Agents and Models

Sidequest runs whichever coding-agent CLIs you have installed. Claude Code is the default and the most deeply integrated; the others run quests through the same embedded terminal.

Supported agents

Agent	Binary	Permission flags Sidequest passes	Plan mode	Model flag
Claude Code	claude	--permission-mode auto (default) or --dangerously-skip-permissions	--permission-mode plan	--model
Gemini CLI	gemini	--yolo	--approval-mode plan	-m
Codex CLI	codex	--full-auto	--sandbox read-only --ask-for-approval untrusted	none (model picker unavailable)
Copilot CLI	copilot	--allow-all	--plan	--model
OpenCode	opencode	none	none	-m

Every agent runs inside tmux via your login shell, so tools installed with nvm, mise and the like are found. Sidequest auto-accepts the “trust this folder” prompts Claude Code and Copilot show on first launch in a directory.

What only works with Claude Code

Several features read Claude Code’s transcripts or rely on its hooks, so they apply to Claude Code quests only:

• The smart router, effort levels and Claude account profiles
• Auto Mode, Bypass Permissions in Home and Chrome Browser Tools settings
• Needs attention status and notifications (via the Claude Code Stop hook)
• Automatic journals and the “quest context” block injected at session start (other agents get the same context as plain prompt text, plus a hint about the sidequest-journal command)
• Resume of an existing conversation after a restart; other agents start fresh
• Token statistics in Live Stats and the Usage Dashboard

Supervisors, Council summaries, journals, handoffs and the router all use the claude binary themselves, so keep Claude Code installed even if your quests run on another agent.

Choosing the agent

• Default: Settings > Agent > Default Agent (“The AI coding agent to use for new quests. Can be overridden per-quest.”)
• Per quest: the Agent select in the Create Quest dialog. The agent cannot be changed after the quest is created.
• CLI paths: each agent has a CLI Path field in Settings > Agent with a Detected / Not found indicator. Leave it blank to auto-detect.

Permission modes (Claude Code)

Setting	Effect
Auto Mode on (default)	--permission-mode auto: Claude Code’s classifier approves routine actions and escalates risky ones
Auto Mode off	--dangerously-skip-permissions: nothing is checked
Bypass Permissions in Home	Only shown when Auto Mode is on. Home quests get --dangerously-skip-permissions because the classifier escalates constantly outside a repo
Start in plan mode (per quest)	--permission-mode plan

Supervisors always run with --dangerously-skip-permissions.

Models and effort

Open the collapsible Model & effort section in the Create Quest, Edit Quest or Recurring Quest dialog.

Model is a searchable combobox. It groups Recently used models (the last five per agent), Presets, and for OpenCode a Discovered list fetched from opencode models. You can also type any model id; Sidequest passes it through as-is and warns that it is not a known preset.

Agent	Presets
Claude Code	Fable, Opus, Sonnet, Haiku
Copilot CLI	GPT-5.4, Claude Sonnet 4.6, Claude Sonnet 5
Gemini CLI, OpenCode	none — type a model id (OpenCode lists discovered models)
Codex CLI	model selection is not available yet

Auto is the default. For Claude Code with smart routing on it reads Auto (smart router); otherwise Auto (agent default), which simply omits the model flag.

Effort applies to Claude Code only and is passed as --effort:

• The global default is Settings > Agent > Effort Level (Low, Medium, High, Extra High, Max). The default is Extra High.
• The per-quest Effort select offers Auto (settings default), Low, Medium, High, XHigh, Max.
• Precedence: quest override > router decision > global setting.

Changing a quest’s model or effort later

Open Edit on the quest and change the values. Effort applies the next time the session starts (including Restart Session). A new model applies only to a fresh session: a resumed Claude Code conversation keeps the model it started with.

Smart model router

With smart routing on (the default), a Claude Code quest whose model is Auto is routed before its first session starts. Sidequest runs a quick, non-interactive claude --print call with the quest’s type, plan-mode flag, title and description, and the routing rules as the system prompt. It returns a model (opus or sonnet) and an effort level, and the quest launches with both. The quest card shows Routing model… during this step.

graph LR
A[Quest: title + prompt] --> B[Router: claude --print, Sonnet]
B -->|JSON model + effort| C[Launch quest session]
B -->|timeout or error| D[Fallback model, medium effort]
D --> C

• The router itself always runs on Sonnet at low effort, so routing costs very little.
• If routing fails or takes more than 20 seconds, the quest launches on the Fallback model (Settings > Routing; Haiku, Sonnet or Opus, default Sonnet) at medium effort.
• Picking a model manually bypasses the router entirely and uses the global effort setting unless the quest overrides it.
• Routing happens only when a new session is created, not when a quest resumes.
• Supervisors and recurring quests never route; their Auto means the agent default.

The same rules file also decides which model Sidequest’s own background jobs use — handoff summaries, journal entries, Council briefings and project analysis. See Smart Router Rules to read or edit the rules.

Turn routing off with Settings > Routing > Enable smart routing. Quests then launch with no model flag, i.e. your Claude Code default.

Claude account profiles

A profile is a separate Claude Code configuration directory (CLAUDE_CONFIG_DIR) under ~/.sidequest/profiles/<name>/. Each profile holds its own credentials, settings and conversation history, so you can run some quests on a work account and others on a personal one. Skills, plugins, commands, agents and CLAUDE.md are shared with your main ~/.claude by symlink.

Adding a profile

1. Open Settings > Agent > Claude Profiles
2. Type a name in Profile name and click Add profile
3. A Log in dialog opens with an embedded terminal running claude auth login for that directory. Finish the login flow there and click Done

Each profile row shows the signed-in email and plan (or Not signed in) with Log in / Switch account, Refresh and Delete buttons. Deleting a profile removes its directory, including that account’s conversation history; quests using it fall back to the default account.

Using profiles

• Per quest: once at least one profile exists, the Create Quest dialog shows a Claude account select with Default account (your ambient ~/.claude) and each profile. The account is fixed once the quest has been activated, because its conversation lives in that directory.
• Per project: Project Settings > Default Claude account sets the account new quests in that project start on. Recurring quests use it too.
• The terminal header shows the profile name while a quest runs on one, and the Usage Dashboard includes every profile’s transcripts.