Adapter source research
This page is a research record: it captures what each vendor's documentation said on the consultation date and the format decisions derived from it.
The vendor sources below were consulted on 2026-07-15. The repository comparison uses baseline commit cbe8c9dd7c48f945711f2cd52b9ef01748fccb1e. Only vendor-maintained documentation was used for vendor contracts.
Evidence boundaries
Official format confirmation means that a vendor source documents the path, syntax, metadata, and invocation behavior recorded here. Local parsing means that the generated artifact was accepted by a real parser for its format. Installation means that the generated bundle passed the repository filesystem lifecycle in a temporary project. External discovery means that the vendor application recognized the installed artifact. Execution means that the external agent actually ran the workflow. Outcome review means that a retained run was reviewed against objective recipe criteria.
These stages are independent. A vendor URL does not prove local serialization, parsing, installation, discovery, execution, or outcome quality. This documentation research alone establishes no external discovery, agent execution, outcome review, or tested agent version. Historical records once established explicit discovery, parsing, and execution for review-pull-request on Claude Code 2.1.209 and Codex CLI 0.144.4 only. Those records were retired after the intentional history reset and remain available only as archived context. They are not active evidence and do not promote the current compatibility status.
Official format matrix
| Adapter | Official source | Confirmed project path and format | Metadata and assets | Invocation control | Decision and limitation |
|---|---|---|---|---|---|
| Generic Markdown | Project-defined, with no external vendor contract | .agentic-workflows/workflows/<recipe-id>/workflow.md in a Markdown bundle | Bundle-owned Markdown files follow the project contract. | No agent invocation behavior exists. | Treat it as documentation and report invocation as not applicable. |
| Claude Code | Skills | .claude/skills/<recipe-id>/SKILL.md | Claude accepts skill-directory support files; name is optional and description is recommended. | Skills are model-invocable by default; disable-model-invocation: true preserves explicit /skill-name invocation only. | Emit the field for sensitive recipes; do not add allowed-tools as a safety control because it grants pre-approval rather than restricting tools. |
| OpenAI Codex | Build skills | .agents/skills/<recipe-id>/SKILL.md | name and description are required; scripts, references, assets, and agents/openai.yaml are supported. | Implicit invocation defaults to true; policy.allow_implicit_invocation: false in agents/openai.yaml retains explicit $skill invocation. | Include the policy file in sensitive bundles; plugins are a distribution alternative, not execution evidence. |
| Cursor | Agent Skills, Rules, and Cursor 2.4 | Prefer .cursor/skills/<recipe-id>/SKILL.md; .agents/skills/ is also recognized. | name and description are required; scripts, references, and assets are supported. | Skills are automatic by default; disable-model-invocation: true requires explicit /skill-name invocation. | Migrate from the baseline agent-requested MDC rule; an MDC rule with a description and alwaysApply: false is not manual-only. |
| Gemini CLI | Custom commands | <project-root>/.gemini/commands/<recipe-id>.toml using TOML v1 | prompt is required and description is optional; @{path} can inject workspace content. | Custom commands are explicitly invoked as /command; no implicit-invocation field is documented. | Do not invent a policy field or emit !{...} shell interpolation; command-local asset resolution is not documented. |
| OpenCode | Commands | .opencode/commands/<recipe-id>.md | YAML frontmatter supports description, agent, subtask, and model; $ARGUMENTS receives arguments and @path can include project files. | Custom commands are explicitly invoked as /command; no implicit-invocation field is documented. | Emit only needed fields, do not emit shell-output interpolation, and do not assume a command-local asset root. |
AWF-006 manual-invocation decision
A recipe is sensitive when it has high or critical risk, any mutable effect, an approval gate, or an operation involving publication, migration, removal, or traffic change.
| Adapter | Required sensitive-recipe behavior | Failure disposition |
|---|---|---|
| Generic Markdown | No invocation policy is available. | Present the export as documentation only. |
| Claude Code | Add disable-model-invocation: true to SKILL.md. | The export is policy-incomplete without the field. |
| OpenAI Codex | Add agents/openai.yaml with policy.allow_implicit_invocation: false. | The bundle is policy-incomplete without the metadata file. |
| Cursor | Export an Agent Skill with disable-model-invocation: true. | An automatic skill or agent-requested MDC rule is policy-incomplete. |
| Gemini CLI | Retain the documented explicit slash-command surface. | Do not invent an unsupported field. |
| OpenCode | Retain the documented explicit slash-command surface. | Do not invent an unsupported field. |
Installation output must state whether invocation is manual-only, model-invocable, or not applicable. The status must be derived from generated metadata rather than inferred from Markdown readability.
AWF-018 contract implications
| Adapter | Required local format contract | External contract when the consumer is available |
|---|---|---|
| Generic Markdown | Verify the entrypoint, complete assets, header, and all relative links. | Not applicable until a consumer is selected. |
| Claude Code | Parse YAML frontmatter, validate documented fields, require the sensitive policy when applicable, and resolve bundle assets. | Confirm discovery and explicit-only behavior, then retain the exact CLI version. |
| OpenAI Codex | Parse both SKILL.md and agents/openai.yaml, validate required fields and the boolean policy, and resolve assets. | Confirm discovery and explicit-only behavior in a fresh session, then retain the exact version. |
| Cursor | Parse Agent Skill frontmatter, require folder and name equality, validate the sensitive policy, and resolve assets. | Confirm discovery and verify that a sensitive skill is not selected implicitly. |
| Gemini CLI | Use a real TOML parser, require prompt, validate description and | Confirm discovery with /commands list, invoke manually, and retain the exact version. |
| OpenCode | Parse YAML frontmatter, validate the Markdown template and $ARGUMENTS, resolve assets, and reject unintended shell interpolation. | Confirm command discovery and manual invocation, then retain the exact version. |
Local YAML and TOML parsing in repository tests validates the serializer contract. It is not a substitute for parsing, discovery, or execution by the vendor application.
Registry facts and evidence boundary
The central registry derives global support from independently named facts. An observed fact must cite at least one repository evidence path, while an untested fact cannot cite passing evidence.
| Adapter | Format | Serializer | Local generation contract | Temporary installation contract | Vendor consumer parse | External discovery | External execution | Outcome review |
|---|---|---|---|---|---|---|---|---|
| Generic Markdown | project-defined and confirmed | implemented | passing | passing | not-applicable | not-applicable | not-applicable | not-applicable |
| Claude Code | confirmed | implemented | passing | passing | untested | untested | untested | untested |
| OpenAI Codex | confirmed | implemented | passing | passing | untested | untested | untested | untested |
| Cursor | confirmed | implemented | passing | passing | untested | untested | untested | untested |
| Gemini CLI | confirmed | implemented | passing | passing | untested | untested | untested | untested |
| OpenCode | confirmed | implemented | passing | passing | untested | untested | untested | untested |
Serializer implementation is retained in packages/core/src/adapters.ts. The local generation contract is retained in packages/core/src/core.test.ts. The temporary-directory installation contract is retained in packages/cli/src/install.integration.test.ts. Official format decisions are retained in this document.
The global status supported is available only when format, serializer, generation contract, and installation contract satisfy the registry rule. It does not imply vendor consumer parsing, external discovery, external execution, a tested agent version, or an approved outcome. The generated compatibility matrix records active recipe counts and therefore reports no current Claude Code or Codex execution coverage. Archived pre-reset records remain visible through the archive link above but cannot promote any current status. No human outcome-review timestamp is recorded because no active outcome review exists.