Chimera 0.5.0 — Five-Strong Coding-Agent Family

Chimera 0.5.0 — Five-Strong Coding-Agent Family

Released: 2026-04-30

The 0.5.0 release ships the full mink / otter / ferret / weasel / shrew coding-agent family on a single Chimera substrate. v0.4.0 was cut for the mink wave-2 milestone; 0.5.0 rolls up otter waves 1–2, mink waves 2–3, and the wave-5 ship + wave-6 cross-CLI verification + wave-7 gap closure for the three new CLIs (ferret / weasel / shrew).

All five share the same Agent / AgentLoop / EventSourcedSession / provider factory, the same 26-event EventBus, and the same tool registry — adding a sixth is an additive walk through the same agent allocation pattern.

Five coding-agent CLIs on one substrate

CLI	Posture	Lines that distinguish it
chimera mink	TUI-first	glm-5.1:cloud defaults, 31 slash commands, 11 benchmark adapters
chimera otter	Server-first	TUI + HTTP serve + ACP serve --acp, share-by-link, 26 slash commands
chimera ferret	Sandbox-first / IDE-flagship	three sandbox modes, three approval presets, IDE-first ACP, cloud bridge
chimera weasel	Minimal harness, four modes	interactive / print / RPC / SDK, four-command slash palette, auto-discovered extensions
chimera shrew	Small-model tuned	local-first provider chain, 11 curated skills, MoE-aware context sizing, Aider Polyglot + GAIA

Composition over rebuild — none of the new CLIs forks Chimera; each is a thin posture on the upstream substrate.

New CLIs

chimera ferret — sandbox-first / IDE-flagship

Three sandbox modes (read-only / workspace-write / workspace-write-network), three approval presets (read-only / auto / full), an IDE-first ACP transport that is a strict superset of otter’s ACP schema with four extra notification kinds (code/diff, editor/open_file, terminal/output, progress/step), and an optional chimera ferret bridge long-poll HTTPS pipe with bearer auth.

chimera ferret --version                                 # chimera ferret 0.5.0
chimera ferret -p "audit the repo"                       # default: read-only sandbox + read-only approval
chimera ferret --sandbox workspace-write --approval auto -p "fix tests"
chimera ferret serve                                     # ACP JSON-RPC stdio (default)
chimera ferret serve --http --port 5173                  # opt-in HTTP + SSE
chimera ferret bridge --remote-url https://bridge.example.com

13 modules, 303 tests, 8 user docs.

chimera weasel — minimal four-mode harness

A single resolver picks interactive / print / rpc / sdk from --mode, -p, or default. Newline-delimited JSON-RPC 2.0 over stdio (four methods: prompt, cancel, get_state, list_models; all five JSON-RPC error codes exported). Embeddable SDK with sync run, async arun, sync / async streaming, multi-turn chat. Auto-discovered extensions from .weasel/extensions/* (project) and ~/.weasel/extensions/ (user).

chimera weasel                                           # interactive REPL
chimera weasel -p "summarize TODOs"                      # one-shot print
chimera weasel -p "..." --json
chimera weasel -p "..." --stream-json
echo '{"jsonrpc":"2.0","id":1,"method":"list_models","params":{}}' | chimera weasel --mode rpc
python -c "from chimera.weasel.sdk import Agent; print(Agent().run('list files').text)"

Four-command slash palette (/help, /exit, /clear, /model) — no /agent, no /share, no /init by design. 9 modules, 164 tests, 7 user docs.

chimera shrew — small-local-model tuned

Layered on top of weasel. Local-first provider chain inversion (probes llama.cpp at $LLAMACPP_BASE_URL and Ollama at $OLLAMA_BASE_URL before any cloud key). Default qwen3.6-35b-a3b on llama.cpp.

11 curated skill markdowns under knowledge/, protocols/, tools/ with stdlib-only frontmatter parsing. Three small-model-fit extensions:

• moe_offload — MoE-aware context-window sizing.
• scaffold_fit — sub-13B prompt wrapping (idempotent).
• tool_filter — sub-9B tool-list trimming.

Aider Polyglot + GAIA benchmark adapters (chimera shrew bench aider-polyglot --bench-limit 5, chimera shrew bench gaia --level 1). --max-steps defaulted to 30, restricted --allowed-tools=Read,Write,Edit,Bash. 14 modules, 193 tests, 7 user docs.

Otter (waves 1 + 2)

• chimera otter introduced as a sibling to chimera mink — streaming tool calls, hooks, sessions, and the same provider abstraction.
• HTTP server (chimera otter serve --port) and ACP server (chimera otter serve --acp) both wired.
• Slash palette grew to 26 entries; persisted-run inspection (sessions list/show); share-by-link (chimera otter share); preset registry; LSP first-class tools.
• Full O-WIRE-1..6 set landed: real provider, MCP runtime, plugin → agent registry, rules → system prompt, custom commands → slash registry, LSP default-on.

