Settings Reference¶

All settings use the sidekick.* prefix. Open VS Code Settings (Ctrl+,) and search for "sidekick".

Provider¶

Accounts¶

When the threshold is greater than 0, the extension host watches multi-provider quota and switches to another saved account for the active provider once utilization crosses the threshold. Account sign-in and switching are available from the status bar menu and the Command Palette (Sidekick: Add Account (Sign In), Sidekick: Switch Account (All Providers)). See Account Management.

z.ai (OpenCode routing)¶

When OpenCode is configured with a z.ai Coding Plan (GLM), Sidekick reads authoritative 5-hour and weekly quota percentages from z.ai's quota endpoint using the z.ai token stored by OpenCode, with fallback support for the official plugin's ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN environment variables.

See OpenCode → z.ai Coding Plan quota for details and limitations.

Model Selection¶

All model settings accept: auto (recommended), a tier (fast/balanced/powerful), a legacy name (haiku/sonnet/opus), or a full model ID.

See Model Resolution for details on how tiers map to models.

Inline Completions¶

Session Monitoring¶

sidekick.enableSessionMonitoring is read once at activation, so changing it prompts you to reload the window. While it is false, the Agent Hub container keeps only the Session Analytics view — which shows a placeholder with an Enable session monitoring button — and hides Mind Map, Kanban Board, Plans, Project Timeline, Latest Files Touched, Knowledge Notes, Subagents, and Event Stream.

Commit Messages¶

Explanations¶

Peak Hours¶

Only applies when the inference provider is Claude Max and the session provider is Claude Code. See Peak Hours for background.

Notifications¶

sidekick.inlineTimeout is deprecated and will be removed in a future release. Use sidekick.timeouts.inlineCompletion instead.

Timeouts¶

Event Logging¶

Pricing & Cost Tracking¶

The catalog is cached at ~/.config/sidekick/pricing-catalog.json with a 3s fetch timeout and stale-cache fallback — if the network is down, the last good cache is used; if there is no cache, the static table ships as a fallback. Unknown models return null cost and render as — (yellow in the CLI; in the VS Code dashboard as — per row with a footer warning, and a * appended to the total when priced and unpriced rows are mixed).

Context window sizes¶

The same catalog supplies context window sizes — it is literally named model_prices_and_context_window.json — so a newly released model sizes its context gauge correctly without waiting for an update.

Lookup is layered, most-trusted first: Claude Code's [1m] marker, then a window a provider actually reported, then the catalog, then the built-in table. Within that order an exact match in any layer beats a prefix match in every layer, so a known model is never sized from a catalog key that merely happens to be a prefix of its ID.

Windows that a provider reports for itself are remembered across runs in ~/.config/sidekick/observed-context-windows.json. Codex reports one on every token count, and it reflects the window your account tier actually gets — often well below the model's published maximum — which is why it outranks the catalog. That file is written locally, never fetched, and needs no setting of its own.

Session Handoff¶