v3.9.2

What's Changed

• deps: bump opentelemetry_sdk from 0.31.0 to 0.32.0 by @dependabot[bot] in #54
• deps: bump polars from 0.51.0 to 0.53.0 by @dependabot[bot] in #56
• deps: bump hmac from 0.12.1 to 0.13.0 by @dependabot[bot] in #58
• deps: bump rand from 0.8.6 to 0.10.1 by @dependabot[bot] in #59
• deps: bump wasmtime from 43.0.2 to 44.0.1 by @dependabot[bot] in #57
• deps: bump serde_json from 1.0.149 to 1.0.150 in the minor-and-patch group by @dependabot[bot] in #60
• deps: bump opentelemetry-otlp from 0.31.1 to 0.32.0 by @dependabot[bot] in #62
• deps: bump opentelemetry from 0.31.0 to 0.32.0 by @dependabot[bot] in #63
• deps: bump opentelemetry_sdk from 0.31.0 to 0.32.0 by @dependabot[bot] in #61
• deps: bump tracing-opentelemetry from 0.32.1 to 0.33.0 by @dependabot[bot] in #64
• deps: bump async-nats from 0.48.0 to 0.49.0 by @dependabot[bot] in #66
• deps: bump the minor-and-patch group across 1 directory with 4 updates by @dependabot[bot] in #65

Full Changelog: v3.9.1...v3.9.2

Converge v3.8.1 — Authority slice, extension-driven foundation

[3.8.1] - 2026-05-06

First release of the v3.8 line. Foundation has been refocused around the
authority slice; ML, analytics, policy, generic provider adapters, and
domain packs now live in dedicated extension repos.

Removed

• BREAKING: converge-knowledge extracted to the mnemos extension repo
  (~/dev/extensions/mnemos). Foundation no longer ships vector storage,
  agentic memory, or learning implementations. Downstream consumers must
  depend on mnemos directly. See ADR-008.
• BREAKING: converge-analytics extracted to the prism extension repo
  (~/dev/extensions/prism). Foundation no longer ships Polars/Burn ML
  pipelines, training agents, or analytics packs. Downstream consumers must
  depend on prism directly.
• BREAKING: converge-domain and examples/* extracted to the
  atelier showcase repo (~/dev/atelier). The four built-in domain packs
  (trust, money, delivery, data_metrics) and 13 worked examples now live in
  atelier. Downstream consumers depend on atelier-domain.
• BREAKING: converge-policy extracted to the arbiter extension
  repo (~/dev/extensions/arbiter). Cedar policy engine, policy suggestors,
  and ed25519-signed delegation tokens now live in arbiter. Foundation
  converge-pack keeps the gate trait and authorization vocabulary.
• BREAKING: converge-provider-adapters (LLM, search, tool, embedding,
  reranker, vector adapters) extracted to the manifold extension repo.
  Foundation no longer ships ready-made vendor adapters; downstream
  consumers depend on manifold directly.
• 3,469 lines of dead code: orphaned consensus/ module, broken billing/
  module, orphaned stress_tests.rs.
• ContextView alias (use Context directly).

Changed

• BREAKING: Provider contract crate renamed from converge-provider-api
  to converge-provider. Generic adapter implementations moved to Manifold.
• BREAKING: ContextView removed — use Context (the trait); the
  concrete implementation is now ContextState. Consumers using
  converge_core::ContextView must change to converge_core::Context.

Added

• Optimization Suggestor adapter: SolverSuggestor<P: Pack> wraps all
  11 optimization domain packs as first-class Suggestors. Every solver
  participates in the convergence loop via register_suggestor_in_pack.
• Policy Suggestor adapters: PolicyGateSuggestor,
  DelegationVerifySuggestor, FlowGateSuggestor bridge Cedar policy
  evaluation into the convergence loop.
• Pack typing primitives: UnitInterval governance primitive,
  Fact::parse_content / ProposedFact::parse_content typed
  deserialization, AgentEffect::push / extend for incremental
  composition.
• Compile-fail contract tests (7): prove at compile time that wrong
  API usage is impossible — no Fact construction without kernel authority,
  no orphan Agent trait, no register_in_pack, immutable facts, closed
  ContextKey enum.
• CI/CD infrastructure: pre-commit hooks, GitHub Actions
  (CI / Security / Coverage / Stability), dependabot auto-merge,
  self-hosted coverage badges across converge, organism, and axiom.
• 700+ new tests across all modules: property tests, negative tests,
  edge cases, serde roundtrips, async paths with mocked LLM backends.
• performance-profile justfile recipe: per-bench targets with optional
  features and explicit PERF_MODE (save|compare).

