arcfluxresearch/dudes is a Rust daemon and CLI that assigns roles and goals to agents like Claude Code, Codex, and Pi, and has the daemon itself execute and verify the completion conditions. It also ships a 3D office screen to be viewed in a browser. The 2026-11-02 commit 2fd4671 fixes, as stated in the title, the problem of failing to reach the 'Ready for you' state in real repositories.
Instruction file monitoring is the feature that stores samples of the state of project rule files such as AGENTS.md that agents read, and compares them between runs. This commit looked at two bottlenecks, and the core fix went into bins/duded/src/instruction_sample.rs. One is the way a row was left for every candidate file in the sample. It stored even files that do not exist, so in real repositories it exceeded the sample budget, and that overrun held up tickets that could have passed for the reason 'cannot determine whether shared project instructions changed'. From now on, rows are left only for files that actually exist and for missing files that are import targets. A file that was absent at the start but appears at the end is still judged as creation, and the opposite as deletion. The other is the cost of traversing subdirectories. It rescanned the tree for every rule that looked below the root, so the limit on the number of sub-items was consumed multiple times within a single directory tree. Now the tree is walked only once per sample, rules that use the same tree share the results, and the headroom that hits the limit is reported again for each agent and context.
This change presents no performance figures or benchmarks. The evidence is the daemon flow test added to bins/duded/tests/instruction_review.rs, which confirms that even in a large repository both an agent alone and an agent with a copy reach the review stage, and that edits, deletions, and creations of instruction files, rule files, and config files remain intact in the shared folder. The fact that old-format samples containing nonexistent rows and new-format samples compare cleanly with each other was also locked in by a test. Two schemas (rpc, journal-record) were regenerated to match the sample input comment in the proto definition. However, since license or release information is not included in this commit, whether it has been distributed and the scale of adoption cannot be known.