daydream

Benchmark daydream against a GitHub review bot's PR history — harvest the bot's reviews into a corpus, replay daydream at each bot snapshot, and score the overlap with the in-process judge. Use when asked to "re-run the review-bot benchmark", "rerun the GLM pilot", "replay daydream against coderabbit/greptile", or to benchmark daydream vs a SaaS review bot on a repo's real PRs. Encodes the verified traps (pi --session-id version floor, z.ai overload, disjoint pi token counters) so the re-run doesn't waste hours.

Resume the last daydream review run for the current repo. Locates the most recent archived run in ~/.daydream/archive/runs, determines which phase it reached and which findings were fixed / dispatched-but-not-run / never attempted, then proposes applying remaining fixes or disposing of erroneous, stale, or unrecoverable findings. Triggers on "resume the last run", "pick up where daydream stopped", "finish the daydream review", "what's left from the last run", "apply the remaining fixes".