routine: curiosity-redirect-driven (2026-08-05) - #131
Draft
jim4226 wants to merge 1 commit into
Draft
Conversation
A frontier item that rolls back twice on the same root re-investigate prompt now escalates to a redirect-driven follow-up instead of reissuing the identical retry a third time. Mirrors the exploration pattern from Anthropic's Frontier Red Team cryptographic-weakness research: when a narrow attempt doesn't work, redirect toward broader/deeper exploration rather than repeating it verbatim. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Rdk9Wn6QAJwjEpG9ZwJFjL
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Adds a "redirect-driven" escalation to
Curiosity: after a frontier item rolls back twice in a row on the same root item, the follow-up queue stops reissuing the identicalre-investigate '...'prompt and instead emits a broaderredirect: '...' kept failing ...; broaden scope instead of retrying the same fixitem. A promotion on that root clears its failure streak.Source: https://www.anthropic.com/research/discovering-cryptographic-weaknesses (Frontier Red Team, published 2026-07-28) — "Discovering cryptographic weaknesses with Claude." Quote: "when initial attempts fail, strategic prompting redirects the agent away from 'low hanging fruit' toward deeper exploration."
Theme: 7 — Curiosity / frontier-item generation (exploration policies, novelty detection, automated red-teaming).
Touches:
csis/curiosity.py(Curiosity.record_rollback/record_promoted/next),tests/test_daemon.py.Size: small (~90 LOC incl. tests). Risk: low — does not touch any cycle-9 chokepoint (
Coordinator.__init__,_BackendTracker,writer_iteration_id, promotion CAS); purely additive to the frontier-item source rotation the daemon already consumes through the unchangednext()/record_rollback()/record_promoted()interface.Test plan
Both pass locally (252 passed, including 2 new regression tests:
test_curiosity_redirects_after_repeated_failure_on_same_rootandtest_curiosity_record_promoted_clears_failure_streak).Generated by Claude Code