Compute cost from stream usage using model-specific pricing rates. Falls back to Claude-like pricing ($3/$15 per million) if no rates provided.

Compute diffs from message step snapshots. Scans messages for step-start and step-finish parts to find the earliest "from" snapshot and the latest "t

Create a child session

Create a ProviderRegistry using the given bootstrap config and auth store. This is the primary entry point when you have a loaded application config w

Create a user message and a synthetic text part via the ToolContext callbacks, matching the TS `Session.updateMessage()` + `Session.updatePart()` patt

Calculate the retry delay in milliseconds. Priority: 1. `retry-after-ms` header (milliseconds) 2. `retry-after` header (seconds as float, or HTTP-dat

Derive a name from a file path relative to the base directory.

Compute diff summary from messages

Execute a command and return the rendered template

Proxy an HTTP request through the plugin's custom fetch. Only valid when `has_custom_fetch()` is true (after a successful `load()`).

Transform a models.dev model into a runtime ProviderModel.

Simple fuzzy string matching: returns up to `max` candidates from `options` that share a common substring with `query`, sorted by edit-distance-like s

Get the status for a session. Returns Idle if not tracked. Matches TS `SessionStatus.get(sessionID)`.

Get the custom loader for a provider ID.

Return a valid access token, refreshing if necessary. Returns `None` if no tokens are stored (the caller should initiate the auth flow).

Aggregate usage across all assistant messages in the session

Walk up from `start` to `stop`, expanding a glob pattern at each level.

Get the global plugin system, creating a default one if not initialized.

Whether the plugin provides a custom fetch that must proxy LLM requests.

Apply semantic highlighting to rendered lines. Recognizes: shell commands, file paths, error keywords, numeric stats.