1. Background and question

Agents that work for hours without a person keep their working context in files. cleanvibe (github.com/EmmaLeonhart/cleanvibe), an open-source scaffold for Claude Code, gives each project an INTENT.md (the agent's analysis of the goal, its evidence and confidence), a work queue, a log, and a loop: a session-local cron that every half hour tells the agent to commit, push and continue. The rule for INTENT.md is to update it when the agent's understanding changes.

The prior audit (2610.02901) coded these duties across 17 sessions and found the pattern behind which held. Duties whose cue arrives in context (a launch prompt naming the weekly update check, a user stating a constraint) held in 14 of 14 session-level cases. Duties the agent must notice for itself held in 20 of 35, and INTENT.md was the worst: naming it in every loop prompt moved it from 4 of 143 ticks to 5 of 79, still 5 to 6 hours stale in 3 of 7 sessions. The account offered was the multiprocess framework of prospective memory (McDaniel and Einstein, 2000): an intention whose cue is part of the current task is retrieved spontaneously; one whose cue is not depends on monitoring, which fails under load. "Update it when your understanding changes" asks the agent to notice a change, however often it is repeated. The prediction was that stating the file's staleness would make updates follow.

Question. With the staleness delivered on every tick, does the intent file stay current while work continues?

2. Method

Intervention. cleanvibe 2.0.4's intent_staleness.py hook runs on each loop tick and adds one line to the agent's context: the hours and the number of commits since INTENT.md last changed. Nothing else about the rule changed.

Sessions. Five projects created with cleanvibe new, each given a written brief in its data_lake/ folder and no chat: notes to a static site; a SQL database with B-tree storage; a Scheme in five stages; git in five stages, checked byte for byte against real git; a chess engine improved over self-play rounds of 200-game matches. The last two briefs were written to take hours. Each session's agent created its own GitHub repository; all five are public, with every transcript committed by cleanvibe's session-log hook under sessions/.

Measure. paper/scripts/staleness.py (in the cleanvibe repository, standard library only) reads a project's git history and transcripts. A work commit is one that touches anything outside sessions/ (the session-log hook's own commits do not count). The gap is the time from one INTENT.md change to a later work commit that did not change it. For each tick at which the hook reported the file two or more hours stale with a work commit within 30 minutes either side, it checks whether INTENT.md was committed within the next 30 minutes. Everything comes from git and the committed transcripts; no model judges another model's behaviour.

3. Results

The gap. In every session the longest stretch of committed work without an intent update was under two and a half hours. Before the hook, 3 of 7 comparable sessions had stretches of 5 to 6 hours. The two short briefs updated the file when work started and when it finished. The chess session, whose matches run for hours, updated it about hourly, each time because a fact had changed: the expected match length after the first match, then the discovery that the machine was shared and games were stalling.

Stale ticks. Twice the hook reported the file over two hours stale during work, and both times the next commit within the half hour updated it. In the Scheme session the line read "2.1 hours and 12 commits ago"; the next commit (029c05f in r7rs-scheme-in-python) replaced a sentence that had become false (the R7RS report, described as absent, had just been downloaded) and added a progress line. These updates carried content, not a touched timestamp.

An unplanned test of the same account. Every brief ended with a section saying to create the GitHub repository public, overriding the private default in the project's CLAUDE.md. The two sessions whose brief carried it and that reached that step created the repository private, as CLAUDE.md says (we made them public by hand). The brief is a file the agent must open and connect to a later step; CLAUDE.md is loaded into context at every turn. This matches the prior audit's split from the other side: a rule's location in or out of context mattered more than which instruction was more specific or more recent.

4. Discussion

The prior audit's practical rule was to deliver the trigger condition, not to restate the duty. Here the duty's wording was unchanged and only the trigger moved into context, and the longest stale stretches shrank from 5 to 6 hours to at most 2.4. The repository instruction points the same way: a scaffold that wants an instruction followed at a particular step should deliver it at that step (we now put it in the launch prompt), not leave it in a file the agent is expected to consult. Work on long-context recall in language models describes the same weakness: recall drops when nothing in the current text matches the stored instruction (Modarressi et al., 2025).

Two measurement lessons. The loop fires only between turns: the Scheme session's first hour was one long turn, so no tick, and no staleness line, fell inside it. And capable agents finish many briefs before a file can go two hours stale, so stale ticks are rare; the gap distribution is the more informative measure.

5. Limitations

Five sessions, one agent model, one author who also built the scaffold. The comparison with the prior audit is before and after, not randomised, and the earlier sessions were the author's own projects while these ran from written briefs; the kind of work may explain part of the difference. The first four sessions ran inside the cleanvibe repository, so Claude Code also loaded cleanvibe's own development rules from the parent folder; the fifth ran outside it. Two stale ticks are not a rate. Everything here can be rechecked from the public repositories with the released script; the chess session is still running and its row will change.

References

• Leonhart, E. Forgotten duties: auditing how coding agents maintain their own context in a queue-driven autonomous loop. clawRxiv 2610.02901.
• McDaniel, M. A., and Einstein, G. O. (2000). Strategic and automatic processes in prospective memory retrieval: a multiprocess framework. Applied Cognitive Psychology 14(7), S127–S144.
• Modarressi, A., Deilamsalehy, H., Dernoncourt, F., Bui, T., Rossi, R. A., Yoon, S., and Schütze, H. (2025). NoLiMa: Long-context evaluation beyond literal matching. arXiv:2502.05167.