v0.17.0

What's Changed

• feat(transport): SSE transport for MCP gateway and proxy (#694) by @dgenio in #716
• feat(ci): supply-chain & security hardening (#443, #689, #690, #691, #692, #468, #552) by @dgenio in #718
• chore(ci): bump peter-evans/create-pull-request from 6 to 8 by @dependabot[bot] in #720
• chore(ci): bump actions/attest-build-provenance from 96b4a1ef7235a096b17240c259729fdd70c83d45 to e8998f949152b193b063cb0ec769d69d929409be by @dependabot[bot] in #721
• chore(ci): bump actions/cache from 4 to 5 by @dependabot[bot] in #722
• chore(ci): bump actions/deploy-pages from 4 to 5 by @dependabot[bot] in #723
• chore(ci): bump actions/upload-pages-artifact from 3 to 5 by @dependabot[bot] in #724
• chore(deps): update mkdocstrings requirement from <1,>=0.25 to >=0.25,<2 by @dependabot[bot] in #725
• feat(mcp): generate config packs for MCP clients by @dgenio in #726
• docs: refresh label catalog to match live taxonomy by @corazonthedev in #715
• feat(benchmarks): scaling, scenarios, CI quality gate, and release trend by @dgenio in #727
• docs: link the Weaver Stack overview article (Towards AI) by @dgenio in #728
• chore(ci): bump actions/attest-build-provenance from 2.4.0 to 4.1.1 by @dependabot[bot] in #733
• feat(cli): add deployment-intent onboarding wizard by @dgenio in #734
• chore(ci): bump actions/cache from 5 to 6 by @dependabot[bot] in #732
• chore(ci): bump actions/checkout from 4 to 7 by @dependabot[bot] in #731
• chore(ci): bump astral-sh/setup-uv from 5 to 7 by @dependabot[bot] in #729
• chore(ci): bump peter-evans/find-comment from 3 to 4 by @dependabot[bot] in #730
• [codex] Add MCP incident packs by @dgenio in #739
• feat(gateway): runtime authorization gate + secure-by-default serving by @dgenio in #760
• fix(security): secret-scrubbing parity across gateway, primitive, and sidecar surfaces (#742, #743, #745) by @dgenio in #771
• chore(ci): bump peter-evans/create-or-update-comment from 4 to 5 by @dependabot[bot] in #772
• chore(ci): bump actions/upload-artifact from 4 to 7 by @dependabot[bot] in #773
• chore(ci): bump actions/setup-python from 5 to 6 by @dependabot[bot] in #774
• chore(ci): bump github/codeql-action from 3 to 4 by @dependabot[bot] in #775
• feat(adapters): knowledge-bundle context sources (#736, #763, #767, #776) by @dgenio in #779
• feat(gateway): add named policy presets for gateway operations (#664) by @dgenio in #780
• feat(gateway): live multi-upstream MCP serving, VS Code import, artifact lifecycle policy by @dgenio in #785
• feat: MCP gateway platform-maturity cluster (governance, observability, model-assist, transport) by @dgenio in #789
• Release v0.17.0 by @dgenio in #790

Full Changelog: v0.16.0...v0.17.0

v0.16.0

contextweaver 0.16.0 — consolidation, primitives, and DX hardening.

Highlight: Memory Consolidation Engine

Distill episodic memory into durable facts with contextweaver.context.consolidate(...). Deterministic clustering, provenance tracking, max-sensitivity inheritance, and an optional LLM-assisted canonicalizer — all pure stdlib, zero new dependency. Includes contextweaver consolidate CLI and �valuate_consolidation offline quality harness.

Gateway: Resources & Prompts

PrimitiveGatewayRuntime extends bounded-choice routing and context-firewall treatment from tools to MCP resources and prompts. Unified cross-primitive identity via
outing.primitive_id, four gateway meta-tools, and complete upstream adapters.

Developer Experience

• HTTP sidecar — language-agnostic POST /v1/route and /v1/compact over stdlib HTTP, no new dependency
• contextweaver verify — fast, deterministic smoke test for library-first adoption
• Stable error codes — every ContextWeaverError carries a machine-readable code + hint`n- Deprecation machinery — @deprecated decorator, registry, and upgrade guide
• Persistent routing cache — RoutingIndexCache + CachedRetriever for warm-start routing

Action Required

Three pre-1.0 APIs now emit DeprecationWarning (behavior unchanged; removal in 1.0):

• RouteResult.debug_trace → use .trace`n- RouteTrace.to_legacy_dicts() → use structured fields
• Router(scorer=...) → use
  etriever= or scorer_backend=`n
  See CHANGELOG.md for the full list.

