Pipeline FSM core #17

New issue

Open

opened 2026-04-07 20:23:45 +00:00 by austin · 0 comments

austin commented

2026-04-07 20:23:45 +00:00

Owner

Implement the finite state machine that drives task execution.

Define states: UNDERSTAND, LOCATE, PLAN, GENERATE, VERIFY, COMPLETE
Each state is a callable that receives the current pipeline context and returns the next state + updated context
State transitions are deterministic from the harness's perspective — the model's output informs but doesn't control transitions
Pipeline context carries: user intent, located symbols, generated candidates, verification results
Support early exit (e.g., UNDERSTAND determines intent is "explain" — skip GENERATE, go straight to a model explanation call)
Timeout/retry limits per state to prevent infinite loops
Logging of state transitions for debugging

This is the central orchestrator. It replaces the "model decides what to do next" agent loop with a predictable, harness-driven workflow.

Implement the finite state machine that drives task execution. - Define states: UNDERSTAND, LOCATE, PLAN, GENERATE, VERIFY, COMPLETE - Each state is a callable that receives the current pipeline context and returns the next state + updated context - State transitions are deterministic from the harness's perspective — the model's output informs but doesn't control transitions - Pipeline context carries: user intent, located symbols, generated candidates, verification results - Support early exit (e.g., UNDERSTAND determines intent is "explain" — skip GENERATE, go straight to a model explanation call) - Timeout/retry limits per state to prevent infinite loops - Logging of state transitions for debugging This is the central orchestrator. It replaces the "model decides what to do next" agent loop with a predictable, harness-driven workflow.