fix(llm): OpenAI-compatible streaming tool calls (#475) + api-key Responses 401 diagnostics (#476) - #481
Draft
TYRMars wants to merge 2 commits into
Draft
fix(llm): OpenAI-compatible streaming tool calls (#475) + api-key Responses 401 diagnostics (#476)#481TYRMars wants to merge 2 commits into
TYRMars wants to merge 2 commits into
Conversation
StreamAccumulator.#finalise dropped any streamed tool call whose deltas
never carried an `id`, because the guard required both `id` and `name`.
OpenAI-compatible backends (Kimi/Ollama) commonly stream
`{index, function:{name, arguments}}` with no `id`, so the real call was
silently discarded while finish_reason stayed "tool_calls" — leaving an
empty assistant message and the tool never running over SSE/WS (the
non-streaming path kept the call, so the same backend worked via
/v1/chat/completions).
Skip only true placeholder slots (no `name`) and synthesise `call_<index>`
when the stream omits an id, matching how the Google provider synthesises
`gem_<index>`. Adds regression tests for the id-less and padding cases.
Fixes #475
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AfgFeWNPZaPcWc8CSDL3H6
completeStream checked `resp.status === 401` before the branch that reads
the upstream body, and refreshIfUnchanged's api_key arm unconditionally
threw "API-key auth cannot refresh". So for JARVIS_PROVIDER=openai-responses
any 401 from api.openai.com/v1/responses (revoked key, wrong project/org,
disabled account) surfaced as that one fixed string — the real
`{"error":{"message":…}}` body and status were discarded. The 401 Response
body was also never consumed (undici kept the socket pinned until GC), and
the message contained none of the tokens fallback.ts scans for
(401/unauthorized/authentication), so it wasn't classified as an auth error.
Guard the refresh branch on `auth.kind === "chatgpt_oauth"` so api-key 401s
fall through to the `!resp.ok` branch, which reports `status 401: <body>`,
consumes the body, and yields a fallback-recognisable message. Updates the
test to assert the surfaced status + body.
Fixes #476
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AfgFeWNPZaPcWc8CSDL3H6
TYRMars
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Two independent
@jarvis/llmprovider fixes, both surfaced as freshly-opened bug reports.Closes #476 (the issue that triggered this change) and #475.
#476 —
responses.ts: api-key 401s discard the real upstream bodycompleteStreamcheckedresp.status === 401before the branch that reads the upstream body, andrefreshIfUnchanged'sapi_keyarm unconditionally threw"API-key auth cannot refresh". So forJARVIS_PROVIDER=openai-responses, any 401 fromapi.openai.com/v1/responsessurfaced as that fixed string — the real{"error":{"message":…}}body and status never reached the operator. The 401 body was also never consumed (socket pinned until GC), and the message wasn't classified as an auth failure byfallback.ts.Fix: guard the refresh branch on
auth.kind === "chatgpt_oauth"so api-key 401s fall through to the!resp.okbranch, which reportsstatus 401: <body>, consumes the body, and yields a fallback-recognisable message.#475 —
openai.ts: id-less streamed tool calls dropped (Kimi/Ollama)StreamAccumulator.#finaliserequired bothidandnameto keep a tool call. OpenAI-compatible backends (Kimi/Ollama) stream{index, function:{name, arguments}}with noid, so the call was silently dropped — leavingfinish_reason: "tool_calls"on an empty assistant message and the tool never running over SSE/WS.Fix: skip only true placeholder slots (no
name) and synthesisecall_<index>when the stream omits an id, mirroring the Google provider'sgem_<index>.Tests
openai.test.ts: id-less deltas get a synthesisedcall_0and survive; padding slots still skipped.responses.test.ts: api-key 401 surfacesstatus 401: …Incorrect API key provided, no refresh retry.All 158
@jarvis/llmtests pass; changed files lint clean. (Pre-existing repo-wide typecheck errors about DOM globals are environmental — the sandboxpnpm installcan't fetch the electron/DOM lib types — unrelated to this diff.)🤖 Generated with Claude Code
https://claude.ai/code/session_01AfgFeWNPZaPcWc8CSDL3H6