Summary

The OpenCode plugin's thinking profile for claude-opus-4-7 only advertises the legacy 5 levels (off | minimal | low | medium | high), so /think xhigh, /think adaptive, and /think max all fail with "not supported", even though docs/tools/thinking.md states Opus 4.7 supports all three.

Steps to reproduce

1. Configure an opencode:default auth profile (api_key)
2. In a session using opencode/claude-opus-4-7, type /think xhigh
3. Server rejects:

   Thinking level "xhigh" is not supported for opencode/claude-opus-4-7. Use one of: off, minimal, low, medium, high.

The same xhigh / adaptive / max directives work on anthropic/claude-opus-4-7 (direct provider).

Expected behavior

Per docs/tools/thinking.md (v2026.4.24):

• xhigh → "ultrathink+" (GPT-5.2+ and Codex models, plus Anthropic Claude Opus 4.7 effort)
• adaptive → ... Anthropic Claude Opus 4.7 ...
• max → provider max reasoning (currently Anthropic Claude Opus 4.7)

OpenCode Zen routes Opus 4.7 requests to Anthropic, so the same level set should be exposed for opencode/claude-opus-4-7 as for anthropic/claude-opus-4-7.

Actual behavior

/think xhigh|adaptive|max are rejected with the legacy 5-level menu, indicating the runtime validation profile for opencode/claude-opus-4-7 doesn't include the dynamic levels.

Root cause hypothesis

• extensions/anthropic/register.runtime.ts declares xhigh/adaptive/max for Opus 4.7 (xhigh referenced in the file).
• extensions/opencode/ has no resolveThinkingProfile and no xhigh references at all, so runtime validation falls back to the default profile (the legacy 5 levels).

This is similar in shape to #71269 — same root pattern of "provider plugin doesn't surface dynamic thinking options to runtime/UI", just on a different plugin.

Proposed fix

OpenCode plugin should mirror the Anthropic plugin's thinking profile for proxied Anthropic models (claude-opus-4-7, claude-opus-4-6, claude-sonnet-4-6). Either by:

1. Implementing resolveThinkingProfile in extensions/opencode/ that delegates to the underlying Anthropic profile based on model id, or
2. Declaring per-model level sets in the plugin's manifest

Environment

• OpenClaw 2026.4.24
• Auth: opencode (api_key)
• Channel: Telegram main session