(test) Built-ins SDK tests: vacuous coerce_tool_configs resolver test and substring-only JSON example check

[mmabrouk]

Two SDK unit tests from the built-ins rework pass without exercising what they claim to check.

1. Vacuous resolver test. test_a_bare_tool_name_string_is_ignored_too (sdks/python/oss/tests/pytest/unit/agents/tools/test_resolver.py, line 243) calls ToolResolver().resolve(coerce_tool_configs(["read"])). coerce_tool_configs returns a ToolConfigParseResult pydantic model, not a sequence of tool configs, and iterating a pydantic model yields (field, value) tuples. So resolve() never sees the coerced BuiltinToolConfig and returns an empty spec list for the wrong reason: the test is green without exercising the coercion path. It should pass coerce_tool_configs(["read"]).tool_configs.

2. Substring JSON check. test_no_json_example_writes_a_builtin_tool_entry (sdks/python/oss/tests/pytest/unit/agents/test_agenta_builtins_reference_files.py, line 168) checks each JSON example block in the builtins reference with '"type": "builtin"' not in block (line 174). Any whitespace variant such as "type":"builtin" or "type" : "builtin" slips through. The test should parse each block as JSON and recursively reject any object with type == "builtin".

Follow-up from review comments on #5651:

• fix(agents): Always activate Pi built-ins, control them with rules #5651 (comment)
• fix(agents): Always activate Pi built-ins, control them with rules #5651 (comment)

[linear-code]

AGE-4060

[mmabrouk]

🤖 Step-by-step instructions for picking this up (written with coding agents in mind, but they work for humans too):

Steps

1. In sdks/python/oss/tests/pytest/unit/agents/tools/test_resolver.py, fix test_a_bare_tool_name_string_is_ignored_too (line 241; the call is on line 243). coerce_tool_configs(["read"]) returns a ToolConfigParseResult pydantic model, and iterating a pydantic model yields (field, value) tuples, so resolve() never sees the coerced BuiltinToolConfig and the empty result proves nothing. Rewrite the test to read the model's fields properly: first assert the coercion itself (that the parse result's tool-config list holds one BuiltinToolConfig with name "read"; check the model's actual field names in coerce_tool_configs's return type before writing the assertion), then pass that list into ToolResolver().resolve(...) and keep the assertion that resolved.tool_specs == [], which is the point of the test: a coerced builtin entry is legacy and produces no spec.
2. In sdks/python/oss/tests/pytest/unit/agents/test_agenta_builtins_reference_files.py, fix test_no_json_example_writes_a_builtin_tool_entry (line 168). The current check on line 174 is the substring '"type": "builtin"' not in block, which any whitespace variant such as "type":"builtin" slips past. Replace it: for each fenced block the regex on line 173 extracts, json.loads the block, then walk the parsed value recursively (dicts and lists) and assert no object has type == "builtin". Also assert that json.loads itself succeeds with a clear failure message, so a malformed example fails loudly instead of being skipped. Only json fences match the regex (jsonc blocks do not), so every matched block should be strict JSON; if one is not, that is a real bug in the reference file worth flagging on the PR.

Validation

Run the two targeted files first, then the full SDK suite. All commands run from the repo root; pass means pytest exits 0 with no failures.

cd sdks/python && uv sync --locked && uv run --no-sync pytest oss/tests/pytest/unit/agents/tools/test_resolver.py oss/tests/pytest/unit/agents/test_agenta_builtins_reference_files.py -q
cd sdks/python && uv sync --locked && uv run --no-sync python run-tests.py

As a sanity check that the rewrites are not vacuous, temporarily break what each test guards (pass the raw parse result again in the first; add "type":"builtin" without a space to a JSON example in the second) and confirm each test fails, then undo.

PR requirements

• Title follows the repo convention test(sdk): <symptom-led title>, and the body follows the repo's PR style: lead with the symptom (two green tests that do not exercise what they claim) and a before/after of what each test now catches.
• The PR description must include a section stating which AI model was used, at what thinking or reasoning budget, and a short summary of the conversation or session that produced the change, so reviewers can calibrate their review.
• Link this issue with Closes #5662.

[Yuvakunaal]

Hi @mmabrouk , could you assign this to me?

I've already picked it up and verified both root causes against the current code:
coerce_tool_configs returns a ToolConfigParseResult pydantic model, so passing it straight to resolve() iterates (field, value) tuples instead of the coerced BuiltinToolConfig — confirmed the resulting tool_specs == [] was true for the wrong reason.

Confirmed the substring check has no current false negative, but reproduced the exact whitespace-variant gap by injecting a no-space "type":"builtin" into a nested tools array — the old check missed it, a JSON-parse + recursive walk catches it.
Fixed both: test 1 now asserts on the coerced tool_configs list directly before resolving; test 2 parses each block as real JSON and recursively rejects any object with type == "builtin", plus fails loudly on invalid JSON instead of skipping it.
Ran the targeted files (46 passed) and the full suite (2354 passed — the only errors are pre-existing acceptance tests that need a live AGENTA_API_URL, unrelated to this change).

Sanity-checked both rewrites by deliberately
breaking what each one guards and confirming they fail, then reverted.

[mmabrouk]

assigned