Repository navigation
The LLM finishes what it starts
Two fixes born from a field report of tasks dying mid-way.
Edit recovery now completes the task
When a SEARCH/REPLACE edit failed to match, Monastery's recovery ladder repaired and applied it — then went silent. The model never learned its edit landed, so every remaining step of the task was abandoned. Recovery now refreshes the live preview and hands the conversation back to the model with a note that the files are current, so it finishes the job. (With auto-continue off, you're told how to resume instead.)
Response discipline (learned from a deep-dive into bolt.diy)
The system context now carries the completeness rules that make bolt.diy reliable, adapted to Monastery's hybrid edit protocol:
- Plan the entire request first, then emit all file changes in a single response — never "continue later".
- Prefer complete-file blocks over SEARCH/REPLACE for small or mostly-changed files — full contents always apply; anchors can miss. This directly reduces "no SEARCH text matched" failures at the source.
- No placeholder ellipses; minimal prose.
Monastery keeps its token-efficient SEARCH/REPLACE for large files and its three-stage recovery ladder — both ahead of bolt.diy's whole-files-only, no-auto-recovery approach — now with the loop actually closed.
Also
- Project deletion fixed: deleting a previously-deployed project failed with a FOREIGN KEY error (deployment-tracking rows weren't cleaned up) and left a half-deleted project. Cleanup is now atomic (DB first, directory after) and includes deployment records — previously stuck projects heal on the next delete click.