GH Copilot Context Size update

[slhad]

GH Copilot has recently updated the context size of all or almost of the models.

After seeing PR #1104 being closed then found out PR #810 , It seems they may have finally sorted out the official context size (no output/input column).

I don't know how/where @rekram1-node is working on it (or anybody else), I just hope that the current limit setup is indicative and don't force compaction when there is a context size window a lot larger available

Here is my contribution from my Business account if it can help :

[d4rky-pl]

I can confirm that at least for Opus 4.6 the new context sizes are not limited to just vsc / copilot cli but also work fine in opencode (I tweaked opencode.json to increase them and been running since yesterday with no issues)

[slhad]

so we do need to tweak them manually for now ? I did not find anything in https://opencode.ai/docs/models/ for that, how did you do it @d4rky-pl ?

Other issues like anomalyco/opencode#16308 seem to indicate there are bugs preventing us to do that

[d4rky-pl]

@slhad the documentation for opencode.json is unfortunately very poorly maintained. The entire schema for the file is available under https://opencode.ai/config.json and includes all the undocumented features as long as you either are OK with analyzing the file yourself or you ask AI to do it for you. Or, in this case, you can try using the config I shared in opencode issue. It works for me but I don't "vibe code" and have auto-compaction enabled disabled (edit: typo, long day) so YMMV.

Please be aware that I give zero guarantees, I do not provide support, I will not keep this trick up to date if anything changes and I'm not a maintainer of this project.

The issue you linked seems to have nothing to do with Copilot (which never provided 1M context window) but it's hard to say because the reporter seems to be very confused about things.

[rekram1-node]

I think for some of this opencode just needs to sync models based on ur account/org froom the github models endpoint.

I'm looking at the model limits from the api and I don't see these values.

[rekram1-node]

This is the script I use to sync:

#!/usr/bin/env bun
import fs from "fs/promises";
import { homedir } from "os";
import path from "path";
import { randomUUID } from "node:crypto";

const authPath = path.join(homedir(), ".local/share/opencode/auth.json");
const authData = JSON.parse(await fs.readFile(authPath, "utf-8"));
const token = authData["github-copilot"].access;

// assuming individual account type otherwise use https://api.${accountType}.githubcopilot.com
const url = "https://api.githubcopilot.com";

const COPILOT_VERSION = "0.32.4";
const EDITOR_PLUGIN_VERSION = `copilot-chat/${COPILOT_VERSION}`;

const headers: Record<string, string> = {
  Authorization: `Bearer ${token}`,
  "content-type": "application/json",
  "copilot-integration-id": "vscode-chat",
  "editor-version": `vscode/1.105.s`,
  "editor-plugin-version": EDITOR_PLUGIN_VERSION,
  "user-agent": `GitHubCopilotChat/${COPILOT_VERSION}`,
  "openai-intent": "conversation-panel",
  "x-github-api-version": "2025-04-01",
  "x-request-id": randomUUID(),
  "x-vscode-user-agent-library-version": "electron-fetch",
};

const response = await fetch(`${url}/models`, {
  headers,
});

if (!response.ok) {
  throw new Error(
    `Failed to fetch models: ${response.status} ${response.statusText}`,
  );
}

export interface ModelsResponse {
  data: Array<Model>;
  object: string;
}

interface ModelLimits {
  max_context_window_tokens?: number;
  max_output_tokens?: number;
  max_prompt_tokens?: number;
  max_inputs?: number;
}

interface ModelSupports {
  tool_calls?: boolean;
  parallel_tool_calls?: boolean;
  dimensions?: boolean;
}

interface ModelCapabilities {
  family: string;
  limits: ModelLimits;
  object: string;
  supports: ModelSupports;
  tokenizer: string;
  type: string;
}

interface Model {
  capabilities: ModelCapabilities;
  id: string;
  model_picker_enabled: boolean;
  name: string;
  object: string;
  preview: boolean;
  vendor: string;
  version: string;
  policy?: {
    state: string;
    terms: string;
  };
}

const data = (await response.json()) as ModelsResponse;

const pretty = JSON.stringify(data, null, 2);

await Bun.write("models.local.json", pretty);

console.log(pretty);

[d4rky-pl]

@rekram1-node I think it may be an issue with your script, I see the same results:

      "capabilities": {
        "family": "claude-opus-4.6",
        "limits": {
          "max_context_window_tokens": 144000,
          "max_non_streaming_output_tokens": 16000,
          "max_output_tokens": 64000,
          "max_prompt_tokens": 128000,
          "vision": {
            "max_prompt_image_size": 3145728,
            "max_prompt_images": 1,
            "supported_media_types": [
              "image/jpeg",
              "image/png",
              "image/webp"
            ]
          }
        },

but at the same time I already confirmed multiple times that I can comfortably hit at least 180K context (as measured by opencode so I may be misunderstanding something). The script you shared makes you pretend to be vscode, maybe the issue is with the version it presents itself as? 🤔

I'll also try to poke at this a bit more

[d4rky-pl]

@rekram1-node I only now realized that VSC changed the way it's displaying the context size between versions which may be why it's so confusing.

edit: and after re-reading the issue I see it was mentioned at the very top 🤦

VSC version 1.109.5

VSC 1.110

They seem to be displaying the "Context Size" as Input + Output now.

At the same time with the override in opencode.json I comfortably hit this (with compaction disabled):

...but I admit I may just as well be misunderstanding what I'm seeing here. With auto-compaction enabled it usually hit the threshold at around 90-100K before but it could be due to other bugs.

For comparison, Copilot CLI:

If I can help in any way let me know, otherwise I'll stop making extra noise

[slhad]

Indeed adding "prompt_tokens" and "output_tokens" seems to be the context value displayed (if we ignore Grok and Gemeni models which seems to have a 19K tokens deducted for what ever reason)

Easier to see with that addition

console.log(JSON.stringify(data.data.map((model) => ({
  id: model.id,
  name: model.name,
  prompt_tokens: model?.capabilities?.limits?.max_prompt_tokens,
  output_tokens: model?.capabilities?.limits?.max_output_tokens,
  max_context_window_tokens: model?.capabilities?.limits?.max_context_window_tokens,
  sum_prompt_and_output: (model?.capabilities?.limits?.max_prompt_tokens ?? 0) + (model?.capabilities?.limits?.max_output_tokens ?? 0),
})), null, 2));

[rekram1-node]

The script isn't wrong it's pulling directly from their api, but yeah I see what u mean. Ill play w/ their limit see what happens

[hugefiver]

@rekram1-node

I am the author of #1104. Here is my command for fetching the model list from the Copilot API, and the file changes were made by LLM.

$T = curl -s https://api.github.com/copilot_internal/v2/token -H "Authorization: Bearer $COPILOT_TOKEN" -H 'User-Agent: GitHubCopilotChat/0.39.0' -H 'Editor-Version: vscode/1.111.0' -H 'Editor-Plugin-Version: copilot-chat/0.39.0' -H 'Copilot-Integration-Id: vscode-chat' -H 'X-GitHub-Api-Version: 2025-05-01'; curl -s "$($T|jq -r '.endpoints.api')/models" -H "Authorization: Bearer $($T|jq -r .token)" -H 'User-Agent: GitHubCopilotChat/0.39.0' -H 'Editor-Version: vscode/1.111.0' -H 'Editor-Plugin-Version: copilot-chat/0.39.0' | jq '[.data[] | select(.model_picker_enabled == true) | select(.id | test("claude|gpt|gemini"))]' | clip

And I guess the context shown in VSCode's "chat: Manage LMs" maybe the lower on of max_context_window_tokens and max_prompt_tokens + max_output_tokens.