fix(api): stop mock tool loop after role=tool observation - #674
Closed
cursor[bot] wants to merge 31 commits into
Closed
fix(api): stop mock tool loop after role=tool observation#674cursor[bot] wants to merge 31 commits into
cursor[bot] wants to merge 31 commits into
Conversation
…closed otherwise Chat history: message-level audio and legacy function_call are null/empty omit no-ops; non-empty fail closed with named errors (including tools passthrough). Tip substrate from #577 assistant refusal/annotations honesty. Local full unit: 940 passed.
…ed otherwise OpenAI fine-tune style message weight is not applied on this gateway. Accept null/0/1 as honest no-ops; reject other types and values with invalid_message_weight. Tip substrate from #578. Local full unit: 943 passed.
…ion role Reject unsupported message keys with named unknown_message_fields (not silent strip or tools-passthrough smuggle). Reject legacy function role with invalid_message_role migration to tool. Tip substrate from #579. Local full unit: 947 passed.
OpenAI partial-assistant prefix flag is not applied on this gateway. null/false are honest no-ops; true and non-booleans fail closed with invalid_message_prefix. Tip substrate from #580. Local full unit: 950 passed.
…therwise Named invalid_max_tool_calls on /v1/chat/completions instead of opaque unknown_fields. Aligns with Responses max_tool_calls honesty; gateway has no multi-step tool loop.
…losed otherwise Legacy /v1/completions treated max_tool_calls as unknown_fields. Accept the key for named invalid_max_tool_calls (null/empty/whitespace omit-equivalent), matching chat/Responses honesty so SDKs get a clear migration path.
SDK clients often send include_usage/include_obfuscation as JSON null. Drop null flag values before validation so null (and null+false mixes) match omit / all-false no-ops on chat, Completions, and Responses. True flags remain fail-closed with invalid_stream_options.
…age honesty Null include_usage/include_obfuscation stay omit-equivalent, but unknown stream_options keys no longer become no-ops just because their value is null. Weight, prefix, refusal, annotations, developer role, empty user/system content, and participant name now use the same named errors on the tools passthrough path as on orchestration. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Hoist stream, required model, stream_options, and temperature/top_p range checks before proxy_completion so a tools or response_format body cannot return a billed JSON completion when the SDK asked for SSE, or silently pick a pool model when model is omitted. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Hoist stream, required model, stream_options, and temperature/top_p range checks before proxy_completion so a tools or response_format body cannot return a billed JSON completion when the SDK asked for SSE, or silently pick a pool model when model is omitted. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Hoist attribution and routing validation before proxy_completion so a tools or response_format body cannot bill a sync completion with an unknown spend dimension or a batch/latency_tolerant hint. Tools passthrough has no batch job plane. Buyer next action: send known sync attribution; omit routing.channel=batch and latency_tolerant=true on tool-calling requests. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Call the orchestration message, max_tokens, attribution, and routing validators before proxy_completion so a tools or response_format body cannot bill a completion with no prompt, crash on a non-object entry, or silently drop unknown spend/routing keys. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Match _validate_messages: tools + user content 123 must 400 invalid_message instead of billing a sync completion. Assistant/tool JSON null stays omit-equivalent. Buyer next action: send user/system content as a non-empty string or a content-parts array. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Match the tools-path honesty contract to the invalid_message behavior so SDK clients send a string or content-parts array. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Passthrough has no batch job plane. Reject routing.channel=batch and latency_tolerant=true before proxy_completion so a tool-calling body cannot bill a silent sync completion. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Hoist the remaining chat request knobs before proxy_completion so an OpenAI SDK tool-calling body cannot bill a sync completion for seed, stop, n>1, logprobs, logit_bias, out-of-range token/penalty values, unsupported reasoning_effort, or a non-default service_tier. Buyer next action: omit those fields on tool-calling requests; they are not applied on this gateway. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Call _validate_messages before proxy_completion so tools + [] / omitted / null / non-list / non-object messages raise invalid_message instead of billing a completion with no prompt. Buyer next action: always send a non-empty messages array of objects on tool-calling requests. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Hoist the orchestration request-knob validators before proxy_completion so a tools or response_format body cannot bill a completion that silently drops seed, stop, n>1, logprobs, logit_bias, or out-of-range penalties. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
stream_chat defaulted to 0.2 even after the HTTP path wrote default_temperature from the request. A streamed invoice summary at temperature=0.8 no longer silently falls back to 0.2. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
… proxy SDK optional defaults serialize omitted fields as JSON null. Accepting those keys without popping them is not omit-equivalent: proxy_completion forwards the body and several providers reject a null JSON Schema object. Pop the keys in place, keep non-null wrong types on invalid_tools, and assert omit-real via mock echo on chat and Responses. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