Fixed

• [lib] bench = false on converge-core and converge-optimization so
  cargo bench --bench NAME no longer also runs the lib unittest binary
  in bench mode (which rejects --save-baseline).
• Stability workflow: install protoc on bench-compile / bench-run / soak
  jobs; remove the dead live-endpoints job; bump cargo-deny-action to
  v2 (matches the modern config schema); allow 0BSD license
  (vec_mut_scan via --all-features).
• Dependabot auto-merge: drop the approval step. Branch protection on
  main does not require reviews, and GITHUB_TOKEN is not permitted
  to approve PRs — the step blocked every dependabot PR.

Dependencies

• firestore 0.45.1 → 0.48.0
• metrics-exporter-prometheus 0.16.2 → 0.18.3
• reqwest 0.12.28 → 0.13.3 (major)
• toml 0.8.23 → 1.1.2+spec-1.1.0 (major; tracks TOML 1.1 spec)
• schneegans/dynamic-badges-action 1.7.0 → 1.8.0

v3.7.3

Patch release for the formation-builder substrate.

Highlights:

• Adds formation template catalog semantics and grouped kernel formation surfaces for downstream builders.
• Records HITL gate decisions as experience events and validates HITL resume gate IDs.
• Adds formation correlation IDs and role-scoped provider backend requirements.
• Publishes Organism handoff guidance for formation compilation, business strategy, and the vendor-selection proof wedge.

Validation:

• cargo check --workspace --all-targets
• cargo test --all-targets --workspace
• cargo fmt --all --check
• just test-layout
• cargo clippy --all-targets -- -D warnings

v3.7.1 — Sealed Promotion Gate

What's Changed Since v3.0.3

This release covers seven minor releases of architectural work: the promotion gate is now fully sealed, the convergence loop is cleaner, and the codebase is significantly leaner.

---

Breaking Changes

• ContextView removed — use Context (the trait) and ContextState (the concrete type). All downstream consumers updated.
• ConvergeResult now requires integrity — Engine::run returns an IntegrityProof (merkle root, Lamport clock, fact count) alongside the converged context.
• Provider capability contracts moved — ChatBackend, ModelSelector, ChatRequest/Response now live in converge-provider-api. Downstream crates depend on the lightweight API crate, not the full engine.

---

Promotion Gate — Now Fully Sealed

The "agents suggest, engine decides" invariant is now enforced at compile time with no escape hatch:

• AgentEffect is proposal-only — with_fact() has been removed
• Fact has no public constructor — only the PromotionGate inside converge-core can promote
• ValidatedProposal and ValidationReport have private fields — no forgery via struct literals
• Proposal<Validated> cannot be deserialized or constructed outside the gate
• 26 compile-fail tests prove the gate is sealed across all public crate boundaries

---

New Capabilities

Formation pattern — multiple heterogeneous suggestors (domain logic, policy gates, optimizers, knowledge) converge in a single Engine::run. See examples/formation-mixed.

Optimization as first-class agents — SolverSuggestor<P: Pack> wraps all 11 optimization domain packs as Suggestor implementations. No separate pipeline.

Policy as first-class agents — PolicyGateSuggestor, DelegationVerifySuggestor, and FlowGateSuggestor bring Cedar policy evaluation into the convergence loop.

Run integrity — every ConvergeResult carries a tamper-evident IntegrityProof: SHA-256 content hashes, Lamport logical clock, and a Merkle root over all promoted facts.

New providers — KongBackend (Konnect Key Auth, retry, tool calling), OpenRouterBackend (100+ models, cost metadata), ResilientChatBackend (retry + format fallback).

