fix: clamp unsupported max variant for non-Opus Claude models by RaviTharuma · Pull Request #2606 · code-yeongyu/oh-my-openagent

RaviTharuma · 2026-03-16T06:50:30Z

Summary

This PR is a narrow compatibility fix, not a general model-fallback system.

The model has already been selected.
The actual problem is that the request can still carry a variant level that the selected model does not support.

For Claude models, variant: "max" is valid on Opus 4.6, but not on Sonnet or Haiku. When that unsupported level is passed through unchanged, Anthropic rejects the request with:

level "max" not supported, valid levels: low, medium, high

So the failure is not “the wrong model was chosen”.
The failure is “the chosen model received a thinking/variant level it cannot accept”.

Problem

When an agent with variant: "max" (for example Prometheus or Atlas) moves from claude-opus-4-6 to claude-sonnet-4-6, the request keeps "max" even though Sonnet does not support that level.

That makes fallback and recovery flows brittle:

the selected model itself is valid
but the request still fails because the attached variant level is invalid for that model

Root cause

The anthropic-effort hook in src/hooks/anthropic-effort/hook.ts handled the Opus case, but skipped non-Opus Claude models:

Opus + variant: "max" -> injects effort=max
non-Opus Claude + variant: "max" -> returns early and leaves message.variant="max" untouched

That allows OpenCode to send the unsupported level straight to Anthropic.

Fix

When the selected Claude model is not Opus and the request carries variant: "max", this PR now:

clamps the effort/level to "high"
mutates message.variant to "high"

That keeps the request compatible with Sonnet/Haiku while preserving the intent to use the highest supported level.

Why this matters

Without this fix, fallback looks broken even when the fallback model itself is perfectly usable.
The system chooses a model that could work, then immediately breaks the request by carrying over a level that model does not support.

This PR fixes that narrow but important compatibility gap.

Scope

This PR intentionally stays small and targeted.

It does not implement a full general-purpose resolver for all model-specific settings such as:

variant
reasoningEffort
thinking
maxTokens
other provider/model-specific options

That broader problem should be solved separately as a dedicated settings compatibility resolver that runs after model selection and independently from model fallback.

Changes

src/hooks/anthropic-effort/hook.ts
- clamps non-Opus Claude variant: "max" to "high"
- keeps Opus behavior unchanged
src/hooks/anthropic-effort/index.test.ts
- adds regression coverage for the clamping behavior

…back When an agent configured with variant: "max" falls back from Opus to Sonnet (or Haiku), the "max" variant was passed through unchanged. OpenCode sends this as level: "max" to the Anthropic API, which rejects it with: level "max" not supported, valid levels: low, medium, high The anthropic-effort hook previously only handled Opus (inject effort=max) and skipped all other Claude models. Now it actively clamps "max" → "high" for non-Opus Claude models and mutates message.variant so OpenCode doesn't pass the unsupported level to the API.

cubic-dev-ai

No issues found across 2 files

Confidence score: 5/5

Automated review surfaced no issues in the provided summaries.
No files require special attention.

_{Auto-approved: The fix correctly handles variant clamping for non-Opus Claude models to prevent API errors, with appropriate test coverage and no apparent side effects.}

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9346bc8379

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

RaviTharuma · 2026-03-16T07:27:29Z

Update: After further testing, we found a cleaner approach that works as a standalone OpenCode plugin (.opencode/plugins/variant-clamp.js) without touching the effort hook at all — it just clamps message.variant in chat.params before the anthropic-effort hook runs:

// Anthropic: "max" is Opus 4.6 only — clamp to "high" for Sonnet/Haiku
if (message.variant === "max" && isClaude(model.providerID, model.modelID)) {
  if (!OPUS_PATTERN.test(model.modelID)) {
    message.variant = "high";
  }
}

However, the fix in this PR is still valuable as a built-in safety net so users don't need a separate plugin. The two approaches are complementary — the PR catches it in the effort hook, the plugin catches it even earlier.

Also filed anomalyco/opencode#17736 for OpenCode itself to validate thinking levels against model capabilities at the provider layer.

chatgpt-codex-connector · 2026-03-17T10:58:18Z

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

cubic-dev-ai

0 issues found across 4 files (changes from recent commits).

_{Requires human review: The change to mutate the raw message object by reference instead of using a copy, combined with potential over-clamping for non-Opus models, introduces possible regressions.}

RaviTharuma · 2026-03-18T19:52:55Z

@acamq please check it again :)

cubic-dev-ai Bot approved these changes Mar 16, 2026

View reviewed changes

chatgpt-codex-connector Bot reviewed Mar 16, 2026

View reviewed changes

Comment thread src/hooks/anthropic-effort/hook.ts Resolved

Comment thread src/hooks/anthropic-effort/hook.ts Resolved

RaviTharuma mentioned this pull request Mar 16, 2026

Validate thinking level/variant against model capabilities before sending to API anomalyco/opencode#17736

Closed

RaviTharuma changed the title ~~fix: clamp variant max→high for non-Opus Claude models on fallback~~ fix: clamp unsupported max variant for non-Opus Claude models Mar 17, 2026

fix(anthropic-effort): clamp variant against mutable request message

71b1f7e

cubic-dev-ai Bot reviewed Mar 17, 2026

View reviewed changes

code-yeongyu added the triage:bug Confirmed bug with repro steps label Mar 24, 2026

code-yeongyu merged commit 63ebedc into code-yeongyu:dev Mar 25, 2026
8 checks passed

RaviTharuma deleted the fix/clamp-variant-on-non-opus-fallback branch September 23, 2026 19:39

Navigation Menu

Repository navigation

fix: clamp unsupported max variant for non-Opus Claude models - #2606

fix: clamp unsupported max variant for non-Opus Claude models#2606
code-yeongyu merged 2 commits into
code-yeongyu:devfrom
RaviTharuma:fix/clamp-variant-on-non-opus-fallback

RaviTharuma commented Mar 16, 2026 •

edited

Loading

Uh oh!

cubic-dev-ai Bot left a comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

Uh oh!

Uh oh!

RaviTharuma commented Mar 16, 2026

Uh oh!

chatgpt-codex-connector Bot commented Mar 17, 2026

Uh oh!

cubic-dev-ai Bot left a comment

Uh oh!

RaviTharuma commented Mar 18, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

RaviTharuma commented Mar 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Problem

Root cause

Fix

Why this matters

Scope

Changes

Uh oh!

cubic-dev-ai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

RaviTharuma commented Mar 16, 2026

Uh oh!

chatgpt-codex-connector Bot commented Mar 17, 2026

Uh oh!

cubic-dev-ai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

RaviTharuma commented Mar 18, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

RaviTharuma commented Mar 16, 2026 •

edited

Loading