Empty Claude API Responses Often Trace to Thinking Tokens Exhausting max_tokens

·2 views

An HTTP 200 response with an empty text block can be more confusing than an explicit error, since nothing is thrown or logged yet parsing still fails. One team found that 4,474 of its 5,369 output tokens went to thinking, leaving only 895 for the actual answer. The article lays out a triage order that separates failures to retry, requests to fix, and conditions to wait out. It advises branching on the error type rather than message text and explains that 429 responses can stem from rate limits or spend caps, distinguished by the retry-after header. It also lists common 400 messages tied to newer models' parameter restrictions.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

How to Benchmark AI Voice-Agent Latency Reliably

Latency measurements for voice systems depend on how the timer is defined, since time to first model event, first audio chunk, audible playback, and full response completion each answer different questions. The guidance recommends measuring the interval from the end of a user turn to the first audible response, while separately tracking first audio chunk and full completion times. Benchmarks should document the device, audio input, network conditions, region, configuration, and whether connections were pre-opened, and should avoid mixing cold and warm turns without labeling. Results should be repeated enough to show variability and reported with sample counts and distributions such as median and p95 rather than the fastest run. Latency should also be checked alongside interruption handling, intelligibility, error rates, and dropped turns, since a fast but incomplete response may not be useful.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Docker Health Checks for Self-Hosted AI Workflows Should Test Real Readiness

Running containers do not guarantee that the applications inside them are ready to serve users. A workflow platform may start while its database is unreachable, and a local model service may only respond after its model has loaded. The article recommends checks that test dependencies that matter to users, such as HTTP endpoints, database connections, or confirmation that a required model is available. It also advises bounded retries, clear timeouts, and reporting which specific check failed, followed by a small safe test of the intended workflow path that avoids production data and credentials. The author stresses that readiness is an application-level condition rather than just container status.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Where human review belongs in AI workflows: designing review points in advance

Human review works best when it is built into an AI workflow from the start rather than added after a system error occurs. The workflow should specify which cases require review, what the reviewer sees, and how processing resumes after a decision. Review triggers can include missing information, unsupported languages, ambiguous categories, conflicting validation rules, or actions with significant consequences. The article advises against relying solely on a model's confidence score unless it has been calibrated for the task, and recommends showing the reviewer the original input, the proposed result, the relevant rule, and the reason for the pause. Recording reviewer decisions, using stable identifiers to avoid duplicates, and tracking review volume and unresolved cases are also recommended.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

French Danfoss panel exposure falls 71% in three days after supermarket cooling attack

Between 11 and 14 October 2026, the number of internet-exposed Danfoss StoreView panels in France dropped from 1,825 to 537 after a cyber-attack on cooling systems at hundreds of supermarkets. Exposure also fell outside France during the same period, though the source does not report figures for every date. A separate eleven-day snapshot taken on 22 October showed 260 exposed Danfoss assets worldwide, with just one in France. Censys reported that panels across all version categories declined by roughly 70%, which suggests operators disconnected systems rather than patching them, though that conclusion rests on one vendor's telemetry. Analysts note that the reduction reflects removal from the internet, not verified remediation.

0 comments Read more at DEV Community