Compatibility

Compatibility and evidence

Portable core

The packages use English Markdown, standard Agent Skills frontmatter, relative resources, plain file handoffs, and semantic tool capabilities. No proprietary SDK, provider configuration, agent home path, or subagent interface is required. Optional descriptive metadata and reasoning hints can be ignored without changing the task contract. Optional UI metadata and icons are bundled separately from the procedure.

Surface	Intended support	Evidence boundary
Agent Skills compatible hosts	Any host able to read the package and provide the needed capabilities	Format conformance does not prove behavior on every host
Sequential execution	One agent performs each companion skill and returns artifacts	Reference workflow; no delegation API needed
Delegated execution	Optional isolated workers with the same artifact contract	Host adapter is outside this collection
Text-only/manual execution	Local procedures and manual checks	Official conformance still requires execution; unavailable required checks block readiness
Local helpers	Node.js 24+ built-ins on an owned, stable workspace	Automated regressions cover exercised environments; no concurrent-adversary confinement claim
Installed package relocation	Bundled resources independent of this checkout and the calling directory; caller-selected outputs	Disposable project/global layouts and directory aliases are exercised by distribution tests; no real consumer installation or universal host-runtime claim
Hosts without helper support	Portable prose and manual checks	Unsupported filesystem primitives fail explicitly; official validation can return through CI

Host integration research

Primary documentation was inspected on 2026-09-12. This table records documented loading conventions, not execution tests in every agent. Each runtime still controls workspace trust, tools, model choice, and installation. No global agent directory is modified by this repository.

Host	Documented core and useful integration	Repository support
Codex	Reads repository .agents/skills; optional agents/openai.yaml supports interface metadata and icon paths. Official guide	Native canonical tree; per-skill interface metadata, source-attributed SVG small icon, and PNG large icon
Codex session index	Optional trusted project hook renders bounded project/global package metadata; the manual command renders the same output	Project-local .codex/hooks.json; no automatic trust, package loading, installation, or route invocation
Installed Codex/Claude plugins	Root manifests map native hooks and MCP to direct Node 24 TypeScript transports	Native discovery, SessionStart and local source rollback passed in the isolated pilot on Codex 0.159.0/Claude 2.1.277. Claude MCP initialized; the separate Codex MCP pilot proved explicit tool calls with caller-selected external storage. Current read adapters have synthetic test coverage at the documented limits; those historical pilots did not exercise the updated mappings.
Claude Code	Reads SKILL.md and local resources under its skill directories; provider-specific frontmatter includes model and invocation controls. Official guide	.claude/skills alias and CLAUDE.md alias; no required Claude-specific fields or hooks
GitHub Copilot	Supports skill directories and resources; VS Code adds invocation/frontmatter and extension-level capabilities. GitHub overview, VS Code guide	.github/skills alias; no invented per-skill manifest or extension package
Copilot CLI plugin MCP	Root legacy manifest maps the shared Node 24 transport; evidence defaults to ~/.agents/skills-usage.db, with explicit I9_AGENT_STATE_ROOT and I9_SKILLS_USAGE_DB overrides. Native plugin-data variables do not select storage.	The native MCP pilot on CLI 1.0.89 exercised ephemeral plugin loading, 13-tool discovery and direct catalog/evidence calls using the earlier storage mapping. No model turn or persistent marketplace installation is claimed. Current storage and read-hook mappings have synthetic test coverage, not new native pilot evidence.
Gemini CLI	GEMINI.md is the default context filename; context.fileName can select other names. Official context guide	Preserve the legacy GEMINI.md alias to canonical AGENTS.md without changing host settings; session hooks are described separately
Antigravity	Current workspace location is .agents/skills; older .agent/skills remains a compatibility path. Official guide	Native canonical tree; no duplicate legacy alias
OpenCode	Documents standard skill metadata and .agents/skills discovery; permission settings belong to host configuration. Official guide	Native canonical tree; no required opencode.json or permission change
Hermes	Reads SKILL.md and resources; supports extra platform and metadata.hermes fields. Pinned authoring guide	Portable packages available to its authorized loader; no unverified project-autoload claim or global symlink
pi	Reads project .agents/skills after project trust; package-level pi.skills is optional. Pinned skills guide	Native canonical tree; no npm package/installer required

