Resolve two locally available Git refs into immutable commits and diff metadata.

The repository path is on the server host. This compares the resolved commits directly; choose the merge base explicitly when a merge-base comparison is required. It does not fetch refs, inspect uncommitted changes, or construct the evidence packet. Unresolvable refs fail; both refs must resolve to different commits with a nonempty diff.

Return one specialist's fixed ReviewRoute checklist: profile, display_name, focus, required_sections, constraints, and escalation_signals.

Use before reviewing one scoped concern when no session is active, or to recover a checklist when the triage response is unavailable; no session is required.

• Standalone review: use get_qa_orchestration_catalog() to match profile focus to changed files and confirmed stack.

• Session primary_review: use selected_review_routes returned by the successful triage advance call. It contains every selected checklist and execution_group; do not make a separate preparation call for the same session.

This local lookup reads bundled definitions, without executing reviews or calling models. Use advance_qa_orchestration to apply a session route. Results depend only on agent_profile. Invalid values fail input validation; retry with a listed enum value.

Return every fixed ReviewRoute checklist and execution wave for one bundle.

Use for standalone bundle lookup or to recover checklists when the successful triage response is unavailable. In a session, the triage advance response includes the selected checklists and execution_groups. Run execution_groups in order; profiles in one group are independent and may run concurrently when the host supports it. For session transitions, follow advance_qa_orchestration's group-reporting rules. For a single profile, use prepare_qa_orchestration. This stateless lookup reads bundled definitions; it does not execute reviews, create a session, or call models.

Read the static profile/bundle catalog: specialists, review scopes, and execution schedules.

Use before creating a session or choosing its triage route; no session or arguments required. Task-type recommendations are shortlists, not restrictions. Choose a bundle for broad work or a single profile for a narrow concern using changed files and confirmed stack.

• For a session, the successful triage response includes selected_review_routes with the selected profile or bundle checklists and execution_groups.

• Standalone or recovery lookup: prepare_qa_orchestration(agent_profile) or prepare_qa_orchestration_bundle(bundle).

• Create a session: start_qa_orchestration; set its route: advance_qa_orchestration.

• Session IDs: list_qa_orchestrations().

Available routes come from the installed server's bundled definitions; active sessions and configured models do not change this catalog.

Read the server-wide model policy used to create new sessions.

The mode is either host_default (use the host's configured model and reasoning) or configured (use the returned per-stage provider, model IDs, and reasoning).

Use to inspect the policy loaded for new sessions; takes no arguments.

• Existing session state and current-stage policy: get_qa_orchestration(run_id).

• Create a session with these settings: start_qa_orchestration.

Reads the loaded in-memory snapshot, not the current policy file; restart the server connection after file changes. No session is created and no provider is contacted; the result does not verify which model the host executed.

Create a QA session and bind code reviews to one immutable review input.

Before calling for ordinary_review or widget_review, complete review preflight: resolve the authoritative task/MR, pin exact base_sha and head_sha for Git input, or use input_kind=diff with no Git revisions for a standalone unified diff; inspect the exact diff, build the host-owned Evidence Packet and stable refs, compute diff_sha256 and evidence_packet_sha256, and provide input_manifest.schema_version=2 with source_inventory_status and source_conflict_status. A task or MR URL alone is not enough. If an exact repository, required Git revisions, diff, or packet identity cannot be pinned, do not start; resolve it or tell the user what is missing. Record unavailable requirement/spec sources with partial or unknown manifest statuses.

Every call creates a separate session. Use prepare_qa_orchestration for a stateless profile checklist. task_type narrows the bundle shortlist; the host still selects the review route. Code review task types require review_context with the exact source and Evidence Packet identifiers plus an input manifest schema v2. The server stores identifiers and submitted report fields, not the raw diff or full Evidence Packet.

• Sessions expire after the configured TTL (1800 seconds by default).

• Capacity is 100 sessions by default and can be configured.

• Expired sessions are purged and retained terminal sessions may be evicted. If capacity can only be freed by removing an active session, the call returns a session limit error.

Record one active QA step's result; return updated state and next_action instructions for the host.

After completed triage, the response also includes selected_review_routes with the exact selected checklists and execution_groups, avoiding a separate preparation tool call.

For finalization, use finish_qa_orchestration at awaiting_host_outcome; after early stops, match the recorded outcome.