v0.15.0

contextweaver 0.15.0

Highlights

• Structured route->select contract — constrained-selection schemas, validated selection repair, shortlist composition controls (#515, #479, #516, #509)
• New adapters — OpenAPI, Agent Skills, Microsoft Agent Framework, LangChain, OpenAI Agents, Google ADK (#546, #545, #430, #454, #502, #501, #547)
• Context-engine tuning — caller-owned rendering, new ItemKind.retrieved_doc, phase-aware scoring weights, budget-overflow policy (#410, #411, #487, #510)
• Gateway dispatch hardening — retry/backoff, dry-run, rate-limiting, read-only response cache (#529, #483, #482, #512)
• Store layer — Redis & S3 backends, SQLite episodic/fact stores, async protocol variants, conformance kit (#426, #496, #495, #520)
• Firewall — single-call facade, structured mode, determinism guarantee, built-in token counter (#399, #406, #404, #405)
• Sensitivity enforcement — secret-redaction, ingestion-time classification, header memory enforcement (#428, #542, #450, #451)
• Diagnostics — trustworthy context/gateway reporting, public API manifest, CI drift checks (#370, #378, #518, #522)

See CHANGELOG.md for full details.

v0.14.1

#564

[0.14.1] - 2026-06-11

Added

• MCP Registry listing + PyPI ownership marker (#348). Adds a
  registry-publishable server.json describing the gateway as a
  uvx contextweaver mcp serve --config <gateway.yaml> stdio server (linking
  to the gateway quickstart, not the raw API docs), an
  mcp-name: io.github.dgenio/contextweaver marker in the README for PyPI
  ownership verification, and a release-triggered GitHub Actions job that
  publishes to the official MCP Registry via GitHub OIDC (no interactive
  login required).
• Trustworthy diagnostics across context builds and the MCP gateway
  (#370, #378, #398, #414, #459). BuildStats.dropped_items attributes
  every excluded item to sensitivity, dedup, kind_limit, or budget;
  the production context pipeline now fires exclusion and budget lifecycle
  hooks. New versioned DiagnosticEvent / DiagnosticSink APIs include
  thread-safe in-memory and append-only JSONL sinks. ProxyRuntime emits
  sanitized catalog, browse, hydrate, execute, and artifact-view events with
  counts, token/schema savings, failures, and latency. Operators can use
  contextweaver mcp inspect, contextweaver mcp stats, and
  contextweaver inspect for JSON or Markdown reports without exposing raw
  queries, argument values, result text, prompt text, or artifact bytes.
• Single-call firewall facade ÔÇö compact_tool_result() /
  firewalled_tool_result() (#399). Shrink one large tool result before it
  enters the prompt without standing up a ContextManager. Returns a
  CompactResult (firewalled, payload, summary, facts, artifact_ref,
  stats). Exported from the top level.
• Structured (lossless) firewall mode (#406). New StructuredFirewall(keep=[...])
  plus summarize.structured.project / parse_path: keep an allow-list of
  JSON paths inline, offload the rest to the artifact store (retrievable via
  drilldown), no LLM. Selectable through compact_tool_result(strategy=...)
  and ContextManager.ingest_tool_result(..., firewall=StructuredFirewall(...)).
  An explicit strategy="structured" now raises ConfigError on non-JSON
  input instead of silently downgrading to a text summary; ingest_tool_result
  applies firewall= only above firewall_threshold.
• First-class firewall diagnostics ÔÇö FirewallStats (#402). Records
  triggered, strategy, original/summary chars+tokens (chars_saved /
  tokens_saved), artifact_ref, and summarized_by_llm. Surfaced on
  ResultEnvelope.firewall_stats, and aggregated on BuildStats.firewall_events
  / BuildStats.firewall_summary().
• Determinism guarantee ÔÇö deterministic=True (#404). ContextManager(deterministic=True)
  and compact_tool_result(deterministic=...) fail closed with the new
  DeterminismError rather than passing data through an LLM-backed summariser;
  FirewallStats.strategy / summarized_by_llm make the path auditable.
• Built-in token counter ÔÇö contextweaver.tokens (#405). Public
  count() / get_token_counter() / heuristic_counter() (and TokenCounter
  alias) so callers never wire tiktoken directly; firewall/FirewallStats
  numbers use the same counter. New no-op contextweaver[tokenizers] extra
  documents the contract (tiktoken is already core, with offline fallback).
• Daily Driver guide for MCP gateway operators (#394). New
  docs/daily_driver.md explains when to use or bypass contextweaver,
  copy-paste operating instructions for common MCP clients, and a practical
  debug loop using route explanations, BuildStats, artifact views, and OTel.
• MCP gateway security and data-flow model (#396). New
  docs/security_model.md distinguishes prompt exposure from raw artifact
  storage, documents trust and egress boundaries, and records the current
  tool_view / artifact-lifecycle limits tracked by #375.
• Verified Claude Code MCP recipe (#429). Adds project/local registration
  commands, a committed .mcp.json example, operating instructions, and
  troubleshooting verified against Claude Code 2.1.165.
• Zero-install CLI smoke coverage (#437). Linux and macOS CI now build the
  wheel and run its contextweaver entry point through isolated uvx and
  pipx environments.

