Participant: codex (AI agent)
Understanding
compareWorkspaceIntent correctly scopes its result directory, but passes the
unchanged ambient T2CConfig to both pipelines. AST and documentation caches
read config.outputDir, so an external --out can still create .intent/cache
inside the analysed checkout. Existing output can also be reported as dirty
input. The comparison then observes artifacts produced by the observer.
Execution plan
- Add a regression with an external output directory and enabled cache.
- Bind each pipeline config’s output directory to its actual pipeline output.
- Exclude only the selected generated directory from the pre-run Git snapshot;
retain every unrelated dirty file.
- Run focused, full, governance and Docker verification.
Actual changes
- Read governance and current implementation, reproduced the defect on clean
Registry and Core worktrees, and recorded explicit approval before edits.
- Bound both pipeline configurations to their actual output directories and
excluded the selected in-repository evidence path from the status snapshot.
- Added regression coverage and passed focused, full Node and Docker checks.
- Transitioned to
IN_PROGRESS / PUBLICATION; no merge is claimed locally.
- Reopened the merged comparison workstream after live
PLF-8091 evidence
showed a 142,557,246-byte generated Platform graph failing the generic
128 MiB JSON ceiling.
- Kept the generic JSON limit unchanged and introduced a comparison-only,
bounded 256 MiB graph ceiling with focused regression coverage.
Current state
IN_PROGRESS / PUBLICATION; the exact 142,557,246-byte Platform graph replay,
426-test suite, governance and Docker gates pass. Protected review remains.