No documented per-skill claude.yaml, copilot.yaml, hermes.yaml, or similar sidecar was found. Nested metadata.hermes is outside the standard's flat string-map profile, and host hooks or permission settings can execute behavior; neither is silently added to the universal core. Useful future adapters must cite the runtime format they actually target. The runtime-field comparison distinguishes Claude skill fields, Codex subagent configuration, Copilot CLI and VS Code, OpenCode V1/V2, and the remaining reviewed hosts.

The four repository symlinks are exact relative aliases: CLAUDE.md → AGENTS.md, GEMINI.md → AGENTS.md, .claude/skills → ../.agents/skills, and .github/skills → ../.agents/skills. Validation does not traverse or count them as a second package tree. Use the canonical path with the strict package helper. Git preserves these aliases; environments that disable symlink checkout must use the canonical tree through their documented loader and must not treat a link's text as a duplicate skill.

Instruction contracts were rechecked on 2026-09-19. Codex uses AGENTS.md. Claude Code explicitly supports a CLAUDE.md symlink to that file. Its newer native AGENTS.md loading is conditional on version and settings; keeping the alias preserves the existing repository entrypoint. Gemini's filename adapter remains useful with its default settings. These instruction-loading claims are structural and documented contracts. The later native plugin pilot does not test every project instruction alias or Windows symlink checkout; no real home settings changed.

The optional Codex session index remains an adapter, rather than a portable package requirement. Its command discovers project and global .agents/skills entrypoints with canonical deduplication and writes bounded context. If a host does not support hooks, or the project is not trusted for them, run node bin/index.mjs context available-skills and use the same compact map manually.

Skills CLI 1.5.26 was exercised locally with skills add . --list against the canonical collection, with telemetry disabled. It found each canonical name once and exited without installing packages or creating an installation lock. This checks discovery, while detached-package tests cover bundled resources and helper execution separately; neither establishes an actual consumer installation. The pinned discovery source explicitly lists .agents/skills; its recursive discovery is bounded rather than an unrestricted scan of every file.

Evidence levels

• Structural checks validate package metadata, license presence, references, source lock consistency, catalog agreement, and basic public hygiene.
• Official skills-ref validate must pass on each new or changed package; the pinned setup and PR workflow make this a separate required check.
• Run checks validate stage order, source constraints, bounded regular-file reads, and artifact hashes. They do not grade the truth of reports.
• Synthetic regression tests exercise concrete tool and failure behavior in disposable directories.
• Behavioral evaluation inspects decisions and artifacts for frozen task cases. The report must state executor, candidate, baseline, independence, limitations, and sample count.
• Named-provider and production compatibility require actual recorded execution on those systems. The initial catalog makes neither a universal model-quality claim nor a production approval claim.

Optional reasoning-effort recommendations are explained in each skill body and do not select models or apply host settings. They are not benchmark claims. An execution environment supporting the file format may still differ in model behavior, available tools, context budget, or permissions. Record those differences rather than hiding them behind a compatibility label. The OpenAI sidecar reference separates verified host fields from collection authoring conventions.

The initial pilot report records the historical six-stage synthetic creation exercise, actual output observations, official validation, and the limits of its same-session manual baseline. It does not establish readiness for the current version 2 authoring contract, which adds mandatory domain research.

The current research pilot exercises version 2 on two frozen synthetic tasks with a separate task-only baseline and independent grading. Both arms met the visible task cases and passed official package validation. Overall pipeline readiness failed: the executor recorded research depth after retrieval despite the prior-planning requirement. The report retains that deviation, exact identities and the follow-up; no universal improvement or successful multi-contributor synthesis is claimed.