feat(eval): replace internal harness with black-box eval

.agents/skills/agent-eval/SKILL.md

@@ -0,0 +1,14 @@
+---
+name: agent-eval
+description: Implement or verify the independent black-box Eval package, versioned datasets/suites, jobs, baselines, Autoevals grading, telemetry and release gates.
+---
+
+# Black-box Agent Eval
+
+1. Work only through `EvalTarget` in `eval/src/contracts.ts`. Never import product runtime, prompts, context, kernel, model client, control plane or databases. `eval/targets/agent-os.ts` is a type-only integration placeholder until separately authorized.
+2. Run `npm ci --prefix eval` then `npm run eval:check`. Tests use local HTTP servers; they are infrastructure verification, never Candidate quality baselines. No browser automation.
+3. Candidate and Judge use explicit independent `EVAL_CANDIDATE_*` and `EVAL_JUDGE_*` configuration. Use Autoevals with an instance client, never globals, provider fallback or product credentials.
+4. Live evaluation is explicit: follow `docs/agent-eval.md`. Never claim API validation without running it. Missing credentials/prices, unknown usage, incomplete jobs, evaluator errors and incompatible baselines must block release.
+5. Version dataset/suite content immutably. Promote eligible completed runs only with a review reason; never replace an existing baseline. Review portable baseline files before using them in CI.
+6. Persist bounded synthetic identifiers, scores, usage and failure codes in telemetry/reports; no raw prompts, answers, secrets or unrestricted error messages. Private local SQLite includes dataset inputs for reruns and must not be uploaded as a public artifact.
+7. Report only owning tests, gate outcomes and material limitations. Agent runtime coverage belongs to its own integration tests, not an internal Eval adapter.

.agents/skills/agent-eval/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Agent Eval"
+  short_description: "Maintain deterministic Agent Eval gates and reports"
+  default_prompt: "Use $agent-eval to update and verify deterministic Agent Eval coverage."

.agents/skills/agent-runtime/SKILL.md

@@ -0,0 +1,23 @@
+---
+name: agent-runtime
+description: Implement, debug, refactor, or review LingxiLoop Agent OS runtime behavior, work leasing, retries, kernels, Host Bridge actions, message delivery, or LLM ledger integration. Use when changes touch server/src/agent-os, server/agent-os, Agent work items, or Agent runtime contracts.
+---
+
+# Agent runtime
+
+Trace the work item from claim through model turn, `ipython`, Host Bridge effects, ledger writes, and final message before editing.
+
+## Preserve contracts
+
+- `ipython` is the only model-visible tool. Do not expose product or network tools directly.
+- Product effects cross the authenticated Host Bridge and repeat server-side authorization using the acting user, company, project, and run.
+- Claims use leases; retries are bounded; attempts are idempotent; stale owners cannot commit results.
+- WuKongIM is the durable message source. PostgreSQL stores Agent work, audit, and LLM ledgers rather than shadow chat history.
+- Every LLM path uses the shared client and records success/failure, model, tokens, latency, tenant scope, and correlation identifiers without prompt secrets.
+- Kernel and Agent home isolation must survive cancellation, timeout, restart, and concurrent runs.
+
+## Verify
+
+Add deterministic integration coverage for changed runtime behavior and update Agent Eval when user-visible decisions or traces change. Run `npm run server:typecheck`, the owning unit and integration files, without coupling tests to the independent Eval package. Never run the complete test or Eval suites for a localized change.
+
+Report lease/retry impact, Host Bridge authorization impact, ledger impact, and the exact failure behavior.

.agents/skills/agent-runtime/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Agent Runtime"
+  short_description: "Change the Agent OS runtime without breaking contracts"
+  default_prompt: "Use $agent-runtime to implement and verify an Agent OS runtime change."

.agents/skills/better-layout/SKILL.md

@@ -69,7 +69,7 @@ Never park a critical action where resizing or scrolling clips it. Keep it in th
 
 **Severity.** `HIGH` blocks content or an action at a supported viewport. `MEDIUM` harms hierarchy, reading order, or adaptability. `LOW` is isolated alignment or spacing polish.
 
-**Verification.** Without a browser: logical properties in place of physical ones, container and media queries against the supported viewport list and DOM order against the intended reading order. With one: every supported width, 200% zoom and the RTL mirror. Report every check you could not run as `Not verified`.
+**Verification.** Review logical properties in place of physical ones, container and media queries against the supported viewport list, and DOM order against the intended reading order.
 
 **Format.** Group findings under the principle each violates, ordered by severity, one row per root cause listing every location it appears in: