Specifying effort
Effort is set using macros—special hashtags that modify agent behavior during a run. Include#mini, #auto, or #max in any of these places:
- Repository custom instructions — Apply a default effort level to all runs in a repo
- Trigger-level instructions — Set effort for specific triggers (e.g., use
#minifor auto-labeling) - GitHub issue body — Set effort when creating an issue
- GitHub comment — Include the macro when mentioning
@pullfrog
#max in a comment overrides #mini set in repo instructions.
If not specified, Pullfrog defaults to #auto.
Effort levels
#mini — Fast, minimal reasoning
#mini — Fast, minimal reasoning
Best for simple, straightforward tasks. Uses smaller/faster models with reduced reasoning to minimize latency and cost.
#auto — Balanced (default)
#auto — Balanced (default)
Good for most tasks. Balances capability and cost with sensible defaults.
#max — Maximum capability
#max — Maximum capability
Best for complex tasks requiring deep reasoning, difficult debugging, or intricate planning. Uses the most capable models with maximum reasoning.
Agent implementations
Each agent implements effort levels differently based on their available models and configuration options.Summary table
| Agent | #mini | #auto | #max |
|---|---|---|---|
| Claude Code | Haiku | Opus + Sonnet | Opus |
| Codex | Mini model + low reasoning | Standard model | Max model + high reasoning |
| Gemini | Flash + low thinking | Flash + high thinking | Pro + high thinking |
| Cursor | Auto (default) | Auto (default) | Opus 4.5 Thinking |
| OpenCode | No effect | No effect | No effect |
Claude Code
Claude Code uses model selection based on effort level.| Effort | Model | Description |
|---|---|---|
mini | haiku | Fast, efficient |
auto | opusplan | Opus for planning, Sonnet for execution |
max | opus | Full Opus capability |
Codex (OpenAI)
Codex combines model selection with themodelReasoningEffort parameter.
| Effort | Model | Reasoning | Description |
|---|---|---|---|
mini | gpt-5.1-codex-mini | low | Smaller model, reduced reasoning |
auto | gpt-5.1-codex | default | Standard model, default reasoning |
max | gpt-5.1-codex-max | high | Largest model, maximum reasoning |
Gemini
Gemini combines model selection with thinking level configuration.| Effort | Model | Thinking Level | Description |
|---|---|---|---|
mini | gemini-3-flash-preview | LOW | Fast model, minimal thinking |
auto | gemini-3-flash-preview | HIGH | Fast model, deep thinking |
max | gemini-3-pro-preview | HIGH | Most capable model, deep thinking |
Cursor
Cursor uses model selection via the--model CLI flag.
| Effort | Model | Description |
|---|---|---|
mini | auto (default) | Let Cursor select optimal model |
auto | auto (default) | Let Cursor select optimal model |
max | opus-4.5-thinking | Claude Opus 4.5 with extended thinking |
OpenCode
OpenCode does not currently support effort-level configuration. The effort parameter is ignored.| Effort | Behavior |
|---|---|
mini | No effect |
auto | No effect |
max | No effect |
To control the model used by OpenCode, use a project-level
opencode.json configuration file or set environment variables (e.g., ANTHROPIC_MODEL, OPENAI_MODEL).