Three Shifts

1. READ shifts/state.json
   → If status is "completed", "cancelled", "idle", or "awaiting_approval": reply HEARTBEAT_OK (no-op)
   → If status is "executing": continue

2. READ shifts/tasks.md
   → Find first step marked [ ] (not started)
   → If none found, check for [!] (blocked) steps — retry if dependency cleared
   → If all steps are [x], [-], or still [!]: shift is DONE → go to Handoff

3. READ shifts/context.md
   → Load rules, constraints, pitfalls

4. CHECK FOR STALE CLAIMS
   → If any step is marked [~] (claimed by a previous cycle that crashed/timed out):
     - Change it back to [ ] — the previous cycle didn't finish
     - This step becomes the next available step to claim

5. CLAIM the step
   → Mark it [~] (in progress) in tasks.md
   → Write current timestamp to state.json

6. EXECUTE the step
   → Do the work (one focused action)
   → Stay within the step's scope — do NOT drift to other tasks

7. WRITE RESULTS back to tasks.md
   → Success: [x] Step N.N: description — DONE (cycle X): brief result
   → Blocked: [!] Step N.N: description — BLOCKED (cycle X): reason why
   → If blocked, check if next [ ] step has no dependency on this one → continue to it

8. UPDATE state.json
   → Increment cyclesRun, update completed/blocked counts, lastCycleAt

9. UPDATE shifts/context.md (if learned something new)
   → New pitfall discovered? Add it.
   → Found a useful command? Note it.
   → This is how cycles teach future cycles.

10. IF time permits in this cycle, take the NEXT available [ ] step
   → But only if the current step took <5 minutes
   → Never take more than 2 steps per cycle

# Handoff: Afternoon → Night | 2026-02-22

## Completed
- Task 1: API migration plan — recommendation written to memory/projects/example/
- Task 2 steps 1-3: Workspace commits (skills, memory reorg)

## Still In Progress
- Task 2 steps 4-5: Script commits and push (2 steps remaining)

## Blocked (needs user)
- Task 1 step 3: Need pricing confirmation for mor-[REDACTED] paid tier

## Lessons Learned This Shift
- GLM-5 handles git operations well but struggles with multi-file diffs
- Brave search rate limit hit on cycle 4 — added 3-second delay to context.md

## Recommendations for Next Shift
- Finish Task 2 push (quick, 1 cycle)
- Review migration recommendation in memory/projects/example/transition.md

Name: three-shifts-plan-morning
Schedule: cron 0 6 * * * (America/Chicago)
Model: your-heavy-model (e.g. Claude, GPT-4)
Session: isolated
Message: >
  Read the three-shifts skill at skills/three-shifts/SKILL.md.
  Execute Phase 1 (Planning) for the MORNING shift (6 AM – 2 PM CST).
  Read shifts/handoff.md, shifts/context.md, shifts/state.json, and recent memory files.
  Generate the shift plan and present for user approval.
  If carryover items exist from last night that were already approved, note them.
  After approval, decompose all tasks into atomic GLM-5-sized steps and write to shifts/tasks.md.

Name: three-shifts-plan-afternoon
Schedule: cron 0 14 * * * (America/Chicago)
Model: your-heavy-model (e.g. Claude, GPT-4)
Session: isolated
Message: >
  Read the three-shifts skill at skills/three-shifts/SKILL.md.
  Execute Phase 1 (Planning) for the AFTERNOON shift (2 PM – 10 PM CST).
  Read shifts/handoff.md, shifts/context.md, shifts/state.json, and recent memory files.
  Generate the shift plan and present for user approval.
  After approval, decompose all tasks into atomic GLM-5-sized steps and write to shifts/tasks.md.

Name: three-shifts-plan-night
Schedule: cron 0 22 * * * (America/Chicago)
Model: your-heavy-model (e.g. Claude, GPT-4)
Session: isolated
Message: >
  Read the three-shifts skill at skills/three-shifts/SKILL.md.
  Execute Phase 1 (Planning) for the NIGHT shift (10 PM – 6 AM CST).
  Read shifts/handoff.md, shifts/context.md, shifts/state.json, and recent memory files.
  IMPORTANT: If shifts/tasks.md has carryover items already approved today, auto-approve them
  and set state.json nightAutoApproved=true. Do NOT ping the user for already-approved carryover.
  Only present new tasks for approval.
  After approval (or auto-approve), decompose into atomic steps and write to shifts/tasks.md.

Name: three-shifts-cycle
Schedule: cron */15 * * * * (America/Chicago)
Model: your-light-model (e.g. GLM-5, GPT-4o-mini)
Session: isolated
Message: >
  You are a cycle executor for the three-shifts system.
  Read skills/three-shifts/SKILL.md Phase 2 (Cycle Execution) for your algorithm.
  Then read shifts/state.json — if status is not "executing", reply HEARTBEAT_OK.
  If executing: read shifts/tasks.md and shifts/context.md.
  Find the first [ ] step, claim it [~], execute it, write result back.
  If blocked, mark [!] with reason, move to next [ ] step.
  Update state.json and context.md as needed.
  Stay focused. One step. Write results. Done.