Pipeline route check and leaner review
- The Unstoppable pipeline (3.4) checks the routing result before anything runs.
SKIPandDIRECTdispatches take a one-agentlocalpath with no orchestration plan; in 5.4.0 a trivial rename made the plan check fail. A stopped route ends the run. - A
capabilitiesprobe runs in parallel with planning, with the execute step's model and permissions.route_checkrefuses a delegated plan the runtime cannot run before anyone approves it. - One integrated reviewer by default. A second provider reviews only when the new
additional_reviewinput names a distinct risk. Required package reviews are unchanged. - Scenario revisions exclude a skill's
adapters/, so adapter-only edits no longer force re-recordings. A transcript that reads an adapter file fails instead. --record-scenariorecords four scenarios at a time.
Skill text is unchanged: /alchemy in Claude Code and Codex behaves as in 5.4.0.
Verification
- 20 performance regression tests passed; all 168 deliberate validator mutations were caught locally; installer tests 81/81.
- Affected behavior scenarios were re-recorded and validated.
- Live pipeline trial on the pilot fixture: a trivial rename failed on 5.4.0 after 46 s ($0.69) and finished on 5.5.0 in 37 s ($0.56) with 20/20 project tests. An audit objective planned multi-agent work, passed the plan check, and paused at plan approval. A no-agent pipeline confirmed the skipped reviewer leaves review, consolidate, and gate running.
- CI passed before publication.
Limits
No pipeline run has yet completed execution, review, and gate, in this or any earlier version. Multi-agent speed is unmeasured: one audit run per arm on plain Claude Code is not a statistical comparison. The local path edits the worktree without a plan-approval step; it never pushes or opens a pull request. The probe costs about $0.13–0.23 per run, including runs that do not need it.
Includes #48.