E.3a: ToolResult.providerExecuted passthrough (phase 1E; stacked on #26)

[cunninghamcard-bit]

Phase 1E unit E.3a (PHASES.md, added 2026-10-11): adds providerExecuted? to ToolResult only. No enum change. Stacked on #26 (E.2) → #25 → #24. The E.3a-only change is commit 5f427aa.

Change

• aimux-core/src/tool.rs: ToolResult gets pub provider_executed: Option<bool> with #[ts(optional)] and serde(default, skip_serializing_if = "Option::is_none"). It sits after dynamic. Sources: docs: full-chain AI SDK alignment design and impact map arcships/aimux#200:2292 and AI SDK generate-text/tool-result.ts:17,32.
• bindings/node/src/types/ToolResult.ts: +6 lines only, which is the field's doc comment plus providerExecuted?: boolean, (git diff --stat: 6 insertions, 0 deletions). gen_ts_types.py --check passes. Hand-written bindings are untouched.

Where ToolResults are produced (all 29 struct literals; the commit adds provider_executed at each)

Every literal is in a provider converter that builds the V4 provider tool-result. In @ai-sdk/provider@4.0.21, language-model-v4-tool-result.ts has no providerExecuted field, so the source carries nothing and each literal writes None. This keeps provider output identical to AI SDK provider output.

I first tried mirroring the paired call's Some(true). That broke openai_upstream_e2e_test::responses_web_search_stream, a port of the AI SDK provider test, because its expected tool-result has no providerExecuted. So I reverted to None.

Everywhere else in core and bindings, ToolResult values move whole. These paths carry the field through unchanged:

• stream_part.rs:357, :395
• generate.rs:472, :792
• response_messages.rs:174, :257
• serde (recordings, bindings, replay)

Literal sites (file:line of the inserted field):

• aimux-providers/src/anthropic/stream.rs:380
• aimux-providers/src/anthropic/stream.rs:400
• aimux-providers/src/anthropic/stream.rs:502
• aimux-providers/src/anthropic/stream.rs:800
• aimux-providers/src/anthropic/stream.rs:835
• aimux-providers/src/anthropic/stream.rs:860
• aimux-providers/src/anthropic/stream.rs:876
• aimux-providers/src/anthropic/stream.rs:897
• aimux-providers/src/anthropic/stream.rs:917
• aimux-providers/src/anthropic/stream.rs:933
• aimux-providers/src/anthropic/stream.rs:952
• aimux-providers/src/google/model.rs:481
• aimux-providers/src/google/model.rs:536
• aimux-providers/src/google/model.rs:839
• aimux-providers/src/google/model.rs:971
• aimux-providers/src/huggingface/responses.rs:455
• aimux-providers/src/huggingface/responses.rs:1044
• aimux-providers/src/huggingface/responses.rs:1070
• aimux-providers/src/openai/responses/provider_events.rs:44
• aimux-providers/src/openai/responses/responses_convert.rs:913
• aimux-providers/src/vertex/model.rs:649
• aimux-providers/src/vertex/model.rs:710
• aimux-providers/src/vertex/model.rs:928
• aimux-providers/src/vertex/model.rs:1076
• aimux-providers/src/xai/responses/mod.rs:162
• aimux-providers/src/xai/responses/mod.rs:280
• aimux-providers/src/xai/responses/mod.rs:319
• aimux-providers/src/xai/responses/mod.rs:1063
• aimux-providers/src/xai/responses/mod.rs:1131

Open for E.3 (needs a decision): AI SDK core sets providerExecuted: true itself when it turns a provider tool-result into step content or a stream part:

• convert-language-model-content.ts:113,129,149,165
• stream-language-model-call.ts:772,788

The matching aimux points are generate.rs:792 and stream_part.rs:395. E.3a does not do this, because it would change core output that existing tests assert. It belongs with E.3's provider-executed variant.

Samples (fixtures/v4/enum-state/)

• tool_result.jsonl adds 6 rows: providerExecuted false/true crossed with dynamic absent/false/true. That includes provider-executed without dynamic. Rows without providerExecuted are the absent case.
• normalized/tool_result.jsonl is now the A7 D25 form plus the A7 correction:
  • client-executed results (providerExecuted not true) always write dynamic as a boolean;
  • provider-executed results pass dynamic through;
  • providerExecuted itself is always passed through.
• Tests:
  • tool_result_samples_match_baseline: all 14 rows round-trip byte for byte (passthrough). The 8 rows that existed at 8455db9 still equal baseline/. The normalized lines are a fixed point. serialize(x) == normalize(x) is E.3's job (the enum).
  • New: tool_result_provider_executed_is_passed_through covers absent/false/true and provider-executed without dynamic, compared byte for byte.
  • baseline_equals_samples no longer lists tool_result, because its new rows cannot exist at 8455db9. The old rows are still checked against baseline/ in the test above.

Acceptance

• (a) Absent/false/true are byte-identical, and old samples are unchanged.
• (b) ToolResult.ts only gains providerExecuted?: boolean (plus its doc comment).
• (c) git diff --stat -- fixtures/ contract-tests/fixtures/ aimux-providers/tests/cassettes/ outside fixtures/v4/enum-state/ is empty.
• (d) No untagged or flatten.

Results (local linux x64)

	R.0	E.2 (#26)	E.3a
cargo test --workspace	2871 / 0 / 7	2882 / 0 / 7	2883 / 0 / 7
fmt / clippy -D warnings	—	ok	ok
boundary script	0	—	0
gen_ts_types --check	—	ok, empty diff	ok, ToolResult.ts +6
Node ava (node 22)	87 pass, 1 fail	87 / same 1	87 / same 1
Python pytest	93 pass, 5 fail	93 / same 5	not re-run