Skip to content

Changelog

All notable changes to this project will be documented in this format.

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

[0.4.2] - 2026-07-10

Added

  • Package rename: privysha → asha on PyPI and in imports (from asha import ...).
  • ANCHOR runtime: mission-aware agent governance: action, memory, chain, and plan guards; human approval; OS sandbox (auto hooks or subprocess).
  • Framework adapters: CrewAI, LangChain, AutoGen, LlamaIndex, MCP, LangGraph, asha.Agent, and generic duck-typed agents via anchor() / anchor_any().
  • ANCHOR extras: asha-si[crewai], asha-si[langchain], asha-si[langgraph], asha-si[llamaindex], asha-si[mcp], and meta asha-si[anchor].
  • Mission-drift gate: tests/fixtures/anchor/mission_drift.jsonl with CI coverage.
  • CLI: asha entry point (replaces privysha).

Changed

  • Environment variables use ASHA_* prefix (ASHA_DISABLE_ML, ASHA_ANCHOR_INTERACTIVE, …).
  • ANCHOR telemetry is opt-in (ASHA_ANCHOR_TELEMETRY=1 / explicit path); isolation docker/bwrap deferred with clear errors.
  • CI optional-dep job installs asha-si[anchor] and runs golden-path adapter + mission-drift tests headless.
  • Memory poisoning bands floored for short benign agent outputs; mission/intent loaders accept trained pipeline dumps.
  • README and docs include adapter support matrix (no unimplemented docker/kernel claims).
  • Productization: SUPPORT.md, issue/PR templates, docs support policy (Phase 4).
  • Pilot checklist for 0.4.2: docs/production-readiness.md (1.0 remains deferred).
  • Document PII masking: process / sanitize / optimize accept file paths/bytes; wrap_llm masks local attachments before send (pypdf + python-docx on base install).

Migration

See docs/migration-v0.4.md (includes package rename). Pin asha==0.4.2 for pilots.

Acknowledgments

  • ANCHOR was designed and led by Ajay Rajan A.
  • The initial module scaffold and partial implementation were contributed by Yeshwanth Raj C.

[0.4.1] - 2026-06-15

Architecture completion (breaking within 0.4.x)

  • Root API frozen - only process, sanitize, optimize, Agent, __version__; all lazy root exports removed.
  • Pipeline, AsyncPipeline, Processor removed - use PromptProcessor and process().
  • runtime/resolve.py - hot-path argument resolution (no compat/ on process() calls).
  • PolicyConfig fields - pii_mode, reversible, preserve_intent, security_level via policy= only.
  • security_fail_closed removed - use mode="strict" / mode="balanced".
  • Integration safety unified - wrap_llm(mode=...); auto_patch(mode="strict") configurable.
  • routing_decision - only in compat/legacy_results dicts, not runtime hot path.
  • Architecture tests - types/ and utils/ must not import compat/.

[0.3.0] - 2026-06-05

Developer Preview

Re-release track as 0.x developer preview. APIs may change before 1.0.0. See docs/status.md.

Added

  • AshaFit - project-aware local LLM advisor (recommend_local_model, asha recommend)
  • Workload fingerprinting from compiled prompts + IR
  • HuggingFace catalog with offline fallback; VRAM fit ranking
  • Router integration: RoutingStrategy.LOCAL_PRIVACY
  • Agent(local_model="auto"), wrap_llm(..., auto_select_local_model=True)
  • Interactive demos via Streamlit: streamlit run examples/showcase/streamlit_app.py

Changed

  • Version 1.0.1 → 0.3.0 (semantic versioning: preview until stable 1.0)
  • PyPI classifier: Alpha / developer preview
  • README status banner, scope table, and feedback links

