>cogtrix v0.5.2

Docs/Support/Changelog

Changelog

Cogtrix follows semantic versioning. Breaking changes bump the minor version while pre-1.0; the patch version is for fixes and additive changes.

This page is the canonical changelog — Cogtrix’s source and per-PR history are no longer public, so this summary of highlights per release is what’s available outside the company.

Releases from 0.4.0 onward are cut as squashed release commits, so the upstream per-release notes are thin. The summaries below are drawn from the code that actually changed between tags.

0.5.2 — current (2026-07-09)

  • Package renamed src/ → cogtrix_core/. The old top-level src package polluted the global import namespace. Every Python import moves from src.<module> to cogtrix_core.<module>. This is the breaking change of the release — see the Code Reference for the current surface.
  • The committed OpenAPI spec was removed from the backend repo. The live /api/v1/openapi.json endpoint on a running server is now the authoritative contract.
  • The release process is documented upstream in docs/RELEASING.md.

0.5.1 (2026-06-28)

  • Orchestration hardening — tool-message kinds, tool-argument correction, and graph-runtime fixes.
  • Shell and file-operation tool fixes.

0.5.0 (2026-06-26)

  • MCP runtime in API mode — MCP server lifecycle is now managed inside the API process rather than only by the CLI.
  • Updates across the messaging and web tools (WhatsApp, Telegram, weather, shell, HTTP).
  • The largest release of the 0.4/0.5 line: 115 files, ~8,200 insertions.

0.4.1 (2026-06-21)

  • Background task queue work.
  • The HTTP layer behind the web tools was split into separate fetch and safety modules.
  • Memory-mode and graph-node fixes.

0.4.0 (2026-06-19)

  • RAG relevance scoring — a new scoring stage in the retrieval pipeline.
  • API route and schema additions.

0.3.0 (2026-06-06)

  • Resolver / dispatcher hardening — tighter contracts around message routing.
  • ThreadPoolExecutor migration — background work (knowledge extraction, LLM-judge classification, summarisation) moved off ad-hoc threads onto bounded pools with timeouts.
  • Agent fleet runner — orchestration entry point for fanning a task across multiple agents.
  • Rate-limit productionization — per-chat and per-channel limits now enforced consistently across REST + WebSocket + assistant mode.

0.2.13 (2026-05-31)

Release bundle, v0.2.13.

0.2.12 (2026-05-27)

Bug-fix release; production release-title guard.

0.2.11 (2026-05-24)

  • REST + WebSocket API: 159 endpoints across 27 route groups.
  • WebSocket auth: session WebSocket now accepts JWT via Sec-WebSocket-Protocol: ["bearer", "<token>"] for browser clients, in addition to the existing Authorization: Bearer header for CLI/SDK.
  • 67 built-in tools.
  • Multi-provider: Ollama, OpenAI, Anthropic, Gemini, DeepSeek, xAI; any OpenAI-compatible endpoint.
  • MCP support: connect any Model Context Protocol server as a tool source.
  • Three memory modes: conversation, code, reasoning.
  • Deep Think: Tree-of-Thought reasoning engine via /think.
  • Delegation: parallel sub-agent dispatch via /delegate.
  • Setup wizard: --setup generates a starter config.
  • Docker: official image with both CLI and API mode in the same container.
  • Assistant mode: headless WhatsApp / Telegram daemon with workflow routing.

Earlier releases

0.1.x and the M-series milestones predate this summary and are not covered here.