From a Healthy Backend to a Useful Answer
Agent Reach’s most demanding test begins after Doctor reports a plausible route. This chapter follows the final mile from backend readiness to meaningful content, using YouTube captions and audio transcription as the central application. The question is not whether a command exits successfully, but whether the requested operation returned the kind of evidence an agent can safely use—and whether a fallback adds new cost, data exposure, or resource limits.
The Acceptance Test Is the Content
A health check and a content operation answer different questions. Doctor may establish that yt-dlp is installed, executable, and accompanied by the JavaScript runtime its route expects. That is useful local evidence. It does not request a particular video, inspect whether that video exposes captions, or determine whether the returned text is complete enough for a summary. Agent Reach’s video guidance therefore treats active_backend: yt-dlp as a starting point, not as proof that a user’s video can be transcribed.
The difference is easy to miss because the first command often succeeds. A version probe is short, predictable, and relatively safe. A caption request crosses into the platform’s current behavior and the target video’s own metadata. The platform may challenge the request, the caption endpoint may return an empty response, the video may have no captions, or an otherwise valid caption URL may have expired between retrieval and download. These causes look similar if the agent checks only the exit code. They require different responses.
The project’s documented YouTube path begins with the primary subtitle operation. The success condition is actual non-empty caption content, not simply a process that returned zero. If the first attempt encounters a bot challenge, an empty caption response, or another content-level problem, an OpenCLI transcript route may be tried when its browser bridge is connected. An expiring caption URL may justify a bounded retry; it does not justify treating an empty response as evidence that the video has no subtitles. If captions are genuinely absent, or if the caption routes remain unsuccessful, audio transcription becomes a different modality rather than another attempt at the same operation.
That distinction between transient failure and absence is a small but important form of reasoning. A transient failure is an observation about the retrieval attempt: the URL may have expired, the network may have returned a challenge, or the external service may have failed temporarily. Absence is a claim about the target: there may be no available subtitles. The agent should not move from the first category to the second without an appropriate retry or alternate check. Nor should it retry indefinitely. The reference guidance bounds retries for the expiring-caption case and defines success by substantive content.
The same principle applies to a web page. Agent Reach’s Jina Reader route is designed to turn a public URL into readable content, but its adapter checks for high-confidence anti-bot or security-verification pages and refuses to present them as the requested article. A challenge page is non-empty, and it may even look structured. It is still not evidence about the target page’s contents.