codeswarm

A real, standalone multi-agent software-engineering swarm. Given a coding task with a verifiable success criterion (a pytest oracle), a small team of LLM agents plans, edits code, runs tests, and iterates until the task passes or a budget is exhausted — all inside an isolated sandbox. Every run emits a structured Trajectory (steps, tool calls, outcomes, failures, the final verified verdict): the dense-per-failure-signature raw material a recovery/training plane feeds on.

See DESIGN.md for the authoritative architecture and interface contract.

• v1 has ZERO Omium dependency. It runs entirely on its own with a LocalExecutor (in-process step execution + naive local retry).
• The ONLY seam where Omium plugs in later is codeswarm/workflow/executor.py::Executor. A future OmiumExecutor (added in a separate module at connection time) routes step execution + failure recovery through the Omium SDK. Nothing else imports or knows about Omium.

cd ~/workspace/Omium/codeswarm
python -m pip install -e .            # core is stdlib-only
# optional extras:
python -m pip install -e '.[test]'      # pytest, for the test suite
python -m pip install -e '.[anthropic]' # the real Claude SDK (real runs)

python -m codeswarm tasks                       # list builtin tasks
python -m codeswarm run --task math_utils --mock
python -m codeswarm batch --tasks all --mock --repeat 2

--mock uses the deterministic MockClient, which drives a real plan → code → test → review → verify loop to a passing verdict — including a deliberate first-attempt failure so the trajectory carries failure/recovery signal. Trajectories are written to runs/<run_id>.jsonl.

Provide credentials (ANTHROPIC_API_KEY, or an ant auth login profile) and drop --mock:

python -m codeswarm run --task math_utils --model claude-opus-4-8

The default model is claude-opus-4-8 (override with --model or CODESWARM_MODEL).

Provider openai_compatible speaks the OpenAI /chat/completions wire format over stdlib (no extra install). Default endpoint is Nebius Token Factory (https://api.tokenfactory.nebius.com/v1/), default model MiniMaxAI/MiniMax-M3.

codeswarm's secret convention is environment variables, never files. The Token Factory key lives in AWS Secrets Manager (omium/staging/nebius-token-factory, JSON {"api_key": ...}); fetch it into the env for local runs:

export NEBIUS_API_KEY="$(aws secretsmanager get-secret-value \
  --profile omium-stg --secret-id omium/staging/nebius-token-factory \
  --query SecretString --output text | python3 -c 'import json,sys; print(json.load(sys.stdin)["api_key"])')"

python -m codeswarm run --task math_utils --provider openai_compatible
# or via env instead of the flag:
CODESWARM_LLM_PROVIDER=openai_compatible python -m codeswarm run --task math_utils

Knobs (all env-overridable): CODESWARM_OPENAI_API_KEY (wins over NEBIUS_API_KEY), CODESWARM_OPENAI_BASE_URL, CODESWARM_MODEL / --model, CODESWARM_OPENAI_MAX_TOKENS (default 8192 — reasoning models truncate below that; don't lower it).

Gotcha (live-verified): Token Factory answers a wrong model id with HTTP 200 and empty choices — the client raises on zero choices instead of returning silent empty text.

python -m pytest -q     # green with NO API key (MockClient path)

codeswarm/
  pyproject.toml  README.md  DESIGN.md
  codeswarm/ __init__.py __main__.py cli.py config.py
    llm/     client.py            # LLMClient Protocol, MockClient, AnthropicClient, OpenAICompatibleClient
    trace/   types.py recorder.py # Event/StepResult/Verdict/Trajectory + recorder
    tools/   base.py fs.py shell.py testing.py
    sandbox/ workspace.py         # ephemeral, path-confined temp dir per run
    agents/  base.py planner.py coder.py tester.py reviewer.py
    workflow/ executor.py engine.py state.py   # executor.py = the Omium seam
    tasks/   spec.py builtin/     # real self-contained pytest tasks
  tests/ test_smoke.py            # full engine on a builtin task via MockClient

Each run yields a Trajectory{task_id, run_id, events[], verdict, failure_signature} serialized as JSONL: a meta line followed by one line per Event (kind ∈ agent|tool|test|checkpoint|failure|recovery|verdict). Ordering is a monotonic ts_index — no wall clock in core, so runs are deterministic and replayable.