Chimera v0.3.0 — coding, intelligence, resilience, interop

Chimera v0.3.0 — the M13 cycle: coding, intelligence, resilience, interop.

One thesis runs through all of it: make a weak/cheap model perform like a frontier one — with proof. Panel fusion is for combining strong models; everything new here is for lifting a single weak one. (Chimera is a self-evolving agent with an LLM-Fusion reasoning core.)

⌨️ Coding — locate, read, edit surgically

• Surgical editing: edit_file (exact, unique-anchored replace) + apply_patch (multiple SEARCH/REPLACE hunks, atomic). Edits in place instead of rewriting whole files; a missing/ambiguous anchor is refused, not guessed.
• Repo-map (solve --repo-map): a structural table of contents (files → top-level symbols, via ast) so the agent jumps to the right file instead of exploring blind.
• read_document: ingest PDF/DOCX/PPTX/XLSX/HTML/CSV/EPUB as Markdown (opt-in [documents] extra).

🧠 Intelligence — the weak-model lift

• Progress ledger (--progress-ledger): a structured per-attempt self-check (Magentic-One's inner loop) turns "it failed" into a concrete next-focus, so a weak model stops re-trying the same dead end.
• Completion contracts (--contract): declared, machine-checkable success clauses as an AND gate on top of verify-or-revert — catches the model narrating success it didn't achieve.
• Dual-ledger re-plan (--replan): on a stall, record why it's stuck and rebuild the plan from that accumulated cause (outer loop) — the retry is fundamentally different, not reworded.
• Self-consistency (fuse --best-of N): cheap single-model fusion — sample one model N times and take the consensus.

🔁 Resilience & interop

• Streaming: token primitive + a typed event stream from the solve loop; solve --stream shows live progress.
• A2A message/stream (SSE): a LangGraph/CrewAI orchestrator sees progress without polling; the agent card advertises streaming: true.
• Durable execution (solve --thread <id>): checkpoint to SQLite; a crash mid-run resumes from the last checkpoint, repeating no verified work.
• Human-in-the-loop (--pause-on-taint, --approve/--deny): a run that consumed untrusted content pauses for sign-off before finalizing — the safety valve for the lethal trifecta.

🌐 Doing, not just answering

• Browser navigation (opt-in [browser] extra): a stateful browser tool drives a real Chromium via the accessibility tree — elements get stable refs, so the model clicks/types by ref, no vision model. Page content is data-fenced and browsing taints the run.

The fusion-as-reasoning core and the governance layer (taint provenance, quarantined reader, red-team) from v0.2 are unchanged — this cycle closed the execution gaps around them.

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

Full Changelog: v0.2.0...v0.3.0