Hive v0.6.1 — Security fix for rust_brain snapshot restore, PyPI distribution, and refreshed dependen...

[0.6.1] - 2026-06-29

Security

• rust_brain snapshot integrity: RustBrain.snapshot_to_file now embeds the
  SHA-256 of the node payload in the file, and restore_from_file verifies it
  before mutating state. Previously the checksum was computed and discarded, so
  a corrupted or tampered snapshot restored silently. A failed check now raises
  ValueError("snapshot checksum mismatch ...") and leaves the existing store
  untouched. Backward compatible: pre-checksum snapshots skip verification.

Added

• PyPI publishing in the release workflow (trusted publishing via GitHub OIDC).
• [full] optional extra: busybee-cpu + honey-comb pulled from PyPI.
• docs/PYPI.md with one-time publisher setup instructions.

Changed

• Dependency pins refreshed across core, dev, observability, and GPU extras.
• hive_api_server reads version from hive.__version__ instead of a hardcoded string.

Tests

• Replaced the catch-all corruption test with deterministic checks: content
  tamper → checksum ValueError, restore atomicity (existing data survives a
  failed restore), and truncated-file framing failure.

Hive v0.6.0 — Real-workload evaluation, API hardening, and a documentation/metadata pass that brings...

[0.6.0] - 2026-06-11

Added

• Real-workload SWE-bench-lite A/B evaluation (scripts/hive_swebench_eval.py) with
  committed baseline vs Hive runs under docs/benchmarks/swebench-lite/.
• Compression sensitivity sweep (scripts/hive_compression_sweep.py).
• Hybrid Logical Clock (HLC) for causal-memory ordering, plus rust_brain concurrency tests.
• Release workflow (.github/workflows/release.yml) and an evaluation-result issue template.
• CITATION.cff validation in CI.

Changed

• README rewritten to reflect the actual current state and module surface: real measured
  evaluation results, accurate public API, the full ~28-module architecture, and corrected
  reproduce/script paths.
• Routing accuracy framed as in-distribution with an explicit out-of-distribution caveat;
  removed the unmeasured ROI/case-study and energy headline claims.
• Packaging metadata aligned to the 0.6.0 release.

Fixed

• pyproject.toml: restored the missing [build-system] table and removed a stray
  top-level version key; the package version now reports 0.6.0 (was 0.5.0).
• Version drift: hive.__version__, the FastAPI server, and the Helm chart now report 0.6.0.
• README: balanced the code fences that previously swallowed the Online Learning / Causal
  Memory and Development sections; corrected the compression label set to
  CORE/DISTILL/COMPACT/DROP/STALE/ESCALATE and the public API examples to the real types.
• .github/citation.cff is now valid CFF 1.2.0 (was JSON); CI validates it as CFF via
  cffconvert instead of json.load.

Hive v0.5.0 — Enterprise-Grade Infrastructure

Enterprise-Grade Infrastructure (Critical → High → Medium → Low)

Critical Tier

• Schema validation (hive.schemas): Pydantic models for AgentState, RouteDecisionOut, MemoryNodeIn with validate_state() / validate_memory(). Enabled via HiveStack(validate=True); backward-compatible default validate=False.
• Multi-tenancy (hive.rust_brain): RustBrain(tenant_id=..., tenant_isolation=True) prefixes internal storage keys per tenant. Cross-tenant reads return None.
• Health / readiness probes (hive.health): HealthServer with /health (liveness, always 200) and /ready (readiness, 200/503). is_healthy() synchronous check for load balancer integration.
• Encryption at rest (hive.encryption): AES-256-GCM transparent encryption for RustBrain values. PBKDF2 key derivation from HIVE_ENCRYPTION_KEY. Disabled by default (backward compatible).
• Production IdP integration (hive.auth): JWT validation with JWKS support + role-based access control (RBAC). JWTValidator.from_env() reads HIVE_JWKS_URL, HIVE_JWT_ISSUER, HIVE_JWT_AUDIENCE.
• Backup / disaster recovery (hive.rust_brain): snapshot_to_file() writes gzip-compressed JSON with SHA-256 checksum. restore_from_file() validates version and rebuilds store.
• Automated offboarding (hive.rust_brain): revoke_tenant() wipes all tenant data for GDPR Article 17 compliance.

