E.3: ToolResult split by source; core marks provider results providerExecuted:true (phase 1E; stacked on #27)

[cunninghamcard-bit]

Phase 1E unit E.3 (A7: "E.3 建模定稿" plus "E.3 补充: core 转换时置 providerExecuted:true"). Stacked on #27 (E.3a) → #26 → #25 → #24. The E.3 commits are 132c301 and the follow-up 0f4aa9e.

Model (aimux-core/src/tool.rs)

• ToolResult keeps its common fields as public fields: toolCallId, toolName, result, isError, preliminary, providerMetadata. isError and preliminary stay independent, as A7 requires. The source becomes pub source: ToolResultSource:

  • Client { dynamic: bool, provider_executed: Option<bool> }: dynamic is always written as a boolean (AI SDK execute-tool-call.ts:169,199). provider_executed is plain data, written as given. There is no hidden "input had false" marker. Client results built by core never set it.

  • Provider { dynamic: Option<bool>, mark: ProviderExecutedMark }: dynamic is passed through (stream-language-model-call.ts:774,789). ProviderExecutedMark has two variants:

    • Unmarked: V4 provider-layer output; providerExecuted is not written.
    • Marked: core output; "providerExecuted": true is written.

    The marker is transitional. Provider converters and core share ToolResult today, so this enum is what tells the two layers apart. Follow-up unit E.3b will switch the provider side to feat(provider): V4 type crate aimux-provider (phase 1 unit 1.6) #16's V4 tool-result type once feat(provider): V4 type crate aimux-provider (phase 1 unit 1.6) #16 merges (the 3.3 split), and the marker goes away. feat(provider): V4 type crate aimux-provider (phase 1 unit 1.6) #16 (b837aac3) is not an ancestor of this branch (git merge-base --is-ancestor exits 1), so this PR is not stacked on it.

• Wire. The private ToolResultWire is the E.3a struct, unchanged, used through #[serde(from/into)] and #[ts(export, as = "ToolResultWire")]. ToolResult.ts has zero diff and gen_ts_types.py --check passes. rg 'pub dynamic:' aimux-core/src/tool.rs hits only RawToolCall, the V4-style DTO that A7 leaves alone.
• Wire to model:
  • providerExecuted: true becomes Provider { mark: Marked }.
  • absent or false becomes Client { provider_executed: None | Some(false) }.
• Provider layer. The 29 provider converter literals now build ToolResultSource::provider(dynamic), which is unmarked. Their JSON is identical to before: no providerExecuted, and dynamic as given. This matches @ai-sdk/provider@4.0.21, whose language-model-v4-tool-result.ts has no providerExecuted field.
• Core conversion points set providerExecuted: true through ToolResult::into_core(), which passes dynamic through:
  • aimux-core/src/generate.rs:794: provider content becomes step content (AI SDK convert-language-model-content.ts:113,129,149,165).
  • aimux-core/src/stream_part.rs:396: into_text_stream_part (AI SDK stream-language-model-call.ts:772,788).
• Accessors: is_dynamic(), dynamic() (the value as written on the wire), is_provider_executed(), provider_executed(), into_core().

Changed assertions

providerExecuted-related changes: none. No existing assertion about a provider-executed result had to change. openai_upstream_e2e_test::responses_web_search_stream passes with its assertions untouched.

D25 normalization changes (two, unrelated to providerExecuted): client rows now gain "dynamic":false, as A7 D25 requires (execute-tool-call.ts:169,199).

1. aimux-core/tests/enum_state_wire_test.rs tool_result_samples_match_baseline, now tool_result_samples_match_normalized. Samples are compared with normalized/tool_result.jsonl, which was committed in E.3a and is unchanged here. The client rows without dynamic gain "dynamic":false. Provider-executed rows are byte-identical, including the one without dynamic.
2. aimux-core/tests/enum_state_wire_test.rs tool_result_provider_executed_is_passed_through. The expected output of the two client inputs (absent, and "providerExecuted":false) gains "dynamic":false. The provider rows are unchanged.

