config: existing model limit override is ignored

[akuendig]

Summary

A configured limit.context override for the existing openai/gpt-5.6-sol catalog model is ignored. This prevents users from lowering the effective context limit to trigger automatic compaction earlier.

Environment

• opencode version: opencode2 v0.0.0-next-15766 (also reproduced on 0.0.0-next-15760)
• OS: Linux 7.0.9-ogc3.2.fc44.x86_64 x86_64
• Terminal: TERM_PROGRAM=tmux, TERM=xterm-256color, COLORTERM=truecolor
• Shell: /bin/bash
• Install/channel: globally installed next channel
• Active plugins: No user-configured plugins or local plugin files found

Reproduction

1. Add the following to ~/.config/opencode/opencode.jsonc:

   {
     "$schema": "https://opencode.ai/config.json",
     "providers": {
       "openai": {
         "models": {
           "gpt-5.6-sol": {
             "limit": {
               "context": 308000
             }
           }
         }
       }
     }
   }

2. Restart the shared service:

   opencode2 service restart

3. Inspect the resolved model:

   opencode2 api get /api/model \
     | jq '.data[] | select(.id == "gpt-5.6-sol") | {id, modelID, providerID, limit}'

4. Observe that the configured context limit is not present.

Expected Behavior

The existing catalog entry should be overlaid with the configured partial limit:

{
  "limit": {
    "context": 308000,
    "input": 922000,
    "output": 128000
  }
}

Automatic compaction should therefore use approximately 308000 - 128000 = 180000 tokens as its preflight threshold with the default compaction buffer.

Actual Behavior

The resolved model retains the original catalog limits:

{
  "id": "gpt-5.6-sol",
  "modelID": "gpt-5.6-sol",
  "providerID": "openai",
  "limit": {
    "context": 1050000,
    "input": 922000,
    "output": 128000
  }
}

The behavior is consistent across service restarts. It reproduced on both 0.0.0-next-15760 and 0.0.0-next-15766.

Additional Context

• The configuration uses the documented V2 providers shape from https://v2.opencode.ai/models.
• Restarting the background service does not help.
• Changing the configured model into a custom alias also did not make the model appear in /api/model.
• No user-configured plugins were found.
• The service remains healthy after restart.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Variants of models config does not allow the change of the context size. #34544: Similar issue where model config limit overrides are not applied (that one is specifically about variants, but the underlying merge logic may be shared)

[akuendig]

It worked on 0.0.0-next-15718.

[akuendig]

Ok, figured out that this was actually because of v2 reading the v1 config and maybe silently failing/skipping the providers config due to unsupported (?) agent field.

[opencode-agent]

This is not a model limit merge bug. The config mixes V1 and V2 fields: V1 uses agent, while V2 uses agents and providers.

The presence of the V1-only agent field causes the loader to classify the entire document as V1. During V1 decoding/migration, the V2 providers field is treated as an unknown field and ignored, so the provider/model override never reaches the catalog. That also explains why a custom model alias did not appear.

The immediate fix is to use one config generation consistently:

• V2: migrate agent to the V2 agents shape and keep providers.
• V1: keep agent, but use the singular provider field.

The model overlay itself already merges partial limits correctly. Separately, we should consider rejecting or warning on mixed V1/V2 documents rather than silently dropping fields.