Signed decision evidence for Haystack Agents.

Install the native Haystack integration.

pip install loopgrid-haystack

Version 0.1.0 is public on PyPI. The package targets haystack-ai 3.3.x, depends on loopgrid 0.8.x, and supports Python 3.10 or newer.

Bind one consequential decision to one Haystack Agent run.

from haystack.components.agents import Agent
from haystack.dataclasses import ChatMessage
from loopgrid_haystack import LoopGridHaystack

loopgrid = LoopGridHaystack(
    base_url="http://127.0.0.1:8000",
    workspace_id="default",
    agent_id="support-agent",
)
loopgrid.enable_tracing()

decision = loopgrid.start_decision(
    decision_type="customer_refund",
    agent={"id": "support-agent", "version": "1"},
    authority={"acting_for": "Example Store", "scope": ["refund:create"], "limit_usd": 100},
    model={"provider": "openai", "name": "gpt-5"},
    context={"prompt_version": "support-v1"},
    proposed_action={"tool": "refund.create", "amount": 25, "currency": "USD"},
    policy={"policy_id": "refund-policy", "version": "1", "decision": "auto_allowed"},
)

agent = Agent(
    chat_generator=...,
    tools=[...],
    hooks=loopgrid.agent_hooks(decision["decision_id"]),
)

result = agent.run(messages=[ChatMessage.from_user("Handle this duplicate charge.")])
loopgrid.flush()
loopgrid.assert_healthy()

# Only after the application observes the authoritative downstream result:
loopgrid.record_outcome(
    decision["decision_id"],
    {"status": "succeeded", "external_reference": "refund_123"},
    observer="billing-webhook",
)

Decision correlation is explicit and request-scoped. Constructing LoopGridHaystack does not change Haystack tracing automatically; enable_tracing() is an explicit application choice.

Use Haystack tracing for model evidence and Agent hooks for the tool boundary.

Keep Haystack ConfirmationHook native — and record only the call that survives it.

ConfirmationHook / other before_tool control
  → reject: no tool_requested evidence
  → modify: surviving arguments continue
  → approve: surviving call continues

application authenticates reviewer
  → loopgrid.record_human_review(...)
  → human_approved

LoopGrid before_tool
  → tool_requested
  → actual Haystack tool executes

LoopGrid after_tool
  → tool_executed

application observes downstream result
  → loopgrid.record_outcome(...)
  → outcome_observed
  → evidence_complete
  → verify valid:true

Pass confirmation/modification hooks through before_tool_prefix. LoopGrid appends its request hook after them, so a rejected call cannot become tool_requested evidence and a modified call is recorded with the final parameters that survive confirmation. Reviewer identity is never inferred merely because a tool executed.

Commit to content without storing it by default.

Framework callbacks queue transport work so evidence delivery does not raise into the Haystack Agent execution path. Call flush() and assert_healthy() when delivery must be confirmed. Explicit human-review and outcome writes drain earlier queued evidence first, preserving lifecycle order under slower transport. If another Haystack tracing backend is required, pass a concrete tracer instance explicitly as the delegate; do not feed the process-level tracing facade back into LoopGrid.

Validated through the real Haystack Agent runtime, native confirmation paths and LoopGrid Core.

The release E2Es use deterministic Haystack test components and sandbox-only refund functions, so no real-money movement is required. Verification establishes the integrity/provenance of captured evidence; it does not prove that the underlying decision was correct, that every upstream statement was true, that capture was complete, or that the system is legally compliant.