[Bug]: Claude Opus 4.8 - thinking.type.enabled not supported (regression of #3477)

[MAST1999]

Prerequisites

• I will write this issue in English (see our Language Policy)
• I have searched existing issues to avoid duplicates
• I am using the latest version of oh-my-opencode
• I have read the documentation or asked an AI coding agent with this project's GitHub URL loaded and couldn't find the answer

Bug Description

OpenCode: 1.15.10
Plugin expected: 4.5.1
Plugin loaded: 4.5.1
Bun: 1.3.14

Selecting anthropic/claude-opus-4-8 (any variant) on any built-in agent returns the same 400 from the Anthropic API that #3477 first reported for Opus 4.7:

"thinking.type.enabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior.

#3477 was closed as fixed for Opus 4.7, but the underlying hardcoded thinking.type: "enabled" is still present in the current dev source and now breaks again on Opus 4.8 (Anthropic flipped 4.8 to adaptive-only on release, same as 4.7).

The error is generated before the request reaches sst/opencode. opencode core already detects Opus 4.7+ correctly via anthropicOpus47OrLater in packages/opencode/src/provider/transform.ts and emits thinking: { type: "adaptive" }. The bug is that oh-my-openagent overrides that variant by setting thinking directly in every built-in agent's config:

File	Line	Code
src/agents/sisyphus.ts	601, 657	thinking: { type: "enabled", budgetTokens: 32000 }
src/agents/oracle.ts	588	thinking: { type: "enabled", budgetTokens: 32000 }
src/agents/momus.ts	415	thinking: { type: "enabled", budgetTokens: 32000 }
src/agents/metis.ts	310	thinking: { type: "enabled", budgetTokens: 32000 }
src/agents/sisyphus-junior/agent.ts	154	thinking: { type: "enabled", budgetTokens: 32000 }

The compatibility layer in packages/model-core/src/model-settings-compatibility.ts (L197-L205) only drops thinking when supportsThinking === false; it never rewrites type: "enabled" → type: "adaptive". The claude-opus family is marked supportsThinking: true in model-capability-heuristics.ts (L14-L19), so the hardcoded object passes through unchanged and hits the 400.

The type union also cannot express "adaptive":

• packages/model-core/src/model-resolution-types.ts L11: thinking?: { type: "enabled" | "disabled"; budgetTokens?: number }
• packages/model-core/src/fallback-model-object.ts L8: same
• packages/model-core/src/model-requirements.ts L9: same

So a coercion can't even be written without widening these types first.

The only model-version detector currently in the plugin is isClaudeOpus47Model in packages/model-core/src/model-family-detectors.ts, which strict-matches claude-opus-4-7 and does not cover 4.8 or future minors.

Steps to Reproduce

1. Set any agent in oh-my-opencode.json to anthropic/claude-opus-4-8 (e.g. sisyphus).
2. Start any task that invokes that agent.
3. Anthropic returns 400 with "thinking.type.enabled" is not supported for this model.

The same reproduces with anthropic/claude-opus-4-7 and was previously reported in #3477.

Expected Behavior

For Claude Opus 4.7+ models, the plugin should emit:

{ "thinking": { "type": "adaptive" } }

and route the variant level (low / medium / high / xhigh / max) via output_config.effort, matching what sst/opencode's transform.ts already does for the same model family. The hardcoded { type: "enabled", budgetTokens: 32000 } should be replaced with model-aware logic, or the compatibility layer should coerce it transparently so existing agent configs self-heal.

Actual Behavior

Every task using Opus 4.7 or Opus 4.8 stops with the 400 above. No retry, no fallback. Older Claude models (Opus 4.6, Sonnet 4.6, Sonnet 4.5, etc.) still work because type: "enabled" is still accepted on them.

Doctor Output

 oMoMoMoMo Doctor

 ⚠ 1 issue found:

1. TUI plugin entry missing from tui.json
   The server plugin is registered in opencode.json, but the TUI plugin entry ("oh-my-openagent/tui") is missing from tui.json. The Roles · Models sidebar section and TUI-only commands will not appear.
   Fix: Re-run the installer (`npx oh-my-openagent install`) to auto-write tui.json, or add "oh-my-openagent/tui" to the "plugin" array in /home/mast/.config/opencode/tui.json.
   Affects: TUI sidebar, TUI commands

(The TUI warning is unrelated to this bug.)

Error Logs

Surfaced as a TUI toast:

"thinking.type.enabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior.

Configuration

{
  "$schema": "https://raw.githubusercontent.com/code-yeongyu/oh-my-opencode/master/assets/oh-my-opencode.schema.json",
  "agents": {
    "sisyphus":  { "model": "anthropic/claude-opus-4-8", "variant": "max" },
    "oracle":    { "model": "anthropic/claude-opus-4-8", "variant": "max" },
    "metis":     { "model": "anthropic/claude-opus-4-8", "variant": "max" },
    "momus":     { "model": "anthropic/claude-opus-4-8", "variant": "max" }
  }
}

Operating System

Linux

OpenCode Version

1.15.10

[EricXu-0805]

Possibly related to this (same model + thinking area), but a different error string — flagging in case it's a second Opus-4.8 thinking failure mode, or downstream of the same hardcoded thinking config.

Env: opencode 1.15.12, oh-my-openagent 4.5.1, Anthropic OAuth (Claude Max) via opencode-anthropic-oauth, macOS (Apple Silicon).

On anthropic/claude-opus-4-8 + variant: max (Sisyphus), I do not hit the thinking.type.enabled is not supported 400 from this issue — the agent runs several steps fine (thinking + tool calls), then on the next tool-use continuation it fails with:

messages.1.content.9: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.

It reproduces in a fresh session within ~1 min (error at content.9, so not a long-context issue) and recurs on every retry.

Clean signal: the same agent/config on anthropic/claude-opus-4-7 + variant: max works fine — ran 40+ agentic (PR-merge) rounds without this error. Only opus-4-8 triggers it. So it looks like the latest assistant turn's Opus-4.8 thinking block isn't being preserved byte-for-byte (signature/text) when the message is reconstructed and resent alongside the tool_results.

Repro:

1. Set sisyphus to anthropic/claude-opus-4-8, variant: max in oh-my-opencode.json.
2. Run any multi-step task that uses tools.
3. After 1–2 tool-use turns → the 400 above.

Workaround: rolled the agent back to opus-4-7 (keeps max thinking); variant: none also avoids it (no thinking blocks at all).

Happy to open this as a separate issue if you'd prefer — I wasn't sure whether it shares the root cause with the hardcoded thinking.type here or is a distinct preservation bug in the message-reconstruction path. On 1.15.12 I seem to get past the type: "enabled" 400, which makes me suspect this is a second layer.