agentic-systems 2.0.0__tar.gz → 2.1.2__tar.gz

{agentic_systems-2.0.0 → agentic_systems-2.1.2}/CHANGELOG.md

@@ -1,5 +1,139 @@
 # Changelog
 
+## 2.1.2 - 2026-10-22
+
+The 2.1 grammar and execution contracts remain stable. This delivery hardens
+publication and aligns versioned notebooks, Studio and validation bundles.
+
+### Changed
+
+- The semantic evaluator exposes identical, flat criterion arguments through
+  its public Tool schema and OpenAI Agents callable, preserving the strict
+  evidence-backed rubric across Frameworks.
+- Public release acceptance is scoped to OpenAI, Bedrock API key and Ollama
+  across four Frameworks, with deterministic Python controls. vLLM/Colab,
+  AWS IAM and ADA IAM require subsequent exact-wheel validation. ADA bundles
+  are separate from the mandatory public release assets.
+- Release inference uses a persistent allocation capped at USD 10 and 1000
+  transport attempts, including pilots, evaluators and retries. Reservations
+  are conservative bounds, not billed costs.
+
+- OpenAI Agents preserves completed function-tool evidence when a subsequent
+  SDK operation fails, in both synchronous and asynchronous execution. Recording
+  is execution-local and snapshots outputs without changing the public API.
+
+- Pydantic is constrained to the supported v2 major (`>=2.7.0,<3`).
+- Publication uses typed, hash-addressed manifests and sealed candidate assets;
+  a matching index is a no-upload repeat, while partial or conflicting files
+  fail closed. No published version, tag or asset is overwritten.
+- GitHub finalization verifies uploaded bytes before closing a draft and supports
+  an exact no-op repeat. OIDC jobs and environments have scoped permissions.
+- Quality CI tests both Python 3.10 and 3.14; Studio notebook tests isolate
+  configuration instead of depending on a developer's private `.env`.
+- Live CI records disabled routes as not-run, never as successful live evidence.
+
+Registration of a Trusted Publisher is not proof of an OIDC upload. Release
+acceptance still requires the actual workflow runs and new exact-wheel evidence.
+
+## 2.1.1 - 2026-08-30
+
+Patch release that keeps the 2.1 grammar stable while aligning the published
+wheel, Studio application, skill and enterprise bundles after the public 2.1.0
+tag.
+
+### Fixed
+
+- Bedrock now preserves an Agent's explicit empty Tool boundary instead of
+  interpreting it as access to every Tool registered on the System.
+- Studio returns human language, removes private reasoning markers, retains a
+  bounded public conversation window and reports normalized processing and
+  usage evidence across Providers and Frameworks.
+- Graph inspection exposes the selected orchestration boundary and available
+  topology without leaking native SDK objects.
+- Lineage Memory retains compact, masked business evidence from Tools, including
+  scalar values nested in the normalized Tool envelope.
+
+### Distribution
+
+- Studio remains a separate portable reference application: the PyPI wheel
+  provides the computational grammar, while the GitHub release provides Studio,
+  notebook, skill and offline enterprise ZIPs pinned to the same wheel hash.
+- The Agentic Systems skill now documents immutable release identity, portable
+  `.env` configuration and the distinction between PyPI and application assets.
+- Release certification is generated from validated live artifacts and records
+  certified/assembly commit equivalence instead of relying on a hand-edited
+  summary.
+
+## 2.1.0 - 2026-08-30
+
+Compatible hardening release. The 2.0 public facade remains stable while three
+blocking gates now certify production readiness, Pydantic contracts and
+provider/framework substitutability.
+
+### Added
+
+- A canonical registry for five Providers, four Frameworks, capabilities,
+  dependencies, environment inputs, six scenarios and all 20 declared pairs.
+- Strict versioned Pydantic schemas for execution limits, runtime discriminators,
+  public specs, normalized outputs, errors and live attestations.
+- Hypothesis boundary/round-trip coverage, Protocol conformance tests, Import
+  Linter layer contracts, architecture branch checks, complexity and benchmark
+  ratchets, dependency/license/secret audits and weekly mutation testing.
+- Protected OpenAI, Bedrock and Ollama live workflows plus an exact-wheel vLLM
+  Colab attestation contract with 24-hour freshness validation.
+- Reproducible skill, Studio and release asset builds with secret auditing,
+  hashes, SBOM generation and PyPI Trusted Publishing.
+- Versioned `ModelArtifact`, `VLLMServerSpec`, `EndpointInfo` and `ServerHealth`
+  schemas plus the `ModelServer` Protocol and explicit `model_server(...)` facade.
+- One Colab tutorial that installs a CUDA-compatible vLLM stack, serves an
+  Unsloth/Qwen artifact and certifies native, LangGraph, OpenAI Agents and Strands.
+
+### Changed
+
+- `SchedulerConfig`, `RuntimeConfig` and `RunPolicy` share `ExecutionLimits`;
+  `max_tool_calls=0` now consistently means that tools are prohibited.
+- `RunResult` uses one public projection across API, CLI, notebooks, Studio and
+  evals. A balanced leading native reasoning block is removed from public text
+  while raw evidence remains available explicitly.
+- Retries are limited to classified transient failures; SDK and tool failures
+  carry structured category, runtime identity, code and retryability.
+- Provider/Framework lists in compatibility, CLI and Studio derive from the
+  canonical registry instead of parallel hardcoded inventories.
+- API checksums ignore methods injected by Pydantic, unwrap declared decorated
+  methods and fingerprint only the public contract.
+- Strands callbacks with typed event parameters are normalized into HookProvider
+  objects, preserving compatibility across Strands 1.29 through the current 1.x.
+- Strands model policy configuration now follows each SDK model's declared
+  shape: OpenAI-compatible request parameters remain nested under `params`,
+  while native models such as Bedrock receive `temperature` and `max_tokens`
+  directly, preventing ignored settings and invalid-parameter warnings.
+- Stateful Strands Agents retain their native SDK conversation while each
+  public `RunResult` projects only the current invocation, preventing historical
+  Tool events from being counted again against `RunPolicy.max_tool_calls`.
+- The OpenAI Agents tutorial bootstraps its optional SDK dependency idempotently
+  in a fresh Jupyter kernel instead of failing with a raw ModuleNotFoundError.
+- Public local-model defaults and examples now use the exact certified Qwen3 4B
+  Instruct 2507 artifacts for Ollama and vLLM; `.env` remains authoritative.
+
+### Compatibility
+
+- No top-level 2.0 export is removed.
+- Existing `text`, `final`, `engine`, `model`, `usage` and `meta` result views
+  remain available; typed projections are additive.
+
+### Certification
+
+- The primary five-Provider by four-Framework matrix passed all 20 routes and
+  all 76 semantic episodes, including the deterministic Python control.
+- Bedrock AWS credential-chain execution passed four Framework routes and 16
+  semantic episodes in SageMaker, then repeated the same 4/4 and 16/16 result
+  in the ADA enterprise sandbox.
+- Across the primary matrix and both IAM environments, 108/108 semantic
+  episodes were manually reviewed for public answer, hierarchical lineage,
+  deterministic evidence, judge verdict, runtime identity and absence of
+  fallback. `final-certification-summary.json` pins the exact wheel and evidence
+  hashes used by the release gate.
+
 All notable changes to Agentic Systems are documented here.
 ## 2.0.0 - 2026-08-19