mirror of
https://github.com/docker/docker-agent.git
synced 2026-11-04 10:02:15 +00:00
Sayt-0
a3dc1a8ed6
When an agent's config names a force_handoff target, the runtime intercepts the agent's natural stop and routes the conversation to that agent deterministically, bypassing the LLM's tool-calling. This makes strict pipelines (extractor -> summarizer) reliable even when the model would forget to call the handoff tool. - config: new force_handoff field on agents (latest schema only), validated for existence, self-reference, and cycles - runtime: forced switch on stop with an implicit handoff notice message; skipped for pinned sessions to avoid infinite loops - hooks: on_agent_switch now reports kind "force_handoff"; turn_end reports reason "force_handoff" - docs, JSON schema, and examples/force_handoff.yaml updated Closes #3078
31 lines
1.2 KiB
YAML
31 lines
1.2 KiB
YAML
#!/usr/bin/env cagent run
|
|
# Deterministic pipeline using force_handoff:
|
|
#
|
|
# extractor ──(always)──> summarizer
|
|
#
|
|
# Unlike `handoffs` (where the LLM decides whether to call the handoff
|
|
# tool), `force_handoff` is enforced by the runtime: whenever the agent
|
|
# produces a final response, the conversation is unconditionally routed
|
|
# to the named agent, carrying over the full context. This guarantees
|
|
# the pipeline order even when the model would otherwise "forget" to
|
|
# hand off.
|
|
|
|
agents:
|
|
root:
|
|
model: anthropic/claude-sonnet-4-5
|
|
description: Extracts the key facts and entities from the user's input
|
|
instruction: |
|
|
You are the first stage of a two-stage pipeline.
|
|
Extract the key facts, entities, and figures from the user's input
|
|
as a concise bullet list. Do not summarize or editorialize - the
|
|
next agent in the pipeline will do that.
|
|
force_handoff: summarizer
|
|
|
|
summarizer:
|
|
model: anthropic/claude-sonnet-4-5
|
|
description: Produces the final summary from the extracted facts
|
|
instruction: |
|
|
You are the final stage of a two-stage pipeline.
|
|
Using the extracted facts from the conversation history, write a
|
|
short, well-structured summary for the user.
|