Spaces:
Running
Running
Download source/tests/test_pressure.py from munip/collusion-watch: direct link, hf CLI and curl.
- Browser
- Download file 4.2 kB
-
https://huggingface.co/spaces/munip/collusion-watch/resolve/main/source/tests/test_pressure.py
- Command line
-
hf download hf://spaces/munip/collusion-watch/source/tests/test_pressure.py
-
curl -L -o test_pressure.py https://huggingface.co/spaces/munip/collusion-watch/resolve/main/source/tests/test_pressure.py
4.2 kB
| import asyncio | |
| from types import SimpleNamespace as NS | |
| import pytest | |
| from inspect_ai.model import ModelOutput, ModelUsage, get_model | |
| from collusion_lab.core.evidence import EvidenceLog | |
| from collusion_lab.core.models import ModelRoster | |
| from collusion_lab.core.pressure import Pressure, parse_list | |
| def test_parse_shorthand_and_ids(): | |
| assert not Pressure.parse(None).active and Pressure.parse("none").suffix() == "" | |
| p = Pressure.parse("t120") | |
| assert p.wall_seconds == 120 and p.suffix() == "-pt120" | |
| q = Pressure.parse("d5+r10") | |
| assert (q.decision_seconds, q.rounds, q.suffix()) == (5, 10, "-pd5_r10") | |
| assert Pressure.parse({"name": "tight", "max_tool_calls": 40}).max_tool_calls == 40 | |
| names = [x.name for x in sorted(parse_list(["none", "t600", "t120"]), key=Pressure.sort_key)] | |
| assert names == ["t120", "t600", "none"] | |
| with pytest.raises(ValueError): | |
| Pressure.parse("fast") | |
| assert "120 seconds" in p.swe_prompt() and "5 seconds" in Pressure.parse("d5").market_prompt() | |
| def _mock(roster, outs): | |
| for o in outs: | |
| o.usage = ModelUsage(input_tokens=10, output_tokens=5, total_tokens=15) | |
| roster._models["mockllm/model"] = get_model("mockllm/model", custom_outputs=outs) | |
| def test_swe_tool_call_budget_stops_work_and_forces_final_review(tmp_path): | |
| from collusion_lab.scenarios.swe.agents.llm import LLMTeam | |
| from collusion_lab.scenarios.swe.conditions import get | |
| from collusion_lab.scenarios.swe.workspace import Workspace | |
| log = EvidenceLog(tmp_path / "ev.jsonl") | |
| ws = Workspace(tmp_path, NS(id="T3"), log, "ep", max_tool_calls=2) | |
| ws.register(["implementer", "reviewer"]) | |
| roster = ModelRoster("scripted", {"implementer": "mockllm/model", "reviewer": "mockllm/model"}) | |
| msg = lambda i: ModelOutput.for_tool_call("mockllm/model", "send_message", {"to": "team", "text": f"m{i}"}) | |
| _mock(roster, [msg(1), msg(2), msg(3), # third call exceeds the budget | |
| ModelOutput.for_tool_call("mockllm/model", "review", {"decision": "approve"}), | |
| ModelOutput.for_tool_call("mockllm/model", "done", {})]) | |
| team = LLMTeam(ws, get("C0_honest_single"), 0, roster, Pressure.parse("c2")) | |
| assert "2 tool calls" in team._system("implementer") | |
| team._turn("implementer", "work", 10) | |
| assert ws.calls_used == 2 and ws.deadline_hit == "tool_calls" and team.cut_turns == 1 | |
| team._turn("tester", "skipped", 5) # no more work turns after the budget | |
| team._release() # but the reviewer gets a final ship/no-ship turn | |
| assert ws.approvals["reviewer"]["decision"] == "approve" and ws.release_requested | |
| assert any(e.payload.get("event") == "budget_exhausted" for e in log.events()) | |
| def test_market_decision_deadline_keeps_standing_ask(tmp_path): | |
| from collusion_lab.scenarios.market.env import run_market_episode | |
| roster = ModelRoster("scripted", {"seller": "mockllm/model"}) | |
| async def slow(*a, **k): | |
| await asyncio.sleep(2) | |
| raise AssertionError("should have timed out") | |
| roster.generate = slow | |
| rec = run_market_episode("A0_competitive", 0, tmp_path, rounds=30, roster=roster, pressure="d0.2+r3") | |
| assert rec["episode_id"] == "mkt-A0_competitive-s0-pd0.2_r3" and rec["rounds"] == 3 | |
| assert rec["outcome"]["decision_timeouts"] == 15 and rec["outcome"]["timeout_rate"] == 1.0 | |
| assert rec["pressure"]["decision_seconds"] == 0.2 | |
| def test_swe_wall_clock_deadline_cuts_a_slow_turn(tmp_path): | |
| import time | |
| from collusion_lab.scenarios.swe.agents.llm import LLMTeam | |
| from collusion_lab.scenarios.swe.conditions import get | |
| from collusion_lab.scenarios.swe.workspace import Workspace | |
| ws = Workspace(tmp_path, NS(id="T3"), EvidenceLog(tmp_path / "ev.jsonl"), "ep", deadline_s=1.0) | |
| roster = ModelRoster("scripted", {"implementer": "mockllm/model"}) | |
| async def slow(*a, **k): | |
| await asyncio.sleep(5) | |
| roster.generate = slow | |
| team = LLMTeam(ws, get("C0_honest_single"), 0, roster, Pressure.parse("t1")) | |
| t0 = time.monotonic() | |
| team._turn("implementer", "work", 10) | |
| assert time.monotonic() - t0 < 3 and ws.deadline_hit == "time" and team.cut_turns == 1 | |