test: add unit and integration tests for stdlib components (#817)

[planetf1]

test: add unit and integration tests for stdlib components (#817)

Type of PR

• Bug Fix
• New Feature
• Documentation
• Other

Description

• Link to Issue: Fixes test: unit tests for stdlib components (react, unit_test_eval) #817

Adds 32 tests across 3 files for stdlib/components/react, stdlib/frameworks/react, and stdlib/components/unit_test_eval. These are part of the broader #726 testing epic to boost unit and integration coverage for pure logic and reduce reliance on e2e tests.

Bug fix: from_json_file list misalignment

Review feedback identified a bug in TestBasedEval.from_json_file() where examples without user messages caused inputs, targets, and input_ids to fall out of sync. The eval runner (cli/eval/runner.py) does positional lookups across these lists, so misalignment would silently score predictions against the wrong reference targets. Fixed with an early continue to skip examples that have no user messages, with a new test to cover the scenario.

Coverage improvement

All three modules went from 0% to near-full coverage without any e2e infrastructure:

Module	Before	After
mellea/stdlib/frameworks/react.py	0%	100%
mellea/stdlib/components/react.py	0%	97% (1 uncovered line: warning log path)
mellea/stdlib/components/unit_test_eval.py	0%	98% (2 uncovered lines)

Approach: ScriptedBackend for orchestrator testing

The react() framework orchestrator is tricky to test — it coordinates a loop of aact() → _call_tools() → check final_answer. Previously the only options were mock-heavy unit tests (patching mfuncs.aact + mfuncs._call_tools, which mirrors internal call structure and breaks on any refactor) or full e2e tests against a real LLM backend (slow, non-deterministic, needs Ollama/API keys).

Instead, these tests use a lightweight ScriptedBackend fake that subclasses Backend and overrides _generate_from_context() to return scripted responses. The key insight is that real aact(), real _call_tools(), real MelleaTool functions all run — only LLM inference is faked. This means the tests exercise the actual pipeline (context management, plugin hooks, tool execution) and are robust to internal refactors of react().

These are marked @pytest.mark.integration (not unit) because they wire multiple real components together and mock only at the outermost boundary.

Test breakdown

File	Tests	Level	What's covered
test/stdlib/components/test_react.py	10	Unit	ReactInitiator format/parts/parse, ReactThought format/parse, finalizer override edge case
test/stdlib/components/test_unit_test_eval.py	16	Unit	Pydantic validation, format_for_llm, set_judge_context branching (0/1/N targets), from_json_file with edge cases (multi-user messages, mixed user messages, non-assistant filtering)
test/stdlib/frameworks/test_react.py	7	Integration	Loop termination, budget exhaustion, non-final tool continuation, model_options tool merge, format branch, multi-tool rejection, context type guard

Testing

• Tests added to the respective file if code was changed
• New code has 100% coverage if code as added
• Ensure existing tests and github automation passes (a maintainer will kick off the github automation when the rest of the PR is populated)

[github-actions]

The PR description has been updated. Please fill out the template for your PR to be reviewed.

[ajbozarth]

I didn't have bandwidth to do a deep review, but I did run the tests and have Claude review, which gave a couple inline comments:

ran uv run pytest test/stdlib/components/ test/stdlib/frameworks/:

100 passed, 1 skipped, 1 xpassed, 15 warnings in 159.23s (0:02:39)

All new tests passed

[ajbozarth]

Changes LGTM and Claude approve so I'm going to give an approve on this since in general more tests are better than none,