Evaluation-backed software engineering skills for Agent Skills-compatible agents.

• deliver-software carries substantial software work from repository discovery through implementation, cleanup, validation, and executable verification.
• deno-software adds current Deno-specific repository, dependency, compatibility, security, quality, publication, and artifact guidance.
• explore-ecosystems gives every in-scope dependency a cheap ecosystem identity check, then verifies monorepos, siblings, adapters, plugins, specifications, alternatives, and exclusions deeply for material decisions.
• build-libraries owns reusable programming models, public APIs, selective adoption, data-flow shapes, explicit resource ownership, data-oriented hot paths, library performance, packaging contracts, and restart/resume contracts.
• build-clis owns command language, configuration, output, interaction, cancellation, installed artifacts, and CLI verification.
• build-web classifies hybrid web surfaces and owns renderer, component, motion, security, and browser contracts used across those surfaces.
• build-sites owns Astro content, marketing, documentation, CMS, feeds, and static/server site delivery.
• build-web-apps owns stateful Solid and TanStack applications, URL/query/local state, sessions, authorization, and product interaction.
• build-apis owns service-module, endpoint, validation, response, middleware, authentication, authorization, and request contracts.
• build-workflows owns durable coordination, workers, queues, timers, signals, checkpoints, recovery, and resumable pipelines.
• build-data owns operational, analytical, search, graph, artifact, query, migration, and projection contracts.
• build-devtools owns toolchains, generators, packaging, releases, performance experiments, and repository hygiene.
• use-okikio provides source-grounded playbooks for Okikio libraries and recurring project patterns without inventing private exports.

The skills are independently installable and deliberately composable. deliver-software owns the general delivery lifecycle and the current cross-project engineering standards for naming, schema/type contracts, TSDoc/comments, documentation, formatting, resource ownership, and completion. Domain skills own their contracts. explore-ecosystems owns dependency topology and evidence. A composed task performs one repository discovery pass, one plan, and one final verdict.

npx skills add okikio/skills

For a version-pinned GitHub CLI installation:

gh skill install okikio/skills [email protected]
gh skill install okikio/skills [email protected]

A skill is not complete because it has a SKILL.md, references, or a pair of frozen routing cases. Every shipped operational reference must be:

1. explicitly routed from the owning SKILL.md;
2. represented in evals/capabilities.json;
3. grounded in at least one registered evidence source;
4. covered by train and valid-seen cases;
5. covered by at least one held-out split (valid-unseen, transfer, adversarial, or test-frozen);
6. connected to decision questions, failure signatures, exclusions, and verification instructions.

deno task validate enforces this reference-to-capability coverage. This prevents a large reference library from looking complete while the agent behavior remains unevaluated.

The repository evaluates five distinct capabilities:

1. discovery: whether the correct skill is selected;
2. routing: whether the skill loads only applicable references;
3. application: whether required behavior appears in the trajectory;
4. outcome: whether the resulting repository or answer passes its verifier;
5. efficiency: token, reference, tool-call, latency, and duplication cost.

Results can compare real no-skill, individual-skill, and composed-skill variants. A no-skill rollout omits the target skill from the installation rather than only hiding its telemetry. An optimized candidate is never promoted solely because a qualitative judge prefers its prose. Deterministic acceptance criteria remain authoritative, and provider/judge/integrity-invalid runs are not treated as ordinary task failures.

Install Deno 2, then run:

deno task validate
deno task check
deno task test
deno task skillopt:matrix
deno task sources:verify --attachments /path/to/uploaded-evidence

The optional source check recomputes the registered SHA-256 values for the uploaded guidebooks and archives and verifies every registered ZIP claim path. evals/capabilities.json then connects those sources to routed references and behavioral cases, so a tool name by itself is not treated as verified coverage.

skillopt:matrix exports and verifies every capability reference, root router, and frozen composition topology. It proves selection and immutability contracts; it does not substitute for model rollouts or behavioral judging.

Cross-model execution uses deno task skillopt:rollout. The repository runner owns fixture isolation, deterministic assertions, qualitative-judge ordering, redaction, change accounting, and workspace integrity. trajectory-rubric and mixed cases require an explicit --judge adapter; target models never receive rubric criteria, and judges never receive fixture or skill-tree paths. Use --without-target for an actual no-skill baseline. After the exact case/run matrix completes, deno task skillopt:report creates a sample-counted aggregate report from verified exported cases and normalized result.json files. deno task skillopt:gate compares paired reports and rejects invalid benchmark runs before comparing model quality. Provider-specific integrations implement the versioned JSON protocol in skillopt/adapter-contract.md; enable only adapters that can truthfully report the telemetry required by their declared request kinds. Credentials remain in explicitly allowed environment variables and never enter fixtures or skill files.

SkillOpt is kept as a separately reproducible optimization layer. See skillopt/README.md. Generated candidates are review artifacts, not source files, until they pass held-out, cross-model, composition, and safety gates.