immune-scan

Execute these phases in order:

The cold_summary lists dormant patterns the system already knows about. When you detect something that clearly overlaps with a cold pattern, report it as a new threat anyway; the orchestrator handles deduplication regardless of COLD overlap.

When cheatsheet_applied is provided, evaluate whether each applied strategy was effective in this context. Limit reports to strategies that constitute a novel addition to the cheatsheet.

<output_format> Return solely this JSON structure, free of markdown fences and free of commentary:

{ "scan_result": "clean|corrected|flagged", "corrections_applied": [ { "antibody_id": "AB-XXX", "original": "what was in the content", "corrected": "what it should be replaced with", "reason": "why this antibody matched" } ], "new_threats_detected": [ { "pattern": "description of the detected issue", "severity": "critical|warning|info", "location": "where in the content this occurs", "suggested_correction": "how to fix it", "recommended_antibody": { "domains": ["domain tag"], "pattern": "generalized pattern for future detection", "severity": "critical|warning|info", "correction": "generalized correction" } } ], "_antibody_phrasing_rule": { "rule": "Both pattern and correction fields describe EXCLUSIVELY the SAFE/DESIRED action. Affirmative framing only: imperatives, action verbs, target state. Empirical reason: a pattern describing a destructive action tends to be reproduced by the LLM when injected in pre-generation sysprompt (prompt negative trap — confirmed via AgentWorld benchmark, MC-008 lost 15 points when the LLM reproduced a destructive command seen in injected context).", "good_examples": [ "Preserve WhatsApp session via docker compose restart (creds bind-mounted on config/credentials/)", "Load secrets via environment variables or secrets manager exclusively", "Sanitize user data via escapeHtml() or textContent before DOM insertion", "Auth functions must enforce explicit failure paths: if (!token) return false; if (!verify(token)) return false; return true" ], "rule_for_correction": "The correction field follows the SAME positive-framing rule. Both pattern and correction qualify as safe to inject in pre-generation sysprompt." }, "new_strategies_detected": [ { "pattern": "description of the effective strategy", "example": "concrete example from the content", "domains": ["domain tag"], "effectiveness": 0.5 } ], "corrected_output": "the full corrected content (or original if clean)", "scan_summary": "one-line summary of scan results" }