fix: allow adaptive thinking and xhigh effort with Claude Opus 4.7 by Raptors65 · Pull Request #9281 · aaif-goose/goose

Raptors65 · 2026-05-16T16:47:00Z

Summary

Claude Opus 4.7 has a few breaking changes with the Claude API: https://platform.claude.com/docs/en/about-claude/models/migration-guide

Currently, for Opus 4.7 with the Anthropic provider:

the displayed extended thinking modes are "Enabled" (with a fixed token budget) or "Disabled" (no extended thinking); adaptive thinking is not displayed as an option
if "Enabled" is selected, any messages fail with a 400 error, because Opus 4.7 doesn't support extended thinking
xhigh is not an option for effort; this is an thinking effort level that only exists on Opus 4.7
we don't emit thinking.display = "summarized", which means we use default behavior which in Opus 4.7 was changed to be to not emit reasoning at all
we pass temperature to the API call, which fails if it's set to any non-default value for Opus 4.7

This PR fixes this by:

adding Opus 4.7 as a "known model" in the Anthropic provider
adding xhigh as an effort level only for Opus 4.7, and making this the default for Opus 4.7 (this is the recommended default in the Claude docs)
disabling "extended thinking" for Opus 4.7 in favor of adaptive thinking
emitting thinking.display = "summarized" for all models with adaptive thinking (not technically needed for Sonnet 4.6 and Opus 4.6, but would rather be explicit there too)
omitting temperature for Opus 4.7

Testing

There are some unit tests, but also tested this manually.

CLI ("adaptive" instead of "enabled" mode displayed, xhigh effort available):

Opus.4.7.CLI.mov

Desktop app:

Screen.Recording.2026-05-16.at.1.06.21.PM.mov

Summarized thinking being displayed with Opus 4.7:

Related Issues

N/A

Screenshots/Demos (for UX changes)

Before:

After:

Raptors65 · 2026-05-16T16:47:13Z

fix: allow adaptive thinking and xhigh effort with Claude Opus 4.7 #9281 👈 (View in Graphite)
main

This stack of pull requests is managed by Graphite. Learn more about stacking.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 184d306a47

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".

chatgpt-codex-connector · 2026-05-16T17:16:11Z

+    } else if is_adaptive_model {
+        ThinkingType::Adaptive
+    } else if std::env::var("CLAUDE_THINKING_ENABLED").is_ok() {


Leave Opus 4.7 thinking disabled unless user enables it

This branch now defaults every adaptive-capable model to ThinkingType::Adaptive, which means claude-opus-4-7 requests get thinking turned on even when users did not configure any thinking mode. For Opus 4.7, that changes baseline behavior from "no thinking unless explicitly requested" to "always adaptive thinking", which can materially increase latency/token usage and alter response shape for existing users who only changed model IDs.

Useful? React with 👍 / 👎.

this is already what we do for other models that support adaptive thinking (Sonnet 4.6 / Opus 4.6), so seems to make sense to do that for Opus 4.7 too

chatgpt-codex-connector · 2026-05-16T17:16:11Z

+    if is_claude_opus_47(model_name) {
+        ThinkingEffort::XHigh
+    } else {


Keep Opus 4.7 effort at API default unless configured

Setting Opus 4.7's implicit effort to xhigh forces a more expensive effort tier whenever users don't explicitly set CLAUDE_THINKING_EFFORT/effort. Since thinking_effort() uses this as the fallback, existing configurations that rely on defaults will silently run at higher cost/latency than the provider default. xhigh should be an explicit opt-in, not the fallback.

Useful? React with 👍 / 👎.

xhigh is the recommended default in the Claude docs for "coding and agentic work", which seems like the most likely kind of work someone would be using goose for, but open to having the default stay at high if we want to do that (high is the API default)

shakoorshkh · 2026-05-17T09:07:13Z

Observed:
Opus 4.7 rejects requests with extended thinking enabled, non-default temperature, or missing thinking.display = "summarized" — all silent 400 failures from upstream breaking changes.
Downside:
Opus 4.7 is now a special-cased model with divergent behavior from the rest of the Anthropic provider. That surface grows if Claude releases further model-specific API contracts.
Tradeoff:
Hardcoded model-specific logic over a generalized capability flag system. Faster to ship; accrues provider complexity with each future model that breaks convention.

