Read Portable review protocol, Backend contract, Orca native mechanics, and Purpose and architecture contract completely.

Accept only an exact 40-hex , intent source, scope, mode and two user-approved reviewer selections. Read selections through bundled when it is available; never pick a fallback model, effort or posture flag. A dirty/mismatched checkout or invalid selection is .

Create both review tasks before launching either worker. Start one Codex and one Claude reviewer (or another explicitly approved vendor-diverse pair) with stable titles . Each receives original intent and candidate, but no peer bytes. The first lifecycle pair uses ; remediation uses with only that reviewer's own prior report and dispositions. is only an explicitly requested standalone/advisory mode. The portable protocol may escalate any mode to .

Each reviewer proves the full SHA and clean status, performs non-mutating review, and returns the entire textual report in authoritative Orca . Its native Dispatch id is the opaque . After accepted delivery, show a brief TUI severity summary linked to that id. A missing summary is a UX defect, but never invalidates a complete ; terminal text never replaces a missing/rejected report. If delivery is not confirmed, display and return .

Wait for both full reports before releasing either to the caller. Preserve the original bodies without merging, ranking, paraphrasing or exposing peer output. Keep remediation reviewers hot. Before the one final same-SHA proof, release only owned old sessions and create two fresh independent sessions with no prior reports. Never close unnamed or foreign tabs.

P0–P2 block settlement. Deliver every P3 for fixing or reasoned rejection, but do not create a separate P3-only round. A substantive slice permits at most five paired review/fix attempts; remediation does not reset that budget. This limit does not weaken the final requirement: settlement still needs two independent reports on one exact SHA, or an honest .

Never use , edit/commit, start another reviewer, or run concurrent full QC. Report E2E as not evaluated unless separately requested.