GitHub Copilot plugin ignores user-configured model limit overrides #21564

Problem

User-configured limit overrides for GitHub Copilot models are silently discarded. The Copilot plugin always overwrites limit.context, limit.input, and limit.output with API-reported values, even when the user explicitly sets them in opencode.json.

This causes compaction to trigger much earlier than expected — around 11-13% of the configured context window — because:

1. The user sets limit.context: 1000000 in config
2. The Copilot plugin fetches model info from the API and rebuilds limits entirely from the response (max_context_window_tokens, max_prompt_tokens, max_output_tokens)
3. The prev model (carrying user config) is passed to build() but only used for family, name, options, variants, capabilities, etc. — not for limit
4. Compaction checks limit.input (= max_prompt_tokens, ~128K from API), not limit.context
5. The UI percentage is calculated against limit.context, so compaction at ~128K tokens shows as ~12.8% of 1M

Config example

{
  "provider": {
    "github-copilot": {
      "models": {
        "claude-opus-4.6": {
          "limit": {
            "context": 1000000,
            "output": 64000
          }
        }
      }
    }
  },
  "compaction": {
    "auto": true,
    "prune": true,
    "reserved": 2000
  }
}

Despite this config, compaction triggers at ~128K tokens (the API-reported max_prompt_tokens), not at ~1M.

Root cause

In packages/opencode/src/plugin/github-copilot/models.ts, the build() function:

function build(key: string, remote: Item, url: string, prev?: Model): Model {
  return {
    // ...
    limit: {
      context: remote.capabilities.limits.max_context_window_tokens,
      input: remote.capabilities.limits.max_prompt_tokens,
      output: remote.capabilities.limits.max_output_tokens,
    },
    // prev is used for family, name, options, variants, etc. but NOT limit
  }
}

Other fields from prev (user config) are preserved with the prev?.field ?? remote.field pattern, but limit always comes from the API response.

Expected behavior

User-configured limit overrides should be respected, consistent with how other model properties (options, variants, capabilities) from user config take precedence. Something like:

limit: {
  context: prev?.limit?.context ?? remote.capabilities.limits.max_context_window_tokens,
  input: prev?.limit?.input ?? remote.capabilities.limits.max_prompt_tokens,
  output: prev?.limit?.output ?? remote.capabilities.limits.max_output_tokens,
},

Additional context

The compaction trigger in overflow.ts prioritizes limit.input over limit.context:

const usable = input.model.limit.input
  ? input.model.limit.input - reserved
  : context - maxOutputTokens(model)

So even if limit.context were respected, users would also need to set limit.input to actually delay compaction. This should be documented, or limit.context should influence the compaction calculation when limit.input is not explicitly set by the user.