refactor(mini-swe-agent)

[FatPigeorz]

Summary

Task 5 of the 7-task batch. Expands the mini-swe-agent integration from a 25-line passthrough to a runner that captures the agent's native v2 trajectory, converts it into a structured Trajectory (the same ATIF v1.2 shape harbor uses in MiniSweAgent.populate_context_post_run), and aggregates token usage + cost for downstream eval / RL consumers.

New return shape

result = await client.remote(
    mini_swe.run,
    task=task,
    workdir=workdir,
    agent=agent,
    trajectory_path="/tmp/mini-swe-agent.trajectory.json",
)
# result: MiniSweAgentResult(
#     exit_status=..., submission=..., workdir=...,
#     raw_trajectory={...},                       # mini-swe-agent v2 native
#     trajectory=Trajectory(...),                 # structured ATIF-v1.2
#     usage={"n_input_tokens": ..., "n_output_tokens": ..., "cost_usd": ...},
# )

Module layout

agentix.agents.mini_swe_agent/
├── trajectory.py   Trajectory + Step + ToolCall + Observation + Metrics + FinalMetrics
│                   + from_mini_swe_agent(raw, session_id=...)
│                   + aggregate_usage(raw)
└── runner.py       run(task, workdir, agent, trajectory_path=None, session_id=None)
                    -> MiniSweAgentResult

The structured trajectory rides back to the host inside client.remote(...)'s pickled return value — no shared filesystem assumed. Tool calls are lifted into structured ToolCalls, tool results attach as observations on the preceding agent step, and cost is apportioned by completion-token share (matching harbor's converter).

Example update: examples/run-mini-swe-agent/main.py prints mini_usage and mini_trajectory_steps lines alongside the existing mini_exit_status / mini_submission.

Test plan

• 8 new tests in plugins/agents/mini-swe-agent/tests/test_mini_swe_agent.py cover the structured result shape, trajectory load + parse, inline trajectory passthrough, exception propagation, aggregate-vs-final-metrics consistency, dict / json round trip, and the string-arguments fallback for tool calls.
• Full repo suite: 184 passing.
• pyright clean.
• ruff clean.

Made with Cursor