Validate agent changes and performance claims

Test three separate things

Assertion	Check	Does not prove
Source renders	render/apply and inspect target	harness loaded it
Client accepts config	bash install/codex.sh check, bash install/cursor.sh check, and focused tests	endpoint reachability or task quality
Task behavior improved	representative artifact + independent verifier	general gain from one sample
Role routed	observed model/effort and inventory	lower total cost or universal correctness
MCP action available	fresh-session inventory + harmless call	broader authentication or write authority

Validate the managed Codex path

bash install/codex.sh sync-config
bash install/codex.sh check
bash install/cursor.sh check
bash install/skills-sync.sh check

Expected result: generated Codex configuration passes; the final skill command reports drift without changing it. Start a new task before inspecting profiles or roles. Use ./tests/ci.sh full for broad repository changes and ./tests/run.sh for Docker bootstrap behavior; write bulky output to an artifact.

Evaluate instruction and skill changes with paired tasks

Keep model/effort, prompt, revision, inputs, tools, acceptance checks, artifact path, and verifier constant. Record configuration fingerprint, routing, artifacts, result, usage, failures, and parent rework.

Wide diagrams scroll horizontally. Focus the figure to use arrow keys.

Do not attribute a different model, task, or unverified output to instruction quality. Smaller prompts and cheaper children are not performance results by themselves. Use writing-skills for the full comparison route.

Retain evidence that a reviewer can inspect

Keep machine-readable large logs, UI screenshots/traces, and a project work log for sustained tasks. Report commands, exit status, artifact paths, observed gaps, and failure evidence. Verifiers must not weaken tests or edit sources to turn failure into pass. For docs-only work, use the book’s rendering/link path; cross-check scripts before documenting code or installer behavior. See validation before publishing.

Report limits plainly

Say what ran: “the generated role parsed and the fixture passed,” “the interaction completed in this local profile,” or “the source supported this setting on its retrieval date.” Do not expand that claim without corresponding evidence.