chap-llama-index

Adapter between LlamaIndex Workflows and the CHAP Coordinator. When a workflow step pauses for human input, 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   (proposed vs returned, as a diff)
reject        decide.reject

The proposed output on the InputRequiredEvent is the artefact under review; the human's HumanResponseEvent is the decision on it. An edit is recorded as an RFC 6902 diff, so the chain captures what changed and why, not just approved/rejected.

Install

pip install chap-llama-index

Depends on chap-coordinator>=0.2.9. LlamaIndex is optional: the adapter reads events structurally, so the bridge and its tests work without it installed. Install the extra to run a live workflow:

pip install "chap-llama-index[llama-index]"

Quick start

Workflows surface human-in-the-loop as events: a step returns an InputRequiredEvent and waits; a second step consumes the matching HumanResponseEvent. The driver that answers the pause is where both events meet, so that is where the decision is recorded:

from workflows.events import InputRequiredEvent, HumanResponseEvent
from chap_coordinator import Coordinator
from chap_llama_index import ChapHitlBridge

bridge = ChapHitlBridge(
    Coordinator(),
    workspace="wsp_payments",
    agent="agent:writer#v1",
    reviewer="human:alice@example.org",
)

handler = workflow.run()
async for event in handler.stream_events():
    if isinstance(event, InputRequiredEvent):
        response = HumanResponseEvent(
            response={"amount": 50, "to": "acct-9"},   # the edited output
            decision="override",
            user_name="human:sam@example.org",
            rationale="over the desk limit; capped to 50",
            tags=["limit-exceeded"],
        )
        bridge.record_decision(event, response, decision=response.get("decision"))
        handler.ctx.send_event(response)
await handler

Decision is explicit

Workflows events are schemaless and carry no approve/edit/reject signal, so decision is a required argument -- the adapter never guesses intent from the response text. proposed, returned, rationale, tags, intent_preserved, and approver default to fields read off the events (event.get(...)) and can each be overridden per call:

bridge.record_decision(input_event, response_event, decision="approve")

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

Approver identity

CHAP has no ambient actor: the decider is whatever from the envelope carries. The bridge uses its reviewer by default, but the HumanResponseEvent's user_name (or an explicit approver=) overrides it, and the adapter joins that approver -- with a participant type taken from the URI scheme -- before recording. Each decision 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 paused step with an edit yields:

seq=3  task.create     agent:writer#v1
seq=4  task.complete   agent:writer#v1
seq=5  review.request  agent:writer#v1   to=human:sam@example.org
seq=6  decide.override human:sam@example.org  diff=[{op:replace, path:/amount, value:50}]

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 drives one paused workflow through approve, a refining edit, a substituting edit, and a reject against real llama-index-workflows, and prints the resulting chain.

Compatibility

• chap-coordinator 0.2.9
• llama-index-workflows 1.x–2.x (optional; verified against 2.22)
• Python 3.10, 3.11, 3.12, 3.13

License

Apache 2.0. See LICENSE.