fix(manifest): drop Agent Plugins $schema so Codex stops truncating skills at 8KB

[tmchow]

Summary

Codex users get whole skills again. Since 3.22.0, Codex >= 0.147 was injecting only the first 8,000 bytes of each SKILL.md — 26 of 33 bundled skills were silently cut mid-body (ce-setup lost the last ~30%, ce-plan ~93%), dropping approval gates, repair steps, and handoffs with only a warning. This PR removes the trigger and adds a guard so it cannot come back unnoticed.

Fixes #1412

Why it happened

• Enforce Agent Plugin runtime boundaries openai/codex#37027 (0.147) classifies a plugin as an Agent Plugin when root plugin.json has a $schema starting with https://agent-plugins.org/schemas/. That manifest wins over .codex-plugin/plugin.json, which becomes an overlay.
• Agent Plugin skills are bounded by MAX_SKILL_PROMPT_BYTES = 8_000 at injection; legacy manifests are exempt.
• 3.22.0 (feat(manifest): add Agent Plugins v1.0.0 manifest support #1345) added exactly that $schema. Nothing else in the plugin changed for Codex.

What changed

• Root plugin.json keeps the Agent Plugins field shape but omits $schema, so Codex resolves the legacy .codex-plugin/plugin.json again and injects full skill bodies.
• tests/codex-skill-prompt-budget.test.ts: measures every SKILL.md as a CRLF checkout would (matches the 11,392-byte ce-setup figure in the issue), fails if a skill outside the known-over-budget set exceeds 8,000 bytes, fails if a listed skill has shrunk under the bound but is still listed (ratchet), and fails if $schema reappears while any skill is over budget. The set does not pin sizes, so ordinary skill edits do not churn it.
• tests/release-metadata.test.ts: $schema is now optional; when present it must still be the v1.0.0 URI.
• docs/specs/agent-plugins.md: records the posture and the conditions for restoring $schema.

Validation

bun run test (3174 pass) and bun run release:validate clean. Not exercised against a live Codex 0.147 install; the mechanism is confirmed from Codex source (utils/plugins/src/plugin_namespace.rs::find_plugin_manifest_path, core-skills/src/injection.rs::bounded_skill_prompt_contents).

Follow-up and caveat

• Durable fix is restructuring the 26 over-budget skills into <=8 KB entrypoints with progressive-disclosure references; the allowlist shrinks as that lands, and $schema returns when it is empty. That work needs its own plan.
• Bound explicit skill prompts and avoid duplicate host injection work openai/codex#37463 (open) proposes applying the 8 KB bound to all skill sources. If that merges, this workaround stops helping and the restructure becomes required.

Security Disclosure

No security-relevant changes.

Agent Disclosure

• Model: Claude Code · claude-fable-5

---

[cursor]

PR Summary

^{Cursor Bugbot is generating a summary for commit 290de6c. Configure here.}