fix(agent): inject local skill context into codex follow-up prompts

[tatoalo]

Problem

Codex cloud runs don't reliably recognize local skills after the first message. The sandbox agent installs uploaded skill bundles to disk and passes the invoked skill's definition via _meta.localSkillContext, but only the Claude adapter consumed that meta. codex adapter dropped it. Since codex bakes its skill catalog into the thread at start, a skill first invoked in a follow-up message was installed to disk but never became visible to the model: the /my-skill text reached codex as meaningless plain text.

Two adjacent gaps had the same effect even for Claude:

• auto-bundled dependency skills (feat(skills): auto-bundle declared skill dependencies in cloud runs #3057) are installed silently, so nothing told the model they exist or where they live when they arrived mid-session;
• a mid-message mention ("use /my-skill for this", or a skill chip anywhere in the text) installs the bundle but produced no skill context at all — the context builder only fired for message-leading /name invocations.

Changes

• update the Codex adapter to consume localSkillContext and localSkillName, prepending inlined skill definitions to the prompt and stripping the /skill invocation only when appropriate. Prompt ordering is prContext → localSkillContext → remaining user input

• include all installed skills and their paths in the agent-server prompt, allowing the model to resolve references between bundled or previously installed skills

• add a fallback for messages that install skill bundles without directly invoking one, inlining referenced skill definitions and listing the remaining installed skills with their paths

[trunk-io]

😎 Merged manually by @tatoalo - details.

[github-actions]

React Doctor found no issues in the changed files. 🎉

_{Reviewed by React Doctor for commit 166536e.}

[greptile-apps]

_{Reviews (1): Last reviewed commit: "fix(agent): announce mid-message skill i..." | Re-trigger Greptile}

[github-actions]

Well-tested fix that extends skill context injection to Codex follow-up prompts, mirroring the existing Claude adapter. No showstoppers: the interface change is backward-compatible, the new fallback path is guarded, and the prompt-building logic is coherent.

New commits pushed (delta classified non_linear_history) — stamphog approval dismissed; re-review running automatically.

[github-actions]

Well-scoped fix that extends local skill context injection to the Codex adapter, mirroring the existing Claude adapter. The Greptile P1 concern about prefix-matching was addressed with a token-boundary regex and confirmed resolved. Interface change is backward-compatible, the fallback path is correctly guarded, and new behavior is covered by tests.

[gantoine]

/trunk merge