Skip to content

Latest commit

 

History

History
85 lines (70 loc) · 4.8 KB

compatibility.md

File metadata and controls

85 lines (70 loc) · 4.8 KB

Compatibility and evidence

SkillWhy's offline comparison, demo, and sharing use Node.js 22 or later. Native capture is a separate compatibility surface.

Native capture admits Codex CLI 0.166.0 on macOS arm64 and Linux x64. Linux evidence covers Ubuntu 24.04 x64, not every Linux distribution. Other CLI versions or platform/architecture combinations produce an unsupported snapshot instead of silently guessing. This is an early alpha, not a promise that every configuration on an admitted combination works.

Actual alpha.2 metadata conformance runs passed on macOS arm64 with Node.js 22.20.0 and 24.8.0. They covered duplicate names, an explicitly disabled temporary skill, a missing-description error, and capture → diff → share. See the validation record for exact fixtures, timestamps, reproduction commands, and limits. Linux x64 was admitted after reviewing both Node versions in the live Ubuntu 24.04 candidate run. The original Linux build still returned unsupported in that run; the ten cases passed through its explicitly labeled temporary candidate copy. The subsequent production run also passed all ten checks under both Node versions using the admitted, unmodified production build.

The adapter requests a fresh process's skills/list response. It does not attach to an existing Desktop chat or read a model transcript. The native result does not prove that a skill's description reached the model, that its body was read, or that it executed. A CLI version change can change discovery behavior.

The official App Server documentation describes initialization, skills/list, and reload behavior. The Skills documentation explains that same-name skills are not universally merged into a single winner.

Validation categories

  • Unit and process tests use synthetic inputs and fake local processes.
  • The default CI suite does not require Codex, a login, or paid API calls.
  • Live observations, when performed, are recorded in the release validation notes with their exact CLI, Node, OS, scope, and limitations.
  • A native live smoke check does not replace the entire compatibility suite or validate model invocation, Desktop state, or every plugin backend.

A complete capture reports completion of the requested native inventory, not strict YAML/skill-specification validation. The host may accept metadata forms that a separate linter rejects. Only errors returned by the host or detected by the capture adapter can affect that status; SkillWhy does not independently lint every SKILL.md.

Do not broaden the support manifest just because a newer version parses. Add schema and discovery-semantic tests for that version first.

The opt-in scripts/native-conformance.mjs --live uses temporary fixtures and a per-process configuration override, without editing user settings. It inherits the current Codex environment and is not a network-isolation or managed-policy test. A passing default CI job validates offline and fake-process behavior; it is not evidence of live Codex support on each CI operating system.

Native verification and candidate history

Production decisions are centralized in src/compatibility.ts. The two verified Linux candidate summaries are preserved in the validation record.

Before admission, the source-checkout script's --candidate linux-x64 mode checked the original CLI's unsupported result and then ran the fixture suite against a temporary copy with one compatibility-predicate override. Now that the tuple is reviewed, that experimental mode refuses it with CANDIDATE_ALREADY_SUPPORTED on Linux x64. Use the ordinary --live command. There is no candidate flag or environment-variable bypass in the shipped CLI. Historical temporary-candidate reports remain labeled as such.

The manually dispatched Native conformance (opt-in) workflow installs exactly @openai/codex@0.166.0 on Ubuntu 24.04 x64 and runs Node 22.20.0 and 24.8.0. Its default supported mode tests the unmodified production manifest and should fail on unapproved hosts. The historical candidate mode is retained but rejects this already reviewed tuple. Only sanitized successful summaries are uploaded, with the source revision.

Review both native job results, their ten fixture outcomes, and any limitations before adding a production tuple. Keep unsupported versions and architectures rejected. Container emulation, a matching response schema, and the default offline CI matrix are useful preliminary checks but cannot substitute for that evidence. Current observations and pending checks are recorded in validation.md.