Context window limit alternates between 128k and 272k for GPT-6 Luna #128

What happened?

Bug description

The context-window limit in Pi Web intermittently switches between 128k and 272k within the same GPT-6 Luna session.

For example:

• Incorrect: 145.9k / 128k (100%, red)
• Expected: 145.9k / 272k (~54%)

Both Pi Code and Pi Web's /api/models endpoint correctly report a context window of 272,000 tokens.

Environment

• Pi Web: @ygncode/pi-web@0.0.1-beta.38
• Model: openai/gpt-6-luna
• OS: WSL2 (Ubuntu on Windows)

Verification

Pi Code reports:

provider  model       context  max-out
openai    gpt-6-luna  272K     128K

Pi Web's /api/models endpoint returns:

{
  "provider": "openai",
  "id": "gpt-6-luna",
  "name": "GPT-6 Luna",
  "contextWindow": 272000
}

Suspected root cause

There appears to be an inconsistency between model display names and model IDs.

In chat-selectors.js, modelDisplayLabel() prefers model.name over model.id.

However, in context-usage.js, buildContextWindows() registers lookup keys using model.id, not model.name.

This produces different results:

const models = [{
  provider: "openai",
  id: "gpt-6-luna",
  name: "GPT-6 Luna",
  contextWindow: 272000
}];

const windows = buildContextWindows(models);

getModelContextLimit("GPT-6 Luna", "openai", windows);
// 128000 (incorrect)

getModelContextLimit("gpt-6-luna", "openai", windows);
// 272000 (correct)

The fallback logic does not explicitly recognize GPT-6 models and defaults to 128,000.

The model selector and worker-status polling can use different model labels, which may explain the intermittent switching.

Relevant source:

• [context-usage.js](https://github.com/ygncode/pi-web/blob/v0.0.1-beta.38/web/src/components/session/chat/context-usage.js)
• [chat-selectors.js](https://github.com/ygncode/pi-web/blob/v0.0.1-beta.38/web/src/session/chat/chat-selectors.js)
• [worker-status.js](https://github.com/ygncode/pi-web/blob/v0.0.1-beta.38/web/src/components/session/chat/worker-status.js)

Relation to existing issue

This may be related to #110, fixed by #122.

However, #110 primarily addressed context-usage flickering caused by incomplete usage snapshots during streaming.

The issue reported here specifically concerns the context-window limit itself alternating between 128k and 272k.

Suggested fix

Use the canonical model ID rather than the display name when looking up context-window limits.

Alternatively, register both model IDs and display names as lookup keys.

A regression test for models whose display names differ from their IDs would also be helpful.

Thanks for maintaining Pi Web!

Steps to reproduce

Steps to reproduce

1. Open a Pi Web session using GPT-6 Luna.
2. Check the context-usage indicator.
3. Observe that the displayed context-window limit sometimes switches between 128k and 272k.

The exact trigger for the switching has not yet been isolated.

pi-web version

v0.0.1-beta.38

OS & browser

WSL2 (Ubuntu on Windows)

Relevant logs

Privacy

• I have not pasted private session content; logs and screenshots are scrubbed of sensitive data.