fix: correct github-copilot model limits to match Copilot API

[Nepomuceno]

Summary

Corrects token limits for 6 models in the github-copilot provider to match the authoritative values from the Copilot API (GET https://api.githubcopilot.com/models).

This PR only fixes limits — no models are added or removed.

Changes

Model	Field	Before	After (Copilot API)	Delta
claude-haiku-4.5	context	144,000	200,000	+56,000
	input	128,000	136,000	+8,000
	output	32,000	64,000	+32,000
claude-opus-4.5	context	160,000	200,000	+40,000
	input	128,000	168,000	+40,000
claude-opus-4.6	context	144,000	200,000	+56,000
	input	128,000	168,000	+40,000
	output	64,000	32,000	-32,000
claude-sonnet-4.5	context	144,000	200,000	+56,000
	input	128,000	168,000	+40,000
claude-sonnet-4.6	input	128,000	168,000	+40,000
gpt-5.2	context	264,000	400,000	+136,000
	input	128,000	272,000	+144,000
	output	64,000	128,000	+64,000

How the correct values were verified

curl -s "https://api.githubcopilot.com/models" \
  -H "Authorization: Bearer $(gh auth token)" \
  -H "Copilot-Integration-Id: vscode-chat"

Each model in the response includes:

"limits": {
    "max_context_window_tokens": ...,   // → limit.context
    "max_prompt_tokens": ...,           // → limit.input
    "max_output_tokens": ...            // → limit.output
}

Impact

The under-reported limits cause tools like OpenCode to trigger context compaction 40K–136K tokens earlier than necessary, significantly degrading the user experience especially for Claude and GPT-5.2 models.

Fixes #858

[rekram1-node]

@Nepomuceno Where are u getting these? What kind of subscription do u have?

I'm testing and this doesnt seem accurate??

    "family": "claude-haiku-4.5",
    "limits": {
      "max_context_window_tokens": 144000,
      "max_non_streaming_output_tokens": 16000,
      "max_output_tokens": 32000,
      "max_prompt_tokens": 128000,
      "vision": {

[Nepomuceno]

Interesting I do have an enterprise subscription with very relaxed limits. Interesting problem on how do we do with that.

[Nepomuceno]

Happy to chat on alternatives on how this can be made more dynamic because those limits do change all the time and based on your subscription

[zakriabilal]

That is very odd, I've just done this on my end and can see the 200K limits. I’m not on the enterprise plan, but we do have the Business Copilot subscriptions with models enabled.

Could it be that the functionality is geo-restricted, or perhaps it is being rolled out gradually?

[rekram1-node]

I think the ideal solution is just having opencode plugin sync from github models list and that way itll use the specifics for ur account

[tobwen]

That would also solve this issue: #1193

[tobwen]

Is anyone working on it already?