fix(translator): six format-conversion defects (#3243, #3247, #3254, #3369, #3492, replaces 6 PRs)

[ntdatt812]

Six translator/handler fixes that were sitting in separate PRs (#3243, #3247, #3254, #3369, #3492/#3528, #3549), rebased onto current master and sent as one review. Each is an independent commit, so any one of them can be dropped without touching the others.

All six are format-conversion defects: a request or response crossed a format boundary and lost a field that the target API requires.

commit	defect
1b042a1c	reasoning.effort was sent flat to openai-responses targets, which nest it
0f23a482	a non-streaming binary-transport reply was returned in the upstream's format, not the client's
8805156e	tool_calls: [] was treated as "has tool calls", producing an empty tool_use block
74658a51	a tool result arriving without an id emitted tool_use_id: undefined; Anthropic rejects the whole request
65339392	a request with no stream key defaulted to streaming, so non-streaming clients got SSE
d43850f6	the CommandCode error chunk opened without a role, so clients that key on the first delta dropped it

Verification

npx vitest@3 run --config tests/vitest.config.js \
  tests/unit/{commandcode-to-openai,nonstream-binary-transport-source-format-3199,openai-format-empty-tool-calls,stream-mode-default-3492,thinking-responses-nested-3154,tool-result-missing-id-3362}.test.js \
  tests/translator/thinking-unified.test.js
→ 7 files, 107 passed

Two notes on running these, because both cost me time:

• The alias @/ only resolves with --config tests/vitest.config.js; without it collection dies on Failed to load url @/lib/dataDir.js.
• vitest 2.x cannot run any test that reaches open-sse/translator/index.js. It fails with TypeError: register is not a function at the top-level register(...) calls. That is not a defect in this repo: index.js deliberately relies on function-declaration hoisting to survive the import cycle (the comment above var requestRegistry says so), and Vite 5's SSR transform turns the named import into a call-time property access that is not yet populated. I confirmed the same failure on clean master, on master as of 15 Aug, and on this repo's own tests/translator/bugs-antigravity.test.js — so it predates these changes and is not caused by them. vitest 3 runs them correctly.

Each claim was checked by reverting it and confirming which test dies; no test passes for a reason other than the fix.

One of the six did not survive that check as it stood. Reverting the call site in chatCore.js for #3492 left stream-mode-default-3492.test.js green: the tests exercised clientRequestedStreaming in isolation and nothing observed the wiring — the same shape of bug the fix is about. handleChatCore takes ~30 collaborators and cannot be driven from a unit test, so 18e67517 pins the call site at the source level instead. Reverting chatCore.js now fails 3 of that file's 10 cases.

[ntdatt812]

Rebased onto master (39e36d3d). Was CONFLICTING, now clean. Two commits came out of the bundle rather than being merged forward, because master has since decided both questions the other way:

• fix(commandcode): open the error chunk with the assistant role — withdrawn. That branch now throws instead of emitting the error as content, so the stream handler marks the stream errored. Giving the emitted chunk a role has nothing left to apply to.

• fix(translator): nest reasoning effort for openai-responses targets — withdrawn, and I would like your read on it. d1de3245 (22 Sep) added tests/unit/thinking-budget-max-level.test.js, which asserts reasoning_effort stays flat on an OPENAI_RESPONSES target:

  const out = applyThinking(FORMATS.OPENAI_RESPONSES, "gpt-5.6-sol", body, "codex");
  expect(out?.reasoning_effort).toBe("max");

  That is the opposite of what the commit did, so it is out. Worth asking anyway: the OpenAI Responses API reads reasoning: { effort } and ignores a top-level reasoning_effort, so for a genuine /v1/responses upstream the flat key is dropped on the floor. If codex is a wire that wants the flat key and only some responses targets want it nested, the split is per provider rather than per format and I am happy to send that separately. If you meant flat everywhere, nothing to do.

What remains is five commits, unchanged in intent:

commit	what
f2688315	non-streaming binary-transport replies are built in the client's format, not the source's (#3199)
d052cd8f	an empty tool_calls: [] is no tool calls, not a tool call with no fields
ea7492dd	a tool result that arrives without an id is matched to its call instead of dropped (#3362)
350c8ce7	a request with no stream key is non-streaming (#3492)
36d4cf18	pins that chatCore reads the mode from streamMode.js

Two conflicts were merges rather than choices, and both keep your side: thinkingUnified.js keeps the new display passthrough, and chatCore.js keeps the token-saver block — only the inlined streaming predicate is replaced by requestedStreaming(body, sourceFormat).

npx vitest@3 run --config tests/vitest.config.js: 89 failures, and they are the same 89 by name as on master at 39e36d3d — compared per test name, not by count.