REPLAY · loop71

Behind the 38 minutes
the fastest PASS, taken apart

The fastest run so far (loop71, v146) built taskboard in about 38 minutes. Straight from the run records, here's what the local LLMs were doing between receiving the spec and the PASS.

37.7 min

from start to PASS

calls to the LLMs

1.08M

tokens read (input)

~$3.2

if it ran on Claude Opus 5.5

Where the time went

Width is time. Color is the model (blue = Qwen3.8-Flash-Next, green = Gemma-4-26B-A4B-it, yellow = Qwen3.8-27B).

What happened

"+N min" is time since the start. The seconds on the right are how long the LLM took to answer.

+0.0 min

1. DesignQwen3.8-Flash-Next19.1 min

Turn the spec into requirements113 s
Design (1st try)212 s
Check the design against the spec → gaps found159 s
Re-check the gaps it found67 s
Try a small design patch15 s
Redo the design (2nd try)188 s
Check the new design → OK159 s
Write the test plan170 s

+19.1 min

2. Test-plan reviewGemma-4-26B-A4B-it0.7 min

Check the test plan against the spec6 s

+19.8 min

3. Test writingQwen3.8-27B2.3 min

Write the pass/fail tests (Acceptance)87 s
Repair the tests once39 s

+22.1 min

4. Test reviewGemma-4-26B-A4B-it0.8 min

Check what the tests mean12 s

+22.9 min

5. Wrap-up checkQwen3.8-27B0.2 min

Take the review result → all good

+23.1 min

6. CodingQwen3.8-Flash-Next14.6 min

OpenHands：Write its own tests (Public) (10 calls, 93% cache hits)209 s
OpenHands：Part 1: task_model.py (8 calls, 85% cache hits)97 s
OpenHands：Part 2: task_store.py (6 calls, 87% cache hits)89 s
OpenHands：Part 3: task_service.py (7 calls, 89% cache hits)70 s
OpenHands：Part 4: app.py (the UI) (8 calls, 90% cache hits)92 s
Final check: does the app match the spec?246 s
Ran 38 tests → 🎉 PASS

Highlights

Half the time went into design (test plan included). The first design failed the check, a small patch didn't fix it, so it was redone from scratch — and passed on the second try. Resilient.
The reviewer is in and out. Gemma-4-26B-A4B-it's two turns each took under a minute, with only 6 and 12 seconds of actual answering. The rest is mostly swapping models.
Coding runs on the cache. OpenHands resends the whole conversation every time, so about 90% of its input hits the cache.
The final spec check takes a while. About 4 minutes just to confirm the finished app matches the spec. Careful.

All numbers come from loop71's run records (escalation-chain-run.json, llm_calls.jsonl, token-summary.json). The Claude API estimate uses Opus 5.5 list prices with the measured cache hits.