Skip to content

HAYSTACK-INTEGRATION-PROPOSAL.md

Latest commit

 

History

History
66 lines (49 loc) · 3.18 KB

HAYSTACK-INTEGRATION-PROPOSAL.md

File metadata and controls

66 lines (49 loc) · 3.18 KB

Proposal: Haystack Integration

Status: ✅ Shipped — Integration package implemented, upstream collaboration active
Author: Agent Governance Toolkit Team (Microsoft)
Created: 2026-03-08
Target: deepset-ai/haystack

Summary

Integration adapter connecting Haystack pipelines with AGT governance. Implements GovernancePolicyChecker, TrustGate, and AuditLogger as Haystack SuperComponents.

External Engagement

Implementation

  • Package: packages/agentmesh-integrations/haystack-agentmesh/
  • Module: src/haystack_agentmesh/

Components

Component File Description
GovernancePolicyChecker governance.py Enforces governance policies on agent actions — tool allowlist/blocklist, content pattern filtering (substring, regex, glob), token limit enforcement, per-agent rate limiting with time-based windows. Outputs decision (allow/deny/audit), reason, and passed flag.
TrustGate trust_gate.py Evaluates agent trust scores (0.0–1.0) and routes decisions based on configurable thresholds. Tracks success/failure with reward/penalty adjustments and time-based trust decay. Outputs trusted flag, score, and action (pass/review/block).
AuditLogger audit.py Append-only tamper-evident audit logging with SHA-256 hash chain integrity. Each entry hashes to the previous for immutability verification. Supports JSONL export for offline analysis. Outputs entry_id and chain_hash.

Internal Data Classes

Class File Description
AgentTrustRecord trust_gate.py Internal state container for per-agent trust data (score, successes, failures, last_update)
AuditEntry audit.py Immutable audit record container (entry_id, timestamp, action, agent_id, decision, metadata, prev_hash, chain_hash)

Graceful Fallback

All three components implement identical fallback logic when haystack-ai is not installed:

try:
    from haystack import component
except ImportError:  # pragma: no cover
    class _ComponentShim:
        def __call__(self, cls):
            return cls
        @staticmethod
        def input_types(**kwargs):
            def decorator(func): return func
            return decorator
        @staticmethod
        def output_types(**kwargs):
            def decorator(func): return func
            return decorator
    component = _ComponentShim()

This ensures components function as regular Python classes even without Haystack installed, while registering as proper SuperComponents when Haystack is available.

References