Aviary agent max_timesteps and fixed test_gather_evidence_rejects_empty_docs

[jamesbraza]

Since its addition in #309, test_gather_evidence_rejects_empty_docs was failing for the wrong reason. It was supposed to fail due to empty docs, but it was failing due to a bad patch:

ERROR    paperqa.agents.main:main.py:237 Agent <aviary.tools.utils.ToolSelector object at 0x10d5a73b0> failed.
Traceback (most recent call last):
  File "/path/to/paper-qa/paperqa/agents/main.py", line 194, in run_aviary_agent
    obs, tools = await env.reset()
                 ^^^^^^^^^^^^^^^^^
  File "/path/to/paper-qa/paperqa/agents/env.py", line 136, in reset
    self.state, self.tools = self.make_initial_state_and_tools()
                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/path/to/paper-qa/paperqa/agents/env.py", line 123, in make_initial_state_and_tools
    self.tools = settings_to_tools(
                 ^^^^^^^^^^^^^^^^^^
  File "/path/to/paper-qa/paperqa/agents/env.py", line 71, in settings_to_tools
    tool = Tool.from_function(
           ^^^^^^^^^^^^^^^^^^^
  File "/path/to/.venv/lib/python3.12/site-packages/aviary/tools/base.py", line 339, in from_function
    raise ValueError(f"Missing description for parameter {pname}.")
ValueError: Missing description for parameter args.

This PR:

• Properly propagates the __doc__ onto the patched gen_answer, so Tool.from_function succeeds and the test fails as expected
• Implements max_timesteps so the test doesn't cycle forever until it times out
  • Which partly resolves Moving timeout and max steps into PaperQAEnvironment itself #478
• Moves from AgentStatus.TIMEOUT to a more general AgentStatus.TRUNCATED to match aviary's general terminology

[mskarlin]

looks good to me -- couple q's on naming

[mskarlin]

jc -- why are we calling the agent steps timesteps? I think of timesteps from physical models (i.e. molecular dynamics) where each iteration is a unit of time, like 4 picoseconds or something. This is more like actionsteps in my mind.

[jamesbraza]

I hear what you're saying.

I think one can also encounter vagaries with "step", for example one can wonder does "step" mean:

• One step = agent selection + environment step
• Two steps = agent selection + environment step

I went with timestep to exactly match ldp.data_structures.Transition.timestep

[jamesbraza]

Let me know if you think we should name it otherwise