Search Documentation

Find and navigate to documentation pages and their content

Release Notes

All notable changes to Codistry, the privacy-first AI coding assistant, are documented here.

[0.6.9] - 2026-08-03

Added

  • Detect models from your endpoint - the Anthropic and OpenAI provider cards can now enumerate the models your configured endpoint actually serves ("Detect models from this endpoint"), and the custom model form gained a matching "Browse available models" action. Pick the models you want from a checklist, and fine-tune each one (API style, reasoning, image input, temperature, context and output limits) in a built-in per-model editor before applying. Works with corporate gateways, including mTLS-protected endpoints.
  • Advanced context tools (beta, off by default) - an experimental setting (Index tab → Experimental Settings) that gives the agent finer-grained tools for exploring your code indexes like a map: a whole-codebase overview, cluster and concept-area browsing, and reverse file lookup. Noticeably improves navigation for smaller models.
  • New models: Claude Opus 5 (Anthropic), and OpenAI's GPT-5.6 family — Sol, Terra, and Luna. The retired gpt-5.1-codex models migrate to the matching GPT-5.6 tier automatically.

Changed

  • Relational Index builds now run in a background worker - building or refreshing the knowledge graph no longer competes with the editor UI, so large builds don't cause stutter.
  • Skills runtime image - rebuilt and simplified, reducing image to ~1.2GB, down from ~1.9GB.
  • The Custom Model form was reorganized: credentials now come before model detection (the scan uses the values on screen), and mTLS settings sit directly after the credentials they secure.

[0.6.8] - 2026-07-22

Added

  • Semantic Index (optional) - Relationship Index builds can now also embed your code symbols for meaning-based retrieval: the right files reach the agent more often, with fewer turns per task. Choose a built-in offline model, OpenAI (works with any OpenAI-compatible endpoint via openaiApiUrl), a local Ollama server (server URL configurable in the Index tab or via the contextEngineOllamaUrl setting), or any OpenAI-compatible /v1/embeddings server — LM Studio, vLLM, llama.cpp and friends — with its own server URL, no API key required, and fully decoupled from your chat provider's endpoint. Rebuilds re-embed only what changed.
  • Query rewriting (optional, off by default) - with the Semantic Index on, a small, fast LLM call at task start rewrites your question to improve accuracy before searching. Advanced settings let you customize the embedding and rewrite models.
  • Test configuration button in the Index tab - runs a real embedding call (and a real rewrite call when query rewriting is on) against your current values, so you can verify keys, models, and endpoints before rebuilding the index.
  • Symbol summaries (experimental, off by default) - index builds can write one-line purpose summaries for the most connected symbols, and retrieval uses them to match your question against what code does, not just what it's named.

Changed

  • Context pack auto-inject is now workflow-aware: the orientation seed is skipped when your request already names specific files or asks for a direct code change, saving tokens where the pack doesn't help.
  • The agent is now told when a knowledge graph is already built, so it no longer wastes turns rebuilding it.

Fixed

  • On very large repositories, capped knowledge-graph builds no longer silently drop whole directories (files are now trimmed evenly across the repo instead of alphabetically), and a partial graph is clearly reported in the build summary and context packs.

[0.6.7] - 2026-07-18

Fixed

  • mTLS client certificates are now reliably presented on every request made from the VS Code extension — previously the extension host's patched networking could silently drop the configured client certificate, causing gateways to reject requests with "No required SSL certificate was sent". All model providers, custom model types, and embedding requests are covered.

Added

  • mTLS now works together with corporate HTTP proxies — requests honor HTTP_PROXY/HTTPS_PROXY/NO_PROXY and present the client certificate through the proxy tunnel.

[0.6.6] - 2026-07-16

Added

  • GPT‑5.6 model family (Sol, Terra, Luna) with reasoning, tools, and images.
  • Enhanced Relational Indexing (optional) — the Relational Index can name code areas with your model for a deeper structural understanding.

Changed

  • Reworked the Context Engine graph visualization — clearer clusters, better labels, resizable sidebar.

[0.6.5] - 2026-07-09

Maintenance release — no user-facing changes.

[0.6.4] - 2026-07-09

Added

  • OpenRouter support — add OpenRouter models with a browsable catalog picker and automatic pricing.
  • mTLS client certificates (.p12) for providers and custom models behind high-security gateways.

[0.6.3] - 2026-07-08

Added

  • Expertise tab — Codistry learns which areas of the codebase you own from git history and defers to you there.
  • Work suggestions on the chat welcome screen, generated from your recent commits.

Fixed

  • Mermaid diagram rendering issues.

[0.6.2] - 2026-07-02

Fixed

  • Security updates across bundled dependencies.
  • Sessions sidebar polish (scrollbar behavior, rail layout).

Changed

  • Restored the Claude Fable 5 model.

[0.6.1] - 2026-06-23

Fixed

  • Improved tool-calling reliability for OpenAI models when Skinny Tools is enabled.

[0.6.0] - 2026-06-18

Added

  • Context Engine — Codistry now builds a local index of your codebase to give the agent more relevant context automatically and on-demand.
  • Chat Sessions — run several conversations at once, and name, pin, reorder (drag), and queue messages per session.
  • Subagents — the agent can delegate self-contained tasks to focused sub-agents.
  • Dynamic Skills — repo-specific skills that live in your project and are discovered automatically.
  • Japanese and Korean language support across the UI and native prompts and notifications.
  • Personalization (renamed from "Context") — Codistry now checks into git architectural insight and learnings to tailor the agent to you and your project.

Changed

  • Default model is now Claude Sonnet 4.6.
  • Skinny Tools is on by default for more efficient tool calling.
  • Tool steps now show descriptive names instead of raw tool names.
  • Settings UI cleanup and general UI performance improvements.
  • When auto-update is off you're notified of new commits; when it's on, you're told so.
  • Removed deprecated models.

Fixed

  • Indexing now works smoothly even when no Anthropic or OpenAI credentials are configured.
  • Improved embedding reliability.
  • Improved conversation summarization and handling of large tool results.
  • Improved stability of "thinking" output for open models.
  • Smoother streaming of step-by-step output.
  • More robust and responsive model streaming.

[0.5.7] - 2026-06-03

Changed

  • Improved Model Context Protocol (MCP) support for connecting external tools and data sources.

[0.5.6] - 2026-05-29

Added

  • Support for the Claude Opus 4.8 model.
  • A dedicated step view for grep searches, so search progress and results appear inline in the conversation.

Changed

  • Faster, more reliable file and text (grep) search.
  • The agent now shows a diff when it updates your Worldview, so you can see exactly what changed.
  • Added loading spinners while diagrams are generated and while your Worldview is updating.

Fixed

  • More accurate handling of context between conversation turns.
  • Improved cleanup of Model Context Protocol (MCP) servers when a session ends.

[0.5.5] - 2026-05-07

Fixed

  • Improved handling of large tool results within the model's context window.