• Triage: select exactly one bundle or profile; default to one profile for narrow work. Set implementation_review=true when implementation behavior is in scope.

• Every call requires run_id, completed_step, and status. Copy completed_step from the latest session's current_step; use status=completed after a successful stage.

• Primary review: after a whole execution group finishes, report its ordered profiles together with completed_profiles. In a review-context-bound session, also supply matching profile_reports. Older hosts may continue reporting completed_profile one at a time. The final group requires risk_signals ({} if none).

• Clean single-profile fast path: set skip_synthesis=true only when the route contains one profile and that review has no findings, no unresolved material questions, and no risk signals. This bypasses only the separate synthesis step; the host still validates and produces the final QA report. The server cannot verify this claim. Bundles, findings, unresolved questions, or any risk signal require the normal synthesis stage.

• Trust boundary: completion fields, execution_metrics, and risk_signals are host-reported. The server cannot confirm each review ran, whether it ran concurrently, or that risk_signals reflect evidence; bound sessions validate report/profile alignment, citation membership, finding-source lineage, and risk-signal reference membership, not report truth.

• Early stop: provide stop_reason; omit route selectors, completion fields, and risk_signals.

• Success saves state and refreshes expires_at.

• Invalid/stale requests and unknown/expired run_ids fail without advancing.

• With request_id and expected_revision, exact retries return current retained state without another transition or TTL refresh. Conflicting payloads and new stale revisions fail.

• Legacy calls without request_id are non-idempotent: never retry them blindly. After a lost response or transition error, read get_qa_orchestration(run_id) and continue from its current_step; if the session is unknown or expired, start a new session.

Extend an active session's TTL while host review work is running.

Pass the retained run_id and, optionally, the latest session revision as expected_revision; a stale revision fails without renewing. Before expires_at, active primary/deep/synthesis and awaiting-host-outcome sessions may be renewed. Terminal or expired sessions cannot be revived. Only the lease changes: no workflow revision advances and no worker or model is launched. Use advance_qa_orchestration to record progress and finish_qa_orchestration for the outcome.

Read one retained run_id's state, audit trace, bound context and structured reports.

Use run_id on the server retaining that session to resume or reconcile a lost advance_qa_orchestration response. Reads do not extend expires_at. Errors unknown run_id or expired session require a new session via start_qa_orchestration. If run_id is lost, use list_qa_orchestrations() to find retained sessions first.

For all-stage server policy, use get_qa_orchestration_model_policy(); for review progress or final outcome, use advance_qa_orchestration or finish_qa_orchestration, respectively. Local stdio access requires no additional credentials.

List non-expired QA session identifiers and brief metadata, including bound review identity.

Recover a lost run_id by matching task type, route, and stage; confirm if several match. Call get_qa_orchestration(run_id) for full state and next action, or get_qa_orchestration_model_policy() for server policy. Takes no arguments.

• Includes retained terminal sessions; expired or evicted sessions are unrecoverable.

• Sorts by expires_at descending, then run_id descending; empty means none are retained.

• After restart, unfinished sessions and finalized sessions with saved review artifacts restored from configured local storage appear.

• Reading does not extend session TTL.

Delete one QA session from memory and configured recovery storage. To retain a final QA outcome, use finish_qa_orchestration.

run_id selects one session at any stage; no prior advance or finish required. Recover a lost ID via list_qa_orchestrations(), matching task type, route, and stage; confirm the target if several entries match.

• Idempotent: deleted=true confirms absence, including unknown or expired IDs.

• Irreversible: get/advance/finish cannot use the ID, even after restart.

• Storage failure: memory stays intact; resolve the error and retry.

Host tasks and model executions are unaffected.

Finalize the QA run and store its structured report when the session is review-bound.

After normal synthesis or an eligible clean single-profile fast path, use the retained run_id at awaiting_host_outcome to choose completed, partial, or blocked. After an early stop recorded by advance_qa_orchestration, outcome must match its partial/blocked status.

• A completed review-bound session requires final_review_report.

• Structured reports survive a restart only with the same configured session store; otherwise they are in-memory only. Legacy finalized sessions without review artifacts are removed at restart.

• Same outcome: returns the retained terminal session.

• Different outcome: conflicting final outcome; status unchanged.

• Premature call: outcome is not ready. Retained sessions expire at the configured TTL (1800s default), or are removed by eviction or explicit deletion. Removed run_ids cannot be deduplicated.