Use adaptive thinking when preserved Anthropic thinking context needs to synthesize a thinking config for adaptive Claude models. This keeps Opus 4.7 out of the fixed-budget path while preserving prior thinking blocks across continuation requests. Signed-off-by: jh-block <jhugo@block.xyz>

Use the Anthropic API default effort when no unified thinking effort is configured. Explicit max still maps to xhigh for Opus 4.7, but an omitted effort now stays at high and is covered by a regression test. Signed-off-by: jh-block <jhugo@block.xyz>

Preserve an unset unified thinking effort through provider formatting and the model switcher. Providers now only serialize effort fields when the user or legacy config explicitly sets an effort, allowing upstream API defaults to apply. Signed-off-by: jh-block <jhugo@block.xyz>

Keep provider request formatting from inventing a thinking effort when none is configured, while making the desktop and CLI model-selection surfaces default explicit thinking effort to high for reasoning-capable models. Signed-off-by: jh-block <jhugo@block.xyz>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0638379368

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".

Signed-off-by: jh-block <jhugo@block.xyz>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 12d93afa06

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".

Signed-off-by: jh-block <jhugo@block.xyz>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 65cda3c96c

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".

Signed-off-by: jh-block <jhugo@block.xyz>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fecbe0c81b

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".

Signed-off-by: jh-block <jhugo@block.xyz>

Raptors65 force-pushed the lucask/fix/support-adaptive-thinking-with-opus-4-7 branch from b1c73dc to 184d306 Compare May 16, 2026 16:55

Raptors65 marked this pull request as ready for review May 16, 2026 17:09

chatgpt-codex-connector Bot reviewed May 16, 2026

View reviewed changes

DOsinga added the needs_human label to set when a robot looks at a PR and can't handle it label May 18, 2026

DOsinga assigned jh-block May 18, 2026

jh-block force-pushed the lucask/fix/support-adaptive-thinking-with-opus-4-7 branch from 184d306 to 29ad7a3 Compare May 20, 2026 11:32

Raptors65 and others added 5 commits May 20, 2026 13:34

fix: allow adaptive thinking and xhigh effort with Claude Opus 4.7

b790909

Default Opus adaptive effort to high

64465da

Use the Anthropic API default effort when no unified thinking effort is configured. Explicit max still maps to xhigh for Opus 4.7, but an omitted effort now stays at high and is covered by a regression test. Signed-off-by: jh-block <jhugo@block.xyz>

jh-block force-pushed the lucask/fix/support-adaptive-thinking-with-opus-4-7 branch from 29ad7a3 to 0638379 Compare May 20, 2026 11:36

chatgpt-codex-connector Bot reviewed May 20, 2026

View reviewed changes

Comment thread crates/goose/src/providers/formats/anthropic.rs Outdated

Preserve thinking context without effort

12d93af

Signed-off-by: jh-block <jhugo@block.xyz>

chatgpt-codex-connector Bot reviewed May 20, 2026

View reviewed changes

Comment thread crates/goose/src/providers/formats/anthropic.rs Outdated

Add xhigh thinking effort

65cda3c

Signed-off-by: jh-block <jhugo@block.xyz>

chatgpt-codex-connector Bot reviewed May 20, 2026

View reviewed changes

Comment thread crates/goose-cli/src/commands/configure.rs

Add xhigh to configure selector

fecbe0c

Signed-off-by: jh-block <jhugo@block.xyz>

chatgpt-codex-connector Bot reviewed May 20, 2026

View reviewed changes

Comment thread crates/goose/src/providers/formats/anthropic.rs Outdated

Gate xhigh to Opus 4.7

387d134

Signed-off-by: jh-block <jhugo@block.xyz>

Conversation

Raptors65 commented May 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Testing

Related Issues

Screenshots/Demos (for UX changes)

Uh oh!

Raptors65 commented May 16, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot May 16, 2026

Choose a reason for hiding this comment

Uh oh!

Raptors65 May 16, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot May 16, 2026

Choose a reason for hiding this comment

Uh oh!

Raptors65 May 16, 2026

Choose a reason for hiding this comment

Uh oh!

shakoorshkh commented May 17, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Raptors65 commented May 16, 2026 •

edited

Loading