High Tier

• Rate limiting (hive.ratelimit): Token-bucket per (tenant_id, operation). HiveStack(rate_limiter=...) returns source="ratelimit" escalation when bucket empty. Backward-compatible default rate_limiter=None.
• Config management (hive.config): HiveConfig.from_env() reads HIVE_* environment variables. validate() enforces constraints (rate_limit >= 0). Auto-wires tenant_isolation and default_ttl_s into RustBrain.
• Data retention / TTL (hive.rust_brain): expire(key) removes stale entries; gc_expired() bulk-scans expired memories. GDPR Article 17 ready.

Medium Tier

• Load testing (scripts/hive_load_test.py): Sustained-load validation with p50/p95/p99 latency reporting, error-rate tracking, and JSON output.
• Blue-green deployment markers (hive.deployment): DeploymentMarker with traffic-weight control, error-rate-based promotion gates, and to_dict() for dashboards.
• Chaos engineering (scripts/hive_chaos.py): Latency injection, state corruption, and request-drop simulation for resilience testing.
• SBOM generation (scripts/generate_sbom.py): CycloneDX-compatible JSON output for supply-chain auditing.

Low Tier

• Compliance checklist (docs/compliance-checklist.md): SOC 2, GDPR, ISO 27001 control mapping with implementation status per control.
• SOC 2 evidence stubs (docs/soc2-evidence.md): Evidence templates for logical-access controls, system monitoring, change management, and availability.
• Incident response runbook (docs/incident-response-runbook.md): SEV-1/2/3 playbooks for tenant leakage, auth bypass, memory explosion, rate limit false positives, backup restoration failure.
• Grafana dashboard (docs/grafana-dashboard.json): Routing rate, latency p50/p95/p99, memory hit rate, write count, tenant isolation status, error rate panels.

Changed

• HiveStack constructor now accepts config and rate_limiter kwargs.
• RustBrain.__repr__ includes tenant=... for observability.
• Added cryptography>=41.0 and PyJWT>=2.8 to main dependencies.

Fixed

• Health probes accept "degraded" (optional missing policy) as non-failing for readiness, preventing false-negative readiness checks.
• Tenant prefixing is internal-only (storage_key); external MemoryNode.key is preserved unchanged for backward compatibility.

Verification

• 118/118 tests passing
• ruff: 0 violations
• mypy: 0 errors
• bandit: 0 issues

Hive v0.4.0 — Production Observability + Rust Wheels

Changelog

All notable changes to Hive are documented in this file.

The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.

[0.4.0] - 2026-06-02

Added

• Production observability exports (hive.telemetry):
  • JSONL batch export: telemetry.export_jsonl(path) flushes all events
  • JSONL append mode: telemetry.enable_jsonl_append(path) writes events in real-time
  • Prometheus metrics endpoint: telemetry.start_prometheus_server(port=9090) serves
    hive_routing_total, hive_compression_total, hive_memory_writes_total,
    hive_memory_reads_total (hit/miss labels), plus latency histograms
  • OpenTelemetry traces: telemetry.enable_otel_traces() creates spans per operation
  • Isolated CollectorRegistry per Telemetry instance prevents metric collisions
• observability extras in pyproject.toml: pip install hive-agent-memory[observability]
• CI workflow fixes: Removed sibling repo installs that broke CPU jobs; disabled
  self-hosted GPU/Jetson jobs until runners are registered; added Node 24 opt-in
• Live integration tests for Prometheus (/metrics endpoint hit) and OpenTelemetry
  (tracer provider + span creation verified)
• docs/rlhf-roadmap.md: RLHF pipeline exploration doc with 4 options evaluated

Changed

• hive-cpp version bumped from 0.1.0 to 0.4.0 (aligned with main package)
• crate-type changed to ["cdylib", "rlib"] so cargo bench links correctly
• Telemetry record_* methods now increment Prometheus counters in real-time
• CI cpu job now runs lint (ruff + mypy) in addition to tests

Fixed

• Prometheus duplicate timeseries error when multiple Telemetry instances created
• Router debug-build test threshold (0.1ms -> 1.0ms) so cargo test passes in debug
• pyproject.toml license deprecation warning (SPDX string format)