Mink (waves 2 + 3)

• chimera mink runs cost — per-run cost rollups and granular token breakdowns (cache, reasoning, per-step).
• Tau-bench (#90) end-to-end wireup + SWE-bench Verified (#84) adapter scaffold with IPythonTool + condensation hooks.
• 11 benchmark adapters under chimera/eval/benchmarks/ (cline, context, dpai, feature, humaneval-plus, livecodebench, math500, mbpp, swe-polybench, swt, tau-bench).

Server hardening

• TLS termination on otter serve --tls-cert / --tls-key.
• Cooperative cancellation propagation: POST /session/<id>/cancel drains the agent loop without killing the worker.
• Server-Sent Events resume-after-disconnect: Last-Event-ID header honored across reconnects.
• Per-step SSE events plumbed through _drive_agent_streaming over async_run_events; legacy fallback preserved.
• GET /runs and GET /runs/cost HTTP routes — eventlog inspection and per-run cost rollups reachable over HTTP without dropping into the CLI.

Event sourcing

• New chimera/events/sourcing/ subsystem — registry, projector, sqlite_store, sink, types, export.
• Append-only event log with file locking, crash recovery, and gap detection; sessions can be reconstructed deterministically from the log alone.
• SQLite store + snapshot fast-resume in EventSourcedSession — resumed sessions skip log replay when a snapshot is current.
• snapshot_after_turn wired into otter’s REPL + server.

SSH / remote execution

• AsyncSSHEnvironment + SFTP transfer + ProxyJump bastion-host support + control-master multiplexing.
• [ssh] extra; chimera mink --remote ssh://user@host[:port]/path.
• $HOME expansion fix in checkpoint / restore tar paths so remote-environment snapshots round-trip correctly.

Bench

• SWE-bench Verified scaffold landed (chimera/eval/benchmarks/swe_bench_verified.py) with IPythonTool, SWEBenchConfig, and condensation hooks.
• IPython tool + LLM condensation plumbed end-to-end through the SWE-bench Verified harness.
• HumanEval (full 164) live from each of the five CLIs against glm-5.1, kimi-k2.6, deepseek-v4-pro.

Live verification — 5×3 matrix

Exercised end-to-end on 2026-04-30 against three Ollama-Cloud models:

CLI \ Model	glm-5.1:cloud	kimi-k2.6:cloud	deepseek-v4-pro:cloud
mink	math + bash	math + bash	math + bash
otter	math (bash flake)	math + bash	math + bash
ferret	math + bash	math + bash	math (bash flake)
weasel	math + bash	math + bash	math (bash flake)
shrew	math (small-model bash limit)	math (small-model bash limit)	math (auto-deny)

Math row 15/15. Bash row 9/15 (failures concentrated on smaller context windows + auto-deny on shrew’s deepseek path; no regressions attributable to wave-7 work). Total wall-clock 813 s across 30 cells.

Quality

• 5654+ tests passing (5651 in the V1 validation pass + 660 ferret / weasel / shrew tests verified independently).
• 553+ mypy source files clean (Success: no issues found in 553 source files).
• uv run ruff check chimera/ clean.
• All five trademark scrubs clean — bash scripts/all_trademark_scrub.sh exits 0 (passed: 5 (mink otter ferret weasel shrew); failed: 0).
• CI green on Python 3.11 / 3.12 / 3.13 with five per-codename trademark-scrub jobs wired (mink-trademark-scrub, otter-trademark-scrub, ferret-trademark-scrub, weasel-trademark-scrub, shrew-trademark-scrub).

Upgrading from 0.4.x

See Migrating from v0.4 to v0.5 for the step-by-step guide, including the AgentPreset.build() → CodingAgent.from_preset(...) rewrite, the chimera code REPL default flip (and the --legacy-react escape hatch), and the deprecation timeline through v0.7.0.

• No breaking API changes. Existing chimera mink users keep every flag and slash command they had at 0.4.0.
• The four new CLIs (otter, ferret, weasel, shrew) are additive subcommands under the same chimera entry point.
• The chimera/events/sourcing/ subsystem is opt-in — existing sessions built on Session continue to work; switch to EventSourcedSession when you want crash-recovery + snapshot fast-resume.
• The [ssh] extra is opt-in; only required if you use --remote ssh://....

Links

• Source: https://github.com/0bserver07/chimera
• Per-CLI quickstarts: mink · otter · ferret · weasel · shrew