---

Quality

• 700+ new tests — property tests, negative tests, soak tests, compile-fail contracts
• Stability testing framework — Criterion benchmarks, chaos tests, soak tests, automated weekly regression detection
• Dead code removed — 3,469 lines of orphaned modules deleted
• Zero clippy warnings — workspace-wide clean pass

---

Workspace

• Deployment and infra moved to the runway repo — converge workspace is now kernel + providers only
• converge-tool / converge-axiom extracted to separate repo

---

Known Limitations

• bipartite_matching() in converge-optimization contains a todo!() — will panic if called
• rsa 0.9.x has a medium-severity advisory (Marvin attack, no upstream fix — transitive via jsonwebtoken)

---

Full Changelog: v3.0.3...v3.7.1

v3.0.3

[3.0.3] - 2026-04-14

Added

• Live endpoint coverage for OpenAI, Anthropic, Gemini, Mistral, Brave, and Tavily, including happy-path and negative credential/model tests
• MistralBackend as a live ChatBackend
• BraveSearchProvider and TavilySearchProvider as WebSearchBackend implementations

Changed

• Unified the live provider path on ChatBackend, with OpenAI, Anthropic, Gemini, and Mistral using the canonical multi-turn request/response types
• Canonical ChatMessage history now carries assistant tool_calls, so real tool loops can round-trip without provider-specific patching
• Anthropic tool-call and tool-result history now round-trip against the live API
• Model selection now filters unavailable providers before choosing a backend and routes chat independently from web search

Fixed

• GitHub Actions now install protoc before protobuf builds
• Provider feature builds now include the async runtime dependency they actually use
• Repo docs and KB pages now describe the canonical chat/search split and Anthropic's provider-native JSON behavior correctly

Removed

• Knowledge lifecycle pack (packs::knowledge) — moved to organism-domain. Includes all knowledge agents (SignalCaptureAgent, HypothesisGeneratorAgent, etc.), invariants (ClaimHasProvenanceInvariant, etc.), and evals (ClaimProvenanceEval, ExperimentMetricsEval).

v3.0.1

Full Changelog: v2.1.2...v3.0.1

Converge v2.1.0 — Kong AI Gateway Provider

New: Kong AI Gateway Provider

KongProvider routes LLM calls through Kong AI Gateway, adding enterprise governance to any upstream model:

• Rate limiting per team/consumer
• PII detection on prompts
• Token usage and cost tracking
• Centralized observability

Usage

converge-provider = { version = "2.1", features = ["kong"] }

use converge_provider::{KongProvider, LlmProvider, LlmRequest};

let provider = KongProvider::from_env("my-route")?;
// Reads KONG_AI_GATEWAY_URL and KONG_API_KEY from env

let response = provider.complete(&LlmRequest::new("Analyze this vendor"))?;

Also available via the provider factory:

let provider = create_provider("kong", "my-route")?;

Full crate list at v2.1.0

All 9 publishable crates updated: converge-traits, converge-core, converge-mcp, converge-provider, converge-experience, converge-knowledge, ortools-sys, converge-optimization, converge-domain.

Converge v2.0.0 — Platform Primitives & Type Safety

Breaking Changes from 1.x

Type System Reconciliation

• ProposedFact: added confidence: f64 and provenance: String — every proposal carries confidence and origin
• AgentEffect: changed from enum to struct { facts, proposals } — agents can emit both facts and proposals in one execution
• Agent::accepts and Agent::execute: take &dyn ContextView instead of &Context
• Agent::dependencies: returns &[ContextKey] instead of Vec<ContextKey>
• TryFrom<ProposedFact> for Fact: type-safe promotion with confidence validation

New: Platform Primitives

• Truth execution: TruthDefinition, TruthKind { Job, Policy, Invariant }, TruthCatalog trait
• Criterion evaluation: CriterionEvaluator trait with four-way CriterionResult (Met { evidence }, Blocked { reason, approval_ref }, Unmet { reason }, Indeterminate)
• Pack-scoped execution: engine.register_in_pack(pack_id, agent) + TypesRootIntent.active_packs
• run_with_types_intent_and_hooks(): single entry point for application-level truth execution
• Honest stopping: StopReason::HumanInterventionRequired { criteria, approval_refs }
• Durable state: ContextStore trait (GAT async) with load_context / save_context
• Event capture: ExperienceEventObserver for run-scoped event observation

Infrastructure

• Owner: Reflective Labs (kenneth@reflective.se)
• License: MIT throughout (79 files migrated from proprietary headers)
• All examples updated for current API
• 1,627 tests passing

Proven in Production

These primitives are exercised by crm.prio.ai with 9 executable truths, 105 tests, and a Svelte/Tauri desktop app with SurrealDB persistence.

🤖 Generated with Claude Code