fix: reset per-attempt cost in SetStatusAgentHook.on_setup_attempt - #1568

Reference Issues/PRs

Fixes #1567

What does this implement/fix? Explain your changes.

Summary

SetStatusAgentHook renders the per-instance status line in the batch progress UI. On a retry attempt it double-counted the previous attempt's spend until the new attempt's first step completed, so the status line showed roughly 2x the actual spend (e.g. Attempt 2 Step 1 ($2.00) when only $1.00 had been spent).

Why (root cause)

sweagent/run/run_batch.py:347 attaches a SetStatusAgentHook to every batch agent. With a retry loop, RetryAgent._next_attempt (sweagent/agent/agents.py:321) re-attaches the same hook objects to a fresh DefaultAgent, whose setup fires on_setup_attempt (sweagent/agent/agents.py:594) at the start of every attempt. In on_setup_attempt (sweagent/agent/hooks/status.py:16-20) the finished attempt's cost was folded into _previous_cost:

# Costs will be reset for the next attempt
self._previous_cost += self._cost

but self._cost was never reset, even though the comment says it would be. Because the status line shows ${self._previous_cost + self._cost:.2f} (sweagent/agent/hooks/status.py:29), every status line of the new attempt displayed the previous attempt's spend twice until on_step_done (sweagent/agent/hooks/status.py:31) overwrote _cost with the new attempt's instance_cost. The previous+current split only yields the true total spend if _cost is zeroed at the attempt boundary.

Changes

• sweagent/agent/hooks/status.py: reset self._cost = 0.0 in on_setup_attempt right after folding the finished attempt's cost into _previous_cost, so _previous_cost + _cost is the true total spend from the first status line of a retry attempt onward. For non-retry agents this is a no-op (_cost is still 0.0 at the single setup call).
• tests/test_agent_hooks.py: new regression test that drives the hook through two attempts and asserts the status line shows $1.00 (not $2.00) at the start of attempt 2, and $1.25 after attempt 2's first step adds $0.25.

Testing

New regression test, failing on main (3ea751c):

$ python -m pytest tests/test_agent_hooks.py -v
>       assert statuses[-1] == "Attempt 2 Step   1 ($1.00)"
E       AssertionError: assert 'Attempt 2 Step   1 ($2.00)' == 'Attempt 2 Step   1 ($1.00)'
1 failed

Passing with this fix:

$ python -m pytest tests/test_agent_hooks.py -v
tests/test_agent_hooks.py::test_status_hook_cost_not_double_counted_on_retry PASSED [100%]
1 passed

Existing tests around the changed module:

$ python -m pytest tests/test_agent.py tests/test_agent_hooks.py -v
12 passed, 2 xfailed            # both xfails pre-existing (test_run_autosubmit, test_successful_submission)

$ python -m pytest -m "not slow" tests/test_run_batch.py tests/test_batch_instance.py -v
8 passed, 3 deselected           # deselected = slow tests that spin up real batch runs

Lint/format (pre-commit config, ruff):

$ ruff check sweagent/agent/hooks/status.py tests/test_agent_hooks.py
All checks passed!
$ ruff format --check sweagent/agent/hooks/status.py tests/test_agent_hooks.py
2 files already formatted

Note: in the sandbox used for this work, the GitHub-API/network-dependent tests (tests/test_run_hooks.py should_open_pr tests, tests/test_env_utils.py::test_get_associated_commit_urls, tests/test_run_replay.py::test_replay) fail or stall identically on unpatched main, so they could not be exercised here; they are unaffected by this change (no code path they use was modified).