90s idle timeout fires on queue wait with a self-hosted backend, and transient failures spend the paid-send budget and park batches permanently (13.35.0)
This Issue has meaningful public discussion or positive reactions, but evidence still comes from one repository.
Project strength and demand confidence are measured separately.
SOURCE EVIDENCE
Start with what users actually said
Reporter context: Summary With a self-hosted OpenAI-compatible backend (Ollama over Tailscale), 13.35.0 parks batches permanently because the 90 s idle timeout fires on time the request spends queued , not on a slow or hung model. Two attempts are then spent and parks the batch, which is never resent. Result on my machine: 14.5 h with…Excerpted from the public Issue. Read the complete thread before interpreting it.
Recruit: Recruit people who can bring a recent literature, evidence-review or source-tracing task.
Guardrail: Measure citation accuracy, coverage and review time; never reward speed at the expense of source fidelity.
01 · Cost & token efficiency
Write the problem hypothesis
For [operator], one completed [workflow] consumes [tokens or spend], exceeding [acceptable budget] because [specific repeated step].
You can name one user, one situation and one measurable consequence without proposing a feature.02 · EVIDENCE INTERVIEW
Interview five researchers, analysts or librarians
Show the cost or token trace for one completed task.
Which step consumes the largest share?
How often is context or work repeated?
What quality level must remain unchanged?
What monthly limit would make this sustainable?
At least three people independently describe the same painful workflow with recent examples.03 · MINIMUM TEST
Run the smallest experiment
Instrument one representative workflow, remove or cache its highest-cost repeated step, then replay the same task set against the original baseline.
Median cost falls by a written target across at least five comparable runs without reducing completion quality.Measure citation accuracy, coverage and review time; never reward speed at the expense of source fidelity.04 · DECISION GATE
Make a build decision
Build: repeated pain and active commitment
Narrow: pain is real but the audience or job differs
Stop: weak frequency or no behavioral proof
Do not let GitHub engagement replace direct validation.
Why this brief exists
Information has value only when it changes action. This page turns one public signal into a bounded validation exercise. It is a research aid, not proof of demand, investment advice or a product recommendation.