ljchang/mecha is a standalone agent harness written in Rust. It is a project that gathers in one place a provider-independent loop, MCP tools, a path jail with prompt injection blocking, a sandbox shell, scheduled triggers, and evaluation tools that score execution traces. The November 7, 2026 GitHub commit d2adb37 added the session kind `driven` while merging PR #848. `driven` denotes a session in which another agent ran on top of the owner's actual work, and it is marked with `MECHA_SESSION_KIND=driven`. The change spans 18 files, with 460 lines added and 53 lines deleted.
The criterion this marker distinguishes by is not 'where it was stored' but 'who ran it'. Driven sessions remain in the owner's own repository, but, like test or experiment sessions, every reader hides them by default and still counts the number hidden. That value is `Corpus::hidden_driven` and the `driven_hidden` counter that list, health, and appraise attach, and it is only read along when `--kind driven` is explicitly specified. The learning paths are closed outright. reflect's transcript pass, distill, the contrast pool, and the persona writer all reject driven sessions through `Scan::admits`, and reflect's outbox pass skips drafts whose staging-session admission was rejected. The design document calls this criterion fitness (actual usage), giving the reason that what wears the harness down to failure is not the owner's own usage.
The remaining limitation is also clear. What this commit covers is the design document's D1 rule, while D2 — giving driven runs a dedicated learning store and keeping them as a marker rather than an evaluation scope — remains in DRIVER-DESIGN.md instead of among this commit's changed files. The place where the judgment shows is the diagnose command. When it could not read a single session, it used to point out only test sessions, and a repository with only driven sessions left would fall through to 'the filter found nothing'. Now it names and reports all excluded markers, test, experiment, and driven, and a regression test confirming this behavior was included as well. The ARCHITECTURE document and the CLI reference document were updated together. However, the structure by which other agents' runs come in inside the owner's repository remains as it is, so this commit is only a choice to block them by default exclusion. How the evaluation tools mentioned in the repository description are affected by this marker is not stated in the given materials.