Repository navigation
Drop the task tool's all-done reminder and trim request checks - #692
Merged
Merged
Conversation
The task tool already makes the model put acceptance criteria in each task and verify them before `complete`, and task state survives compaction. The reminder added in ea46581, appended when the last task is completed, asked for a second sweep over the whole request without knowing what was already verified, so after compaction the model redid all of it. Remove it; `all-tasks-done?` stays for `auto-clear-completed!`. The code agent's "Meeting the request" section also asked for the most literal reading of ambiguous requirements, a final requirement by requirement sweep and checks of failure paths in general, so models read loose instructions literally and handled edge cases nobody asked for. Keep its core: implement exact stated behavior as stated, check stated requirements as existing tests only show old behavior still works, and say what was verified. Beyond stated behavior, ask for the edge cases the code will realistically meet, not hypothetical ones: this guards against doing too much and too little without limiting open-ended requests, and without asking the model to judge how likely a case is to be tested. Also tell it not to bend tests. The cleanup sentence reworded in 9d98b88 goes too, as it was only needed for the instruction to write check scripts. Stated requirements now go into the acceptance criteria of the task that implements them, so they are checked once, on `complete`. The shell_command note that background jobs stop with ECA stays.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
ea46581 and 9d98b88 made the
codeagent check more than it needs to: thetasktool's all-done reminder asks for a second check of the whole request without knowing what was already verified, so after compaction the model redoes everything, and the prompt asks for the most literal reading of loose instructions and for handling edge cases nobody asked for. This removes the reminder (tasks are already verified oncomplete) and keeps the useful core of the prompt: implement exact stated behavior as stated, don't trust existing tests alone, don't bend tests, handle realistic edge cases rather than hypothetical ones, and say what was verified. Stated requirements now go into each task's acceptance criteria, so each one is checked