feat: rework tool calling

[maksymbuleshnyi]

Note

High Risk
Touches core agent LLM invocation, tool schema generation, multi-provider strict transforms, and streaming parsers—behavior changes for all inference modes and can alter model outputs and tool validation.

Overview
This PR reworks agent tool calling end-to-end: schemas, provider transforms, parsing, and streaming alignment.

Schemas & prompts: Structured-output action_input is now a JSON object (finish uses {"answer": "..."}); generator-level strict flags are removed in favor of per-provider handling. ReAct/XML prompts add stricter single-turn rules; function-calling loops default to tool_choice: required when tools are present (unless overridden).

Provider strict mode: BaseLLM.strict_tools (bool or allowlist) drives transform_tool_schemas on OpenAI and Anthropic—schema cleaning, optional→nullable encoding, free-form dict fields as JSON strings, and strict: true where supported. Anthropic adds LiteLLM patching to forward strict and caps strict tools per request.

Agent post-processing: After tool calls, inputs are coerced (stringified dicts, including nested models) and protocol nulls stripped before Pydantic validation; truncated tool JSON can be repaired via repair_truncated_json. Structured-output finish reads action_input["answer"] when present.

Streaming: Structured-output initialization avoids locking onto the raw action_input object for answers (fixes leaking {"answer": ...} in streams); integration FSM assertions account for strict null defaults in streamed tool input.

^{Reviewed by Cursor Bugbot for commit 6f4e213. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

Coverage Report •

File	Stmts	Miss	Cover	Missing
dynamiq/callbacks
streaming.py	644	61	90%	22, 189, 202–203, 211–212, 228–229, 267–268, 276, 303–309, 313, 321–326, 334, 342–343, 395, 399, 403, 445, 448, 502, 566, 570, 623, 663–666, 694–695, 778–781, 903–904, 990, 1017, 1097, 1154–1159, 1195, 1199, 1230
dynamiq/nodes/agents
agent.py	957	187	80%	83–87, 330–335, 340–347, 352–359, 380, 382–393, 413, 418–425, 440–443, 448, 464–465, 483–484, 506, 513–514, 521–522, 528–529, 531, 590–592, 802–804, 819, 828, 854, 870, 889, 902–905, 907, 933–934, 936, 956, 968, 973, 988, 996–997, 1007, 1010, 1020, 1060–1062, 1068–1070, 1179, 1184, 1235, 1269, 1271, 1286–1287, 1294, 1297, 1324–1325, 1341, 1380, 1407–1411, 1478, 1484, 1488, 1491, 1525, 1553, 1556, 1561, 1577, 1651–1652, 1687, 1692, 1716, 1772, 1794–1796, 1799–1800, 1810, 1813, 1823, 1830, 1880–1883, 1903–1904, 1908, 1910–1912, 1914, 1916, 1949, 1957–1958, 1960, 1963, 1966, 2017–2020, 2036, 2045, 2110, 2116–2122, 2128, 2130–2139, 2141–2142
dynamiq/nodes/agents/components
schema_generator.py	164	38	76%	61, 64, 139, 252, 256, 261–266, 269–271, 276, 280–281, 283–291, 293, 308, 310, 321–322, 356–357, 359, 364–365, 401, 418
dynamiq/nodes/llms
anthropic.py	144	92	36%	32–34, 38–39, 43–44, 47, 50–57, 59–60, 118–119, 121–125, 130–134, 137–142, 147–155, 158–160, 162, 166–171, 173–175, 177, 217, 227–233, 240, 242, 244, 284–285, 287–306
base.py	416	45	89%	34, 36, 153, 159, 351, 397–400, 409–415, 442, 488, 531–532, 536, 634–643, 664–665, 680, 718, 720, 754, 756, 823, 886, 985, 1061, 1064, 1095–1096
openai.py	172	99	42%	43–54, 69–70, 72–81, 83–91, 96–100, 102, 105–107, 109–110, 115–116, 120–135, 153–156, 158–165, 167, 173–175, 177–179, 300–312, 317–321
dynamiq/utils
json_parser.py	134	40	70%	28–29, 47, 90–91, 135–136, 201–202, 204–206, 208–224, 226–228, 230–234, 236–237, 239
TOTAL	32824	9732	70%

Tests	Skipped	Failures	Errors	Time
2376	2 💤	0 ❌	0 🔥	2m 42s ⏱️

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 1e2e9b1. Configure here.}