plan_execute: new two-phase setting (plan then execute)

[akhatua2]

Summary

Adds --setting plan_execute: a two-phase variant of coop where agents first plan (coordinating via messaging + git, writing free-form plan.txt files) and then a fresh set of executor agents runs against each agent's own plan verbatim — no feature spec, no teammate plan, no Phase 1 conversation log.

The Phase 1 prompt explicitly frames the goal as "produce a plan such that your patches don't merge-conflict", making coordination the load-bearing skill being tested. Phase 2 keeps the full coop toolset (messaging + git stay on) so we measure whether good plans are enough or whether mid-execute coordination still matters.

Stacked on #69 (the docker backend + patch.txt flow this builds on).

What's new

• runner/plan_execute.py — new orchestrator. Shared infra setup → Phase 1 with submission_artifact = plan.txt → Phase 2 with each agent's plan as task → eval against Phase 2 patches.
• runner/coop.py — _run_pair_phase() + setup_pair_infra() extracted. execute_coop() composes them once, execute_plan_execute() composes them twice. _spawn_agent() gained task_override / extra_config / log_dir_override / setting_subdir kwargs.
• agents/openhands_agent_sdk/adapter.py — _plan_submission_instructions(is_coop) helper + config-based submission_template / submission_artifact overrides so the same adapter handles both phases.
• cli.py, runner/core.py — plan_execute setting wired into CLI + dispatch.
• eval/runs.py, eval/evaluate.py — discover plan_execute/ log dirs, treat the setting like coop for eval (two patches, merge, test).

Log layout

logs/<run>/plan_execute/<repo>/<task>/<pair>/
  phase1/
    agent1.plan, agent2.plan, agent*_traj.json, conversation.json, result.json
  agent1.patch, agent2.patch       # phase 2 at canonical path → auto-eval works as-is
  agent*_traj.json, conversation.json
  result.json                       # rolled-up cost/steps + phase 2 statuses
  eval.json

Test plan

• CI green (ruff / format / mypy / pytest — 385 pass, 63 skip)
• Smoke on llama_index_task/18813 [1,2]: phase 1 produced a coordinated plan with explicit scope split ("Teammate (agent2) will handle AudioBlock.resolve_audio…"), 8 inter-agent messages. Phase 2 trajectory verified to contain ONLY the plan (feature.md title strings absent). Both patches applied, merge clean, eval ran end-to-end.
• Full 50-pair flash subset run is in flight; will post per-pair comparison against openhands_sdk: local docker backend + patch.txt submission #69's baseline (50/50 Submitted, 9 both_passed = 18%) when it completes.

Scope limit (v1)

Only openhands_sdk is supported. CLI rejects other adapters with a clear NotImplementedError. mini_swe_agent_v2 etc. would need to honour the same config["submission_template"] and config["submission_artifact"] overrides.

🤖 Generated with Claude Code