Cad sim review revise loop

CAD Sim Review Revise Loop

Operate the post-candidate loop. In an ordinary project, the calling Meta-Agent is the combined Manager and primary author. It launches independent review, revises current artifacts, regenerates evidence, validates the route, and gives the user the terminal next move.

Dispatch

• Missing intake/project_brief.yaml: use $cad-sim-project-intake.
• Brief exists but no current handoff.json plus canonical annotation_view/: use $cad-sim-design-orchestrator.
• A current candidate needs review, revision, continuation, evidence closure, or maturity routing: use this skill.
• A fresh role-topology benchmark: use $cad-sim-threaded-rehearsal-workflow.

Do not bypass intake or use this loop to invent the first candidate.

Roles

• Direct Meta-Agent is the ordinary user-facing combined Manager and primary author. It owns state, reviewer launch, revision, deterministic validation, route promotion, revision budget, settle checks, and user interaction. It must not fabricate reviewer judgment.
• External Controller exists only for an explicit threaded benchmark, monitoring, or recovery harness. It supervises but does not author CAD or reviews.
• Main Reviewer independently reads the frozen packet and owns one route-bearing reviews/review_round_NNN.yaml.
• Up to two narrow read-only sub-reviewers may advise the Main Reviewer; they never own the route.
• A Tooling Scout may recommend evidence-producing tools. An optional Thread-Behavior Watcher may recommend a mode shift. Neither grades the packet.

Preserve real role and runtime UUID provenance. A controller-written review is not independent; local_probe_allowed is only for declared contract smokes.

Direct-Mode Self-Orchestration Invariant

The Direct Meta-Agent launches each fresh reviewer, integrates its judgment, revises directly when permitted, launches fresh re-review after material change, and continues to a validated terminal route. ready_for_independent_review, review_pending, revision_complete_pending_review, and ready_for_fresh_re_review are internal transitions, never user-visible terminal routes. Do not delegate manager duty back to the user or require an External Controller in an ordinary project.

Current State

Read the current instruction, brief/review contract, handoff.json, canonical candidate and primary surface, reconstructive source, authoring_status.yaml, review_loop.yaml, source boundary, target maturity, stage authority, evidence, and revision budget.

Use the first-review fast path: freeze the smallest trustworthy current packet and get an engineering judgment before compiling optional mirrors. Existing mirrors remain controlling when the brief activates them. Create or refresh only mechanisms triggered by the contract or a material finding:

• intake/active_review_rubric.yaml;
• reviews/current_evidence_workspace.yaml;
• Quality Harness V2 quality/quality_harness.yaml and generated quality/quality_harness_results.json;
• quality/configured_project_integrity.yaml;
• quality/comparison_manifest.yaml;
• the primary artifact registry/brief;
• a sanitized reviewer input packet.

An untriggered mirror is not a readiness defect. New Quality Harnesses use schema_version: 2, acceptance_profile, and mandatory profile rows.

Loop

1. Reconcile current truth. Validate the handoff, one candidate, one primary surface, registered secondaries, current evidence, and source boundary. Repair ambiguous current pointers before review.
2. Assimilate the instruction. Update review_contract.open_issues and regression guards. For fuzzy feedback, write an autopilot_prelude with concrete hypotheses; do not ask the user to diagnose self-contained defects.
3. Reconcile stage and profile. Continue approved in-stage work. Route unapproved scope, heavy/commercial tools, long runs, purchases, physical work, or a new objective to needs_user. Profile inference shadow output is telemetry only and cannot change validators, the brief, or route preflight.
4. Prepare and launch one neutral reviewer. Follow the lifecycle invariants below and the reviewer launch prompt. The reviewer applies $cad-sim-reviewer-lenses, derives compact governing_analysis_rows, and audits current engineering before surface polish.
5. Integrate and route. Activate only contracts justified by the judgment. If approved self-contained work remains, write one sanitized Revision Brief and revise directly. In split topology only, send that brief to the designated author.
6. Regenerate material proof. Refresh affected source, exports, handoff, verifiers, runtime/visual evidence, harness results, comparison hashes, primary surface, registry, and route state. Old review evidence is no longer current proof.
7. Freshly re-review material change. Schema-only repair stays in the same reviewer thread, round, and file; any CAD, model, evidence, analysis, or user-facing change requires a new frozen state and fresh reviewer.
8. Promote and close. Atomically promote only allowlisted route state, run full deterministic route preflight and every declared project verifier, and reproduce the validated user_visible_next_move or final_response_directive without alteration.

Default to two substantive revisions, or three for explicit raw_codex_parity, unless the brief sets another bound. Schema repair is not a substantive round. After the bound, preserve the candidate and unresolved findings; do not launch a rescue author for the same material state.

