Check whether the next action is allowed under the configured limits. Call this BEFORE making an LLM call.

Split text into chunks by paragraph boundaries, merging small paragraphs to stay near `target_size` characters per chunk. Tracks line numbers.

Classify HTTP errors from model providers into actionable error messages. Detects context overflow, billing issues, rate limits, and auth failures.

Compute cooldown duration using exponential backoff. Auth/billing errors use longer base cooldowns than transient errors.

Execute a shell command with timeout and output capture.

Register a handler for a general event type. The handler will fire for all events of this type regardless of action.

Repair tool_use / tool_result pairing after message pruning. If an Assistant message contains `[tool_call:...]` markers but the corresponding Tool re

Validate and resolve a requested path within the workspace. - Rejects absolute paths - Rejects `..` components - Resolves to an absolute path inside

Calculate the available token budget for input context given a model definition.

Remove all handlers for a given event key.

Mark the most recently used key as having failed. Applies exponential backoff cooldown.

Parse a single NDJSON line into a `JsonRpcRequest`.

Pop a specific style from the stack (not necessarily the top, for robustness).

Build the standard HTTP client used by channel adapters.

Split text into chunks of at most `limit` characters, preferring newline boundaries.

Complete a run and notify the parent.

Derive a context-specific subkey using HMAC-SHA256. Each context string (e.g., "session", "config", "media") produces a unique 256-bit subkey from th

Scan plugin directories for `plugin.json` manifests. Directories are scanned in order; first match for a given plugin id wins. Each entry in `dirs` i

Estimate the token count for a string using a chars/token heuristic.

Extract auth credential from HTTP headers.

Extract a URL from a line of text, starting from the given prefix.

Scan an async reader for a line containing a URL-like prefix. Returns the extracted URL (from the prefix to the next whitespace).

Fire a hook event asynchronously (fire-and-forget).

Look up a model by ID and return its known metadata. Returns `None` for unknown models (caller should use defaults).

Process a batch of media attachments through available providers. Returns understanding outputs for each attachment that has a matching provider. Att

Read recognized workspace bootstrap files and return them as a combined string. Files that don't exist are silently skipped. Content is sanitized befo

Record a tool call. Returns an error if the repeat threshold is exceeded.

Record a failed authentication attempt.

Record a successful call.

Returns the workspace path, or an error if no workspace is configured.

Compute the backoff delay for a given attempt number. Uses exponential backoff (1s → 2s → 4s → 8s …) with ±25% jitter. Minimum delay is 100ms. - Att

Remove top-level keys with null values from a JSON object. LLMs often emit `null` for optional parameters, but many MCP servers reject explicit nulls

Serialize this config to a pretty-printed TOML string.

Record token usage.

Build a deterministic cache key from a completion request. Hashes the model, messages, max_tokens, temperature, and system prompt via SHA-256. Two re

Wrap `text` in SGR escape sequences for the given color code.

Dispatch an RPC method to the appropriate handler.

Retrieve pending outbound messages for one web recipient.

Enrich a model definition with catalog metadata if available. Preserves the original ID but fills in context window, costs, etc.