WIP: heartbeat 'challenge' action — anti-drift audit of shared memory

[BobbyZhouZijian]

Status: WIP / draft

Skeleton only. Prompt + registration + tests. No active counter-attempts — review-half only, per maintainer scoping decision.

Refs #67 (shared-memory drift / memory poisoning)
Refs #49 (Coyote / mandatory adversarial dissent)

Why

#67 raises that shared notes/skills can quietly homogenize agents over long runs (one agent's hallucination becomes group doctrine). #49 proposes a heavier governance stack but the one well-grounded primitive in it — a mandatory adversarial pass — fits CORAL's existing heartbeat machinery without buying the rest of that framework.

This PR adds that primitive and stops there.

What this PR does

1. New prompt template coral/hub/prompts/challenge.md instructing an agent to:

   • Identify the highest-traffic notes/skills.
   • Adversarially try to falsify each (evidence check, generalization check, staleness check, counter-search against top attempts that didn't follow it).
   • Re-classify, do not delete — set status: validated | hypothesis | stale | disputed so future agents see what was once believed.
   • Append a dated entry to challenge_log.md summarizing the audit.
   • Hand back without changing the agent's current strategy.

2. Register challenge in coral/hub/heartbeat.py:

   • DEFAULT_GLOBAL["challenge"] = True — population-level concern, one pass per run is enough.
   • DEFAULT_TRIGGER["challenge"] = "plateau" — drift matters when scores stop improving, not on a fixed clock.

3. Add to default heartbeat list in coral/config.py: every=10, is_global=True, trigger="plateau".

4. Tests asserting registration and default-config wiring.

Why these defaults

• Plateau trigger, not interval — adversarial review pays off when the system has stopped finding gains. On a healthy upward trajectory it would just burn turns.
• Global, not per-agent — sycophantic convergence is a population property. One challenger pass across the run is the right granularity.
• Re-classify, not delete — preserves the institutional record of what was once believed, so future agents can re-evaluate rather than relearn from scratch.
• Distinct from lint_wiki — lint_wiki does janitorial work (merge duplicates, fix orphans). challenge questions whether the surviving content is true. The prompt explicitly calls this out.

Out of scope (deliberately)

The "active" half of the anti-drift design — running attempts that deliberately violate consensus (a agents.challenger_fraction knob, or a sharing-disabled agent slot) — is not in this PR. Heartbeats interrupt an existing agent rather than spawn a divergent one, so that piece belongs to agent-spawning, not heartbeat. Tracked for a follow-up.

Execution plan (for reviewers / next steps)

This PR is the skeleton; landing the full feature needs:

• Prompt + registration (this PR)
• Default-config entry (this PR)
• Tests for registration + config defaults (this PR)
• Field test on a real run — verify the prompt produces useful re-classifications, not noise. Best validated on a long-running task config from examples/.
• Decide whether challenge_log.md should be schema'd (YAML frontmatter per entry) so the UI can render it; currently free-form markdown.
• Consider a coral notes --status disputed filter so disputed notes are easy to find.
• Active half: agents.challenger_fraction config knob to spawn a fraction of agents in sharing-disabled "challenger" mode. Separate PR.
• Eval whether every=10 plateau threshold is reasonable across task types or needs to be task-specific.

Test plan

• uv run pytest tests/test_heartbeat.py tests/test_config.py -v — 33 passed
• uv run pytest tests/ — 119 passed
• uv run ruff check on touched files — clean
• Smoke test on an actual coral start run to confirm the action surfaces in coral heartbeat listing and fires after a plateau.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
coral	Ready	Preview, Comment	Apr 25, 2026 1:43pm