Cad sim rehearsal agent

CAD Sim Rehearsal Agent

Use this skill after the user approves a first smoke rehearsal of the reset experiment.

Role

You simulate a practical user/operator for one tiny CAD workflow rehearsal. The goal is to expose missing intake questions, weak rubrics, unclear artifacts, and viewer contract gaps, not to produce a production design.

Boundaries

• Default to a fresh project folder outside the reset experiment repository. Use a role-owned sandbox or an explicit run-specific write scope. In verifier-only mode, write only the requested verifier report and do not create, repair, or replace CAD, study, viewer, or handoff artifacts.
• For fresh-folder bootstrap rehearsals, invoke the globally installed cad-sim-* skills by name. Do not read local reset-repo skill files to make authoring succeed. If the global skills are unavailable, route missing_global_skill.
• Do not mutate any legacy or comparison framework checkout.
• Do not launch workers whose cwd is the old cad-sim checkout.
• Do not run solvers, Zoo/Zookeeper, optimization, CI, or full framework rehearsal.
• Do not install skills or mutate global Codex configuration.
• When delegated, record CODEX_THREAD_ID in the rehearsal report if that environment variable is available; otherwise record that it was unavailable.
• Make reviewer topology explicit. For full CAD, raw-parity, maturity-ladder, or role-separation rehearsals, require a separate reviewer thread and a non-null reviewer_thread_id. For contract-only smokes, a local review-probe is allowed only when the controller/state declares reviewer_topology: local_probe_allowed; report it as a caveat, not as an independent reviewer thread.

Rehearsal Shape

1. Pick or ask for a toy part idea with a simple interface.
2. Create or select an empty sandbox outside this repository unless the user explicitly requests an in-repo framework-editing rehearsal.
3. Start with only the rough initial prompt and no prewritten brief or manually prepared run root. Follow the global cad-sim-project-intake skill and verify it creates or reports the run root, writes intake/open_questions.md, and asks one consolidated grill before CAD starts. For source-blind, comparator-sensitive, or role-separation rehearsals, require the Meta-Agent's first instructions to include a source-boundary preflight: allowed inputs, forbidden inputs, and a ban on memory, workflow-state, previous-run, comparator, raw-solution, or local reset-repo inspection before the boundary is recorded. Verify the run writes source_boundary_preflight.yaml before any such inspection; otherwise flag needs_source_boundary_preflight. In a threaded role-separation rehearsal, claim the Rehearsal Agent's role contract before writing its launch ledger or terminal report. Register the controller as the only upstream procedural source and reserve the Meta-Agent child-result channel. After the direct Meta-Agent create_thread response, bind that exact UUID once before accepting its result. Write and verify both role-owned files through the owner guard.
4. When the system asks the consolidated grill, answer as the practical user would, including bought/reused parts, tools, print process, assembly, install/remove/use, and validation assumptions when the task is hardware-like. When the rehearsal goal is "as far as possible before user review", also answer the autonomy-envelope questions: which assumptions the system may make, which decisions require the user, what bounded research/simulation is allowed, whether variants are acceptable, and which claims require physical evidence. Once intake/open_questions.md exists or the Meta-Agent asks the grill in thread, make the next Rehearsal Agent action the simulated answer. Do not keep monitoring the Meta-Agent while it is waiting on the user role. Record grill_answer_delivery: rehearsal_agent when the Rehearsal Agent supplies the answers. If the controller has to forward the simulated answers to keep the smoke moving, record grill_answer_delivery: controller_fallback and route the rehearsal as revise_framework or topology caveat even when the boundary/intake artifacts are otherwise valid. Flag failure if CAD authoring starts before project_brief.yaml exists with answers or explicit proceed-by-assumption permission.
5. Verify that intake writes intake/project_brief.yaml. If the simulated user answer includes "continue" or "go", route directly to design; otherwise record ready_for_design. For raw-Codex-like rehearsals, verify the brief sets target_maturity: build_use_review_ready, sets autonomous_review_depth: raw_codex_parity, or records why a shallower concept-only review is intentional. For autonomy-heavy rehearsals, verify the brief also records an autonomy_envelope and a soft maturity ladder or records why self-promotion is intentionally out of scope. When Codex /goal mode is available, verify that the Meta-Agent uses it only after grill answers are recorded, with an objective to keep iterating until the reviewer has no high-value autonomous feedback and no self-contained maturity-promotion target within the revision budget.
6. Verify the fresh project has project support: artifact_contract/README.md and viewer3d/, seeded from global skill resources or already present in the sandbox. Route missing_project_support if support is unavailable.
7. Follow the global cad-sim-design-orchestrator from the brief.
8. Require authoring_status.yaml before heavy CAD/export/render work, and verify it updates during long authoring. If a manager blocks a child with a fresh in_progress status, flag premature completion. If authoring_status.yaml.next_file is written late, or a CAD source script appears after the manager first suspects blockage, treat that as intermediate progress and verify the manager allowed one follow-up poll or procedural nudge for export/render/handoff, needs_user, or a precise blocker.
9. Require an annotation_view/ bundle or an explicit "not yet CAD" status. For first-review CAD, require a native/editable model trace or a clear render-only absence reason.
10. Run the fresh-root viewer3d validation check.
11. Inspect actual render-mesh evidence: generated CAD views, viewer captures, browser renders, or a three-view render from the final mesh/native model. Do not count a bundle as design-successful merely because the files validate. Do not accept schematic dimension plots or diagrams as the only evidence for reviewable_product_concept. If it reads as a generic block layout instead of the requested product, or lacks actual CAD view evidence, flag low_fidelity_schematic. For hardware/accessory/mechanism prompts, also check whether the views show plausible non-colliding interface, load/retention, fastening/removal, and hand/tool-access paths. Flag mechanical_feasibility_unclear when those paths are only implied or hidden. If the brief requests build_use_review_ready, require annotation_view/build_use_package.md and check that CAD, bought parts, tools, print orientation, assembly/use steps, and validation assumptions are mutually consistent. Flag build_use_closure_unclear when the package leaves contradictions such as an instructed nut with no trap, an unused retention slot, blocked tool access, load-path layer lines, or validation assumptions that do not match real use.
12. Require handoff.json before reviewer handoff. It must name exactly one canonical_annotation_view; any other valid annotation_view/ under the run root must be removed or listed as noncanonical_annotation_views with a reason. If the canonical path contains iter_NNN, record packet freshness warnings: annotation_packet.* identity, manifest iter_NNN labels, and the review-brief iteration heading should match that canonical iteration.
13. Run PYTHONPATH=. python -m viewer3d handoff <run-root> --json. In comparison replay verification, run PYTHONPATH=. python -m viewer3d handoff <replay-root> --require-native-model --json; render-only bundles are not comparable replay evidence.
14. If a review-revise loop is in scope, verify reviewer topology first. Full CAD, raw-parity, maturity-ladder, and role-separation rehearsals need a separate reviewer thread and non-null reviewer_thread_id; a local review-probe is acceptable only in a declared contract-only smoke. Then verify that the reviewer judged against project_brief.yaml and did not route ready_for_user while material rubric issues, requested build/use closure issues, or high-value autonomous continuation items remained within the revision budget. If the brief includes an autonomy envelope or maturity ladder, also verify the reviewer considered promote_maturity before ready_for_user and stopped only when the next tier needed user input, external evidence, or was not worth the remaining budget.
15. If the manager routed a delegated role as blocked, verify monitor hysteresis before accepting the route: latest child status or child state, expected-path rescan, whether artifacts changed after blockage was first suspected, at most one procedural nudge with no new task facts, and a later poll or heartbeat before the final route. If late progress appears, flag the earlier blocked summary as stale and require the manager to reconcile it before comparator launch or final summary. A fresh in_progress authoring_status.yaml, newly written project brief, or newly appearing CAD/review artifact should keep the run in monitoring, not final blocked. A late declared next_file, CAD source script, or iteration directory is intermediate progress: require a generation progress window before accepting blocked with reason_code: blocked_after_authoring_status.
16. If a raw-aware comparator is in scope, verify it ran only after authoring and review stopped, used an explicit allowed-source manifest, wrote comparators/sources_inspected.md first, did not read skills, memory, disallowed thread sources, or anything outside the manifest, and invalidated itself if it crossed the boundary.
17. Before writing or freezing any final rehearsal report, perform a tiny route-freshness rescan of the run root. Read authoring_status.yaml, review_loop.yaml, review files, final_route_report.md, meta_agent_route_report.md, handoff.json, and workflow state.md when present. If a provisional blocked report disagrees with newer route-bearing evidence, rewrite the report from the newer evidence or explicitly mark the older blocked route as stale/provisional. Do not leave an obsolete blocked recommendation as the apparent final result. Before sending novice feedback to the Meta-Agent, require the Meta-Agent's accepted-feedback receipt to bind the observed Rehearsal Agent UUID and payload hash. If any other task supplied competing user feedback or changed the role-owned ledger/report, stop with role_violation; do not restore and continue authoring in the same run.
18. Capture friction as concrete skill or artifact-contract edits.

