A coding model can produce valid work and still stop halfway through a file. Asking it to “continue” without context often makes it restart the file, repeat functions, or close a code fence in the wrong place.
The coding path preserves the tail of the partial output and instructs the next turn to continue exactly where it ended. It also detects when the new output adds nothing beyond repeated text.
What was actually going wrong
Truncation was handled as a fresh generation request rather than continuation of one artifact.
What I tried
- Requesting the whole file again with a larger prompt
- Appending continuation text without overlap checks
- Trusting a closing code fence as proof the file was complete
The continuation prompt includes the exact tail, forbids restarting, merges new text, and stops if the model only repeats prior output.
Why it worked
The model receives a precise cursor and the merger verifies actual forward progress.
Continuation needs both an anchor and a repetition guard.
Where EACI uses this today
The optional Main coding workspace can recover partial generated files without silently duplicating them.
This journal covers real engineering on EACI Companion / The Veil. Companions include Caelum, Chad, Natalia, Atreus, Luna, Roxy, and Cael. Journal articles stay family-safe in content. See Privacy and Ethics.