Finishing tasks
An agent run is done when the work you asked for is done, not when the model stops typing. In Agent mode, Ingot checks three things before a turn ends: that the agent did not stop halfway, that every file you named actually changed, and that the project's tests pass. Long files are changed in small search-and-replace steps, so a big file is never cut short.


Keep going when the agent stops early
Smaller models often end a turn too soon: they describe the change instead of making it, ask "Would you like me to implement this?", or update the code but not the readme you asked for. When a reply calls no tool and the task is not done, Ingot sends the agent a short hidden note and the run continues:
| The reply… | The agent is told |
|---|---|
| changed no file, though you asked for a change | make the change now, then run the tests |
| asks permission for work you already asked for | do not ask, continue with the next step |
| skipped a part you named (a file, the readme, the changelog, tests) | which parts are still missing, by name |
was empty, only markup like <br>, or one line repeated over and over | answer the request now (asked once, and the stuck reply is hidden) |
A question ("how does slugifyWithCounter work? Do not modify any files") is never nudged. Files you only point at ("look at api.ts") or protect ("do not modify config.json") are not required.
The number of nudges per turn is capped (3 by default). If parts are still missing after that, the chat says so in plain words, for example Not finished: readme.md was not updated, so a reply never reads as done when it is not.


Run the tests before "done"
After the agent changes files for any request that asks for changes, Ingot runs the project checks as Project check cards in the chat: the package.json gate script, or test, lint and typecheck, or pytest, make test and similar for other projects. These runs do not wait for the approval card.
- A failing check goes back to the agent with the failures (file, line and message), up to 3 times. After that the run stops and says what still fails.
- The agent never deletes, skips or weakens a test to get green.
- The agent is told to say the tests pass only when it ran them after its last edit.
A fix or feature request does not commit. The agent commits only when you ask for a commit, on a feature branch, and only when the checks pass.
node --test and the test script
Projects that use the built-in Node test runner work out of the box. Test discovery reads your test script:
test script | What runs |
|---|---|
node --test (also node --import tsx --test) | the Node test runner, with results per test |
xo && node --test | npm test; failures name the failing test, or the lint step when every test passed |
| any other real script | npm test (or your package manager), pass or fail with the output |
Node's TAP and spec output are both read, so a failure shows up as test.js:12 › maxLength cuts at a separator: Expected values to be strictly equal, the same in the Test Explorer and in the agent's check results.
Long files are edited in small steps
A model that rewrites a 400-line file often drops lines or writes "// ... rest of the file unchanged". For files longer than 120 lines the agent sends REPLACE blocks instead: each one finds the exact current lines and swaps in the new ones.
<<<REPLACE readme.md
<<<<<<< SEARCH
### slugify(string, options?)
=======
### slugify(string, options?)
#### maxLength
>>>>>>> REPLACE
>>>END
A whole-file edit that keeps less than 60% of a long file, or leaves a "rest unchanged" placeholder, is refused: the chat says Did not apply the edit to readme.md, the file stays as it was, and the agent is asked for REPLACE steps. A request to delete or shorten content is allowed to shrink the file.
Settings
All four are in Settings › Tools › Agent tools › Finishing tasks and in settings.json.
| Setting | Default | What it does |
|---|---|---|
ingot.agent.autoContinue.enabled | true | Keep going when the agent stops before the task is done |
ingot.agent.autoContinue.maxNudges | 3 | Most "keep going" nudges in one turn (0 to 10); after that the chat says what is not finished |
ingot.agent.finishCheck.enabled | true | Run the project tests after edits for any request that asks for changes (off: only fix and feature requests) |
ingot.agent.edits.replaceAboveLines | 120 | Files longer than this many lines are changed with REPLACE steps (20 to 5000) |


.ingot/artifacts/ ignore themselves in git, and a test run is only written to .ingot/frozen.json when that file already exists.