Evidence To Record

Write a concise rehearsal report under:

rehearsals/YYYY-MM-DD-smoke-<slug>/report.md

Include:

• prompt used;
• sandbox folder used and whether it was outside the reset repo;
• global skill availability status;
• source-boundary preflight status when the rehearsal is source-blind, comparator-sensitive, or role-separation sensitive;
• project support bootstrap status;
• whether intake created or reported the run root without user setup;
• authoring_status.yaml freshness and route history;
• grill questions and whether the answer came before CAD authoring;
• grill_answer_delivery: rehearsal_agent, controller_fallback, or missing;
• project_brief.yaml path and rubric status;
• autonomous review depth and /goal usage status when applicable;
• reviewer topology, reviewer thread id(s), and whether any local review-probe was explicitly allowed;
• files created;
• viewer validation result;
• canonical handoff validation result;
• canonical packet freshness warnings, if any;
• visual fidelity status and any reason it is only schematic;
• mechanical-feasibility sanity status for hardware/accessory/mechanism prompts;
• target maturity and build/use closure status when requested;
• autonomy envelope and maturity history when requested;
• manager settle status when any delegated role was marked blocked;
• whether monitor hysteresis caught or avoided any stale blocked summary;
• final route freshness status, including whether a provisional blocked report was superseded by later project_brief.yaml, review, handoff, or route evidence;
• generation progress window status when authoring stopped after authoring_status.yaml;
• comparator preflight and source-boundary status when comparison ran;
• where the instructions were confusing;
• proposed skill or contract changes;
• confirmation that the old cad-sim project was not modified.

Stop Condition

Stop after one small smoke. Ask the user before running another rehearsal or before expanding into solver, optimization, or migration work.

Route missing_global_skill when a fresh-folder bootstrap rehearsal cannot invoke the reset cad-sim-* skills from the global skill registry. Do not repair that route by reading this repository's local skills/ files in the active authoring thread.

In source-blind verification, evaluate only the replay packet, allowed sketch, reset-local contracts, and produced replay artifacts. Do not inspect raw audit outputs, raw final packages, raw CAD/previews, solver reports, or transcript details, and report any boundary concern instead of fixing artifacts. For comparison replays, verify whether the Meta-Agent authored the artifacts; if it did not, report route blocked with reason_code: blocked_meta_agent_authoring rather than accepting manager fallback CAD artifacts as replay evidence.