Native provider wire audit — incremental findings

Revision 2, 2026-09-17: all five findings below are FIXED in current source. Re-verified line by line; per-finding status is inline. Retained as a record of what was found and where, not as an open work list. The startup/close refinement is also addressed: both sessions now document a setup deadline covering the initial send. Status reconciliation is summarized in the defects report.

Reviewed 2026-09-17 01:45 UTC (September 16 local). Owner: transcription-adapters. Read alongside the coordinator-owned conversational-voice-verification.md. No production edits, provider calls, or new test execution. Findings below are source-derived; proposed fixture sequences are not claimed as executed tests. Fable owns corrections; coordinator owns acceptance. Sources are changing during review, so recheck before acting.

Paths below are relative to packages/native/apple/HudsonKit/Sources/.

Findings (all fixed in current source)

Verified fixes: explicit BLOCKING/NON_BLOCKING serialization; delegation now requires a backend model and rejects double-declared tools instead of replacing them; discovery follows nextPageToken with dedup and loop guards; unfamiliar bidi models get .unknown capabilities; round start is idempotent per response ID with correlated terminal events.

1. P1 — FIXED — Base Gemini blocking tools were serialized as the opposite default. HudsonConversationGemini/HudGeminiConversationSession.swift:79–83 omits behavior for .blocking. The base model page says default is NON_BLOCKING and explicitly requires BLOCKING for synchronous execution. Emit the explicit value for both behaviors. Fixture: inspect setup for one base blocking and one non-blocking declaration; Extended Thinking continues rejecting blocking declarations.

2. P1 — FIXED — GPT tools without explicit delegation produced an invalid setup. HudsonConversationOpenAI/HudGPTLiveSession.swift:106–128 synthesizes type: responses with an empty responses object, then adds tools but no backend model. The delegation guide requires that model at creation. Reject missing backend configuration before opening a socket, or require an explicit host-provided backend model; do not guess one. Malformed authored delegation JSON currently falls into the same fallback. Also, assignment at line 118 replaces existing authored tools despite the comment claiming a merge (for example, web_search disappears). Fixture: tools plus absent/invalid delegation must fail locally; configured backend tools must be preserved or rejected explicitly, not dropped.

3. P2 — FIXED — Gemini discovery truncated at the first page. HudsonConversationGemini/HudGeminiConversationAdapter.swift:58–85 makes one pageSize=200 request and ignores nextPageToken. The models API is paginated. A valid Live model on page two disappears from the picker. Follow tokens, deduplicate IDs, reject token loops, and avoid publishing a partial refresh if a later page fails. Fixture: first page has no bidi model plus a token; second page contains the model.

4. P2 — FIXED — Discovery invented capabilities for unfamiliar Gemini models. Same adapter, lines 75–84 and 129–142: bidi support implies tool support, while an ID suffix determines thinking support; every other model gets base-3.8 proactive/audio claims. Discovery does not establish these facts. Preserve discovered IDs, but use .unknown for capabilities without specific evidence; documented metadata for known variants is not a catalog allowlist. Fixture: an unfamiliar bidi model remains listed without supported/unsupported thinking assertions or base-3.8 notes.

5. P2 — FIXED — Duplicate response.created destroyed GPT pending-call state. HudsonConversationOpenAI/HudGPTLiveSession.swift:47–52,283–285 always resets a round, even when its response ID equals currentResponseID. The ID is stored but never checked on terminal events either. Fixture: created R1, call C1, duplicate created R1, result C1 currently becomes unknownToolCall; a stale completion for R1 can also mark a later R2 boundary. Make repeated creation idempotent and correlate response-bearing terminal events. This is additional to the coordinator's already-listed per-round continuation issue.

Startup/close refinement of existing checklist item

Both session start() methods send the initial frame before installing the setup timeout (Gemini:87–95; GPT:87–95). A suspended initial send is therefore unbounded. An initial send error escapes adapter open() without closing its new socket (Gemini adapter:123–126; GPT adapter:83–86). Cancellation during the continuation wait has no cancellation handler, and may return a ready session after the caller cancelled. Both finish() methods launch unawaited socket cleanup; close completion does not establish transport closure.

Use a cancellation-aware, whole-start deadline and awaited cleanup, including initial-send failure. Fixtures should suspend send, throw from send, cancel before acknowledgment, then deliver a late acknowledgment. This supplies concrete provider-level failure paths to the existing lifecycle finding, not a separate host-lifecycle claim.

Rechecked, not reported again

Gemini invalid result validation now precedes pending-call removal (HudGeminiConversationSession.swift:132–143); the earlier loss-of-call issue was corrected during this review. Known-call ID sets now suppress repeated tool announcements. Neither source change establishes provider or runtime acceptance.