Claude Code harness: 203 of 239 AI tasks passed first review

The harness rests on a small set of rules the engineer states at the end: never hand everything to one run, split the stages and keep each stage's permissions minimal; use the same session for fixes that need context but a different session and model for independent judgment; force AI verdicts into a fixed shape rather than parsing prose; classify cutoffs by reason and don't let failures that aren't the AI's fault burn through retries; and keep state outside the process so work can resume wherever it stopped.

Several design choices follow from specific incidents. The AI is not allowed to push, merge or switch branches because it had failed several times pushing while the remote was ahead. Design runs read-only and review is limited to Read, Grep and Glob, on the reasoning that a reviewer who fixes things itself leaves no record of what it flagged. Design is deliberately not a mandatory gate: an empty design output does not fail the task, so a stalled auxiliary stage can't halt the main work.

Recovery is treated as a first-class concern. Resume points are recorded after implementation, after review approval and just before merge, so an interruption doesn't restart from scratch. Each launch also reconciles unfinished tasks that have pull requests against their real GitHub state, marking already-merged ones complete so the same task isn't implemented twice. Of completed tasks, 217 finished on the first run and 22 completed only after re-execution.