Not an assertion change: aimux-providers/tests/data_loss_regression_test.rs:266 destructures through tool_result.dynamic() instead of the removed field. The asserted values are the same.

New tests

• enum_state_wire_test:
  • tool_result_absent_and_false_dynamic_are_the_same_variant
  • provider_layer_result_is_marked_by_core_conversion: V4 JSON is unmarked, and into_core adds "providerExecuted":true.
• aimux-core/tests/provider_executed_marking_test.rs (new):
  • generate_marks_provider_tool_result: a mock model returns a V4 tool result, and generate_text step content has providerExecuted: true with dynamic absent.
  • text_stream_marks_provider_tool_result: into_text_stream_part marks the result, while the StreamPart's own JSON stays unmarked.

Per-provider evidence (new aimux-providers/tests/provider_executed_core_conversion_test.rs, 6 tests)

Each test runs the provider's real model through generate_text (step content, generate.rs:794) and stream_text (into_text_stream_part, stream_part.rs:396). It asserts that every tool-result / tool-error part has "providerExecuted": true, and that at least one exists.

Provider	generate input	stream input	E.3a sites exercised
anthropic	cassette anthropic/test_pause_turn_web_search_vcr_1.json (web_search)	cassettes anthropic/test_anthropic_code_execution_tool_stream.json and anthropic/test_anthropic_web_fetch_tool_stream.json	stream.rs parse_anthropic_content and stream_parts_for_result_block
google	cassette gemini/test_code_execution.json	cassette gemini/test_code_execution_stream.json	google/model.rs generate + stream code-execution results
vertex	the same gemini cassettes replayed against Vertex (same wire)	same	vertex/model.rs generate + stream
openai responses	cassette openai/test_openai_responses_model_web_search_tool_1.json	the responses_web_search_stream upstream events (the recorded *_stream_1 cassette holds a non-streaming body)	provider_events.rs web_search pair; streaming web_search result
xai	tool_results_preserve_actions_and_sources body (web_search)	synthetic file_search_call SSE	xai/responses/mod.rs generate web_search + stream file_search
huggingface	should_handle_mcp_tools_with_annotations body (mcp_call)	should_stream_tool_calls events (function_call with output)	huggingface/responses.rs:1042 (generate) and the stream output site

Not covered here: the image_generation result paths (openai responses_convert.rs partial image; xai image_generation_result) and the remaining anthropic result kinds (tool_search, advisor, mcp, bash/text-editor). They go through the same two core conversion points but have no direct test.

Also new: enum_state_wire_test::client_tool_result_provider_executed_is_plain_data. Client samples with providerExecuted absent and false (rows already in tool_result.jsonl) round-trip byte for byte as Client { provider_executed: None | Some(false) }.

Fixtures / cassettes with a diff

None. git diff --stat exp/e3a-toolresult-provider-executed -- fixtures/ contract-tests/ aimux-providers/tests/cassettes/ bindings/ is empty, and no test rewrote a fixture. That includes fixtures/v4/enum-state/, because E.3a already committed the E.3 normalized form. The only intended JSON change, "providerExecuted":true on provider results after core conversion, appears in no committed fixture or cassette.

Results (local linux x64; new test names confirmed in the log)

	R.0	E.3a (#27)	E.3
cargo test --workspace	2871 / 0 / 7	2883 / 0 / 7	2894 / 0 / 7 (with the follow-up)
fmt / clippy (ci.yml flags)	—	ok	ok
boundary script	0	0	0
gen_ts_types --check	—	ok	ok, zero .ts diff
Node ava (node 22)	87 pass, 1 fail	same	87 pass, same 1 fail
Python pytest	93 pass, 5 fail	not re-run	not run (no venv found)

New code has no untagged or flatten. Hand-written bindings are untouched. Go/Java/Kotlin/Swift/Flutter were not run.