Replies that finish the job

Models stop short, loop, or write a tool call where it can't run. Since 0.2.14 QuorX catches each of these and sends the model back to finish, or stops it with a clear note, and 0.2.15 closes the cases that still ended a long build early.

1 November 2026QuorX v0.2.14modelstests

Four ways a model stops short, each paired with what QuorX now does about it.

Some models are great at the work and sloppy at the routine around it. They announce a step and never take it, write a tool call inside a code block, or repeat themselves until the reply runs out. Before 0.2.14, a reply like that just ended, and you had to say "go on". Now QuorX notices and sends the model back.

What QuorX catches

What the model doesWhat QuorX does now
Says what it will do, then makes no callSends it back to make the call. A model that keeps stopping short ends with a clear note instead.
Writes a call inside a code block, or right after a sentenceAsks for the call again, instead of showing it as raw text
Loops, writes the conversation's next turns itself, or runs far too longStops it, keeping every call it made before
Ends with a final answer that reads like a next stepKeeps it as the answer, with no false warning

4 steps, 1 loop back.

try again1The modelanswersone step2QuorX checks thestepis it what it says?3Sent back tofinishwith what was wrong4Or stopped, witha notework kept

Fixed in 0.2.15

A few cases still ended long builds early. They're closed now:

  • Calls in DeepSeek's own format, or copied from earlier ones without this step's key, are asked for again instead of ending the reply.
  • File writes with quotes in their text (CSS font names, for example) and calls with flat arguments run, instead of failing again and again.
  • A long build no longer ends on a third blank step. A blank answer gets its retries again after any step that made calls.

Safe as well as stubborn

Asking a model to "try again" must never make something run that shouldn't. Models that write their calls as text (the apps on your PC, and some others) only get a call run when it:

  • carries the key for that step, which changes every step;
  • sits on a line of its own;
  • is outside code, quotes and plans.

A web page or a file the agent read can't trigger a call, and a plan or a mention of a call is never treated as one.