team mode: structural forcing for task-list coordination (5.6× lift in claim+update events)

[ProKil]

Summary

Adds a host-side gate to the team harness that turns the shared task list into a hard precondition for agent progress. With this gate active, a weak model (Qwen3.5-9B) goes from emitting coordination events on 10% of pairs to 56% — without distillation from a stronger teacher.

How it works

Two helpers in DefaultAgent.execute_actions intercept bash before env.execute:

• _team_required_actions(cmd) — pure decision function. Two rules:
  1. unclaimed-own-task: auto-claim any of my tasks at status=open (unless cmd is itself a claim).
  2. submit-with-in-progress: when cmd contains the mini_swe submit sentinel and any of my tasks is status=in_progress, auto-update it to status=done.
• _team_apply_prefix(cmd) — applies those actions server-side via the host TaskListClient and synthesizes a $ coop-task-claim …\n[auto] claimed: <title> line to prepend to the agent's observation. Server-side application sidesteps the in-container coop-task-* shell CLI which needs the redis python module — not always present in task base images, and the install snippet silently no-ops if pip can't reach the index.
• _team_blocking_reason(cmd) — returns a refusal string when a rule fires that the harness can't auto-fix. Currently one case: a lead (task title starts with "Lead-only") trying to submit while a peer's task isn't status=done. Lead has to wait.

scripts/convert_team_to_coop.py also replaces the misleading P2P {from, to} shape in conversation.json with a broadcast {sender, sender_role, owner} shape. Task-list events aren't directed messages — they're broadcasts on a shared log. bench-runner is sender_role: system, not a peer; create events carry owner (the pre-assigned assignee); claim/update events have no recipient and no more self-loops.

Validation

Qwen3.5-9B + --setting team --team-no-protocol + CooperData v1 (345 has_conflict pairs across 26 repos):

metric	unforced	forced	lift
pairs with ≥1 claim	100 (29.0%)	343 (99.7%)	3.4×
pairs with ≥1 update	49 (14.2%)	194 (56.4%)	4.0×
pairs with both events	35 (10.1%)	194 (56.4%)	5.6×
both-agent test pass	21 (6.1%)	23 (6.7%)	≈ unchanged

Both runs are published side-by-side on HF at CooperBench/team-coop, each with README.md + per-pair index.csv (token counts + coordination-signal columns). Report PR at cooperbench/CooperData#109.

The gate doesn't help or hurt task completion — it only adds coordination structure to trajectories that would otherwise be silent. That's exactly the load-bearing column for downstream training that wants to teach "communicate via the task list."

Test plan

• 28 new unit tests in tests/agents/mini_swe_agent_v2/test_team_coord_gate.py (fakeredis, no docker). Covers all three rules + no-team-mode fallthrough + idempotence + broken-redis tolerance.
• Full suite: 413 passed, 63 skipped.
• ruff check, ruff format --check, mypy clean.
• End-to-end validation on the 345-pair sweep above; trajectories published.

🤖 Generated with Claude Code