[mattapperson]

Description

Related to #574

Root cause of #574 is server-side. OpenRouter's backend does not honor eager_input_streaming: false on tools; this package already forwards the field correctly (tools[i].eager_input_streaming), so the invalid streamed Claude tool arguments can't be fixed here. This PR fixes a provider-side bug that hid that failure.

Bug: In doStream, a tool-call part was emitted as soon as the accumulated arguments happened to parse as JSON. With an "early root close" from the upstream (a valid JSON prefix followed by more fragments), e.g. fragments '{"a":[1]}', '</a>\n<b>', '[2]}', the provider emitted a truncated-but-valid tool call with input {"a":[1]} and then kept emitting tool-input-delta parts after it. The model's real (malformed) payload never reached the AI SDK's tool-call error handling / experimental_repairToolCall (#575).

Fix: Tool calls are no longer finalized mid-stream. They are finalized in flush() using the full accumulated arguments:

• finish reason tool-calls (including the existing missing/other -> tool-calls inference from Kimi K2.5 returns undefined finishReason after tool calls, breaking agentic loops #420) or stop: every tool call is forwarded, invalid JSON as-is ('{}' only for empty arguments, as before).
• other finish reasons (length, error, content-filter): only tool calls with parsable arguments are forwarded, matching previous behavior for truncated output.
• every tool call is still emitted exactly once, with the full tool-input-start -> delta... -> end -> tool-call lifecycle, including single-chunk arguments and streams that end without a finish_reason.

The early emission dates back to the original AI SDK v5 port (#77), inherited from the openai-compatible provider; nothing relies on mid-stream emission beyond event ordering. Because arguments are no longer parsed per delta, the JsonCompletenessScanner optimization from #586 is no longer needed and is removed (it was internal, not exported).

Behavior change: tool-input-start/tool-input-delta still stream as arguments arrive (including single-chunk arguments), but tool-input-end and tool-call now arrive at the end of the stream instead of as soon as the JSON closes, so a tool's onInputAvailable and the UI tool-input-available state fire at end of stream. With ai@7, tool execution timing is unchanged (tools already run after the model call ends). For whitespace trailing a complete object, the input now includes it (still valid JSON).

Delta merging hardening (also in this PR):

• Tool-call deltas without an index (optional in the schema since chore: make index optional #99) are now handled. Before, the first one landed at toolCalls[-1] and was never emitted. Now a delta with a new id (or, without an id, a function name) starts a new tool call and other deltas continue the most recent one. A repeated name with no id only splits off a separate call when the previous arguments are a complete JSON object/array and the delta starts a new, different one. The first delta at a new index binds to an indexless call it continues.
• A delta that reuses an existing index with a different non-empty id starts a new tool call instead of being concatenated onto the previous one under the wrong name. Provider indexes map to dense internal slots.
• An identical resend of already complete JSON object/array arguments ({"a":1} sent twice) is ignored instead of producing {"a":1}{"a":1}. This only triggers when the accumulated arguments are a complete JSON container and the delta is byte-identical, so appending it could never produce valid JSON.
• Tool calls dropped at flush because their arguments are incomplete (length, error, ...) now get a closing tool-input-end (no tool-call), so UI parts don't stay in input-streaming.

Checklist

• I have run pnpm stylecheck and pnpm typecheck
• I have run pnpm test and all tests pass
• I have added tests for my changes (if applicable)
• I have updated documentation (if applicable)

Changeset

• I have run pnpm changeset to create a changeset file

Tests added (mocked SSE streams), in src/chat/index.test.ts plus e2e/issues/issue-574-streamed-tool-call-finalization.test.ts (which also checks experimental_repairToolCall receives the full arguments): early root close, multi-chunk args with braces in strings, single-chunk args, interleaved parallel tool calls, stream without finish_reason, invalid args with stop, truncated args with length (plus tool-input-end for the dropped call), no per-delta parsing, deltas without index (single call, two calls, multi-chunk continuations), a different id reusing an index, repeated id at the same index, identical full-args resend, and a fragment equal to incomplete accumulated args (not dropped).

---