mirror of
https://github.com/EveryInc/compound-engineering-plugin.git
synced 2026-10-02 19:57:22 +02:00
Trevin Chow
ddf730e9d6
The ordinary refresh judges accuracy, so a learning whose reasoning the repo already states in a test, a comment, or the instructions file was a clean Keep and the store could never be brought to the capture bar that #1624 gave ce-compound. A cleanup, cull, or upgrade intent in the arguments now turns on a worth lens: the run states that reading back, confirms once, and only then reads a reference that adds the same counterfactual as a second test. Recoverability needs a named artifact quoted per claim; fully recoverable routes to Delete, partly to Update with the recoverable content cut. Non-interactive runs never delete on inferred intent and record those verdicts as recommendations. The bar is byte-duplicated from ce-compound's Preconditions inside a delimited block and pinned by a parity test. Two blocks moved verbatim from the refresh body to the references that own them to stay under the Codex prompt budget. Three eval-cell scenarios cover intent confirmation, the plain refresh, and the confirmed path; all passed on Claude and Codex. Claude-Session: https://claude.ai/code/session_01MoVcBFERvTuVqdyKjx9aEC