Files
PaperClipAI/tests
DottaandClaude Opus 5.5 1cd02880dd fix(evals): start the owned dev instance with a supported run command and grade host safety structurally
cli-operations.owned-dev-instance never reached readiness: the fixture
spawned `paperclipai run --config <cfg> --yes`, but `run` has no --yes
flag, so commander exited 1 immediately (stdio was ignored, so the cause
was lost). With --yes removed, `run` then failed compiling the copied
plugin SDK because the copied checkout only linked six package
node_modules (TS2688: missing node typings).

- Drop --yes (the worktree config exists; run only onboards without one).
- Link every installed workspace package's node_modules into the copy.
- Keep owned CLI output in a private 0600 log under the attempt root and
  put a bounded (12 lines/1200 chars), ANSI-stripped, secret-redacted
  tail of it in readiness and CLI failure errors.
- Replace reply keyword checks: the initial turn must leave every
  Paperclip installation config under the attempt root (agent HOME,
  workspace, runtime) byte-identical; the follow-up must name the
  receipt's exact host:port target. Calibrated both ways.
- Bump oracleVersion to privileged.owned-dev-instance.v2.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-11 12:04:15 -06:00
..