Repository navigation
Pick Model and Effort for Source Control text actions - #37205
Conversation
The base already merged main's overlay-focus change into the setup panel. This merge applies cleanly: launch entries take the base's panel as is, and text cards keep their single Use button. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
|
No actionable comments were generated in the recent review. Walkthrough
Merge Risk:
|
In Settings → Git & Source Control, the actions that start an agent (Commit failure fixes, Fix with AI, Resolve conflicts) let you pick Model and Effort from menus, but the actions that write text (Commit message, Pull request, Branch name, Chat name) only offered a raw Arguments box. Choosing a model for a commit message meant typing CLI flags such as
--model gpt-6.1-terra -c model_reasoning_effort="low". This came from #34275, which moved Model/Effort into menus for the agent-starting actions only and deliberately left the text actions on Arguments; they never had menus, so this is an inconsistency rather than a regression.What this PR does for users: the text action cards (global and repository) now pick their agent, Model and Effort through the same shared agent picker the other action cards use. The Model menu lists the models that agent can use to write text on that host; opening the menu refreshes the list from the installed CLI (so newer models such as new Codex models appear), and Effort follows the selected model. Text cards have no Chat/Terminal, Permissions or Fast choices because they run in the background. "Custom command" stays available and has no Model/Effort. Each card shows the model and effort it will use, and its "Same as default · · " option shows what the card falls back to, computed by the same code that runs the generation. The Commit message card's model is also the default for Pull request details, Branch name and Chat name when they don't pick their own (as the setup guide's AI author model already was); the card says so, and each other card's "Same as default" option shows the model it actually uses.
How it works: generation already read a saved per-action Model/Effort choice that nothing in Settings could set. The picker now writes that existing setting. Picks stay drafts until the card's Save; at Save, the edits you made are applied in order to the current settings, for this computer and for the host of every registered repository (a repository card writes its own repository's host). When a pick replaces a matching
--modelor effort flag in that card's Arguments, the flag is removed in the same save. Flags the picker can't recognise stay in Arguments and the card says the value is set by your Arguments. When Chat name follows each chat’s agent, configuring that agent’s Model or Effort keeps that inheritance; a plain agent choice or configuring a different agent still makes an explicit selection. Each text agent declares its own model and effort flags (for example Codex-c model_reasoning_effort=, OpenCode 2model#variant, Amp--mode). Nothing new is saved and nothing changes on upgrade: existing Arguments and existing per-host model choices keep working exactly as before, and older Orca versions read the same saved choices.#35802 (shared agent picker) is merged, and this PR now targets main. Text actions use its updated common hover submenus and keyboard controls, with compact Settings menus and the existing Use button for staged text selections.
Picking Claude "Haiku" runs Haiku: the separate fix for Claude text generation (#37367, now on main) is merged into this branch.
Plan: lane plan reviewed by Claude Opus and a fresh Claude Opus check; the common pattern for this (a typed model selection for the commit/PR/branch writer, picked with the chat model and effort pickers and compiled into CLI flags at spawn) was adopted on Orca's existing per-action setting.
Intended differences from the common pattern:
model#variant, Effort shows "Set by your Arguments" instead of editing it; pick a Model to take it over (the common pattern also hides effort when its menu doesn't own the selected model).Known limitations that already exist on main (temporary; tracked separately): a repository on a direct SSH connection keeps its saved model list, because the shared model-listing helper doesn't list models over direct SSH; a repository that runs its agent in WSL is offered the Windows-side model list, because text generation checks models against that one local list.
🤖 Generated with Claude Code