mantishack: Skill for Codex

The mantis_findings MCP server owns finding state (PRD section 9, FR-9.*). Findings live in an append-only event log at .codex/findings/events.jsonl, not in your message history -- so a run is reconstructable and nothing is silently dropped. Never track confirmed/rejected findings only in prose; write them through this service so ids, lifecycle, severity, evidence, and grade are authoritative.

Lifecycle (PRD section 5): candidate -> confirmed | rejected -> exploited -> fixed -> verified. rejected is terminal. Detect is generous, Validate is ruthless.

When to call each tool:

• finding_create: the moment Detect/Recon surfaces a plausible weakness, register it as a candidate. Do NOT self-censor false positives at this stage -- high recall is the point; killing false positives is Validate's job. Pass vuln_class, a one-sentence claim, and location. You get back a stable id -- use it in every later step.
• finding_update to confirmed: only after attacker-simulation. The service refuses to confirm without reachability evidence -- pass reachability_note (how attacker-controlled input provably reaches the sink) or attach evidence first. "No proof -> no confirm."
• finding_update to rejected: requires rejected_reason naming the SPECIFIC roadblock (auth gate, sanitizer/parameterization at the sink, provably-unreachable path, self-only harm, framework auto-protection). The service rejects "seems safe" by requiring the field -- but you must still make the reason specific and true.
• finding_update with grade: pass the 5 axes (impact 0-30, proof 0-25, severity_accuracy 0-15, chain 0-15, report_quality 0-15). The service computes the total and the SUBMIT (>=40) / HOLD (20-39) / SKIP (<20) disposition for you.
• finding_update with evidence / poc / patch: append bounded evidence refs, a PoC descriptor (gated exploit stage only), or a patch descriptor as the finding advances.
• finding_list: before you write a report, list findings to get the authoritative queue and a by-status / by-severity summary. Report from this, not from memory.

Evidence discipline (PRD section 9 / 11): the service refuses payloads that look like they contain a raw secret, token, key, or JWT. Store a redacted reference or a hash, never the raw credential -- in findings, evidence, or your report. Keep evidence bounded: file:line refs, hashes, redacted samples, not full response bodies.

Severity = demonstrated outcome, never bug-class. A candidate from a scanner is not a severity; only a confirmed, reachable, attacker-simulated finding gets a real severity driven by what the exploit actually achieves.