Agents were skipping the pre-flight curl/gh check for in-progress work,
causing duplicate task runs. Root causes:
1. Two-document indirection (meta prompt → scheduler) caused the pre-flight
instructions to get lost when the agent opened the second file
2. No proof-of-execution gate — agents could assert "no conflicts" without
actually running the commands
3. curl|jq pipeline silently fails when jq is unavailable
4. "Step 0" labeling felt like skippable preamble
Changes:
- Make the commands the literal first instruction ("run NOW, paste output")
- Use gh CLI as primary (authenticated, no jq needed), curl as fallback
- Require raw terminal output pasted before proceeding (output gate)
- Rename "Step 0" to "PRE-FLIGHT GATE" to signal it's not optional
- Eliminate duplication between meta prompt and scheduler files
- Add explicit "STOP HERE" checkpoint after command execution
- Add skip-ahead clause for agents arriving from meta prompt
https://claude.ai/code/session_01S7jaQb9ctJ4bzqRhEcm28Q
Replace shared CSV files (AGENT_TASK_LOG.csv, WEEKLY_AGENT_TASK_LOG.csv)
with per-entry files in docs/agents/task-logs/{daily,weekly}/. Each log
entry is now a separate file named YYYY-MM-DD_HH-MM-SS_<agent>_<status>.md,
eliminating git merge conflicts and race conditions when multiple agents
run concurrently.
Updated files:
- daily-scheduler.md: rewritten for directory-based log format
- weekly-scheduler.md: rewritten for directory-based log format
- META_PROMPTS.md: updated pre-flight scan and claiming instructions
- AGENTS.md: updated scheduler reference from CSV to directory-based
Migrated all existing CSV entries to individual files in the new
directory structure.
https://claude.ai/code/session_01P62EWUXwupgxPhpnRejUuz
- Move `CONTEXT.md`, `TODO.md`, `DECISIONS.md`, `TEST_LOG.md` to `context/`, `todo/`, `decisions/`, `test_logs/` directories with timestamps.
- Update `AGENTS.md` and `CLAUDE.md` to define the new workflow: agents must read the latest timestamped file and create a new one when updating.
- Update `daily-scheduler.md` and `weekly-scheduler.md` to initialize tasks using the new file structure.
- Update all daily agent prompts to reference the new directory structure.
- This change prevents merge conflicts on persistent state files by treating them as an append-only log of snapshots.
Co-authored-by: PR0M3TH3AN <[email protected]>
Agents were repeatedly picking up content-audit-agent because:
1. CSV only updated after task completion — crashes left no trace
2. PR claim check (Step 1.5) was buried and easy to skip
3. No "started" status meant failed runs were invisible to rotation
4. Pre-flight scan was too narrow (checked one agent, not all)
Changes:
- Add Step 0 (mandatory pre-flight) that scans ALL open PRs and CSV
for in-progress work before selecting any task
- Introduce "started" CSV status logged immediately upon claiming,
before task execution begins — ensures rotation advances on crash
- Promote claim check from sub-step 1.5 to full Step 2 with explicit
MANDATORY warnings
- Add 24-hour stale claim timeout for abandoned "started" entries
- Add fallback: if gh is unavailable, CSV "started" rows still work
- Strengthen META_PROMPTS.md to front-load deduplication as Step 0
with explicit "this has happened before" warning language
- Add Quick Reference table summarizing all 5 deduplication layers
https://claude.ai/code/session_017QCJvunQJ7ATr6uZBHV5iA
When multiple agents (Claude Code, Codex, Jules) run schedulers concurrently,
they can pick up the same task because neither has committed yet. This adds a
"Draft PR Lock" protocol: agents must check for and create a draft PR before
starting work, acting as a distributed lock visible to all platforms.
Changes:
- AGENTS.md §12: new "Task Claiming Protocol" subsection with check/claim/race
handling steps
- daily-scheduler.md: new Step 1.5 with concrete gh CLI commands
- weekly-scheduler.md: same Step 1.5 addition
https://claude.ai/code/session_01Vnih5ZARbEmnh3EGTDQAXX
Update `docs/agents/prompts/weekly-scheduler.md` to include 12 missing agents that were present in the `docs/agents/prompts/weekly` directory but absent from the scheduler list.
This ensures the weekly agent rotation includes all available weekly agents:
- bug-reproducer-agent
- dead-code-agent
- event-schema-agent
- fuzz-agent
- interop-agent
- perf-optimization-agent
- pr-review-agent
- refactor-agent
- smoke-agent
- telemetry-agent
- test-coverage-agent
- weekly-synthesis-agent
Verified that the daily scheduler list is already up-to-date.
Co-authored-by: PR0M3TH3AN <[email protected]>
Add a weekly race condition agent that systematically audits the codebase
for concurrency hazards, timing bugs, and initialization races. The agent
covers relay connection ordering, signer availability, playback fallback
negotiation, cache state consistency, and DOM-ready assumptions.
https://claude.ai/code/session_01AudwR75Y4Bi4WLSYEAJPVW
- Add weekly-scheduler.md with rotation logic for 3 weekly agents
(changelog-agent, frontend-console-debug-agent, perf-deepdive-agent)
- Add separate WEEKLY_AGENT_TASK_LOG.csv to keep weekly and daily
rotations independent
- Mirrors the daily-scheduler.md structure with the same authority
hierarchy, error handling, and log format
https://claude.ai/code/session_017PuKrVuAUbo1seo7jBTbw6
Claude