chap-langgraph

Bridge between LangGraph and the CHAP Coordinator. Wraps the human-in-the-loop interrupt boundary so every review decision becomes a hash-linked, replayable audit entry, without changing the inside of your existing graph nodes.

LangGraph event              CHAP envelope
-----------------            -------------------------
interrupt() called           task.complete + review.request
Command(resume=...) sent     decide.approve | decide.reject | decide.override

The bridge adds three lines to your existing LangGraph node. The override path also turns reviewer edits into JSON Patches automatically, so the audit chain captures what changed and why, not just approved/ rejected.

Install

pip install chap-langgraph

Depends on chap-coordinator>=0.3.3. LangGraph itself is optional: the bridge accepts any dict-shaped state and any opaque decision payload, so it works in environments without LangGraph and is trivial to unit-test.

Quick start

from chap_coordinator import Coordinator
from chap_langgraph import ChapBridge, hil_review

coord = Coordinator(default_profiles=["core/1.0", "review/1.0"])
bridge = ChapBridge(
    coord,
    workspace="wsp_drafter",
    agent="agent:drafter#v1",
    reviewer="human:alice@example.org",
)

# Inside your existing LangGraph node:
def drafter_node(state):
    draft = produce_draft(state)
    chap_state = hil_review(bridge, draft, kind="draft_response")
    return chap_state            # spread into LangGraph state, then interrupt()

# After the interrupt resumes with Command(resume=decision):
def resolve_review(state, decision):
    applied = bridge.apply_decision(state["chap_task_id"], decision)
    return {"draft": applied or state["chap_artefact"], **state}

The decision payload mirrors what a reviewer typically returns:

"approve"                                           # decide.approve
"reject"                                            # decide.reject
{"action": "reject", "rationale": "..."}            # decide.reject (rich)
{"diff": [...], "rationale": "...", "tags": [...]}  # decide.override

What you get in the audit chain

After one hil_review + apply_decision cycle, calling bridge.audit() returns the full chain:

seq=0  workspace.create
seq=1  participant.join     agent:drafter#v1
seq=2  participant.join     human:alice@example.org
seq=3  task.create          kind=draft_response
seq=4  task.complete        agent:drafter#v1
seq=5  review.request       to=human:alice@example.org
seq=6  decide.override      diff=[...] rationale="..." tags=[...]

Every entry carries prev_hash, so you can verify the chain externally or anchor it to a SCITT transparency service with the audit-scitt/1.0 profile.

Examples

See examples/ for runnable end-to-end demos:

• examples/01-basic-review.py: single drafter node, one human approval.
• examples/02-override-cycle.py: agent draft, human edits, audit chain.

Compatibility

Tested against:

• chap-coordinator 0.3.3
• langgraph 0.2.x (optional)
• Python 3.10, 3.11, 3.12, 3.13

License

Apache 2.0. See LICENSE.