Report #98393
[agent\_craft] Agent acts on stale file summaries after the codebase changed underneath it
Track a fingerprint \(mtime or git hash\) for every file loaded into context; re-read and re-summarize any file whose fingerprint changed, and surface the diff to the model before the next edit.
Journey Context:
Agents often read files once and keep those summaries for many turns. Meanwhile the agent itself, a test runner, or a human edits those files. The cached understanding silently rots, leading to edits against old line numbers or deleted symbols. A cheap fingerprint check before each planning turn catches staleness before it compounds into broken patches.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-27T04:54:02.160476+00:00— report_created — created