sync(v2.12): preserve native Responses tool calls, history, schemas, and result metadata end-to-end #139

New issue

Open

opened 2026-09-09 14:58:55 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-09 14:58:55 +00:00

Owner

Upstream

Agent Zero v2.12: https://github.com/agent0ai/agent-zero/releases/tag/v2.12

Gap

a0-symbolics already has custom Responses/XML-tool recovery, but v2.12 adds a broader canonical native-function-call/history contract. Current Agent.hist_add_ai_response() still accepts llm_result: LLMResult | None = None, so the v2.12 required-result/sentinel contract is not yet adopted.

Required work

Preserve canonical native function calls through Context Doctor without text repair rewriting them.
Store canonical function-call content in assistant history/masking and repeated-response detection.
Stream native tool arguments through the existing live tool display with alias normalization and masking.
Add non-strict argument schemas for bundled tool implementations used by Responses.
Preserve policy-filtered guidance in native function descriptions while converting the A0 tool envelope to native args.
Keep Responses-native IDs/encrypted reasoning in prepared history.
Keep Responses output-item reconstruction authoritative in the parser.
Skip Chat-only replay/input preparation for Chat Completions where it is unused.
Make llm_result required for AI-history writes, using an explicit LLMResult.non_llm() sentinel for greetings/interventions/non-provider turns while preserving intentional legacy call compatibility.
Keep encrypted reasoning markers out of readable assistant output.

Acceptance

Native tool calls survive parse -> repair/context doctor -> history -> next Responses request unchanged except intentional policy filtering/masking.
Existing XML/text tool recovery still works and does not hijack canonical function calls.
Non-LLM history writes carry an explicit sentinel, never ambiguous None state.
Repeated-response detection understands native function-call content.
Regression coverage includes canonical tool calls, streamed args, aliasing, masked args, encrypted reasoning, legacy history calls, and mixed Responses/text-repair turns.

## Upstream Agent Zero v2.12: https://github.com/agent0ai/agent-zero/releases/tag/v2.12 ## Gap `a0-symbolics` already has custom Responses/XML-tool recovery, but v2.12 adds a broader canonical native-function-call/history contract. Current `Agent.hist_add_ai_response()` still accepts `llm_result: LLMResult | None = None`, so the v2.12 required-result/sentinel contract is not yet adopted. ## Required work - Preserve canonical native function calls through Context Doctor without text repair rewriting them. - Store canonical function-call content in assistant history/masking and repeated-response detection. - Stream native tool arguments through the existing live tool display with alias normalization and masking. - Add non-strict argument schemas for bundled tool implementations used by Responses. - Preserve policy-filtered guidance in native function descriptions while converting the A0 tool envelope to native args. - Keep Responses-native IDs/encrypted reasoning in prepared history. - Keep Responses output-item reconstruction authoritative in the parser. - Skip Chat-only replay/input preparation for Chat Completions where it is unused. - Make `llm_result` required for AI-history writes, using an explicit `LLMResult.non_llm()` sentinel for greetings/interventions/non-provider turns while preserving intentional legacy call compatibility. - Keep encrypted reasoning markers out of readable assistant output. ## Acceptance - Native tool calls survive parse -> repair/context doctor -> history -> next Responses request unchanged except intentional policy filtering/masking. - Existing XML/text tool recovery still works and does not hijack canonical function calls. - Non-LLM history writes carry an explicit sentinel, never ambiguous `None` state. - Repeated-response detection understands native function-call content. - Regression coverage includes canonical tool calls, streamed args, aliasing, masked args, encrypted reasoning, legacy history calls, and mixed Responses/text-repair turns.

nsaspy referenced this issue

2026-09-11 04:53:38 +00:00

[P0][symbolics-worker] Define canonical Agent Zero tools + clean synthwave themes + Prolog-RLM sync #198

nsaspy referenced this issue

2026-09-15 14:10:21 +00:00

[P0][symbolics-worker] Define canonical Agent Zero tools + clean synthwave themes + Prolog-RLM sync #199