Research that finishes the job
Ask a model to dig deeper and sometimes it would run one search, write "Next I'm checking independent sources", and then stop. The reply ended there, with a promise and no follow-up, and you had to push it again yourself. The model had described its next search instead of running it, and nothing told it that once it stops writing, nothing else happens. Two changes close that gap. First, every model with web search on is now told plainly that the reply ends when it stops writing. So when a question needs several searches, it runs them one after another in the same reply, and when the first results are weak (thin, off topic, or mostly vendor marketing) it searches again with a sharper query instead of narrating what it would do. After each search it also knows how many searches it has left. Second, Portail now notices when a reply that has already searched ends by announcing more research, in English or Russian, and sends it back so the model either runs that search or wraps up properly. You see the reply carry on under its last line. Questions and offers like "Want me to look into pricing next?" are left alone, since those are a fine way to end, and replies that never searched are never sent back. Each reply can also run up to five searches now instead of three, which gives research questions room to check more than one angle. Credits are charged the same way as before, and a reply is sent back at most twice, so it can never loop. The same fix applies in Telegram and in scheduled tasks.
648 words
Replies that say they will keep researching now actually do it, in the same reply.
Weak or marketing-heavy results lead to another, sharper search instead of a stopped reply.
Up to five web searches per reply, up from three.
Portail v2.6.27 shipped on 2026-10-15 as a patch release focused on research that finishes the job. Ask a model to dig deeper and sometimes it would run one search, write "Next I'm checking independent sources", and then stop. The reply ended there, with a promise and no follow-up, and you had to push it again yourself. The model had described its next search instead of running it, and nothing told it that once it stops writing, nothing else happens. Two changes close that gap. First, every model with web search on is now told plainly that the reply ends when it stops writing. So when a question needs several searches, it runs them one after another in the same reply, and when the first results are weak (thin, off topic, or mostly vendor marketing) it searches again with a sharper query instead of narrating what it would do. After each search it also knows how many searches it has left. Second, Portail now notices when a reply that has already searched ends by announcing more research, in English or Russian, and sends it back so the model either runs that search or wraps up properly. You see the reply carry on under its last line. Questions and offers like "Want me to look into pricing next?" are left alone, since those are a fine way to end, and replies that never searched are never sent back. Each reply can also run up to five searches now instead of three, which gives research questions room to check more than one angle. Credits are charged the same way as before, and a reply is sent back at most twice, so it can never loop. The same fix applies in Telegram and in scheduled tasks. The work is recorded here because small releases are still product history. They change how people sign in, read responses, pay for access, share conversations, open files, or trust the system when something goes wrong.
The main changes were direct: replies that say they will keep researching now actually do it, in the same reply, weak or marketing-heavy results lead to another, sharper search instead of a stopped reply, and up to five web searches per reply, up from three. That list is intentionally plain. It names what changed without turning the release into theatre. When a release fixes a build, a redirect, a storage path, or a UI edge case, the value is often that users stop noticing the problem. The archive keeps that invisible work visible enough for future debugging and for anyone checking whether Portail has cared about the details over time.
This version also fits into the wider Portail contract. One account, one chat history, and one model picker only work if the smaller pieces keep their promises. Authentication has to fail clearly. Billing has to count honestly. Uploads have to arrive in a form models can read. Admin tools have to show the operator what happened instead of hiding it behind a blank screen. v2.6.27 moved that contract forward in its own narrow lane.
The screenshots attached to this entry show the public release record as it now appears on portail.cc. For older releases, this archive was rebuilt from the git history, package version changes, release modal text, and commit subjects. That means the page is honest about its source: it documents the shipped change and preserves the record going forward, even when the exact old production screen cannot be recreated from a single saved image.
The practical result is simple. Someone landing on this page can see the version number, date, release tier, headline, highlights, and a real rendered screenshot of the archive page. Search engines can index the post through the canonical URL, sitemap entry, and structured data. Future releases now have a standard to meet before deploy, so the history should get cleaner from here instead of fuzzier.