CrewAI Integration¶

AevumCrewHooks and AevumTaskCallback integrate Aevum governance into CrewAI's task callback system. Every task execution is Cedar-evaluated and recorded in the sigchain. Consequential, irreversible tasks can be gated by a GOVERN checkpoint.

What it captures¶

Based on packages/aevum-core/src/aevum/core/adapters/crewai.py:

Event	Cedar check	GOVERN check	Sigchain record
`before_task()`	Yes — raises `PermissionError` on deny	Yes, if `consequential=True` and `reversible=False`	Context dict (in-memory)
`after_task()`	No	No	Yes (if kernel): `agent_role`, `success`, `output_hash`
`AevumTaskCallback.__call__()`	No	No	Yes via `after_task`

Cedar resource ID format: crewai:<task_description[:50]>.

GOVERN path: if consequential=True and reversible=False, before_task runs a GovernCheckpoint. If the checkpoint is vetoed, PermissionError is raised and the task does not execute.

Installation¶

pip install "aevum-core[crewai]"

Minimum version: crewai>=0.80.0. The callback and hook API changed significantly in 0.80.0. Check installed version:

python3 -c "import crewai; print(crewai.__version__)"

Minimal example¶

from aevum.core.adapters.crewai import AevumCrewHooks, AevumTaskCallback

# kernel=None: Cedar evaluation only, no sigchain writes.
# Pass an aevum.core.Engine instance to enable full recording.
hooks = AevumCrewHooks(kernel=None)

# Attach to a CrewAI Task via callback:
# from crewai import Crew, Task, Agent
#
# task = Task(
#     description="Summarize the quarterly report",
#     callback=AevumTaskCallback(hooks=hooks, consequential=False),
#     agent=...,
# )
# crew = Crew(agents=[...], tasks=[task])
# crew.kickoff()

# For consequential, irreversible tasks — triggers GOVERN checkpoint:
# task = Task(
#     description="Send billing email to all customers",
#     callback=AevumTaskCallback(hooks=hooks, consequential=True, reversible=False),
#     agent=...,
# )

# Manual test — before_task:
ctx = hooks.before_task(
    task_description="Summarize the quarterly report",
    agent_role="analyst",
    consequential=False,
    reversible=True,
)
print(ctx["cedar_permitted"])   # True
print(ctx["consequential"])     # False

# Manual test — after_task:
hooks.after_task(ctx, task_output="Summary: ...", success=True)

AevumTaskCallback¶

AevumTaskCallback wraps AevumCrewHooks in the callable interface that CrewAI passes the task output to after completion:

AevumTaskCallback(
    hooks=hooks,         # AevumCrewHooks instance
    consequential=False, # whether this task has consequential effects
    reversible=True,     # whether this task is reversible
)

CrewAI calls callback(output) after the task completes. AevumTaskCallback.__call__ calls hooks.after_task() and returns the output unchanged.

before_task must be called separately before handing control to CrewAI if you want Cedar pre-evaluation and optional GOVERN gating.

crewai>=0.80.0 requirement¶

CrewAI's callback mechanism was redesigned in 0.80.0 to accept callables directly on Task(callback=...). Earlier versions used a different hook attachment pattern. If you see TypeError or AttributeError on task callback registration, verify the installed version.

GOVERN checkpoint behavior¶

When consequential=True and reversible=False, before_task constructs a GovernCheckpoint with session_id=f"crew-{agent_role}" and calls checkpoint(action). The review_callback is None by default — implement and pass a callback to route the review to a human approval interface.