TORCH/src/prompts/scheduler-flow.md

Scheduler Flow (Single Source of Truth)

Use this document for all scheduler runs.

Numbered MUST Procedure

1. Set cadence variables before any command:

   • cadence = daily or weekly
   • log_dir = task-logs/<cadence>/
   • branch_prefix = agents/<cadence>/
   • prompt_dir = src/prompts/<cadence>/

2. Run preflight to build the exclusion set:

   if daily: `npm run lock:check:daily`; if weekly: `npm run lock:check:weekly`
   

3. Read policy file(s). This step is conditional: if AGENTS.md is missing, continue without failing.

   test -f AGENTS.md && cat AGENTS.md || echo "No AGENTS.md found; continuing"
   

4. Bootstrap log directories before listing files:

   mkdir -p <log_dir>
   

5. Find latest cadence log file, then choose next roster agent not in exclusion set:

   ls -1 <log_dir> | sort | tail -n 1
   

   • An empty <log_dir> is valid and means there are no prior cadence logs yet (first run).
   • On first run, select the first eligible roster entry that is not in the exclusion set.

6. If every roster agent is excluded, write a _failed.md log with: All roster tasks currently claimed by other agents and stop.

7. Claim selected agent:

   AGENT_PLATFORM=<platform> \
   npm run lock:lock -- --agent <agent-name> --cadence <cadence>
   

   • Exit 0: lock acquired, continue.
   • Exit 3: race lost/already locked, return to step 2.
   • Exit 2: lock backend error, write _failed.md with reason Lock backend error, stop.

8. Execute <prompt_dir>/<prompt-file> end-to-end.

9. Create exactly one final status file (_completed.md or _failed.md), run repo checks, commit, and push.