v0.6.1 — AGT 4 Dependency Consolidation

v0.6.1 — AGT 4 Dependency Consolidation (hotfix)

Optional AGT extras now resolve to the consolidated Microsoft Agent Governance Toolkit 4.x packages, with public extra names preserved and the optional-only contract intact (core remains pyyaml).

• agent-os / agent-mesh / agent-hypervisor → agent-governance-toolkit-core>=4.0.0,<5.0
• agent-sre → agent-governance-toolkit-cli>=4.0.0,<5.0
• full / dev → union of both consolidated packages

No public API, endpoint, or payload changes. SRE REST surface (/sre/snapshot, /sre/events, /sre/fleet) unchanged.

Governance: plan docs/plan-qor-phase6-agt-4-dependency-consolidation.md; META_LEDGER Entry #9 (SESSION SEAL); Merkle seal a476775ece21fd833df7bc519a3861f346465a46f88a7e411b7f155f6db0d16f.

v0.6.0 - SRE v2 Fleet Health

What's New

SRE v2 Fleet Health Dashboard

Full fleet health visibility with per-agent metrics, circuit breaker states, and trust stage progression.

New Endpoints

• GET /sre/snapshot - Expanded with v2 fields: slis, auditEvents, fleet
• GET /sre/events - Recent governance audit events with limit parameter
• GET /sre/fleet - Per-agent health status with circuit breaker state

New Module

• agent_metrics.py - AgentMetricsRegistry for per-agent operational metrics
  • Circuit breaker: closed → half-open → open
  • Trust stage derivation: CBT → KBT → IBT based on success rate
  • Thread-safe with configurable thresholds

New Types

• CircuitBreakerConfig - Configurable circuit breaker thresholds
• TrustDimension / TrustScoreV2 - Multi-dimensional trust scoring
• AuditEvent - Governance event for SRE panel
• FleetAgent - Agent health snapshot
• SliMetric - Standard 7-SLI dashboard metrics

Enhanced

• FailSafeComplianceSLI.get_slis() - Returns 7 SliMetric objects
• FailSafeAuditSink.get_recent_events() - Query recent audit events
• DecisionCallback now includes latency_ms parameter
• integration.py wires AgentMetricsRegistry into _on_decision callback

Test Coverage

• 357 tests passing
• 46 new tests for SRE v2 components

---

🤖 Generated with Claude Code

v0.4.0 — Security & Performance Hardening

Breaking Changes

• HMAC key now required — FailSafeAuditSink(hmac_key=...) no longer defaults to a dev key. Pass hmac_key=b"failsafe-dev-key" explicitly for development or set a production key.

Security Hardening

• DecisionRequest.__post_init__ validates action, agent_did, and normalizes artifact_path via os.path.normpath
• Exception messages sanitized — no str(exc) leak in fail-open responses
• Artifact paths in webhook events reduced to os.path.basename only
• SLI signal reasons truncated to 200 chars max
• Escalation queue eviction now runs under lock

Performance

• Persistent SQLite connections with WAL mode and busy_timeout=5000 in both LocalFailSafeClient and FailSafeAuditSink
• LocalFailSafeClient.close() method for explicit connection cleanup
• MCP client subprocess cleanup on handshake failure

Reliability

• fail_open: bool = True parameter on FailSafeInterceptor and GovernancePipeline
• Fail-closed mode raises on governance errors instead of silently allowing
• Trust score diminishing returns: 0.8 ** consecutive_successes decay prevents score inflation

Stats

• 18 source files, 3,387 lines
• 305 tests passing, clean lint

Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

v0.3.0: Shadow Genome, Heuristic Patterns, Trust Dynamics

What's New

Shadow Genome (Failure DNA)

• ShadowGenomeStore Protocol + InMemoryShadowGenomeStore with thread-safe FIFO eviction
• 10-variant failure classifier (classify_failure_mode)
• Negative constraint generation for learning injection (generate_negative_constraint)
• Per-agent constraint retrieval (get_constraints_for_agent)

CWE-Referenced Heuristic Patterns

• 10 default patterns: SQL injection (CWE-89), command injection (CWE-78), hardcoded secrets (CWE-798), PII exposure (CWE-359), deep nesting (CWE-1121), weak hash (CWE-328), basic auth (CWE-522), pinned deps (CWE-1104)
• match_content() scanner with severity-sorted results
• classify_risk() with composable extra_l3_triggers parameter

CBT/KBT/IBT Trust Dynamics

• TrustConfig frozen dataclass with 9 tuning fields
• Pure functions: determine_stage, apply_outcome, is_probationary, calculate_influence_weight, score_to_mesh_trust
• FailSafeTrustMapper.update_trust() and get_trust_stage() integration methods

Stats

• 15 source files, 2587 lines
• 174 tests passing
• Zero new external dependencies

Full Changelog: v0.2.0...v0.3.0