systematic-debugging: check for self-ingested output in agent/LLM pipelines#1967
Open
CucukovskiNikola wants to merge 1 commit into
Open
systematic-debugging: check for self-ingested output in agent/LLM pipelines#1967CucukovskiNikola wants to merge 1 commit into
CucukovskiNikola wants to merge 1 commit into
Conversation
…elines In multi-turn agent systems the model's own prior output (a reply, status line, or cached summary) can loop back into a downstream stage's input and fabricate "impossible" state. Add a Phase-1 step directing the debugger to trace what each stage READS, not just what the user sent, plus a matching Red Flags entry.
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
In multi-turn agent systems the model's own prior output — a reply, a status
line, or a cached summary from an earlier turn — can loop back into a
downstream stage's input and fabricate "impossible" state the user never
supplied.
This adds a new Phase-1 step (step 6, "Check for Self-Ingested Output") that
directs the debugger to trace what each stage actually READS, not just what
the user sent, and to check the provenance of the exact string the failing
stage matched on. A matching Red Flags entry catches the "symptom points at
feature X, so the bug is in X" reflex before it skips that check.
Both additions slot into the existing Phase-1 data-flow section and Red Flags
list; no other content changes.