Methodology: stalk, wait, strike, hold

The four-phase Mantis methodology mirrors how a real praying mantis hunts: motionless stalk, patient calculus, instantaneous strike, quiet grip on the catch.

DISCOVER · Stalk

Map the surface before you touch it. The mantis is invisible until the prey is in range.

Implemented by RECON + AUTH.

Subdomain enumeration (subfinder)
Live-host probing (httpx)
Archived-URL crawl (Wayback / OTX / crt.sh)
Tech fingerprinting from response headers + body
Nuclei templates for known-CVE coverage
JS endpoint extraction
4-tier signup flow: API POST → browser auto-signup → assisted manual → token paste

Output: attack_surface.json grouped by priority, plus auth.json with attacker + victim profiles when available.

REASON · Wait

Pick the strike-path before the forelegs move. Vendor-aware bypass selection, kill-chain hypotheses.

Implemented by the planning portion of HUNT + CHAIN.

The hunter agent, on spawn, calls mantis_read_hunter_brief and receives:

Its assigned surface (one per agent, never more)
The vendor-aware bypass table matched to the surface's tech_stack + edge vendor
Exclusions filtered to its surface (dead ends, WAF-blocked endpoints, scope exclusions)
A specialist mode signal (webapp / api / identity / network)

CHAIN is the second half of REASON: the chain-builder looks for A→B exploit chains that elevate severity. CHAIN reads only findings.jsonl and per-hunter markdown notes; it has no HTTP access.

TEST · Strike

Three contact points confirm the catch. Re-prove every finding three times against the live target.

Implemented by the execution portion of HUNT + VERIFY (three rounds).

This is the Multi-Step Evidence contract. A finding is not eligible for SUBMIT until all three rounds agree:

Round	Verifier	Default posture	Survives if
1	`brutalist-verifier`	"This isn't real, prove me wrong"	Re-running the PoC produces the same evidence
2	`balanced-verifier`	Catches false negatives the brutalist over-rejected	A finding the brutalist killed is actually exploitable
3	`final-verifier`	Fresh HTTP requests, fresh context, on only survivors	The bug is reproducible from scratch

Each round writes through mantis_write_verification_round and reads through mantis_read_verification_round. No markdown is on the critical path.

LEARN · Hold

Three crystals in the grip. Stillness, not celebration.

Implemented by GRADE + REPORT + (optional) EXPLORE + fleet-intelligence write-back.

Activity	Output	Who reads it
5-axis scoring	`grade.json`	Report writer + orchestrator
SUBMIT / HOLD / SKIP verdict	`grade.json.verdict`	Orchestrator (drives next phase)
Submission-ready report	`report.md`	You / the triager
Optional patch suggestion	`patch.md`	The vendor's engineer
Optional gated email send	Gmail draft	The vendor's security contact

The grader leads with the Evidence, Not Alerts contract: a finding earns SUBMIT only if Multi-Step Evidence is present.

How this overlays the 7-phase FSM

DISCOVER  →  RECON, AUTH
REASON    →  HUNT (planning), CHAIN
TEST      →  HUNT (execution), VERIFY ×3
LEARN     →  GRADE, REPORT, EXPLORE, fleet write-back

Internally, the FSM is the truth. Externally, Mantis is the story.