pi-retry-model

Pi Agent extension to auto-retry model execution when an empty response or thinking-only response occurs

Package details

extension

Install pi-retry-model from npm and Pi will load the resources declared by the package manifest.

$ pi install npm:pi-retry-model

Pi manifest JSON

{
  "extensions": [
    "./src/index.ts"
  ]
}

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

An extension for Pi Coding Agent that automatically retries execution when a model returns an incomplete or empty response.

Empty Assistant Response Handling: Detects when an assistant message yields no output text and no tool calls.
Thinking-Only Response Recovery: Triggers auto-retry when a model emits reasoning/thinking blocks but stops before producing output text.
Tool Result Only Recovery: Handles edge cases where execution halts right after tool execution without an assistant follow-up.
Aborted Run Awareness: Automatically skips retries when a run is manually cancelled (e.g., via ESC key).
Max Retry Threshold: Caps retries per occurrence to avoid infinite retry loops (default: 3). Configurable via command.

/retry-model-config [number]: View or update the maximum retry count. Accepts any valid non-negative integer (e.g., /retry-model-config 5). Set to 0 to disable auto-retries.

Install directly via Pi:

pi install npm:pi-retry-model

Or test temporarily in a single session:

pi -e npm:pi-retry-model

The extension automatically creates and reads its configuration from ~/.pi/agent/retry-model-config.json:

{
  "maxRetries": 3
}

maxRetries: Maximum number of auto-retry attempts per empty response occurrence (default: 3, set to 0 to disable).

Listens to the agent_settled event when an agent run finishes.
Inspects the final message of the active branch.
Evaluates if the response was truncated, empty, or halted prematurely.
If an empty response is detected (and retries remain), it displays a UI warning notification and automatically dispatches a follow-up message to prompt the model to continue generation.

Pi provides two lifecycle events when an agent run completes:

agent_end: Emitted immediately when a single low-level agent loop finishes. At this point, Pi core may still perform automatic retries, context compaction, or process queued follow-up messages.
agent_settled: Emitted only when the agent run is completely idle and no automatic retries, compaction, or follow-ups remain.

This extension uses agent_settled to:

Prevent race conditions: Ensures Pi core finish any built-in compaction or retry tasks before triggering a new turn.
Avoid duplicate retries: Guarantees that retries only run when Pi has genuinely stopped and will not automatically continue on its own.