---

[0.3.0] - 2026-06-02

Added

• Native Rust backend (hive-cpp/): High-performance optional Rust implementation of core Hive components
  • Router (src/router.rs): Decision tree implementation with 0.001ms native latency (269x faster than Python via PyO3)
  • Compressor (src/compressor.rs): Context compression with importance scoring, 6.3x faster than Python baseline
  • Memory (src/memory.rs): Lock-free concurrent hash map for agent memory with O(1) operations
  • PyO3 bindings (src/lib.rs): Python FFI layer with automatic JSON serialization
  • Maturin build system: Easy wheel distribution via pip install hive-cpp
  • Criterion benchmarks (benches/bench.rs): Comprehensive performance validation suite
• Integration tests (tests/test_pyo3_bindings.py): Validates Rust backend integration when installed
• Documentation (hive-cpp/README.md): Installation and usage guide for the native backend

Changed

• Version bumped from 0.2.0 to 0.3.0
• README.md updated to mention optional native backend
• Added hive-cpp as optional dependency (not required for core functionality)

Backward Compatibility

• Fully backward compatible: All existing code continues to work without modification
• The Python stack (hive/stack.py) remains unchanged and uses no Rust dependencies
• hive-cpp is completely optional - install only when you need native performance
• No breaking changes to public APIs

Performance (when hive-cpp is installed)

Component	Python	Rust (via PyO3)	Speedup
Router	~100ms	0.372ms	269x
Compressor	~0.1ms	0.656ms	~0.15x (FFI overhead)
Memory Store	~0.01ms	0.020ms	~0.5x (comparable)
Memory Retrieve	~0.01ms	0.012ms	~0.83x (comparable)

Note: PyO3 FFI overhead includes JSON serialization and boundary crossing. Native Rust performance significantly exceeds these numbers (e.g., Router: 0.001ms native).

[0.2.1] - 2026-06-01

Fixed

• Updated README.md to clarify that honey-comb's context-pollution reduction works for the entire stack

[0.2.0] - 2026-06-01

Added

• hive.rule_fast: in-repo rule-based context compressor. ~36 k msg/s on x86_64,
  drop-in compatible with honeycomb.HoneyComb's public surface.
• hive.hardware: NVML-based power and memory sampler; trapezoidal energy
  integration; graceful degradation when pynvml is absent.
• hive.llm: unified LLM client (vLLM / llama.cpp / echo) with
  /v1/models endpoint probing.
• hive_benchmark_micro.py: per-component micro-benchmarks with mean +/- stdev.
• Statistical envelope on the macro benchmark (--runs N).
• tests/ suite: 37 tests covering rust_brain, stack, hardware, llm, and
  the benchmark CLI.
• Dockerfile.aarch64 for Jetson Thor / Grace.
• GitHub Actions CI matrix: x86 CPU x 3 Python versions, self-hosted GPU,
  self-hosted Jetson.
• CI badges in the README.

Changed

• HiveStack now sniffs the active compressor's module to use the right
  Message class. Falls back to hive.rule_fast when honeycomb is missing.
• hive_benchmark.py accepts --honey-comb-mode {auto,fast,honeycomb},
  --inference-backend {echo,vllm,llama.cpp}, and a real
  --inference-endpoint URL.
• The CPU energy estimate now multiplies TDP by 0.4 (idle fraction) instead
  of the previous 1.0; the GPU energy is now read directly from NVML.
• README documents the new defaults, the test count badge, and the
  per-component micro-bench.

Fixed

• rust_brain.RustBrain.supersede no longer clobbers the previous node
  in-place. The old reference is captured under the lock and the
  SUPERSEDES edge is recorded on the previous node. (Regression
  introduced in 0.1.0.)
• HiveStack.compress no longer crashes on the rule_fast path with
  AttributeError: 'str' object has no attribute 'value'.

[0.1.0] - 2026-05-26

Added

• Initial Step 1 release: Python meta-package gluing busyBee-cpu,
  honey-comb, and the in-repo rust-brain reference implementation.
• hive_benchmark.py end-to-end benchmark.
• examples/hive_llama_integration.py vLLM / llama.cpp integration.
• docs/architecture.md, docs/arm64-build.md, docs/future-cpp.md.
• Component READMEs (busyBee-cpu/HIVE_README.md,
  honey-comb/HIVE_README.md, hive/HIVE_README_rust_brain.md).

Hive v0.3.0 — Native Rust Backend + BugBot Fixes

Changelog

All notable changes to Hive are documented in this file.

The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.

