[summer-tt]

Summary

Qoder streams tool calls as OpenAI-style deltas with id as the call's identity and
index as a continuation hint; in observed streams, different explicit IDs can reuse one
index. This implementation therefore keys identity by explicit ID and uses index only
as a correlation hint — a conservative provider rule, not a gateway guarantee.
Previously, index-keyed aggregation merged two explicit IDs and a catch {} fallback
published the survivor with {}.

Reproducer (synthetic wire deltas)

Two explicit IDs reusing index: 0:

{ tool_calls: [ { index: 0, id: "call_A", function: { name: "read", arguments: '{"path":"/a"}' } } ] }
{ tool_calls: [ { index: 0, id: "call_B", function: { name: "read", arguments: '{"path":"/b"}' } } ] }

Before, both deltas shared one index-keyed state and the survivor was published with {};
after, each call keeps its own ID and arguments.

What changed

• ID is identity; index is a hint. Explicit IDs are tracked in a map; a new ID binds
  only to a unique provisional call and never overwrites another identity. Reusing one
  index across different IDs is the supported case; only an ID-less continuation with
  several candidates fails (Ambiguous tool-call continuation) instead of guessing.
• Provisional and retired calls. Argument-first fragments buffer until the block
  opens; a completed call retires only after a strict JSON-object parse; an
  explicit ID can re-associate with a retired call; a conflicting later name fails.
• Strict arguments, no {} fallback. An incremental scan finds only a possible
  root-object boundary; JSON.parse remains the grammar authority. Empty bytes stay legal
  and mean {}; non-empty malformed, non-object, or trailing-garbage input throws.
  Required-schema enforcement stays with the SDK, so {} to a tool with a required field
  becomes a validation error, not an execution.
• Batch, ordering, lifecycle. Calls are validated before any toolcall_end; if one is
  bad, none gets a success event. Blocks open in first-identifiable-event order and keep
  direct references through the thinking-tag splice. finish_reason: "length" is
  preserved; reader cleanup runs in finally; new fragments are scanned incrementally,
  each call parsed once, with Map/Set lookups — no complexity proof for arbitrary inputs.

Validation

• CI-covered suite: 360/360 tests pass (previous 311 retained); lint, typecheck, and
  build pass on the production revision. These run in CI for normal changes.
• Manual offline SDK script (verify-tool-call-flow.ts): 33/33 across SDK 1.0.4,
  1.2.0, and 1.3.0 (11 each; original nine fixtures preserved). It drives the built
  package through the real host SDK with a synthetic credential and real dispatch; its
  guards are Node API guards, not an OS sandbox. New zero-byte and "{}" fixtures for a
  required-path tool yield {} in the parser, an SDK validation error on the correct
  tool-call ID, zero implementation calls, and error replay in the next request. Its
  startup isolation fix also has a real RED/GREEN.
• Live SDK continuation smoke: One real SDK run used Pi SDK 1.3.0 with
  qoder/DeepSeek-Flash:max through the normal provider factory and SDK authentication,
  using the same unchanged normal bundle as offline validation. Across two synthetic
  in-memory fixtures and two chat requests, tool results were returned for the correct
  calls and the interaction reached the expected final stop.
• Bound evidence: consumer/CLI checks (27 SDK cases, 6 CLI cases) ran on the
  production revision and were not re-run for the script-only commit; stream.ts and
  bundle hashes are unchanged.
• Run manually (not wired into npm test/CI or the published package):

node node_modules/tsx/dist/cli.mjs scripts/verify-tool-call-flow.ts \
  --package-root /path/to/package \
  --sdk-root /path/to/node_modules \
  --output /path/to/report.json

Performance

Full npm test wall-clock: +1.895716% against the ≤2% gate. Parser scenarios pass the
current rule — 22/25 under 30 ms, the other three within +0.193%, +2.634%, and +3.556%
relative (per-invocation group-average median; not B−A or a per-sample claim). The
original zero-tolerance FAIL/exit 1 record is unchanged and is not relabeled as a new
pass. Local bounded workloads; no universal claim.

Risks and non-claims

• Live verification covers only this one model/SDK path. It does not establish all-model
  coverage or general gateway-protocol behavior, and it does not cover pi-subagents'
  detached-launcher/provider path. Reused-index cases have deterministic synthetic
  coverage; this smoke does not independently verify the response delta shape.
• Not a gateway protocol guarantee. Reusing one index across explicit IDs is the supported
  normal case; failures cover ID-less continuations with several candidates, identity
  conflicts, and invalid/incomplete or non-object JSON.
• Earlier empty-argument or dropped-call reports can have several causes; this change does
  not attribute all of them to this defect.
• CI runs the 360 unit tests; the SDK integration script is manual, so that flow is not
  automatically guarded.

Scope

Only three files change: src/protocol/stream.ts, src/__tests__/stream.test.ts, and
scripts/verify-tool-call-flow.ts; no dist, dependency, lockfile, build/release
configuration, or unrelated protocol changes.