[1.0.1] - 2026-05-24

Added

  • security_fail_closed opt-in on process() / sanitize() for regulated workloads
  • OTEL pipeline wiring - stage_span() in all pipeline stages via StageBase
  • Benchmark P95/P99 latency percentiles in harness and CI output
  • ~60 new tests - optimize, wrap_llm negatives, phone/address PII, FastAPI, fail-closed, edge matrix, framework smoke, hybrid PII
  • Expanded extras - integrations, all, [llamaindex] in pyproject.toml
  • Publish workflow gates - benchmark compare, twine check, fresh wheel smoke

Fixed

  • process_async input validation parity with sync process()
  • Docs - fail-open/fail-closed in security.md; Flask/Django/LlamaIndex/OTEL in integrations.md; Python 3.10+ in troubleshooting
  • Debug transparency - fallback_reason and original_error in debug=True responses

Changed

  • CI coverage gate documented at 40% (matches --cov-fail-under=40)

[1.0.0] - 2026-05-24

Stable Release

First stable release. Policy modes, PII detection, and drop-in wrappers behave consistently; benchmarks and CI enforce quality gates.

Added

  • Policy gate: mode="off" and disabled security/optimization flags bypass pipeline stages via passthrough (pipeline/policy_gate.py)
  • Canonical PII patterns: security/patterns.py shared across detectors, verification, and masking
  • Benchmark harness: tiktoken metrics, expanded sample_prompts.json (10 cases), false-positive and fail-safe rate gates
  • OpenTelemetry extra: pip install asha-si[otel] with enable_otel() for optional stage span export
  • Docs site: full MkDocs Material config, GitHub Pages deploy workflow
  • CI hardening: coverage gate (55%), docs build, benchmark quality gates, dependabot, publish workflow test gate
  • Tests: policy modes, auto_patch mocks, PII secrets/FP, threat scoring, async streaming, agent helpers, backward compat, otel

Fixed

  • mode="off" still masked PII: early passthrough + post-pipeline scrub skip
  • Agent token metrics: use prompts.compiled structure and tiktoken counts
  • wrap_llm: forwards privacy and token_budget; nested OpenAI streaming
  • auto_patch: replaces last user message only; Anthropic v1 Messages.create; proper unpatch via _restore_original()
  • PII false positives: test@example.com and teaching placeholders skipped
  • Secrets: API keys/JWT/bearer patterns verified and masked with [REDACTED]
  • run_security: honors enable_pii_detection / enable_injection_detection

Changed

  • Version unified to 1.0.0 across pyproject.toml, __init__.py, README, docs
  • Publish trigger: GitHub Release (published) or manual dispatch - not tag-only

Release checklist (manual)

  • Configure GitHub pypi environment for trusted publishing
  • Create GitHub Release v1.0.0 after CI is green to trigger PyPI publish

[0.2.0] - 2026-05-23

Production-Stable Release

This release completes the modular pipeline refactor and stabilizes the public drop-in API (process, wrap_llm, optimize, sanitize) for general use.

Added

  • Opt-in reversible masking: reversible=True on process()/sanitize(), unmask() helper, MaskingVault
  • Observability: TraceContext.log_warn(), LatencyBudgetEnforcer wired in pipeline, DebugTracer deprecated
  • CI hardening: mypy, wheel build + twine check, regression benchmark baseline, slow test job
  • Docs site: mkdocs.yml with navigation to existing docs
  • Test suite expansion: reversible masking, scoring, logging security, masking collision, performance, adapter mocks
  • Domain-owned PII detector: security/pii_detector.py (compatibility shim at utils/pii_detector.py)
  • Pipeline contracts: pipeline/contracts.py for shared types
  • Privacy helpers: utils/dropin_privacy.py (build_security_summary, extract_pii_types)
  • Reproducible benchmarks: benchmarks/run_benchmarks.py and benchmarks/sample_prompts.json
  • Lazy public API: Advanced symbols load on first access (PEP 562) for faster imports
  • CI lint step: flake8 F401/F824 checks in GitHub Actions
  • Integration test marker: API-key tests skipped by default in CI

