Codex desktop pre-injects too much skill/plugin/tool metadata into every new chat #26845

Codex desktop starts fresh chats with very high fixed context usage even when the first user message is trivial, for example just hello.

After local investigation, the main issue appears to be prompt assembly at chat startup:

• Runtime includes a skills_instructions block with ## Skills, ### Available skills, and ### How to use skills.
• Skill metadata for many local, bundled, and plugin-provided skills appears to be injected into every chat.
• Tool schemas for many unused capabilities also appear to be injected by default, including heavy app/tool bundles unrelated to the current task.
• Project AGENTS.md and local custom skills matter, but they are not the primary source of the fixed overhead.

What I verified locally:

• Removing many low-frequency local skills reduced fresh-chat usage only from about 31k to 28k.
• Disabling several plugins also helped only slightly.
• This suggests the dominant cost is still host-level prompt and tool injection, not repo instructions alone.

Evidence sources:

• Archived local sessions contain a developer prompt block with skills_instructions and Available skills.
• Public issues already suggest runtime has a distinct Available skills injection path:
  • Local skills under ~/.codex/skills are not injected into runtime "Available skills" #15136
  • Codex App briefly shows bundled system skills, then hides them after initialization #20629
• Official docs confirm plugins can include skills and workflow guidance, which likely contributes to runtime metadata injection:
  • https://openai.com/academy/codex-plugins-and-skills/
  • https://help.openai.com/nl-nl/articles/20001256-plugins-in-codex

Requested improvements:

1. Do not pre-inject the full skill list into every chat.
   Use lazy discovery via tool_search or a runtime index when needed.

2. Do not expose heavy tool schemas by default in coding-only chats.
   Especially app/tool bundles unrelated to the current task.

3. Add runtime profiles such as:

• coding-light
• coding-full
• research
• design

4. Shorten injected skill metadata aggressively.
   If full lazy loading is not possible yet, reduce descriptions to minimal summaries.

5. Show a context usage breakdown in the app.
   For example: system prompt, developer prompt, tools, skills, and project instructions.

Expected outcome:

A fresh coding chat with a trivial first message should not start with such a large fixed context cost. Most of this appears avoidable through lazy loading and profile-based capability exposure.