01
Start with a goal
A user, application, scheduled task, or chat channel submits a goal along with the active system instructions and session context.
// Pipeline
A simple agentic loop
The core loop stays deliberately legible: accept a goal, ask the model what to do, execute validated tools, feed observations back, and stream the result.
PromptReasonActObserve
02
Let the model decide
The selected provider returns text, tool calls, or both. PasClaw normalizes provider-specific responses into one execution flow.
03
Execute and observe
Validated tool calls run under the configured policy. Their structured results become observations for the next turn.
04
Finish or continue
The loop continues until the model produces a final result, reaches a configured limit, or is redirected by an operator.
From the documentation
Technical details
Implementation notes drawn from the PasClaw repository documentation.
Dispatch and agent loop
The executable routes commands into agent, gateway, and TUI modes. RunToolLoop then coordinates provider responses, tool execution, memory, and session persistence.
- Provider-specific payloads are normalized before entering the loop.
- Independent read-only tool calls can execute concurrently.
- Lifecycle hooks wrap turns, tool calls, results, and errors.
Context that survives
Sessions serialize after every turn and keep a compact working-state snapshot available after compaction or restart.
- Working state tracks the last eight edited files, last shell command, and latest tool error.
- A summarizer compacts long conversations at the configured threshold.
- Queued steering instructions are drained into a running loop between steps.
Source documentation
Common questions
Frequently asked questions
Keep exploring
See how the rest of PasClaw fits together.