Releases

Operator cockpit TUI (pc-tui) for the polychrome control plane: fleet, transcript, approvals, and tools panes.

Polychrome SDK umbrella: re-exports the core polyc-* libraries (edge adapter, wire types, agent loop, and LLM provider) so a consumer can depend on one crate.

Vulkan API bindings generated entirely from vk.xml, with a complete safe RAII wrapper covering compute and graphics: instance/device/queue, buffer, image, sampler, render pass, framebuffer, graphics + compute pipelines, swapchain, a VMA-style sub-allocator with TLSF + linear pools and defragmentation, sync primitives (fences, binary + timeline semaphores, sync2 barriers), query pools, and optional GLSL/WGSL/HLSL→SPIR-V compilation via naga or shaderc. Supports Vulkan 1.2.175 onward — swap vk.xml and rebuild.

Browser session establishment (ADR 0007): one-time login challenges, the shared session cookie, and both the persona-passkey and wallet-passkey ceremonies that mint a polyc_crypto::session token. Consumed by polyc-control-plane; carries no ceremony-consumer naming of its own.

功能完整的 A 股数据获取库，完全对标 pytdx

Core primitives for thingd, an object-shaped local memory engine for apps and agents.

Read layer over the event log: a DataFusion engine for SQL over replayed partitions, and a per-conversation Parquet projection for participation-scoped search.

The durable session-family store: rotation, reuse detection, and expiry for long-lived sessions rooted in a one-time proof. Knows nothing about how a family's root was proven, or about the transport that carries its grant.

Shared projection artifact format, admission, and realm-typed exact-read contract

Post-turn persona orchestration: memory extraction and the shared pass-agnostic supervisor (#459).

Per-persona memory: types + a pure fold over the shared eventlog primitive's events.

Persona directory: durable principals for edge callers on a commonware-storage authenticated database.

Machine Payments Protocol (MPP/Tempo) integration for polychrome: the control-plane composition/glue layer over the standalone outbound, inbound, wallet-delegation, egress, and spend-policy primitive crates, plus the payment proxy/wallet views.

OpenAI-compatible HTTP serving for the memra CUDA inference engine - single-GPU multi-model step-interleave scheduling on RTX 50-series

Every memory and recall backend for harness-rs behind one door: open_memory(url) / open_recall(url) pick the store from the URL. JSONL files with no features; SQLite FTS5, SurrealDB (embedded or remote, BM25 + HNSW hybrid) and CortexDB each behind a feature, so you compile only the store you run.

Outbound-mail transport: a ConnectRPC client the control plane uses to send a fresh (non-threaded) email through this deployment's web Worker (EmailRelayService).

From-scratch CUDA LLM inference engine for NVIDIA RTX 50-series (sm_120a) and Hopper (sm_90a) - custom kernels, no frameworks

polychrome A2A edge: serves a domain-signed Agent Card and drives message/send tasks onto a turn.

Thin connectrpc client over the generated AgentServiceClient, shared by the CLI and Slack receiver.

State plane transport adapter: capability-specific Connect clients and server-trait glue mapping the generated wire types onto the polyc-state kernel — typed outcomes, per-call admission, and the conformance surface the authenticated shell proves itself against (docs/proposals/separated-planes.md).

polychrome turn-runner: run one agent turn from a wire request against an injected provider + tool executor.

The shipped Agent catalog (manifests/base/agents.yaml) as one typed artifact.

Conversation CRD + kube reconciler for the polychrome control plane.

The in-process tool core for polychrome agents: local executors (coding, web fetch, wallet, ...), the tool registry, and MCP composition. The networked connectors live in polyc-connectors.

The agent turn loop: provider + tool-call routing, shared by the control plane and harness.

Durable turn-persistence host: a dedicated-thread Commonware event-log bridge for the tokio control plane (#459).

LlmProvider backend over the OpenAI-compatible /v1/chat/completions API (Ollama, LM Studio, vLLM, llama.cpp, OpenAI).

Offline PackageURL validator using a prebuilt FST of known packages.

Append-only conversation event log on a commonware-storage journal.

Storage-agnostic event, integrity, trust, and navigation model for Polychrome journals.