feat(core): move ChatGPT context caps into a separate plugin - #58194

Issue for this PR

None

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Problem: The ChatGPT context caps were hardcoded inside the OpenAI plugin. They only covered the legacy Codex sign-ins, so Sign in with ChatGPT (token sharing) got each model's full 1.05M window. They also couldn't be turned off and weren't documented.

Change: Move the caps into their own plugin, opencode.provider.openai-context-caps. It applies to all three ChatGPT sign-in methods. API-key connections are untouched.

 packages/core/src/plugin/
 ├── provider.ts                    # registers the new plugin
 └── provider/
     ├── openai.ts                  # keeps filtering, cost, headers; no longer sets limits
+    └── openai-context-caps.ts     # only lowers limits

when the active openai connection is chatgpt-browser, chatgpt-headless, or chatgpt-token-sharing
  for each openai model
    cap = 500k if the model is gpt-6.1 or newer (or not a gpt-N name), otherwise 272k
    if the model's window is already <= cap: leave it alone
    else: input = cap, context = cap + output
on credential switch: reload so the caps follow the active connection

The 272k and 500k values are the context_window defaults returned by the live ChatGPT model list. That list also reports a 872k max_context_window. When probed directly, the backend accepted 921,522 input tokens and rejected 921,922, so it enforces only the model's real ~922k window. The caps are a usage choice, not a backend requirement.

Users can opt out with "plugins": ["-opencode.provider.openai-context-caps"], or override one model's limit in config, which applies after the plugin.

Docs: a new ChatGPT section in providers.mdx covers the sign-in methods, why the limits exist, the limit table, and both ways to change them.

Out of scope: model sync for token sharing is still commented out in chatgpt.ts. When it returns, it can record the server's default and this plugin can apply it instead of its fixed values.

How did you verify your code works?

• bun run script/test.ts test/plugin/provider-openai-context-caps.test.ts test/plugin/provider-openai.test.ts test/plugin/provider-chatgpt.test.ts in packages/core: 31 pass. The new test covers all three sign-in methods, the 6.1+ rule, an unchanged smaller model, and an API key that keeps the full window.
• --rerun-each 10 on the new test file: 40 pass.
• bun run check from the repo root: 34/34 tasks successful.

Screenshots / recordings

N/A

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

Added by the opencode bot (AI-generated).

Related: #47646, #46527

Public API: adds the internal plugin ID opencode.provider.openai-context-caps, which config can disable (backwards compatible)