feat: flattened tool call schema

[mykhailobuleshnyi]

Note

Medium Risk
Changes the FC wire/schema contract and streaming parser for all tool calls; legacy action_input is shimmed but models and clients must align with flat args.

Overview
Function-calling tool arguments are flattened: thought is a top-level field alongside each tool’s real parameters, replacing the nested { thought, action_input } wrapper. Schemas from generate_function_calling_schemas put thought first, drop action_input, and update ReAct FC instructions (including top-level delegate_final).

Parsing and execution use ToolCallArguments with extra="allow" and to_action_input(); parse_as_tool_call still normalizes the legacy wrapper shape for replay/back-compat. Tool-only calls with no other params yield an empty input dict instead of a hard error.

Streaming in FC mode adds JSONInnerThoughtsExtractor to split streaming argument JSON into REASONING (thought) vs TOOL_INPUT (params without thought), with held-buffer flush when thought is missing or late. provide_final_answer still uses the existing JSON field parser.

Tests cover the extractor, flat schemas, streaming behavior, and updated FC recovery (malformed JSON vs missing action_input).

^{Reviewed by Cursor Bugbot for commit 4382ecd. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 3 potential issues.

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

^{Reviewed by Cursor Bugbot for commit dfb36c2. Configure here.}

[cursor]

Unreachable condition in _handle_quote nested thought routing

Medium Severity

The condition if self.is_inner_thoughts_value and self.depth == 1 on line 149 is inside a block guarded by if self.depth >= 2 on line 147. Since depth >= 2 and depth == 1 are mutually exclusive, the inner branch can never execute — it's dead code. This means quotes inside nested thought values (if thought were ever an object or array) would be incorrectly routed to _emit_main instead of _emit_thought. Other methods like _handle_open_object, _handle_close_object, and _consume_value_char correctly handle nested thought routing, but _handle_quote does not — both the opening quote path (line 149) and the closing quote path (line 170-171) unconditionally route to main at depth >= 2.

Additional Locations (1)

• dynamiq/callbacks/inner_thoughts_extractor.py#L169-L171

 

^{Reviewed by Cursor Bugbot for commit dfb36c2. Configure here.}

[cursor]

Streaming deltas contain irrevocable dangling comma when thought is last

Medium Severity

When wait_for_first_key=False (the default) and thought is the last field, _handle_close_object retroactively strips a trailing comma from main_buffer but returns "}" as the delta. The comma was already returned as a delta by _handle_comma in a previous call and cannot be revoked. Concatenating all deltas from process_fragment produces invalid JSON like {"a":"x",}, even though main_buffer is correctly patched to {"a":"x"}. The cumulative buffer and the sum of streaming deltas are inconsistent.

 

^{Reviewed by Cursor Bugbot for commit dfb36c2. Configure here.}

[cursor]

Legacy shim silently discards non-dict action_input values

Low Severity

The backward-compatibility shim in parse_as_tool_call silently converts non-dict action_input values to empty params. When action_input is a JSON string that decodes to a non-dict (e.g., a list or scalar), or is invalid JSON, inner becomes {} and the original parameters are lost. The old code would have raised a recoverable ActionParsingException, prompting the LLM to retry, but the new shim silently proceeds with an empty tool call — a change from "error + retry" to "silent data loss."

 

^{Reviewed by Cursor Bugbot for commit dfb36c2. Configure here.}

[github-actions]

Coverage Report •

File	Stmts	Miss	Cover	Missing
dynamiq/callbacks
inner_thoughts_extractor.py	168	12	92%	41, 71–72, 94, 150, 167, 185, 194, 200, 208, 215, 230
streaming.py	676	80	88%	23, 190, 203–204, 212–213, 229–230, 268–269, 277, 304–310, 314, 322–327, 335, 343–344, 399, 403, 407, 450, 453, 461–462, 528, 592, 596, 649, 689–692, 720–721, 804–807, 920, 922, 948, 974–975, 1061, 1088, 1101, 1127–1128, 1130, 1163, 1187–1192, 1228, 1232, 1248, 1250–1258, 1263
dynamiq/nodes/agents
agent.py	884	143	83%	105–106, 110–111, 357–360, 365, 381–382, 400–401, 423, 430–431, 438–439, 445–446, 448, 507–509, 719–721, 736, 745, 771, 787, 806, 818–821, 823, 843–846, 848, 863, 875, 880, 897–898, 908, 911, 921, 961–963, 969–971, 1080, 1085, 1136, 1170, 1172, 1187–1188, 1195, 1198, 1225–1226, 1242, 1281, 1308–1312, 1379, 1385, 1389, 1392, 1426, 1454, 1457, 1462, 1478, 1540–1541, 1576, 1581, 1605, 1661, 1683–1685, 1688–1689, 1699, 1702, 1712, 1719, 1769–1772, 1792–1793, 1797, 1799–1801, 1803, 1805, 1838, 1846–1847, 1849, 1852, 1855, 1906–1909, 1925, 1934, 1999, 2005–2011, 2017, 2019–2028, 2030–2031
dynamiq/nodes/agents/components
schema_generator.py	188	41	78%	61, 64, 142, 252, 256, 261–266, 269–271, 276, 280–281, 283–291, 293, 308, 310, 321–322, 331, 336, 338, 341, 345, 379–380, 382, 387–388
TOTAL	32728	9508	70%

Tests	Skipped	Failures	Errors	Time
2411	2 💤	0 ❌	0 🔥	2m 48s ⏱️