v0.2.0

Chat overhaul

• New chat flow: POST /chats/{id}/send + GET /chats/{id}/events (Server-Sent Events), with new per-chat endpoints /cancel, /retry, /regenerate, and /export (JSON or Markdown), edit-and-resend (replace_from_message_id), and server-side title search (?q=).
• In the UI: a Stop button while a reply is generating, regenerate, edit-and-resend, copy message / copy code block, chat export, search in the chat switcher, entity hover cards with live previews, and automatic stream re-attach after a page refresh.
• Tool approval: with chat.tool_approval set to always-ask or ask-on-write, chat tool calls pause for your approval in both the web UI and the CLI (fail-closed after chat.tool_approval_timeout_seconds, default 120). The CLI chat was rebuilt on the same engine as the web UI.
• Citation fixes: mixed-reference markers render correctly, hallucinated/unresolved citations are scrubbed, and long multi-hop prompts no longer silently truncate on Ollama.

⚠️ Breaking change

• The legacy POST /chats/{id}/stream endpoint is removed — use /send + /events. Settings keys llm.thinking_auto_detect, llm.chat_interactive_streaming, and chat.enable_response_validation were removed; an older settings.yaml containing them is cleaned up automatically on startup (no action needed).

Fixes & improvements

• Default Ollama URL is now http://localhost:11434 for standalone (pip) installs, overridable via CHAOSCYPHER_OLLAMA_URL. Docker images keep reaching the host's Ollama through host.docker.internal automatically. Existing installs are unaffected.
• chaoscypher health and chaoscypher doctor exit non-zero when checks fail, so they work in scripts and CI.
• CLI audit: dead code removed, ~120 stale help strings corrected, and correctness fixes across chat --tag, db delete, the import pipeline, serve validation, the setup wizard, and Lexicon error handling.
• extract_sync(text=...) / chunk_sync(text=...) now work as documented.
• VRAM preset GPU recommendations corrected (and now pinned to the preset definitions by tests).

Documentation

• A full accuracy audit fixed 300+ inaccuracies across the docs, blog, and READMEs — every documented endpoint, command, flag, and settings key now matches shipped behavior — and added coverage for MCP maintenance mode, parallel workflow DAGs, LLM spend caps, and benchmark v2 scoring. The changelog is now version-anchored.

Install: docker pull ghcr.io/chaoscypherinc/chaoscypher:0.2.0 · pip install chaoscypher-cli==0.2.0

v0.1.1

0.1.1

Redesigned homepage — a "knowledge you can see, trust, and own" framing, a data-crystal hero, a lifecycle screenshot strip, and clickable ecosystem badges. The cited "Why GraphRAG" research now lives in the Search docs.

Model benchmark v2 (chaoscypher-cli) — composite "Overall" scoring, a single models registry as the source of truth, regenerated model cards, and refreshed configs.

Fixes

• core: recover full raw spans for chunk offsets (previously clipped fuzzy windows)
• interface: OUTPUT feeds now load fully (send page_size)