Reviewer Lifecycle Invariants

Detailed commands and recovery state live in the manager protocol; deterministic enforcement lives in scripts/reviewer_lifecycle.py.

• Run prepare-review to compute the canonical material state, close over transitive current evidence, create the immutable mission packet, and publish the launch preclaim. Never trust a caller-supplied material id.
• Bind evidence_escalation_autopilot to that state and round, then consume guard-launch once. Its read-only freeze gate rejects Harness drift, stale escalation, prior-review framing, packet drift, or unregistered inputs.
• Give the reviewer only active_mission.reviewer_input_packet.allowed_input_root; it must not read the broad run root, prior reviews, feedback receipts, topology, or transcripts.
• Record the actual runtime UUID from the direct create response with record-create-response. Never use agentPath, task discovery, or a create retry. A UUID-less response settles fail-closed at blocked_before_review.
• Command verifiers run only through run_reviewer_verifier.py against an isolated material copy.
• The reviewer writes an in-progress checkpoint, then a complete same-directory .review_round_NNN.yaml.pending. reviewer_lifecycle.py publish rechecks identity, hashes, receipt references, and lifecycle state before os.replace.
• Resume a quiet reviewer with non-interrupting settle checks. Do not replace it merely for elapsed time.
• Delegated feedback is actionable only after an accepted-feedback receipt binds the source runtime UUID and payload hash.

Launch prompts may name keys, allowed inputs, permissions, routes, and literal checker errors. They must not prescribe passing values, empty lists, conclusions, or a preferred route.

Revision And Route Invariants

Reviewer criteria and conditional output fields belong to $cad-sim-reviewer-lenses; do not duplicate its engineering checklist here. For master's work, the canonical review still emits route_usefulness, target_outcome_status, and governing_analysis_rows. The required architecture_and_behavior_integrity row must close before ready. A physical boundary belongs in a separate row.

When Quality Harness is active, preserve quality_harness_used, quality_harness_audit, failed_or_missing_harness_rows, harness_evidence_promoted, and quality_reset_recommended. Exemplar work preserves exemplar_calibrated_handoff_status, reusable_reviewer_tests, and integrated_exemplar_judgment.

Promote route state only through the exact allowlist in the route-preflight contract. Never recursively scan the run root for route pointers. An author must not read work/workflows, controller state, or controller messages during closeout. --review-contract-only, viewer health, and manual notes do not prove terminal settlement.

The allowed terminal routes are:

• continue_autonomous: useful approved self-contained work remains;
• revise: bounded correction before fresh review;
• promote_maturity: a pre-approved next stage should begin;
• needs_user: a targeted choice, approval, measurement, credential, upload, purchase, physical result, external run, or broader stage controls progress;
• ready_for_user: the current goal and concrete next action are ready;
• ready_for_bench_test: digital work is closed and the next truth is physical;
• ready_for_review_with_boundary: the bounded review deliverable is complete;
• max_revisions_reached, blocked, or invalid_boundary: explicit terminal failure/boundary routes with coherent evidence.

Quality Reset is the user-facing phrase; deterministic mode is clean_restart. An unresolved recommended reset blocks ready.

Next Move And Resume

A ready directive names a concrete action such as measure, print, bench-test, calibrate, deploy, run, record, upload, buy, order, or build. compare, inspect, choose, and reply are not ready-route actions. A user choice or stage approval is needs_user with options and a resume point. Do not finish with a broad feedback question or require audit-only/superseded side files.

Resume Answer Assimilation

Classify a resumed answer before work:

• selected_option;
• selected_option_plus_objection;
• instruction_completed_plus_complaint;
• new_blocker;
• scope_change.

Apply the stored choice/result, update the brief and active issues, then resume at needs_user_resume. Clear it only after durable assimilation; never ask the user to repeat settled intake facts.

Conditional References

Do not preload every reference. Open only the direct contract activated now:

• state, Revision Brief, settle, recovery, route promotion: references/manager_and_revision_protocol.md;
• reviewer launch boundary: references/reviewer_prompt.md;
• final directive: references/next_move_gate.md;
• schemas and deterministic commands: references/route_preflight.md;
• escalation/Quality Reset: references/evidence_escalation_autopilot.md;
• stage authority: references/stage_frontier.md;
• profile inference: references/rubric_profile_autopilot.md;
• raw/comparator boundary: references/comparator_preflight.md;
• engineering review and output: $cad-sim-reviewer-lenses.

Completion

Stop only when the same current state has: one complete independent YAML review; coherent allowlisted route pointers; current hashes, Harness results, and project verifier receipts; fresh audits of every active user-facing artifact; full route preflight ok: true; settled child tasks; and a validated terminal directive. The final response states that route and reproduces the directive exactly.