[devin-ai-integration]

Summary

Adds hermetic unit tests (123 tests, 5 new files, test-only — no src/ changes) for the small agents and agent helpers that had little or no coverage. The LLM is replaced by a fake BaseModelArgs dataclass whose make_model() returns a scripted callable, so retry/cost_tracker_decorator/asdict(chat_model_args) run for real. No network, browser, or API keys are used.

File	Tests	Covers
tests/agents/test_agent_utils.py	16 fns (+params)	draw_mouse_pointer / draw_click_indicator (pixel-level: changes stay inside the expected box, center gap untouched, input not mutated), draw_arrowhead geometry, zoom_webpage with a MagicMock page (incl. <= 0 → ValueError)
tests/agents/test_most_basic_agent.py	10	args → agent wiring, prompt contents, CoT toggle, retry on a missing code block, ParseError after 3 tries, experiment_config
tests/agents/test_debug_agent.py	9	AgentArgs base defaults (set_benchmark no-op, set_reproducibility_mode → NotImplementedError), DebugAgentArgs.set_benchmark html toggle, obs_preprocessor (patched browsergym flatteners, deepcopy), get_action via patched input
tests/agents/test_visual_agent.py	37	make_instructions (goal/chat/extra instructions; multi-user-message warning), History, Observation (tabs/error visibility, screenshot vs SoM selection + detail), MainPrompt examples + _parse_answer, VisualAgent.get_action (n_retry inference, busted-retry path, history carry-over), agent_configs, tmlr_config (get_vision_agent/get_som_agent don't mutate BASE_FLAGS)
tests/agents/test_visualwebarena_agent_unit.py	24	image_data_to_uri (ndarray/bytes, alpha drop, bad format), parser, args naming + set_benchmark deepcopy, few-shot building for all 3 observation types (incl. embedded PNGs and the bad-format ValueError), get_action message layout, tab string, PREVIOUS ACTION, SoM overlay

Coverage (target modules)

Module	Before (main, full suite)	After (these files only)
agents/agent_utils.py	0%	100%
agents/most_basic_agent/most_basic_agent.py	0%	100%
agents/debug_agent.py	0%	95% (lines 21-22 are the except AttributeError around an f-string assignment, which can't raise)
agents/visual_agent/visual_agent.py	0%	100%
agents/visual_agent/visual_agent_prompts.py	0%	100%
agents/visualwebarena/agent.py	44%	100%
agents/agent_args.py	71%	100%
agents/generic_agent/tmlr_config.py	50%	100%

Bugs / xfails

I didn't find any source bugs, so nothing is marked xfail. Some behavior is pinned as-is:

• draw_* helpers always return an RGBA image, even when given RGB.
• DebugAgentArgs.set_benchmark only ever turns use_html on. It never turns it off.
• retry appends the assistant reply to the message list in place, so AgentInfo.chat_messages (and MostBasicAgent's prompt_length stat) include the answer and any retry turns.

Run

uv run pytest tests/agents/test_agent_utils.py tests/agents/test_most_basic_agent.py tests/agents/test_debug_agent.py tests/agents/test_visual_agent.py tests/agents/test_visualwebarena_agent_unit.py -p no:cacheprovider -q        # 123 passed, ~4s
# also passes with -n 4 and with pytest-randomly (seeds 1-3)
uv run --with pytest-cov pytest <same files> -p no:cacheprovider -q --cov=src/agentlab/agents --cov-report=term-missing

Coverage used a path-based --cov=src/agentlab/agents. With module-name --cov=agentlab.agents.… flags, pytest-cov imports torch/numpy twice in this env and collection crashes.

Devin-Org: engineering

Link to Devin session: https://devin-gtm.devinenterprise.com/sessions/aa5b999c334747bca5b49bbc615e638f
Open in Devin Desktop: https://devin-gtm.devinenterprise.com/desktop/session/aa5b999c334747bca5b49bbc615e638f?variant=devin
Requested by: @danigrim

---

Devin Review

Status	Commit
⚪ Not started	—

Run Devin Review

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.