Harness test plan T1 — Gemini (free tier) and the automation around it¶
No memory needs freeing: the model runs at Google; the laptop runs git and pytest only.
| # | Test | How | Pass when |
|---|---|---|---|
| T1.1 | Set-up check | python -m mmdep.cli harness check *> C:\Projects\harness-check.txt (Claude reads the file) |
pytest ok, GitHub ok, gemini-flash-lite model found |
| T1.2 | Dry run | python -m mmdep.cli harness run --dry-run *>> C:\Projects\harness-check.txt |
Issue #14 listed as "would work this issue" |
| T1.3 | Real run on pilot #14 | Scheduled task at :16 (or python -m mmdep.cli harness run --issue 14) |
Branch auto/issue-14 on GitHub; harness/LATEST.md and ledger.csv on the control branch; tokens recorded; full suite shows no new failures |
| T1.4 | Review and merge | Claude reviews the diff, opens the PR; Atul approves the merge | PR merged; issue #14 closed |
| T1.5 | Failure path | Claude files issue #15 with a deliberately unclear task and a strict test | Status failed, nothing pushed, log kept, ledger row written |
| T1.6 | Refusal path | Issue #16 routed ctx: restricted |
Skipped with the reason; no model call |
| T1.7 | Quota fallback | Issue #17 starts on gemini-flash (20/day) after its quota is used |
Falls back to Flash-Lite; ledger shows "gemini-flash > gemini-flash-lite" |
| T1.8 | Wake / reconnect trigger | Laptop sleeps and wakes, or Wi-Fi off/on | A run within about 2 minutes (Task Scheduler history; LATEST.md time) |
| T1.9 | Stop switch | Create C:\Projects\_mmdep-control\harness\STOP |
Next run does nothing; delete the file to resume |
| T1.10 | Cap | Record a test purchase harness spend --item test --usd 0.01, then delete the row |
Month-to-date includes it |
Later (T2): local model on an old machine (Ollama, Qwen2.5-Coder), same issues, compare pass rate and minutes.