[FEATURE] Agent-Specific Reasoning Level Configuration

[NamedIdentity]

Feature Request: Agent-Specific Reasoning Level Configuration

Summary

On modern OpenCode (v1.0.210+), there is no clear or working method to configure specific agents to use specific reasoning levels (e.g., reasoningEffort: "high"). The documentation and codebase were reviewed by both the user and an AI assistant (Claude Code), and neither could determine a working approach.

Environment

• OpenCode version: 1.1.12 (modern, v1.0.210+)
• Plugin version: opencode-openai-codex-auth@4.4.0
• Platform: Windows

Problem Description

Goal

Configure agents in opencode.json to always use a specific reasoning level when invoked as subagents (where TUI variant selection is not available).

Example use case:

"principal-chatgpt": {
  "model": "openai/gpt-5.2",
  "prompt": "{file:./agent/principal-chatgpt.md}"
  // Need this agent to ALWAYS use reasoningEffort: "high"
}

What We Tried

1. Created separate model entry with options block (per docs/configuration.md Pattern 2):

   "gpt-5.2-high": {
     "name": "GPT 5.2 High (OAuth)",
     "limit": { "context": 272000, "output": 128000 },
     "options": {
       "reasoningEffort": "high",
       "reasoningSummary": "detailed"
     }
   }

   Then referenced in agent: "model": "openai/gpt-5.2-high"

   Result: OpenCode reports model "is not valid"

2. Added id field to map custom model to base model (similar to working Google provider pattern):

   "gpt-5.2-high": {
     "id": "gpt-5.2",
     "name": "GPT 5.2 High (OAuth)",
     ...
   }

   Result: Still reports model "is not valid"

3. Reviewed config/opencode-legacy.json which defines separate model entries like gpt-5.2-high with options blocks - but documentation states this config is for OpenCode v1.0.209 and below only.

Documentation Findings

The documentation provides conflicting or unclear guidance:

1. config/README.md states:

   • Modern OpenCode (v1.0.210+) should use opencode-modern.json with variants
   • Legacy OpenCode (v1.0.209-) should use opencode-legacy.json with separate model entries
   • "Use the config file appropriate for your OpenCode version"

2. docs/configuration.md shows "Per-Agent Models" example:

   "agent": {
     "commit": { "model": "openai/gpt-5.1-codex-low" },
     "review": { "model": "openai/gpt-5.1-codex-high" }
   }

   But doesn't specify which OpenCode version this works on.

3. docs/configuration.md Pattern 2 shows per-model options overrides, but these don't work on modern OpenCode when referenced by agents.

4. The variants system works for interactive use (--variant=high or TUI selection), but agents/subagents cannot access TUI and there's no variant field in agent configuration.

Codebase Review

Reviewed the following files:

• index.ts - Plugin loader extracts userConfig from provider config
• lib/request/request-transformer.ts - getModelConfig() merges global + model-specific options
• lib/request/helpers/model-map.ts - Maps model names including gpt-5.2-high → gpt-5.2
• lib/types.ts - UserConfig structure supports per-model options

The plugin code appears to support per-model options, but OpenCode's validation rejects custom model names before the plugin can process them.

Request

One of the following:

1. Documentation clarification: If there IS a working method to configure agent-specific reasoning levels on modern OpenCode, please document it clearly with a complete example.

2. Feature addition: If no method currently exists, please add support for configuring reasoning levels per-agent. Implementation approach is at developer discretion.

Impact

This limitation affects users who:

• Use multi-agent configurations with different reasoning requirements
• Run agents as subagents (no TUI access for variant selection)
• Want consistent, predictable reasoning behavior for specific agents
• Are on modern OpenCode and cannot downgrade to legacy

Additional Context

The Google provider's authentication plugin (opencode-google-antigravity-auth) appears to work with custom model entries that have an id field mapping to base models (e.g., gemini-3-flash-high with id: "gemini-3-flash"). The same pattern does not work for the OpenAI provider.