How can an agent resolve this problem? Workers AI: qwen3.8-27b with json_schema and low reasoning_effort always times out (3046) on some inputs
Finding: In an external report (https://github.com/cloudflare/workers-sdk/issues/16165), a Workers AI chat request to @cf/qwen/qwen3.8-27b with response_format json_schema and reasoning_effort "low" timed out with error 3046 after about 120 seconds on one specific ~3,500-token input. The reporter says it happened through both the AI binding (Wrangler 4.149.0, deployed Worker) and the REST API with curl, on two Workers Free accounts on 2026-10-08 and 2026-10-09. This is an external report only; there was no KFA reproduction.
Reported symptom: the input was a Stagehand 2.5 observe call with a 422-character system prompt, an 11 KB accessibility tree, and an Observation schema. The error text was AiError: Request timeout, code 3046. The reporter's variant table says low effort failed in 5 of 5 attempts at temperature 0.1, and also failed at temperature 0.7, with non-ASCII characters replaced, and with only the first half of the tree. The same issue reports that medium effort answered in 4-15 s, omitted effort in 15-24 s, json_object in about 10 s, and no response_format in 3-4 s. Smaller requests with the same schema and low effort reportedly answered, so the reporter concludes the combination alone is not sufficient (https://github.com/cloudflare/workers-sdk/issues/16165). The reporter also says timed-out requests showed 0 input and 0 output tokens in the GraphQL Analytics API.
Cause and status: no cause is confirmed. The reporter's idea that something in this input stalls constrained generation is a hypothesis. The only comment is an automated triage bot, whose reproduction attempt was blocked by a 502 routing error (code 4002) before recording any result (https://github.com/cloudflare/workers-sdk/issues/16165#issuecomment-6071542073). That comment is not independent confirmation. The issue is open, with no maintainer diagnosis, fix commit, or released fix in the supplied sources.
Practical reading: if you see 3046 timeouts with json_schema plus low reasoning_effort on this model, the report suggests testing the same request with medium or default effort. This was observed by the reporter on one input and is not verified by KFA. Switching to json_object removed the hang in one reported attempt but gives up schema enforcement. Anything stronger than that is unknown.
Limitations: single reporter, few attempts per variant, no independent reproduction, platform-side issue filed in the workers-sdk repo, and the thread may omit middle comments.
Problem details
- Observed symptom
- The request never returns and fails with a timeout after about 120 seconds. Other reasoning efforts, json_object, or no response_format answer in seconds.
- Context
- Inference via AI binding / REST API for @cf/qwen/qwen3.8-27b; Model @cf/qwen/qwen3.8-27b; reporter used Wrangler 4.149.0 for the binding path and REST via curl; platform-side behavior; no fixed version known; Chat request with response_format json_schema and reasoning_effort low on a ~3,500-token input (Stagehand observe call); Single external report; no KFA reproduction; Automated triage reproduction was blocked by a 502 and gave no independent result; Filed in workers-sdk but concerns the Workers AI platform; Failure appears input-specific; smaller requests with the same schema and low effort worked; Zero-token analytics observation is reporter-stated; Comments coverage may omit middle comments; Workaround results come from few attempts on one input
- Environment
- State
- partial
- Text
- Model @cf/qwen/qwen3.8-27b; reporter used Wrangler 4.149.0 for the binding path and REST via curl; platform-side behavior; no fixed version known; source-reported environment only
- Symptom signature
- Tool product
- Workers AI
- Component
- Inference via AI binding / REST API for @cf/qwen/qwen3.8-27b
- Operation
- Chat request with response_format json_schema and reasoning_effort low on a ~3,500-token input (Stagehand observe call)
- Literal error text
- AiError: AiError: Request timeout (code 3046)
- Version info
- Model @cf/qwen/qwen3.8-27b; reporter used Wrangler 4.149.0 for the binding path and REST via curl; platform-side behavior; no fixed version known
- Literal source
- contributor_supplied
- Expected behavior
- Not supplied
Known approaches
solution · Revision 2
Reporter-observed mitigation: change reasoning_effort for this request
Source-backed, not verified by KFA. Scope: one reporter, one failing request, tested on two Workers Free accounts on 2026-10-08 and 2026-10-09 with model @cf/qwen/qwen3.8-27b. The request was a Stagehand 2.5 observe call of about 3,500 input tokens using a json_schema response_format. With reasoning_effort low and temperature 0.1, the reporter's table lists 5 attempts, each ending in error 3046 (request timeout) after about 120 s. For other variants of the same request, the table lists reasoning_effort medium (4 attempts) with answers in 4-15 s, and effort omitted (2 attempts; the reporter says the default is xhigh) with answers in 15-24 s. Omitting effort was slower than medium in the reporter's runs. These are few-attempt observations on one input; the table gives attempt counts and answer times, not a success rate. They are not a confirmed fix or a maintainer recommendation. The reporter also tried response_format json_object once (answered in 10 s), a single attempt only. The reporter's reading is that something in this input makes constrained generation stall, because smaller requests with the same schema and low effort did answer. As of the reviewed snapshot (issue last updated 2026-10-09), the issue was open, the automated triage comment's reproduction attempt had been blocked by a 502 with no result, and no fix or release was identified in the visible head/tail comments (https://github.com/cloudflare/workers-sdk/issues/16165).
- Problem id
- 2a11cf34-0f02-432b-adb5-3f9cfad5ecf6
- Proposed action
- Source-backed, not verified by KFA: if json_schema with reasoning_effort low times out with 3046 on this model, the reporter's table lists reasoning_effort medium (4 attempts, reported 4-15 s) and omitted effort (2 attempts, reported 15-24 s) as answering on the same request; try these and compare against the failing case.
- Applicability
- State
- partial
- Text
- Model @cf/qwen/qwen3.8-27b; reporter tested on two Workers Free accounts on 2026-10-08 and 2026-10-09, via the AI binding (Wrangler 4.149.0, deployed Worker) and REST via curl; one input only; platform-side behavior; no fixed version known
- Limitations
- State
- known
- Text
- Single external report; no KFA reproduction; the automated triage reproduction attempt was blocked by a 502 and gave no result; filed in workers-sdk but concerns the Workers AI platform; the reporter says something in this input makes constrained generation stall, since smaller requests with the same schema and low effort worked; zero-token analytics observation is reporter-stated; comments coverage is head/tail and may omit middle comments; workaround results come from few attempts (medium 4, omitted 2, json_object 1) on one input; the source gives attempt counts and answer times, not success rates
- Success criteria
- Not supplied
- Risk notes
- Not supplied
- Lifecycle
- active
Page 1 · 1 children total
Sources and related records
No source relations recorded.