Home

I-9 Skills documentation

I-9 Skills is an experimental collection and toolset for creating, validating, preparing distribution, and evolving focused agent skills.

Start here

• Architecture: collection boundaries, specialist handoffs, and lifecycle.
• Collection review: dated responsibility, overlap and icon decisions for the meta-skill collection.
• Authoring standards: package design and contribution requirements.
• Compatibility: portable contract and host-specific support.
• Validation: local checks and the official Agent Skills validator.
• Collection maintenance: audit, reviewable plans and explicit catalog synchronization with verified recovery.
• Skill change reports: pinned observations, evidence-based bump reports and a runnable installed onboarding guide.
• Release management: version preparation and Changesets.
• Lifecycle policy: portable maturity evidence and promotion decisions.
• Entry-path map: native skills, the explicit CLI/MCP, and repository-root plugin hooks.
• Visual guides: the published landing page for interactive maps.

Deeper references

• CLI distribution: immutable Git-source npx calls, preparation, packed CLI verification and remaining registry publication gates.

• Plugin preparation: inspect the repository-root host manifests and marketplace entry, and learn the post-merge installation path.

• Native plugin pilot: measured Codex/Claude installation, session hooks, Claude MCP, local update/rollback and cleanup, with reproducible isolation boundaries.

• Codex MCP pilot: native tool discovery, catalog access and explicit external evidence storage without a model turn.

• Copilot MCP pilot: native legacy-plugin discovery and direct MCP calls with explicit external data and an ephemeral mount.

• Public readiness: dated source/history and remote-surface audit with explicit evidence gaps.

• Host hooks: host-specific session configuration and context output, with explicit untested runtime boundaries.

• Positioning: public communication and demonstrations.

• Skill MCP: bundled skill search, Markdown retrieval and overview plus explicit observed-read evidence and rankings.

• MCP Inspector: a pinned checkout walkthrough for tool discovery, catalog search and selected Markdown reads, with expected output and cleanup.

• Local skill telemetry: typed session, attempt and read observations, period trends and bounded diagnostic logs.

• Lifecycle evidence: explicit route, activation and outcome events, catalog history, cohort denominators and coverage-aware inactivity through CLI/MCP.

• Skill memory inspection: compact evidence summaries and optional cutoff counts from the same read-only store, with explicit identity and retention limits.

• Planning protocol: durable planning for material work.

• Pilot runbook: a bounded, evidence-first procedure for the first external collection pilot.

• Pilot evaluation: historical six-stage pilot evidence and limitations.

• Research pipeline pilot: current seven-stage task comparison, exact evidence and the failed process-fidelity criterion.

• Research chronology pilot: a fresh, independently inspected planning-before-retrieval case with a separate incomplete-research result.

• Federated collections pilot: two-owner synthetic discovery, qualified identity and bounded artifact handoff, including explicit refusal cases.

• Upstream research: reviewed sources, adoption decisions, and provenance.

• Security policy: disclosure boundaries and security reporting.