Repository navigation
feat(agent): ask for the command after a reply without one (plan #11) - #22
Merged
Merged
Conversation
… none A reply that ends without a command and wasn't cut off now gets a nudge that asks for the command itself, and the next request runs at 0.7. At 0.2 the model repeated the same one-line preamble until the agent gave up. Cut-off replies and the give-up rule are unchanged.
ruoshi/feat-act-nudge at 31e2499 on the 10-task dev slice, LLM_MAX_TOKENS=8192, concurrency 2. Exported with scripts/export_trials.py. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…shi/feat-act-nudge
This was referenced Oct 10, 2026
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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Stacked on #21 (retargets to main once #21 merges).
When a reply ends without a command and wasn't cut off ("Let's start by exploring the environment."), the agent now asks for the command itself and samples the next reply at temperature 0.7. Cut-off replies and the give-up rule are unchanged.
On the 10-task slice, 3 trials:
Also records the 16k run (plan #12): halves cut-off give-ups, doesn't fix thinking loops; to be rerun on top of this. Details in docs/experiments.md (2026-10-09).