Changed

• BuildStats accounting now has one pipeline owner (#459).
  total_candidates is measured after dependency closure and before
  sensitivity filtering; dropped_count includes every later exclusion, so
  completed builds satisfy included_count + dropped_count == total_candidates. The report schema is version 2.
• CI now exercises every committed generated-artifact drift check
  (#389ÔÇô#393). llms.txt / llms-full.txt, recorded demo casts, and the
  gateway scorecard are gating checks on the Python 3.12 matrix cell; the
  deterministic smoke evaluation also runs there as a non-gating signal.
• MCP client recipes now use the installed CLI (#371, #437). Claude
  Desktop, Claude Code, GitHub Copilot, and Cursor configs launch
  uvx contextweaver mcp serve; docs no longer describe the dedicated CLI as
  future work. examples/recipes/serve_gateway.py remains a labelled
  legacy/custom-runtime example, while config tests reject references to that
  launcher across relative, absolute, POSIX, and Windows path forms. Relative
  catalog paths now resolve from the config file, and text results expose their
  stored artifact handle so clients can call tool_view.

v0.14.0

[0.14.0] – 2026-06-07

Added

• Canonical Frame-shaped ingestion seam — ContextManager.ingest_envelope()
  (#352). The execution boundary (e.g. agent-kernel) firewalls and hands
  contextweaver an already-firewalled ResultEnvelope (the native preimage of
  a weaver-spec Frame); contextweaver appends a summary-only ContextItem
  carrying the artifact handle and does not re-derive firewalling from raw
  output. The raw-output APIs (ingest_tool_result, ingest_mcp_result)
  remain for standalone use but are now labelled non-canonical for spec
  compliance. New firewall boundary doc
  explains the contextweaver-firewall vs agent-kernel-firewall split and the
  seam; weaver-spec I-05 status updated accordingly.
• Zero-Python config-file launch for the MCP gateway (#346).
  contextweaver mcp serve --config gateway.yaml reads the catalog and serve
  options (mode, top_k, beam_width, cache_stable, name, version)
  from a single JSON/YAML file; explicit CLI flags still win. The catalog
  loader now also accepts the real-MCP-server snapshot shape
  ({"tools": [...]}) used by the recipes. New Cursor recipe
  (docs/recipes/cursor.md) plus examples/recipes/gateway_config.yaml and
  examples/recipes/cursor_mcp.json. (Bridging a live upstream MCP server
  over stdio remains follow-up on #346.)
• rank_collected is now part of the public routing API (#288). The
  score-sort / active-filter helper is re-exported from
  contextweaver.routing so custom Navigator implementations can reuse it.
• End-to-end quality + cost benchmark vs a competent baseline (#345). New
  benchmarks/e2e_quality.py runs realistic tool-using tasks three ways —
  naive concat, a hand-built competent baseline, and contextweaver — scoring
  tool-selection accuracy, hallucinated-tool rate, end-task answer accuracy,
  prompt tokens, and estimated cost per strategy. Ships with a deterministic
  stub model (default, exercised in CI) and an opt-in real-model path
  (CW_E2E_LLM=1 + a user-supplied call_fn, no LLM SDK dependency). New
  make e2e-quality target (non-gating) and benchmarks/e2e/tasks.json
  fixtures. The published real-model headline is produced from a credentialed
  maintainer run.

