed/needs_correction/blocked/failed/verified with explicit owner roles.
5. Output quality-gate metadata: commands run, exact test output, placeholder scan, secret scan, deferred items; easy reject/follow-up for skeleton patches.
6. Secret hygiene: redact user-provided keys before prompts, ledgers, logs, summaries, or worker handoffs.

Acceptance:

• One filtered call tells Codex all worker state.
• Codex can decide local validation vs Antigravity vs Jules follow-up without reading a full diff.
• User is notified only on full-plan completion or true blocking decision.
• MCP never stores/displays secrets.

[SpadesZ]

Antigravity CLI regression observed during rickie-second-brain Phase 7E verification.

Evidence:

• agy.exe --version => 1.0.12
• run_antigravity_handoff with --print waited 900s, stdout/stderr empty, no result file.
• Retry with -p waited 120s, stdout/stderr empty, no result file.
• Result path absent: .dev-triangle/antigravity-results/antigravity-20260627153339-5999909b-result.md.
• Conversation DB directory latest write time is from 2026-06-26, not the 2026-06-27 run, so CLI did not record a usable new conversation.
• Many Antigravity IDE.exe processes were running, but no closed-loop result was produced.

Required MCP behavior:

• Smoke test must prove a fresh non-interactive run can write stdout or RESULT_PATH, not just detect the executable or stale conversation DB.
• run_antigravity_handoff should classify this as NO_RESULT_PRODUCED / CLI_HUNG with concise diagnostics.
• Add fallback strategy: shrink handoff, IDE chat mode, or mark Antigravity unavailable after N retries without requiring user intervention.

[SpadesZ]

Additional confirmation: a minimal smoke handoff also failed.

• Handoff: antigravity-20260627155928-0606e534
• Objective: only write PASS to result mailbox, no repo inspection/modification.
• Result: timed out after 180s via agy --print.
• stdout/stderr empty; result file not produced.

This confirms the issue is closed-loop Antigravity execution/result capture, not handoff size or project content.

[SpadesZ]

Root cause narrowed further on Windows:

1. With repo cwd on D:\Rickie-second-brain, agy --print logs errors like:
   open /Users/Franky Kuo/.gemini/antigravity-cli/brain/<id>/.system_generated/logs/transcript.jsonl: The system cannot find the path specified.
   On Windows, /Users/... resolves against the current drive, so under D: it becomes D:\Users\... while real state is under C:\Users\....

2. Running agy --print from C:\Users\Franky Kuo removes the path error and the model response is written to transcript:
   source=MODEL type=PLANNER_RESPONSE content="AGY_OK"
   but stdout is still empty even though exit code is 0.

Therefore dev-triangle needs two fallbacks:

• Run agy from the user profile / C: drive or ensure a Windows junction maps the current drive /Users/... path to the real profile.
• Do not rely on stdout only. If stdout is empty and exit code is 0, inspect the fresh antigravity-cli/brain/<conversation>/.system_generated/logs/transcript*.jsonl and extract the latest MODEL content or result marker.

[SpadesZ]

Fix update pushed to PR #1 in commit a6b67f2. Root cause confirmed: agy on Windows requires the prompt immediately after --print for --log-file to work; when prompt was moved later, the log stayed empty and long handoffs timed out. The fix preserves that ordering, runs agy from USERPROFILE, uses --add-dir for the target repo, records --log-file, recovers ready results from transcript JSONL or conversation DB, and handles timeout-with-ready-fallback as COMPLETED. Smoke validation passed locally.