Skip to content

MiaoDX/roboclaws / skills/eval-harness

eval-harness

Verified

Select and run Roboclaws validation, product, eval-suite, and live-agent eval rows from a plan, diff, or explicit capability request.

Eval Harness

Use this skill when a Roboclaws plan, diff, PR, or agent-facing change needs one maintainer proof surface. The skill answers:

  1. which deterministic gates, product runs, eval suites, and live-agent evals are relevant;
  2. why each row was selected, skipped, run, failed, or blocked;
  3. where the resulting reports and regression-promotion evidence live.

It is an orchestration skill, not a robot behavior skill. Keep household-world task strategy in household-world; keep reusable robot capability semantics in MCP tools and capability profiles.

Use Just only as the human-facing facade shown below. Frozen rows execute their package owners directly: eval rows use python -m roboclaws.evals.cli, and product rows use python -m roboclaws.cli.main run surface. Keep the subprocess boundary for row timeout and isolation; do not route an executing row back through Just or add a second command registry. The eval CLI grammar is an optional documented kebab-case tool name followed by key=value arguments.

Commands

Recommend rows without running them:

just agent::eval recommend plan=docs/plans/example.md budget=focused

Execute relevant rows for a plan or diff:

just agent::eval execute plan=docs/plans/example.md budget=focused
just agent::eval execute since=origin/main budget=focused
just agent::eval execute profile=baseline-core budget=focused max_parallel=4

Pass scene=<source>/<index>,... to expand rows whose catalog uses scene_scope=selected; the current scene-portable rows are the MapBuild product rows. Case identity, dependencies, commands, and result schema are resolved before local execution. Scheduling respects shared backend locks and writes one result per frozen benchmark case.

Refresh the current baseline at the appropriate cost tier:

just agent::eval execute profile=baseline-core budget=focused
just agent::eval execute profile=baseline-live-default budget=focused
just agent::eval execute profile=baseline-refresh budget=focused

Run one versioned suite directly as a lower-level row/debugging path:

just agent::eval suite=cleanup_capability budget=smoke

Promote failed, blocked, or inconclusive eval evidence:

just agent::eval promote-regression \
  eval_results=output/evals/<suite>/<stamp>/eval_results.json \
  source_sample_id=<sample-id> \
  regression_sample_id=regression.<name>

Profiles And Budgets

  • recommend: never executes rows; it lists commands and preflight needs.
  • profile=adaptive: default mode; select rows from plan text, diff paths, and explicit axes.
  • profile=baseline-core: select deterministic gates, current eval suites, direct local-simulator product rows, and DINO product rows. It excludes all live-provider rows and is the normal broad local refresh.
  • profile=baseline-live-default: select baseline-core plus the current MiniMax live-agent capability rows and the direct same-run map-build-to-consumer proof. Its open-task and cleanup consumers require the canonical prior; it excludes the alternate-provider fixed-prior matrix.
  • profile=baseline-refresh: select the catalog baseline set directly: deterministic gates, all current eval suites including long-horizon tasks, direct product rows, DINO rows, all default live rows, and the explicit alternate-provider matrix. Prior-consuming rows resolve the canonical prior from assets/eval-priors/runtime_map_prior_catalog.json by default; an explicit runtime_map_prior=<path> is allowed for a controlled refresh. A missing prior remains selected and records an explicit blocker instead of silently removing the row. This is the release/nightly full refresh.
  • profile=baseline-ci: select only deterministic rows inherited from baseline-core; provider-backed rows and internal/external provider egress are rejected at selection time. Ordinary dependency downloads remain valid workflow network access.
  • All named baseline profiles run selected rows or record explicit blocked evidence; their rows are not converted to skipped_by_budget.

Baseline prior policy is explicit per row. Prior consumers use prior_policy=required and must consume a valid canonical or explicitly bound prior. Map-build producers and unrelated contract rows use prior_policy=not_applicable. No-prior controls are separate suites and are never a baseline consumer result.

  • Live trials do not retry by default. Set live_retry_limit=<N> only when the evaluation contract explicitly calls for audited fresh-directory retries.
  • execute budget=smoke: deterministic confidence only; selected expensive or live rows are recorded as skipped by user budget in profile=adaptive.
  • execute budget=focused: default maintainer mode; selected required live rows must run or record explicit blocked evidence.
  • execute budget=full: run required and recommended selected rows unless environment, network, provider, hardware, or guard preflight blocks them.

Never downgrade a selected live-agent eval into deterministic-only success. Missing provider keys, provider 5xx/429, and model-service failures are model_or_provider_unavailable. Missing simulator/runtime, DINO sidecar, Python env, or live-session capacity is environment_blocked.

Outputs

Each run writes:

output/eval-harness/<stamp>/
  eval_harness.json
  eval_harness.md
  rows/<row-id>/
  evals/<suite-id>/<stamp>/

The manifest schema is roboclaws_eval_harness_manifest_v1. Rows use roboclaws_eval_harness_row_v1 and may be deterministic_gate, product_run, eval_suite, live_agent_eval, regression_promotion, or manual_review. Provider rows also carry a fail-closed provider_network_scope and allowed_execution_targets contract. All current provider routes execute locally; external routes remain restricted to local execution by the same contract.

Selection Rules

The selector is deterministic and rule-table based over the row catalog, plan text, git diff paths, and explicit overrides. It does not use an LLM classifier. Row policy lives in this skill and catalog/rows.json; Python scripts only load rows, expand paths/profiles, select, execute, and write manifests/reports.

Important signals:

  • Eval harness, eval CLI, eval reports, or regression promotion select eval unit tests and smoke_regression.
  • Runtime Metric Map, map-build, actionability, or waypoint files select map_build_quality; provider consumer rows use map_consumer_fixed_prior only when an explicit canonical runtime_map_prior is available.
  • Cleanup skill, prompt, MCP policy, checker, or done-readiness files select cleanup contract gates, cleanup_capability, and focused/full live eval rows.
  • Agent SDK, provider profile, or live-runtime files select route/preflight checks and affected live-agent eval rows.
  • Visual grounding, DINO, camera labeler, or RAW-FPV files select perception or camera product rows.
  • Docs-only command taxonomy or skill guidance changes should select docs and route checks, with eval rows only when the docs claim behavior changed.

Do not use or recreate a user-facing agent-validation-matrix route. Historical plans may mention it, but maintained guidance should point at @eval-harness and just agent::eval recommend|execute|suite|promote-regression.

reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 469kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 321kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 321kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 321kUpdated Claude CodeCodexVerifiedInstall