Skip to content

Chimera v0.2.0 — security, adoption, intelligence, interop

Latest

Choose a tag to compare

@brcampidelli brcampidelli released this 06 Jul 02:03

Chimera v0.2.0 — the security → adoption → intelligence → interop cycle.

An open-source, self-evolving AI agent whose reasoning core is an LLM-Fusion engine (panel → judge → synthesizer). This release makes the security story measured, the setup out-of-the-box, memory & skills measurable, and adds the two winning agent protocols.

🔐 Security — measured, not asserted

  • Quarantined reader (dual-LLM / CaMeL): a tool-less model extracts only schema-validated fields from untrusted content, so injected instructions never reach the privileged agent.
  • Taint provenance on memories & learned skills — a skill distilled during a tainted run is held for review (the "Zombie Agents" anti-poisoning defense). A taint-adaptive allowlist narrows dangerous tools once a run is tainted; fetched content is data-fenced.
  • Red-team suite (chimera redteam) reports the injection attack-success rate with vs without defenses: 100% → ~14% on the built-in corpus — and it names the remaining leak (exfiltration via an allowed tool) instead of claiming 100%.

🚀 Adoption — works out of the box

  • chimera init — one command: create .env, set a key, verify, point at a real example.
  • Consumer recipes: email_triage, research_brief, repo_watchdog + an MCP guide.
  • Task-oriented docs site (mkdocs-material).

🧠 Intelligence — measurable

  • chimera memory-bench — recall@k vs scale, lexical vs paraphrase. Surfaces the honest keyword ceiling (lexical 1.00 holds at 1000 facts, paraphrase 0.00).
  • Opt-in semantic recall — embeddings bridge paraphrases; FTS fallback on any embedder error. Measured lift: paraphrase 0.00 → ~1.00.
  • Skill retirement — under-performing skills proposed-with-review (reversible, never a silent delete).

🔌 Interop — speaks both winning protocols

  • Chimera as an MCP server (chimera serve --mcp): solve / fuse / memory-search as MCP tools for any client (Claude Desktop, IDEs, other agents).
  • A2A adapter (chimera serve --a2a): Agent Card + JSON-RPC task lifecycle, so a LangGraph/CrewAI/AutoGen orchestrator can delegate a task to Chimera. (Synchronous core; streaming/push not yet implemented.)

691 tests, mypy --strict clean, ruff clean. Apache-2.0. Full detail in CHANGELOG.md.

Full Changelog: v0.1.0...v0.2.0