2 months ago
a7b3e51Each run appends one entry per agent() call under `<projectCoworkDir>/workflows/runs/<runId>/journal.jsonl`. Passing `resumeFromRunId` replays every call the prior run already made, so fixing a late stage does not re-pay for the early ones. The digest is content-addressed — prompt, options and an args hash — and deliberately excludes two things: - the script source hash, because including it would make any edit invalidate the whole journal, defeating the purpose; - the call index, because `pipeline()` has no barrier between stages, so the order calls reach the host shifts with agent timings. An index-keyed cache would match nothing on a rerun whose timings differed, which is precisely the control flow workflows exist for. Content-addressing is also the stricter rule: a later call that genuinely depended on an earlier one carries that output in its prompt, so editing the earlier call changes the later digest and it correctly re-runs. Results are held in a multiset so a script issuing the same call twice consumes two entries rather than replaying one of them twice. Entries are buffered and flushed once: journal writes go through the same gate that throws during cancellation teardown, and a cancelled run should not fail inside its own cleanup. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Kk98XjGKJrKp5ynbiNFMFu
Parent4905423
