chap-ag2

Adapter between AG2 (AutoGen) and the CHAP Coordinator. At an AG2 human-input turn, the human's decision -- approve, edit, or reject -- becomes a hash-linked, replayable CHAP audit entry.

decision      CHAP envelope
--------      ---------------------------
approve       decide.approve
override      decide.override   (message vs reply, as a diff)
reject        decide.reject

The agent message the human is responding to is the artefact under review; the human's reply is the decision on it.

Install

pip install chap-ag2

Depends on chap-coordinator>=0.3.3. AG2 is optional: the adapter reads the message and reply as plain values, so the bridge and its tests work without it installed. Install the extra to run a live conversation:

pip install "chap-ag2[ag2]"

The decision is explicit

AG2's human turn is a weak signal: the same get_human_input loop carries plain dialogue, edits, approvals, and "exit". Inferring intent would write decisions the human never made, which is the one thing this record exists to avoid. So record_turn takes the decision explicitly, and makes only one inference:

• an empty reply means "use the agent's output" -> decide.approve
• any non-empty reply with no explicit decision records nothing -- ending a chat is not a rejection, and dialogue is not an edit

bridge.record_turn(message, reply, decision="override",
                   rationale="over the limit", tags=["capped"])

intent_preserved defaults to true on an override; set it false for a substituting edit -- a different decision, not a refinement.

Where it hooks

The message under review and the reply meet inside get_human_input, so that is where a turn is recorded. self.last_message() gives the message; the return value is the reply:

from autogen import UserProxyAgent
from chap_ag2 import ChapTurnBridge

bridge = ChapTurnBridge(Coordinator(), workspace="wsp_support",
                        agent="agent:assistant#v1", reviewer="human:alice@example.org")

class RecordingUser(UserProxyAgent):
    def get_human_input(self, prompt, **kw):
        reply, decision = capture_from_ui(prompt)   # your UI supplies the intent
        bridge.record_turn(self.last_message(), reply, decision=decision)
        return reply

Approver identity

CHAP has no ambient actor: the decider is whatever from the envelope carries. The bridge uses its reviewer by default; pass a per-turn approver (a human: URI) to override it. The participant type is taken from the URI scheme and the approver is joined before recording. Each turn is its own task whose review is addressed to that approver, so the record satisfies the Coordinator's authorisation rules.

What you get in the audit chain

One reviewed message with an edit yields the following, after the workspace and the three joins at seq 0 to 3:

seq=4  task.create     agent:assistant#v1
seq=5  task.complete   agent:assistant#v1
seq=6  review.request  agent:assistant#v1   to=human:sam@example.org
seq=7  decide.override human:sam@example.org  diff=[{op:replace, path:, value:"refund $50 to Alice"}]

Every entry carries prev_hash, so the chain verifies externally or anchors to a SCITT transparency service with the audit-scitt/1.0 profile.

Example

examples/01-approve-edit-reject.py runs a real AG2 conversation (no LLM needed) through approve, an edit, and a reject, and prints the resulting chain.

Compatibility

• chap-coordinator 0.3.3
• ag2 0.9+ (optional; verified against 0.14)
• Python 3.10, 3.11, 3.12, 3.13

License

Apache 2.0. See LICENSE.