07-security

07 · Security — secure

Two modes. 07-security sprint N audits one sprint's surface (post-ship or pre-ship of that slice); 07-security full audits the whole codebase (pre-release / periodic). 07 is the deep security pass: a bounded parallel read-only OWASP panel whose findings a single sequential synthesizer reduces into docs/security/security-audit-sprint-NN.md with a machine-readable PASS / REMEDIATE / BLOCK verdict that 06-release's G6 gates the ship on. Your graded value is not "finds vulnerabilities" — a strong reviewer does that too (and cries wolf 68–97% of the time). It is the isolated, de-duplicated, false-positive-controlled, machine-verdicted audit a release gate can trust: a complete OWASP partition (no area silently dropped), each finding carrying a source_quote + a re-derived severity, read-only (the audit touched no code), routed to the owning skill.

Operating principle — partition completely, read blind, synthesize sequentially, verdict on severity

• Panel — bounded, parallel, read-only, blind. A fixed complete partition of the OWASP areas across 4 readers (+ a conditional 5th for LLM/agentic), each seeded with only its area-slice + the code paths + a neutral, evidence-required prompt. Leading "find the vulnerability here" framing inflates false positives; neutral "analyze this code; cite file:line + the tainted source" is what controls them. Each reader is blind to the others and writes no code. Cap 3–5 (shared/subagent-protocol.md). Isolation is real only because the spawner is fresh.
• Deterministic scanners under the panel. Where a runtime exists, run npm audit / osv-scanner, gitleaks, semgrep and feed results to the owning reader — the AI panel complements deterministic tools (it catches IDOR, authz gaps, business logic they miss). Their absence is recorded, never silently skipped.
• Synthesizer — one sequential reduce, never parallel. De-dupe by target, take max severity, preserve every source_quote, build the risk matrix, and run the completeness lens ("which area returned zero findings — clean, or under-looked?"). A parallel merge would re-amplify error.
• Verdict on severity, routed; honest about limits. PASS / REMEDIATE / BLOCK keyed to Critical/High presence (below), each finding routed to its owning skill. An AI audit is not a substitute for third-party penetration testing, and live-infrastructure hardening is out of the repo audit's reach — say both, don't pretend otherwise.