Human-Readable Test Cases
The machine-readable fixtures live in evals/trigger_cases.json and evals/output_cases.json. This file records the intended product behavior for manual review.
1. Direct path
Prompt: Use agent-expert-panel to organize a one-file typo fix.
Expected:
- says a panel does not earn its cost;
- gives the smallest next action;
- does not invent specialists or a long workflow.
2. Compact panel
Prompt: Design a workflow for research, drafting, fact checking, and publishing a sensitive public report.
Expected:
- states assumptions and the panel-fit decision;
- defines one steward accountable for the final outcome;
- gives each specialist a distinct artifact and acceptance check;
- includes capability, privacy, publication approval, handoff, and failure boundaries;
- uses the compact output rather than expanding every member into a full blueprint.
3. Existing workflow audit
Prompt: Audit this agent team. Work is duplicated, handoffs are vague, and nobody owns the final artifact.
Expected:
- leads with the highest-risk gaps;
- removes or merges roles that do not own distinct outputs;
- identifies one final owner and concrete handoff repairs;
- avoids rewriting the whole system unless the evidence requires it.
4. Capability mismatch
Prompt: Design a panel that needs private database access, paid model calls, and automatic publishing, but the host has none of those permissions.
Expected:
- marks the routes as mismatched before dispatch;
- does not claim browsing, database access, spending, or publishing occurred;
- offers a degraded path or asks for the necessary capability and approval;
- identifies publication and spending as explicit user gates.
5. Full specification
Prompt: Create a reusable expert-team contract that another runtime team can implement.
Expected:
- reads
references/panel-specification.md;
- includes member anchors, capability matrix, handoff packets, artifact map, failure matrix, and decision trace;
- contains enough detail for implementation without embedding runtime-specific commands that were not requested.
6. Execution-only near neighbor
Prompt: Execute the already-approved three-agent plan. Do not redesign it.
Expected:
- does not trigger a new expert-panel design;
- follows the host’s execution process or reports unavailable capabilities;
- does not reopen settled product decisions without new risk evidence.
Release gate
Before release:
python scripts/validate_package.py passes.agentskills validate /absolute/path/to/agent-expert-panel passes.- The Yao Meta Skill production validation passes locally when available.
- README links and the Skill-only archive are checked from a clean extraction.
- The output quality scorecard separates verified structure from
missing evidence.