Fixed

  • Unified StageContext: Single canonical class in pipeline/components/stage_context.py
  • process() metrics: pii_detected populated correctly; debug_trace no longer breaks result assembly
  • SecurityResult handling: Dict and dataclass representations supported across drop-in utilities
  • Integration adapters: Delegate to process() instead of composing core services directly
  • Async paths: process_async, optimize_async, sanitize_async use shared security helpers
  • Duplicate ThreatType in __all__
  • ~109 unused import lint issues across the codebase
  • comprehensive_test.py: No longer checks for deleted pipeline.py / stages/ paths
  • Integration tests: No longer call sys.exit(1) when API keys are missing

Changed

  • PyPI classifier: Development Status :: 5 - Production/Stable
  • Version unified to 0.2.0 across pyproject.toml, __init__.py, README, and docs
  • README architecture: Updated to reflect pipeline/ modular layout
  • Public API: Core surface limited to process, wrap_llm, optimize, sanitize, Agent, Pipeline, AdapterFactory
  • Test suite: tests_v2/ included in default test paths
  • Performance claims: Benchmark docs reference reproducible harness output

Release checklist (manual)

  • Configure GitHub pypi environment for trusted publishing (see .github/workflows/publish.yml)
  • Tag v0.2.0 and run publish workflow after CI is green

[0.1.7] - 2026-04-28

PROGRESSIVE ENHANCEMENT RELEASE

Added

  • Progressive Enhancement Architecture: ML features now opt-in with pii_mode parameter
  • Lightweight Default: No automatic model downloads, instant startup
  • Multi-stage PII Pipeline: Advanced detection in core/pii_pipeline/
  • Hybrid PII Detector: Optional ML-enhanced detection via asha-si[ml]

Changed

  • Modular pipeline layout: pipeline/ package replaces monolithic pipeline.py
  • Drop-in API: process, wrap_llm, optimize, sanitize as primary entry points

[0.1.2] - 2026-04-23

Added

  • Universal Adapter System: Support for OpenAI, Anthropic, Grok, HuggingFace, and Ollama providers
  • Smart Routing: Task-based intelligent model selection with fallback logic
  • Security Layer: Advanced injection detection and threat analysis
  • Prompt IR (Intermediate Representation): Structured prompt compilation
  • Model Router: Automatic provider selection based on task complexity
  • Debug Tracing: Complete pipeline visibility and performance metrics
  • Documentation: Full ReadTheDocs integration
  • PyPI Publishing: Ready for production distribution

Enhanced

  • Agent Class: Complete v2 functionality with v1 compatibility
  • Pipeline Architecture: Full 7-stage processing pipeline
  • Optimizer Engine: Advanced token reduction (30-70% cost savings)
  • Universal Model Adapter: Retry logic and automatic failover
  • Factory Pattern: Enhanced adapter creation with fallback support

Fixed

  • Missing Dependencies: Added accelerate>=0.20.0 for HuggingFace compatibility
  • Test Framework: Fixed test functions to use proper assertions
  • HuggingFace Adapter: Resolved device_map compatibility issues
  • IR Builder: Enhanced to handle both object and dict formats
  • Package Metadata: Complete README.md with all required sections

Infrastructure

  • Build System: Proper setuptools configuration with src layout
  • Testing: 51 passing tests (excluding HF adapter tests requiring API keys)
  • Documentation: Complete README with installation, quick start, and API reference
  • Distribution: Both wheel and source distribution passing Twine checks

[0.1.0] - 2026-03-13

Added

Initial development release of ASHA.

Core features:

  • Prompt sanitization
  • PII masking
  • Prompt AST parsing
  • Token optimization
  • Context injection
  • Prompt compilation pipeline

Adapters:

  • OpenAI adapter
  • Ollama adapter
  • HuggingFace adapter

Developer features:

  • Pipeline trace debugging
  • Modular stage architecture
  • Python package distribution
  • PyPI installation support

Testing:

  • Sanitizer tests
  • Optimizer tests
  • Pipeline tests