Pipecat voice agent integration

Voice agents built with Pipecat have constraints text agents do not:

• Latency. Anything injected into a spoken turn must be short or the agent sounds robotic.
• Slot repair. When a user corrects a misheard value ("No, I said Tuesday, not Thursday"), the agent must repair that one slot cleanly.
• Post-call mining. Voice sessions produce transcription traces that can be mined for lessons after the call ends.

lessonweaver only produces reviewed guidance; it does not run the voice pipeline and is never a Pipecat dependency.

Pre-session skill loading

Load short skills into the system context before the session starts. Keep them terse to protect the latency budget by using a name-only or summary inclusion level and a small character budget:

from lessonweaver import FileSystemRegistry, SkillLoader

loader = SkillLoader(registry=FileSystemRegistry())
context = loader.load_for_task(
    task="Book a service appointment from spoken input",
    agent_type="voice",
    budget_chars=400,
    inclusion_level="summary",   # or "name_only" for the tightest budget
)
system_context = context.snippet

Per-turn hints

If you must add guidance mid-session, inject at most a one-sentence reminder at the start of a turn and keep budget_chars around 100 so spoken latency stays low.

Post-call trace mining

After the call ends, record the session transcript as a TraceBundle and run detection on it:

lessonweaver detect examples/traces/voice_slot_correction.json

examples/traces/voice_slot_correction.json is a concrete voice trace: the agent mishears both the time and the location, the user corrects it, and detection produces one human_correction candidate.

Conversational repair skill

examples/skills/voice_slot_repair.md is an example reviewed skill for slot correction. It repairs only the corrected slot and confirms in a single short sentence to respect the latency budget.

What lessonweaver does not do

• It does not run the Pipecat pipeline, TTS, or STT.
• It does not auto-inject unreviewed skills; only reviewed skills are loaded, and you choose the inclusion level and budget.