feat(agents): Add MiMo Code (mimo) ACP agent

[Yiminnn]

Summary

Adds mimo (Xiaomi MiMo Code, npm @mimo-ai/cli) as a first-class ACP agent, layered on the v0.6.0 agent stack (#665).

MiMo Code is an OpenCode fork that ships a native mimo acp JSON-RPC stdio server — its initialize handshake reports agentInfo.name="OpenCode" — so this lands as a registry-only change mirroring the existing opencode entry, exactly as the registry docstring prescribes ("adding a new agent is a registry-only change"). No core edits, no shim, no if agent == "mimo" special cases.

Changes

• src/benchflow/agents/registry.py — AGENTS["mimo"]:
  • install_cmd=_js_agent_install("mimo", "@mimo-ai/cli@0.1.0") (pinned; isolated /opt/benchflow node prefix)
  • launch_cmd=_js_agent_launch("mimo", "acp"), protocol="acp", acp_model_format="provider/model" (models.dev ids, OpenCode lineage)
  • env_mapping maps both BENCHFLOW_PROVIDER_BASE_URL→OPENAI_BASE_URL and BENCHFLOW_PROVIDER_API_KEY→OPENAI_API_KEY (codex-acp precedent) so the non-proxy path needs no core edit
  • web tools disabled via ~/.config/mimocode/mimocode.json (tools.webfetch=false, opencode precedent)
• tests/integration/configs/mimo.yaml — integration config; defaults to mimo/mimo-auto, MiMo's free no-account channel (works headless in-sandbox; the runtime skips the LiteLLM proxy and delivers the native model id via ACP set_model)

Evidence (live on a GCP VM, Daytona sandbox, v0.6.0rc6)

• tests/test_registry_invariants.py + tests/test_agent_registry.py: 171 passed (the new entry is auto-covered by the executable registry schema, incl. the JS-isolation invariants)
• ruff check / ruff format --check / ty check: clean; generated install_cmd passes dash -n
• Stdio ACP handshake verified: initialize → protocolVersion 1, agentInfo "OpenCode"
• Live rollouts on mimo/mimo-auto (SkillsBench task.md tasks): 4/4 healthy (tools>0, tokens>0, reward non-None), 2 solved — jax-computing-basics r=1.0 (8 tools, 26k tok), threejs-to-obj r=1.0 (6 tools); data-to-d3, weighted-gdp-calc healthy fails

Notes for reviewers

• The free mimo-auto channel needs no credentials; for MiMo's flagship models, xiaomi/mimo-v2.5-pro routes through the existing xiaomi provider (XIAOMI_API_KEY/XIAOMI_BASE_URL).
• Proxy-alias models (e.g. openai/benchflow-…) are rejected by the opencode family's models.dev validation (ProviderModelNotFoundError) — this affects opencode identically and predates this PR; the integration config therefore defaults to the native channel.

How to test

uv run python -m pytest tests/test_registry_invariants.py tests/test_agent_registry.py -q
uv run bench agent show mimo
uv run bench eval create --tasks-dir <tasks> --include jax-computing-basics \
  --agent mimo --model mimo/mimo-auto --sandbox daytona --skill-mode no-skill \
  --concurrency 1 --jobs-dir jobs/mimo-smoke

[xdotli]

Reviewing for the v0.6 release — the agent registry change itself looks good, but a review (incl. a tracing of the integration-suite wiring) found 3 gaps in tests/integration/run.sh that would keep mimo from actually running in the default suite and cause a config-audit mismatch. Flagging so this can land cleanly:

1. mimo missing from ALL_AGENTS (run.sh:49-58) — a no-arg run.sh never runs the new agent, so the shipped mimo.yaml is dead config in the normal suite path.
2. model_for_agent has no mimo case (run.sh:41-47 + check_results.py:728-730) — run.sh ignores the YAML model: and falls to the default gemini-3.1-flash-lite-preview, while config.json records xiaomi/mimo-v2.5-pro from the YAML → guaranteed model-mismatch audit failure.
3. has_creds_for has no mimo case (run.sh:95-109) — it falls to has_gemini_key/GEMINI_API_KEY, but mimo needs XIAOMI_API_KEY/XIAOMI_BASE_URL (per the YAML comment), so the suite green-lights launching mimo with the wrong credential gate and the eval then fails closed in resolve_agent_env.

Each is a small run.sh/model_for_agent/has_creds_for addition. Happy to help wire these if useful — leaving the PR to you since it's yours.

[Yiminnn]

@xdotli all three gaps closed, plus two deeper blockers your model-spec fix surfaced — and the lane is now live-verified end to end. Also rebased onto latest main (001da3e) and retargeted the PR there, so the real CI gate runs.

Your 3 run.sh gaps (ea92d688):

• mimo added to ALL_AGENTS
• model_for_agent emits xiaomi/mimo-v2.5-pro verbatim (check_results compares by string equality)
• has_creds_for gates on XIAOMI_API_KEY and XIAOMI_BASE_URL (both required — resolve_agent_env fails closed without either)

Blocker 1 — non-proxy modelId mangling (ea92d688): strip_provider_prefix removed the registered xiaomi/ prefix and the models.dev heuristics fell through to anthropic/mimo-v2.5-pro (nonexistent route, silent except a log warning). Fixed with a ("mimo","xiaomi") heuristic entry; both shapes pinned in test_litellm_hardening (xiaomi/... passes through; proxy aliases still format to openai/benchflow-*).

Blocker 2 — fixed-endpoint auth (f51d340c): the CLI pins its native xiaomi provider to the standard platform endpoint, so token-plan/regional keys fail with "Invalid API Key". Shipped via the existing CredentialFile hook: when XIAOMI_API_KEY is set, benchflow writes ~/.config/mimocode/mimocode.json overriding provider.xiaomi.options with {env:XIAOMI_BASE_URL}/{env:XIAOMI_API_KEY} references (resolved by the CLI at runtime — no secret materialized in the file). mimo.yaml runs with usage_tracking: "off" since the CLI rejects LiteLLM proxy aliases (openai/benchflow-*) against its catalog — the native id must go through directly.

Live evidence (Daytona, citation-check, legacy layout):

• mimo + xiaomi/mimo-v2.5-pro: PASS, reward 1.0, 12 tool calls (model id arrives intact, token-plan endpoint authenticated)
• control openhands + deepseek/deepseek-v4-flash: PASS, reward 1.0, 27 tool calls

Also added the conventional per-agent tests (web-policy setup_cmd writes mimocode.json tools.webfetch=false + snippet coverage) and the docs updates (integration-tests.md 8→9 agents + creds row).

One heads-up beyond this PR: skillsbench upstream just converted all tasks to native task.md (64395c1, skillsbench@1.2 draft) — main's legacy-only discovery can't see those packages, which makes landing the v0.6 task.md line more pressing. The live proofs above used the pre-migration legacy layout.