# Initial repository-agent reliability baseline

Observed on 2026-10-10 in the reliability implementation worktree, based on commit `44863cb`.

| Observation                                                           | Evidence                                                                                             | Limit                                                                                     |
| --------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------- |
| Existing clean-install local gate passed                              | Frozen install and `pnpm verify`, including module/browser, static-preview and built-consumer checks | Technical verification is not human task acceptance                                       |
| Guidance described a planning-only repository and old starting branch | Previous `docs/agents/parallel-workloads.md` against implemented workspace and examples              | Context drift is demonstrated; its historical rework cost is unknown                      |
| Dev/static browser commands used fixed shared-machine ports           | Previous package scripts and static Playwright config                                                | Collision risk is demonstrated by configuration; prior failed real tasks were not counted |
| Exact-commit handoff requirements already existed                     | `docs/agents/afk-implementation.md`                                                                  | Previous compliance and handoff reproducibility rates are unknown                         |

Acceptance misses, reviewer correction counts, regressions, accepted completion rate, and AFK throughput have no established measured baseline. Keep those fields unknown until actual issue/PR review supplies observations. Start recording the reliability implementation task as reviewable only after its final evidence passes; update to accepted only following human acceptance.

Compare subsequent real tasks by scope and complexity. Evidence of reliable simultaneous servers proves the isolation mechanism; it does not establish an agent productivity improvement. Existing product-roadmap decision tickets remain human-led.