skills

CircleCI managed plugins for AI Agents

Organization

There's a toplevel codex compatible marketplace meant to facilitate local testing in codex.

The actual plugin in plugins/circleci is compatible with existing codex marketplaces.

Coverage

Skills include:

• general CircleCI build debugging
• usage of the chunk cli and cloud based modes
• circleci cli usage
• circleci config management and optimization

Manual eval runs

Run from the repository root:

1) CI-safe routing evals (frontmatter + static routing)

evals/circleci/scripts/run_routing_evals_ci.sh

This runs:

• SKILL frontmatter checks for plugins/circleci/skills/*/SKILL.md
• quick_validate.py checks (when available)
• routing eval cases from evals/circleci/cases/skill-routing-cases.json

Routing case purpose (evals/circleci/cases/skill-routing-cases.json):

• circleci-builds cases: ensure failed-build, flaky, and root-cause prompts route to circleci-builds (explicit + implicit).
• circleci-cli cases: ensure CLI/auth/rerun/command-line prompts route to circleci-cli (explicit + implicit).
• circleci-config cases: ensure .circleci/config.yml, caching, workspace, and runtime optimization prompts route to circleci-config (explicit + implicit).
• chunk cases: ensure Chunk setup and chunk-cli prompts route to chunk (explicit + implicit).
• negative-control cases: ensure non-CircleCI prompts route to null.

2) Local invocation smoke (codex --json, not in CI)

evals/circleci/scripts/run_invocation_smoke_evals_local.sh

Default local mode is non-strict and still writes artifacts even if preflight or a case fails.

Strict mode (non-zero exit on failures, local-only):

STRICT=1 evals/circleci/scripts/run_invocation_smoke_evals_local.sh

Artifacts are written to:

• evals/circleci/artifacts/invocation-smoke/latest/report.json
• per-case JSONL/stderr files under evals/circleci/artifacts/invocation-smoke/latest/<case-id>/

Invocation smoke case purpose (evals/circleci/cases/skill-invocation-smoke-cases.json):

• builds-explicit-smoke: validate explicit $circleci-builds prompt selects circleci-builds.
• chunk-explicit-smoke: validate explicit $chunk prompt selects chunk.
• cli-implicit-smoke: validate CLI intent prompt selects circleci-cli without explicit skill mention.
• negative-control-smoke: validate unrelated prompt reports none.

3) Trace evals (full codex --json capture + grading, optional)

evals/circleci/scripts/run_trace_capture_evals_local.sh

Trace case purpose (evals/circleci/cases/trace-cases.json):

• validate codex capture preflight and JSONL artifact generation for explicit and implicit prompts.
• validate grading behavior for expected skill mention and negative controls.

If codex preflight fails with a network error, verify that codex exec can reach OpenAI endpoints from your environment.