native-agent-evals
Skill5 starsChanged 2 months agoMITView on GitHub
Test native Codex agent behavior or audit rollout evidence for lifecycle, tool use, and skill invocation. Use for agent-runtime claims, not ordinary task validation.
In this repository10 other files it gives its agents.
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 12
Before running native agents, load native-subagents-first for the lifecycle and worker contract.
- Distinctive: line 20
Read structural rollout audit before inspecting JSONL, classifying invocation, deduplicating resumed history, or reconstructing existing …
- Distinctive: line 24
Never attach a source rollout or hand-author an extract presented as independently auditable.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Native agent evals
Define one observable claim with pass, fail, and inconclusive outcomes. Choose a small probe that can falsify it. Keep production mutations outside the probe unless explicitly requested.
Run a probe
Before running native agents, load native-subagents-first for the lifecycle and worker contract. Exercise only the behavior under test: tool availability, completion, waiting, fanout, interruption, resume, or skill invocation. For example, fanout needs independent scopes; a lifecycle check may need only one harmless child.
For comparisons, hold model, effort, role, tools, prompt, permissions, and artifact state fixed. Change only the named treatment. Mark mismatched conditions as confounded. Repeat independent behavior comparisons before promoting a claim.
Consume the complete child answer and verify its relevant evidence. Lifecycle status alone does not prove that the requested task succeeded.
Audit or export evidence
Read structural rollout audit before inspecting JSONL, classifying invocation, deduplicating resumed history, or reconstructing existing evidence. For the schema and checks used by scripts/export_structural_rollout.py, read structural evidence export.
Parse records structurally; text search locates candidates but cannot prove invocation. Exclude the current audit family from retrospective usage counts. Explicitly selected descendants of a predeclared probe or study may supply experimental evidence; do not count the evaluator's activity as the behavior under test. Keep loaded, read, announced, and applied skill evidence distinct.
Never attach a source rollout or hand-author an extract presented as independently auditable. Use the typed exporter with pinned source hashes and bounded selectors, then verify the extract against the current source. Do not relax a failed evidence gate to obtain a passing result.
Report
State the claim, worker condition, probe or comparison, repetitions, result, and parent verification. For rollout claims, include exact source identities, selected calls and outputs, complete answer, deduplication and invocation classifications, and export verification. Disclose confounds, mutations, untested branches, and whether the result supports promotion.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

