Fix two PR #1543 tests with wrong assertion targets blocking main CI

[JSv4]

Summary

Two tests added in #1543 have been failing on main and blocking pytest (and therefore codecov) on every PR. Both tests assert against the wrong target:

1. test_callable_resolves_via_from_function (test_llms_typing_coverage.py) patches opencontractserver.llms.api.ToolAPI.from_function, but _resolve_tools calls CoreTool.from_function directly. The mock never intercepts, so _resolve_tools returns real CoreTool instances and assertEqual(result, [sentinel]) fails. Patch CoreTool.from_function instead.

2. test_add_document_note_tool_passes_none_when_corpus_absent (test_llms_typing_behavior_guards.py) does a global file scan with assertNotIn("requires the agent to be scoped to a corpus"), but that string legitimately exists for add_exact_string_annotations (a different tool that does require a corpus). Scope the assertion to the body of add_document_note_tool by slicing from its def to the next async def at the same indent level.

Test plan

• pytest opencontractserver/tests/test_llms_typing_coverage.py::TestResolveToolsCallerForms::test_callable_resolves_via_from_function passes
• pytest opencontractserver/tests/test_llms_typing_behavior_guards.py::TestAddDocumentNoteToolCorpusOptional::test_add_document_note_tool_passes_none_when_corpus_absent passes
• Backend CI on this PR is green (was red on main and every downstream PR before this fix)

[claude]

Code Review

Overview

This is a targeted two-test fix that unblocks CI after #1543 introduced tests with incorrect assertion/patch targets. No production code is changed — both changes are test-only. The root cause analysis in the PR description is accurate.

---

Fix 1 — test_callable_resolves_via_from_function (correct)

Root cause confirmed. api.py:713 shows _resolve_tools calling CoreTool.from_function(tool) directly:

resolved.append(CoreTool.from_function(tool))

ToolAPI.from_function is a thin wrapper that ultimately calls CoreTool.from_function, but _resolve_tools bypasses ToolAPI entirely. Patching ToolAPI.from_function was therefore a no-op at the real call site.

The fix is correct. Patching opencontractserver.llms.api.CoreTool.from_function intercepts the right symbol, and from_fn.assert_called_once_with(my_tool) validates the call arguments too — good.

---

Fix 2 — test_add_document_note_tool_passes_none_when_corpus_absent (correct, minor fragility concern)

Root cause confirmed. pydantic_ai_agents.py:2497 contains:

"add_exact_string_annotations requires the agent to be scoped to a corpus"

A global file scan for that substring false-positives on the other tool. The fix is correct in concept.

The body-extraction logic works for the current code and the fallback (if next_def != -1 else text[body_start:]) handles the last-function case.

Minor concerns:

1. Hardcoded indentation string ("\n async def " = 8 spaces). This works because add_document_note_tool lives inside a method that nests it at that indent level. If someone reformats that file or refactors the nesting, this search will silently return a wider/narrower body than intended, without a test failure alerting to the breakage. A comment noting why 8 spaces are expected would add future-reader value (the PR already has an inline comment, which is good).

2. Quote change in assertNotIn — the old search was 'requires the agent to be scoped to a corpus"' (note trailing "), clearly an attempt to match a string literal's closing quote. The new version drops that trailing quote. This is strictly more robust (won't break if the quote style changes), so the change is a net improvement.

3. Approach fragility in general — this class of test (scanning source text rather than exercising runtime behavior) is inherently brittle. That said, it's the right trade-off when the behavior to guard is "does this function contain a rejected guard pattern", and there's nothing better to do here without a larger refactor.

---

Test Coverage

The PR description's test plan is minimal but appropriate — the tests themselves are the coverage. Both assertions in Fix 2 remain in place (the assertIn for the correct ternary and the assertNotIn for the guard string), so regression protection is preserved.

---

Summary

Both fixes correctly identify and address the actual test bugs. No production code is changed. The body-extraction logic in Fix 2 has acceptable fragility for this kind of source-scan test; a one-line comment about the 8-space indent (which the existing comment partially covers) would be the only improvement worth considering.

Approved to merge.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!