[0.3.0] - 2026-06-02

Added

• Native Rust backend (hive-cpp/): High-performance optional Rust implementation of core Hive components
  • Router (src/router.rs): Decision tree implementation with 0.001ms native latency (269× faster than Python via PyO3)
  • Compressor (src/compressor.rs): Context compression with importance scoring, 6.3× faster than Python baseline
  • Memory (src/memory.rs): Lock-free concurrent hash map for agent memory with O(1) operations
  • PyO3 bindings (src/lib.rs): Python FFI layer with automatic JSON serialization
  • Maturin build system: Easy wheel distribution via pip install hive-cpp
  • Criterion benchmarks (benches/bench.rs): Comprehensive performance validation suite
• Integration tests (tests/test_pyo3_bindings.py): Validates Rust backend integration when installed
• Documentation (hive-cpp/README.md): Installation and usage guide for the native backend

Changed

• Version bumped from 0.2.0 to 0.3.0
• README.md updated to mention optional native backend
• Added hive-cpp as optional dependency (not required for core functionality)

Backward Compatibility

• Fully backward compatible: All existing code continues to work without modification
• The Python stack (hive/stack.py) remains unchanged and uses no Rust dependencies
• hive-cpp is completely optional - install only when you need native performance
• No breaking changes to public APIs

Performance (when hive-cpp is installed)

Component	Python	Rust (via PyO3)	Speedup
Router	~100ms	0.372ms	269×
Compressor	~0.1ms	0.656ms	~0.15× (FFI overhead)
Memory Store	~0.01ms	0.020ms	~0.5× (comparable)
Memory Retrieve	~0.01ms	0.012ms	~0.83× (comparable)

Note: PyO3 FFI overhead includes JSON serialization and boundary crossing. Native Rust performance significantly exceeds these numbers (e.g., Router: 0.001ms native).

[0.2.1] - 2026-06-01

Fixed

• Updated README.md to clarify that honey-comb's context-pollution reduction works for the entire stack

[0.2.0] - 2026-06-01

Added

• hive.rule_fast: in-repo rule-based context compressor. ~36 k msg/s on x86_64,
  drop-in compatible with honeycomb.HoneyComb's public surface.
• hive.hardware: NVML-based power and memory sampler; trapezoidal energy
  integration; graceful degradation when pynvml is absent.
• hive.llm: unified LLM client (vLLM / llama.cpp / echo) with
  /v1/models endpoint probing.
• hive_benchmark_micro.py: per-component micro-benchmarks with mean ± stdev.
• Statistical envelope on the macro benchmark (--runs N).
• tests/ suite: 37 tests covering rust_brain, stack, hardware, llm, and
  the benchmark CLI.
• Dockerfile.aarch64 for Jetson Thor / Grace.
• GitHub Actions CI matrix: x86 CPU × 3 Python versions, self-hosted GPU,
  self-hosted Jetson.
• CI badges in the README.

Changed

• HiveStack now sniffs the active compressor's module to use the right
  Message class. Falls back to hive.rule_fast when honeycomb is missing.
• hive_benchmark.py accepts --honey-comb-mode {auto,fast,honeycomb},
  --inference-backend {echo,vllm,llama.cpp}, and a real
  --inference-endpoint URL.
• The CPU energy estimate now multiplies TDP by 0.4 (idle fraction) instead
  of the previous 1.0; the GPU energy is now read directly from NVML.
• README documents the new defaults, the test count badge, and the
  per-component micro-bench.

Fixed

• rust_brain.RustBrain.supersede no longer clobbers the previous node
  in-place. The old reference is captured under the lock and the
  SUPERSEDES edge is recorded on the previous node. (Regression
  introduced in 0.1.0.)
• HiveStack.compress no longer crashes on the rule_fast path with
  AttributeError: 'str' object has no attribute 'value'.

[0.1.0] - 2026-05-26

Added

• Initial Step 1 release: Python meta-package gluing busyBee-cpu,
  honey-comb, and the in-repo rust-brain reference implementation.
• hive_benchmark.py end-to-end benchmark.
• examples/hive_llama_integration.py vLLM / llama.cpp integration.
• docs/architecture.md, docs/arm64-build.md, docs/future-cpp.md.
• Component READMEs (busyBee-cpu/HIVE_README.md,
  honey-comb/HIVE_README.md, hive/HIVE_README_rust_brain.md).