Add unified thinking effort control across all providers#9242
Add unified thinking effort control across all providers#9242jh-block wants to merge 33 commits into
Conversation
|
Documentation preview deployed: https://pr-9242.goose-pr-previews-poc.pages.dev |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e82c782f40
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".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 19144daad0
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".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a2ec5b9344
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".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 58da4d7d21
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".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a051baa586
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".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 52c2ff265a
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".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: aa8ac91b82
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".
Adds a single ThinkingEffort enum (off/low/medium/high/max) to ModelConfig that works across all providers, replacing the fragmented per-provider thinking controls. Backend: - Add ThinkingEffort enum to model.rs with FromStr/Display, read from request_params['thinking_effort'] or GOOSE_THINKING_EFFORT env var - OpenAI format: maps effort to reasoning_effort (low/medium/high) - Anthropic format: maps effort to thinking_type and thinking_effort - Databricks format: maps to reasoning_effort for OpenAI models and budget_tokens for Claude models - Google format: maps to thinking_level (low/high) for Gemini 3 - All providers fall back to legacy config when unified effort is unset UI: - Replace model-specific controls (Gemini thinking level, Claude thinking type/effort/budget) with a single 'Thinking Effort' dropdown - Show dropdown for any model that supports thinking (Claude, OpenAI reasoning models, Gemini 3) - Persist setting via GOOSE_THINKING_EFFORT config key
Remove all legacy configuration paths: - CLAUDE_THINKING_TYPE, CLAUDE_THINKING_EFFORT, CLAUDE_THINKING_BUDGET - CLAUDE_THINKING_ENABLED - GEMINI3_THINKING_LEVEL - Model name suffix effort encoding (e.g., o3-mini-high) All providers now use only the unified GOOSE_THINKING_EFFORT config (via request_params or env var). The CLI configure command now shows a single thinking effort prompt for any model that supports thinking. Also simplifies the CLI configure flow from model-specific multi-step thinking config to a single dropdown.
When a ModelConfig is created with a reasoning model name that has an effort suffix (e.g., 'o3-mini-high'), normalize_effort_suffix() now strips the suffix and sets thinking_effort in request_params. This ensures existing saved configurations with suffixed model names continue to work transparently. Only applies when no explicit thinking_effort is already set in request_params (explicit config always wins).
…t_params, avoid panic in configure
…to anthropic/databricks providers
…thinking to anthropic/databricks
…st, clarify type checks - Remove redundant local ThinkingEffort enum in anthropic.rs; use crate::model::ThinkingEffort directly since it has Display/FromStr - Simplify test_create_request_enabled_thinking_with_budget using cfg_with_effort and assert budget > 0 - Replace !!name checks with typeof === 'string' in SwitchModelModal Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: jh-block <jhugo@block.xyz>
Signed-off-by: jh-block <jhugo@block.xyz>
Signed-off-by: jh-block <jhugo@block.xyz>
Signed-off-by: jh-block <jhugo@block.xyz>
Signed-off-by: jh-block <jhugo@block.xyz>
Signed-off-by: jh-block <jhugo@block.xyz>
Signed-off-by: jh-block <jhugo@block.xyz>
Signed-off-by: jh-block <jhugo@block.xyz>
Signed-off-by: jh-block <jhugo@block.xyz>
jh-block
force-pushed
the
jhugo/unified-thinking-effort
branch
from
c800e5c to
db5e545
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: db5e54527a
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>
Signed-off-by: jh-block <jhugo@block.xyz>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 00fddd28f7
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".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ac8d84177d
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".
jh-block
force-pushed
the
jhugo/unified-thinking-effort
branch
2 times, most recently
from
f2e2910 to
713ae75
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 713ae756ec
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>
jh-block
force-pushed
the
jhugo/unified-thinking-effort
branch
from
713ae75 to
c0b4bce
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c0b4bce830
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>
Signed-off-by: jh-block <jhugo@block.xyz>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ab1783305f
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>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ef49673db6
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>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6aae898ce0
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>
|
Hey @jh-block 👋 — I was independently investigating how thinking/reasoning parameters flow across provider types in Goose and was about to start work on exactly this gap. Really glad to see this PR land first! |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b494213ec8
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>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c8f0cacb4b
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".
| let claude_thinking_providers = ["anthropic", "databricks"]; | ||
| let supports_thinking = model.to_lowercase().starts_with("gemini-3") | ||
| || (model.to_lowercase().contains("claude") | ||
| && claude_thinking_providers.contains(&provider_name.as_str())) | ||
| || goose::model::ModelConfig::new(&model) |
There was a problem hiding this comment.
The new CLI gating excludes Claude models unless the provider is anthropic or databricks, so goose configure skips the unified GOOSE_THINKING_EFFORT prompt for OpenRouter Claude selections even though this commit adds OpenRouter reasoning support via apply_reasoning_config when a model is reasoning-capable. In practice, configuring openrouter with a reasoning Claude model (for example claude-sonnet-4*) leaves users unable to set thinking effort through the CLI path, making the new feature unavailable from the primary setup flow.
Useful? React with 👍 / 👎.
|
Great work on this — the One gap worth flagging: local providers aren't wired up yet.
Both of these are listed as first-class providers on the goose docs providers page, so community users running local models would benefit from the same unified knob. The infrastructure is now in place — |
2 participants
Replaces fragmented provider-specific thinking controls with a single unified Thinking Effort setting across supported providers.
Supersedes #7628 with the following changes:
Also updated some providers to support reasoning where they did not (or did not fully) before: