| name | demo-loop |
| description | Use to run a zero-dependency sandbox loop that proves the loop machinery works end to end. Triggers on "run the demo loop", "test the loop", "quick loop test", "demo loop", "make the sandbox tests pass". Fixes intentionally-broken code until `python -m unittest` passes — no GitHub/CI/MCP needed. |
Demo Loop — Zero-Dependency Sandbox
A self-contained loop for watching the machinery converge. It fixes the three
intentional bugs in loops/E-demo-sandbox/sandbox/calc.py until the tests pass.
No connectors, no worktrees — just the local filesystem and Python's stdlib
unittest. This is the irreducible core of a loop: automation + skill +
maker/checker + memory + a verifiable stop condition.
Stop condition
python -m unittest test_calc (run inside loops/E-demo-sandbox/sandbox/) exits 0.
That exit code is the whole backpressure signal — RED means keep going, GREEN means stop.
Workflow
- Read memory — open
loops/E-demo-sandbox/state/demo-state.md to see prior runs and any already-known fixes.
- Get the RED signal — dispatch the demo-checker sub-agent to run
python -m unittest test_calc -v in the sandbox dir and report the failing tests + assertion messages. (If already GREEN, stop — nothing to do.)
- Fix — dispatch the demo-maker sub-agent with the failure output. It edits ONLY
sandbox/calc.py to fix the bugs (never the test file).
- Re-check — dispatch demo-checker again. If GREEN → go to step 5. If still RED → return to step 3. Cap at 5 attempts; if still RED after 5, stop and report what's stuck (do not thrash).
- Update memory — append the run (date, result, attempts) and the fixed symbols to
demo-state.md.
The six blocks in this loop (and the two it drops on purpose)
- Automation — run once with
/demo-loop (Claude) / $demo-loop (Codex), or on a cadence with /loop 10m /demo-loop to watch it idle GREEN then react after a reset.
- Skill — this file (the workflow + stop condition).
- Sub-agents —
demo-maker (edits code) vs demo-checker (runs tests, no edit). The maker never grades its own work.
- Memory —
state/demo-state.md.
- Worktrees — ➖ dropped: a single agent fixes one file, so there is no collision to isolate.
- Connectors — ➖ dropped: everything is local; this loop deliberately touches no external tool. (Real loops add connectors here — that is the only difference between this and Loop A/B/C/D.)
Anti-rationalisation
| Excuse | Rebuttal |
|---|
| "I can see the bug, I'll just fix it without running the tests." | The RED→GREEN exit code IS the loop. Run the checker first and last, or you have no proof and no stop signal. |
| "I'll edit the test so it passes." | Never edit test_calc.py. The test is the spec; fix calc.py. |
| "It passed once, we're done forever." | Memory records the run; a reset re-breaks it. The loop must re-derive GREEN from the current RED, not from last time. |
Reset (to run again)
pwsh loops/E-demo-sandbox/reset.ps1 (Windows) or bash loops/E-demo-sandbox/reset.sh restores the broken state.