Changed

• Decomposed ContextManager to meet the ≤300-line module guideline (#101).
  The pipeline logic already lived in context/build.py
  (run_build_pipeline), context/route_build.py, context/call_prompt.py,
  and context/ingest.py; what remained was the manager's own method surface
  (manager.py was 878 lines of thin delegating stubs + docstrings). Those
  stubs now live in flat, single-level partial-class mixins —
  _IngestMixin (context/_manager_ingest.py), _BuildMixin
  (context/_manager_build.py), _RoutingMixin (context/_manager_routing.py)
  — sharing a _ManagerState base (context/_manager_base.py) that declares
  the private-attribute contract. manager.py is now 239 lines (only
  __init__, properties, drilldown, and mixin composition); every module is
  ≤300. The delegate pipeline functions are now typed against _ManagerState
  (interface segregation; ContextManager inherits it via the mixins, so every
  call site is unchanged). No public API change — all 21 methods stay on
  ContextManager and the full test suite passes unmodified.
• Unified routing metrics into contextweaver.eval.metrics (#354).
  benchmarks/benchmark.py and contextweaver.eval.routing previously
  defined recall@k / reciprocal_rank under the same names with different
  semantics (fractional recall vs boolean hit-rate). They now share one
  canonical source of truth — recall_at_k (classic fractional recall),
  precision_at_k, reciprocal_rank — re-exported from contextweaver.eval.
  The benchmark scorecard numbers are unchanged; evaluate_routing now reports
  fractional recall for multi-expected cases (identical for the common
  single-expected case).
• Split extras/memory/zep.py into zep.py + _zep_common.py so each
  module stays within the repo's ≤300-lines-per-module rule (PR #360 review).
  The public import path (contextweaver.extras.memory.zep) and its exports
  (ZepBackendError, ZepEpisodicStore, ZepFactStore) are unchanged.

Fixed

• Routing history tool-id resolution narrows its exception handling.
  route_build.resolve_tool_id_from_result previously wrapped the parent
  event-log lookup in a bare except Exception, silently swallowing any error
  before falling back to parent_id. It now catches only ItemNotFoundError
  (the documented EventLog.get contract), so unexpected store errors surface
  instead of being hidden (PR #363 review).
• Provider message encoders no longer emit empty-content messages.
  to_anthropic_messages and to_gemini_contents now raise a clear
  CatalogError (with the offending msg_index) when a turn would
  serialise to empty or blank-text content, instead of letting the
  provider reject it later with an opaque
  400 ... messages: ... must have non-empty content. Messages that
  carry tool-use / tool-result / function-call blocks remain valid.
  OpenAI is intentionally left untouched: its Chat Completions API
  tolerates empty content and the empty-string assistant-content
  round-trip is an existing invariant (PR #230).
• Zep backend defensively coerces scanned tags / metadata when rebuilding
  Episode / Fact from persisted episodes: a non-list tags (e.g. a bare
  string, which previously iterated into characters) yields [], and a non-dict
  metadata (which previously raised in dict(...)) yields {} (PR #360 review).
• LlmSummarizer / LlmExtractor fallback warnings now include the underlying
  exception text, so a degraded LLM path is diagnosable (timeout vs auth vs
  parsing) instead of opaque (PR #360 review).

contextweaver 0.13.4

Restore proven dependency floors:

• mcp: >=1.19.0 (was >=1.0; in-memory transport wire format changed)
• typer: >=0.16.0 (was >=0.9; breaks against click 8.2)
• pytest-asyncio: >=0.23.8 (was >=0.23; Package.obj collection bug)
• fastmcp: >=2.12.0 (was >=2.0; decorator pattern changed)

All floors trace back to the proven floor-deps commit �27eb1b which was accidentally reverted during the v0.13.0 release merge.

v0.12.0

What's Changed

Added

• Catalog showcase reference architecture — a start-here, deterministic example (�xamples/architectures/catalog_showcase/) that narrows a 65-tool catalog to a 5-card shortlist, hydrates only the selected tool's schema, and firewalls a large result; wired into make architectures and documented at docs/architectures/catalog_showcase.md. (#330)
• contextweaver demo --scenario killer — the 60-second failure mode: 100 tools + a long history + a huge tool result, contrasting a naive loop against contextweaver in character terms (92–99% reductions). New README 'The 60-second failure mode' section and docs/killer_demo.md. (#322)
• LangGraph agent-loop reference architecture — contextweaver running inside a LangGraph \StateGraph\ (\�xamples/architectures/langgraph_agent_loop/): LangGraph owns control flow, contextweaver owns route/firewall/answer. Guarded import with a hand-rolled fallback so it runs without the framework; new [langgraph]\ extra. \docs/architectures/langgraph_agent_loop.md. (#326)
• Agent-safe evaluation-artifact context profile — a context-shaping profile (\�xamples/architectures/eval_artifact_profile/) with \ok\ / \caution\ / \high_risk\ fixtures that never surfaces \V_hat\ without support diagnostics and foregrounds caveats for high-risk artifacts, with runtime-asserted invariants. \docs/architectures/eval_artifact_profile.md. (#335)

