---
source_hash: acc5654f088e13d158334a26a24f0e1898d638ae026cf51b5d2bcac59f42ee6f
post: /posts/workflows/540-commits-one-human
shape: doctrinal
---

# Context Pack: 540 Commits. One Human.

Agent reference for the rewritten June 2026 retrospective. Only quoted excerpts are verbatim. All counts describe the reported eleven-day period, not the current repository or a general benchmark.

## Verbatim excerpts

> In the eleven days behind this June 2026 account, 540 commits landed.

> Another model audited the result and returned pass, fail or escalate before it landed.

> More retrieved material had not meant better help.

## Reported figures

- 540 commits in 11 days; peak day 106.
- Six overnight tasks, independently audited, about two hours in total, without Ari intervening during the run. She supplied the plan beforehand.
- Graph precision 54% to 84%, then a 9.5-fold speed improvement in a day.
- A signed macOS app with 165 passing tests; 600 passing tests on the memory engine.
- A runaway loop stopped by a hard cap after about £2.
- A supposedly smarter retrieval approach fetched 29% more context and produced worse results than the simpler baseline, so Ari cut it.

Commit and test counts do not alone establish product quality or complete coverage. Do not turn the no-intervention overnight run into a claim that no human planned or governed the work.

## Model assignments and review

Fable held orchestration; Codex heavy code; Kimi long video; MiniMax M3 looping builds; local models memory; Opus 4.8 additional capability where needed. Preserve these assignments without inventing a distinct-model count from the original inconsistent five-model label.

Independent model audit was required before landing. It reduced routine human checking but did not guarantee infallibility or remove human-owned decisions. The surrounding system handled planning, routing, verification, memory and structural cost limits.

The combined system was called ARI-OS, with Cortex as memory. The practical lesson is to define the input, responsible executor and checkable result for one part of a workflow before stepping back from it.

## Related

- [The Codex loop](/documents/discipline-and-process/the-codex-loop).
- [The model is the replaceable part](/documents/tools-and-plugins/the-model-is-the-replaceable-part).
- [Stop prompting, start defining outcomes](/documents/thoughts-and-scribbles/stop-prompting-start-defining-outcomes).