v2.7.8: The Context Counter Tells the Truth, Mail Shows Up Again, and the Mysterious 'court' Is Gone

The little percentage next to the model picker has one job: tell you how full the current conversation is. For a while it was quietly answering a different question, how full the last conversation was. v2.7.8 makes it say what it means, in every situation, and clears out a handful of other small lies the interface was telling along the way.

A cleared conversation starts empty

Clear a session, start fresh, and the counter would still show the previous conversation's percentage. You could open a completely empty chat and watch it sitting at 14% before a single word was exchanged. Not harmful, but unnerving, like a fuel gauge that reads a quarter tank on a car you just filled.

• A new conversation shows nothing until it has actually used something.
• Resuming a conversation shows its real size straight away, instead of making you wait for the first reply to find out.
• The counter now updates on every moment that changes the window: when a reply starts, when it ends, after a compaction, when a session is cleared, and when one is resumed. It used to refresh on exactly one moment, the end of a reply, which is why it drifted the rest of the time.

Compacting drops the counter immediately

Run /compact and your conversation genuinely gets smaller. The counter, though, kept showing the old size until you spoke again, which defeated the entire point of watching it before deciding to compact. Now it falls the moment the compaction lands, so the number matches the reality you just created.

Changing the context window reaches existing employees

If you run custom LLMs, this is the fix that has been quietly costing you. The instance-wide context window in Admin, Settings was copied into each employee when its provider was first chosen, and then never read again. Raising it changed nothing for anyone. Existing employees stayed on the size they were created with, and the only escape was to reopen each one and pick its model again, one at a time.

• Set it once, it applies everywhere. Each employee picks up the current size the next time its session starts. Clear the session and carry on.
• Nobody is interrupted. Employees already mid-task keep their current size until they restart, so changing the setting never cuts a live conversation short.

Incoming mail shows up in the chat again

Mail arriving for an employee had stopped appearing in the conversation. No badge, live or after a refresh, so unless you went looking you would never know something had landed. The cause was a filter: the notice was being treated as an internal system message and stripped out on its way to the browser. Incoming mail, ended phone calls, and messages from every connected channel are visible in the chat again, the way they should be.

No more stray word before a tool runs

This is the one people had been reporting for weeks. Employees sometimes dropped a single meaningless word into the conversation right before running a command, most often the word "court", appearing from nowhere with no relationship to anything around it. It was never a typo or a hidden message. It was a fragment that slipped out ahead of the real work, which always happened in the command that followed.

Those one-word fragments are no longer displayed. And to be clear about what did not change: the short confirmations you actually rely on, the internal TRIFORCE_OK and HEARTBEAT_OK signals, a quick "Je regarde." before an employee digs in, are all untouched. We stripped the noise, not the signal.

Under the hood

The reason the counter misbehaved is worth a sentence, because it explains why it took real work to fix. The percentage is computed per CLI, and Claude, Codex, custom LLMs, and the internal bridge each track usage differently. All four were reading whichever conversation had been touched most recently, so a brand-new chat would happily inherit the figure from an old one. Each one now ignores anything belonging to a conversation you have already cleared, which is why the counter finally stays honest no matter how you switch around.

Why bother with any of this

None of these are headline features. A counter that reads correctly, a mail badge that appears, a random word that stops showing up. But this is the layer where trust in a product is actually won or lost. A number you cannot believe is worse than no number at all, and a stray "court" in the transcript makes a capable employee look confused for no reason. Fixing them is how the whole thing keeps feeling like something you can rely on rather than something you have to second-guess.

v2.7.8 is live now. If you run custom models, take a moment to confirm your context window is where you want it, then clear a session to pick it up. And if you had been wondering about that word, wonder no more: it is gone.