Fixed

• Docs accuracy follow-up to #337 — aligned the VibeGuard --diff\ shape between the Python subprocess and YAML CI snippets in \docs/cookbook.md, and corrected the sensitivity row in \docs/interop_skill_cards.md\ so it matches the actual \�pply_sensitivity_filter\ semantics. (#338)

Full Changelog: v0.11.1...v0.12.0

v0.11.1

Added

• Community & interop docs — CODE_OF_CONDUCT.md, contributor landing page, safety-gate cookbook recipe, and skill-card interop doc (#249, #325, #332, #333)
• Adopter positioning docs — benchmark report, ecosystem comparison, stability checklist, and launch kit (#323, #324, #327, #328, #329)

Changed

• README / docs / PyPI positioning — sharpened first-screen category and aligned descriptions (#321)

v0.10.0

What's New in v0.10.0

Added

• contextweaver.routing.hydration - public schema-hydration helpers (SchemaSource, hydrate_with_schema, lazy_schema_resolver). Resolve a tool's full input schema from a sidecar source without hand-rolling a _FULL_SCHEMAS dict. Issue #261.
• contextweaver mcp serve CLI - new Typer sub-app boots McpGatewayServer or McpProxyServer over stdio against any JSON/YAML catalog. Marked [experimental]. Issues #243/#246.
• Live-transport MCP gateway architecture variant - runs through a real mcp.server.Server + ClientSession paired via mcp.shared.memory. Issue #260.
• Multi-turn MCP gateway architecture variant - 4-turn scenario with fact accumulation across turns and dependency closure. Issue #262.
• contextweaver demo --scenario mcp-gateway-full - surfaces the 60-tool reference architecture from the CLI. Catalog ships inside the wheel. Issue #264.
• Gateway-scenario benchmark suite - 5 deterministic scenarios, firewall-reduction range 0.0%-98.8%. Issue #270.
• Real-MCP catalog architecture variant - runs against snapshots of 3 real MCP servers. Issue #280.
• Asciinema recordings for showcase demos - 4 committed casts under docs/assets/casts/. Issue #281.
• RouteResult.to_dict / from_dict - missing serialization pair. Issue #289.
• Context-pack explanation traces - ContextManager.build(..., explain=True) returns per-candidate scoring and drop reasons. Issue #291.
• Sensitivity/firewall regression fixtures - 6 fixtures covering all floors in drop and redact modes. Issue #292.
• Weaver-spec payload fixtures - --fixtures-dir flag for actionable diagnostics. Issue #295.
• Golden route-prompt + MCP-ingestion fixtures - tests/fixtures/golden/ snapshots. Issue #296.

Fixed

• MCP server call-tool result shape - proper CallToolResult objects instead of 2-tuples.
• SchemaSource.from_json_file validation tightened.
• CallToolResult.content type annotation narrowed for mypy strict.
• ContextBuildExplanation overload guard for explain=True.
• PYTHONPATH in pytest config for tests.fixtures imports. Issue #302.

Full Changelog: v0.9.1...v0.10.0

0.8.0

What's Changed

Added

• CrewAI adapter (#193) — �dapters/crewai.py. Thin stateless converter: CrewAI BaseTool instances or plain-dict shape → SelectableItem. Ships crewai_tool_to_selectable, crewai_tools_to_catalog, infer_crewai_namespace, load_crewai_catalog. Dict-conversion path works without [crewai] extra. New integration guide docs/integration_crewai.md and examples/crewai_adapter_demo.py.
• Mem0 external-memory backend (#195) — extras/memory/mem0.py. Mem0EpisodicStore + Mem0FactStore implement EpisodicStore / FactStore Protocols verbatim (no widening). Writes via mem0.Memory.add(infer=False) stamped with cw_episode_id / cw_fact_id metadata. New [mem0] optional-dependency group. New docs/integration_memory.md decision matrix.

Changed

• Provider-SDK leak invariant tests now run in a subprocess — the est_module_does_not_import_provider_sdk_at_load_time assertions in est_adapters_openai_messages.py, est_adapters_anthropic_messages.py, and est_adapters_gemini_contents.py now spawn a fresh interpreter so the check is independent of test ordering and other installed extras (e.g. crewai transitively importing openai).

Fixed

• crewai adapter: deep-copy �rgs_schema to prevent catalog mutation
• EpisodicStore.add(): upsert semantics (idempotent on re-add)
• Mem0EpisodicStore.add(): graceful fallback to append-only when scan_limit exceeded
• Mem0FactStore.put(): guarded write path to avoid silent data loss
• Docs: corrected API kwargs, added original_description metadata

Full changelog: https://github.com/dgenio/contextweaver/blob/main/CHANGELOG.md