OpenAI SDK tool callers send stream=true. Returning 400 was honest but left every streaming tool client failing. Pipe a single pool agent as chat.completion.chunk frames so concatenated mock content matches the non-stream JSON body, and pipe live provider text so tool_calls deltas survive. Keep include_usage, empty messages, and unsupported knobs fail-closed. Buyer next action: send stream=true on tool-calling requests; omit stream_options.include_usage. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Keep tools/response_format SSE proxy from this branch and take #601's null tool.function omit-pop, request-knob HTTP cases, and streamed route temperature. Conflict resolution keeps stream=true as SSE, not 400. CHANGELOG records the combined unreleased slice. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Unknown assistant tool_calls entry/function fields, non-boolean include_orchestration_trace, and unknown mode used to bill a JSON or SSE completion on the tools proxy. They now return the same named 400s as the orchestration path before the first byte. Buyer next action: send only id/type/function/optional index on assistant tool_calls; send a boolean or omit include_orchestration_trace; send auto/route/conduct or omit mode. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
mock:// lookup_balance now returns assistant tool_calls and streams delta.tool_calls with finish_reason=tool_calls so offline SDK clients match the live _stream_raw path. Invoice identifiers bind from the user text (default INV-9). Cite Toolformer, ReAct, and OpenAI function calling (APA 7th) in docs/papers and the honesty test. Buyer next action: send stream=true on tool-calling requests and include the invoice id in the user message. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
stream_chat recorded temperature but omitted the request-scoped top_p, presence_penalty, and frequency_penalty that chat() already forwards. A streamed invoice summary at top_p=0.1 no longer drops those knobs. Buyer next action: send the nucleus and penalty values you want on stream=true route requests; streaming matches the JSON path. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Streamed top_p / penalties now match chat(). Record the ICLR 2020 nucleus-sampling paper (APA 7th) so the buyer can see why those knobs must not silently drop on stream=true. Buyer next action: send the nucleus and penalty values you want on streamed route requests. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
The HTTP validator already accepted " none " / tab-newline none as none but discarded the stripped token. Mock selection required an exact "none", so a padded none still emitted tool_calls after a 200. Strip in the selector and write the token back so live providers see none. JSON and SSE honesty tests cover exact and padded none. Buyer next action: send tool_choice=none to skip tools; incidental padding stays omit-equivalent. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
Buyers type "invoice 4419" more often than "INV-4419". The mock binder only matched an INV- prefix, so a realistic prompt still defaulted to INV-9. Accept invoice / invoice id / invoice # plus digits. Buyer next action: put the invoice number in the user text; prefixed and bare forms both bind. Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
A buyer who posts the OpenAI function-calling continuation used to get another lookup_balance tool_calls turn on mock://, spinning the official SDK. After a role=tool observation, JSON and SSE stay content/stop. Also lock padded tool_choice write-back via echo.tool_choice == none on HTTP and proxy_completion, and cover bare invoice 4419 on the HTTP/SSE path (Yao et al., 2023; OpenAI, 2024). Co-authored-by: Seongho Bae <seonghobae@users.noreply.github.com>
This was referenced Aug 16, 2026
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
Author
There was a problem hiding this comment.
Do not merge this in parallel with #660 / #682. Prefer #682 #682 for the stream-path tip.
This tip stops the mock loop on any role=tool observation. #660/#682 require a matching tool_call_id, so an unmatched observation cannot masquerade as a completed action. #682 also binds clerk aliases (invoice no. / nr / inv#4419 → INV-4419).
Buyer next action: use #682. After the first tool_calls hop, POST the tool output as role=tool with the same tool_call_id and read content / stop.
Sent by Cursor Automation: Fix Issues
Author
|
Bugbot is not enabled for your account, so this pull request was not reviewed. Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs. |
seonghobae
enabled auto-merge (squash)
August 16, 2026 17:06
auto-merge was automatically disabled
August 16, 2026 18:22
Pull request was closed
Contributor
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.
Successor to #639. Same padded-
none+ bare invoice bind + #620 stream tip, plus the mock tool-loop hole.Unique vs #639 (
2356202)role=toolobservation, mocklookup_balancestayscontent/stopon JSON and SSE. The official SDK continuation no longer spins anothertool_callsturn (Yao et al., 2023; OpenAI, 2024)._passthrough_upstreamwrites strippedtool_choice(none/auto/required) soproxy_completionecho and live hops seenone, not" none ".echo.tool_choice == "none"for exact and padded none, and bindsinvoice 4419→INV-4419on the HTTP path (not onlyproxy_completion).Buyer next action
Send
tool_choice=nonewhen you want no tool call. Put the invoice number in the user text (INV-4419orinvoice 4419). After the firsttool_callsturn, post the result asrole=toolwithtool_call_idand read the answer — this gateway has no multi-step tool loop. Do not merge #639 / #620 in parallel — this tip already contains that stack plus the observation fix.Do not merge in parallel
Prefer this over #639 / #620 / #606 / #617 / #609 / #610 / #613 / #615 for the stream-path tip. Keep KV/token #644, mode/trace #640, verify #634, and Responses stream #628 on their own tips. Independent non-author review required.
Local checks printed
ok: SSE tools honesty (including observation + padded-none echo lock + HTTP invoice bind), paper/API/self-check/conventions, tool_call_id, assistant tool_calls, OpenAI passthrough, tool_choice strip.