Speed modes
Five orchestrator variants built on the same 7-phase FSM. Pick the one that matches your cost / time / depth tradeoff.
At a glance
| Mode | Use when | Cost vs. std | Wall clock | Pipeline |
|---|---|---|---|---|
/mantis-fast | Triage an unknown target cheaply | ~10-15% | ~10 min | recon → triage → 1 hunter wave → brutalist → grade |
/mantis (standard) | Default. Balanced quality and cost. | 100% | 30-60 min | full 7-phase FSM |
/mantis-ultra | High-value target, deadline pressure | ~200-300% | similar | full FSM, 3× wider waves, parallel verifier dispatch |
/mantis-loop | Long mission, find rare bugs | budgeted | up to N hours | repeated full FSM via EXPLORE iterations |
/mantis-fullsend | Auto-disclose after a verified find | ~110% | +5-10 min | adds PATCH + DISCLOSE phases |
/mantisplan | Pre-HUNT sanity check (no HTTP) | <1% | ~1-2 min | reads recon, writes plan.md only |
/mantis-fast: triage screen
You don't yet know if this target is worth a full run. Fast mode does recon, triage-agent scoring of every surface (promote / defer / kill), one hunter wave on promoted surfaces only, one round of brutalist verification, and a grade. No chain-building, no balanced/final verification, no report.
If fast finds something interesting, escalate to standard or ultra. If fast finds nothing, you've spent ~10-15% of a standard run to learn that.
/mantis: standard
The default. Full 7-phase FSM. ~30-60 minutes wall clock. ~$10-40 typical cost on Opus pricing.
/mantis-ultra: wide parallel
Triples the per-wave hunter count, dispatches verifiers in parallel where their inputs are independent, and uses Opus on every grading-adjacent role. ~2-3× the tokens. Use when target value justifies the spend or a deadline forces wall-clock priority.
/mantis-loop: long mission
Repeats /mantis via EXPLORE iterations until a findings/time budget is hit:
/mantis-loop https://target.com --findings 3 --budget-min 240Each iteration starts from the previous session state, runs an EXPLORE wave on unexplored or low-priority surfaces, then re-runs the verification + grade pipeline. Stops when --findings N SUBMITs have accumulated or --budget-min M minutes have elapsed.
/mantis-fullsend: auto-disclose
Same as /mantis but adds two phases after REPORT:
- PATCH:
patch-writergenerates a suggested code-level fix. - DISCLOSE:
disclosure-senderfinds the verified security contact (security.txt, bug-bounty program URL, then well-known emails), drafts an email with report + patch, and sends through Gmail MCP after user confirmation.
The disclosure step is hard-gated: it never sends to an unverified email and never bypasses user review.
/mantisplan: pre-HUNT sanity check
Reads attack_surface.json from a prior recon run, writes plan.md with proposed wave assignments and reasoning. No HTTP. ~1-2 minutes, <1% the cost of a hunt. Useful when you want to inspect the orchestrator's hunting plan before committing to a full run.
Cost matrix (approximate)
| Phase | Standard | Ultra | Fast |
|---|---|---|---|
| RECON | ~$0.50 | ~$0.50 | ~$0.50 |
| AUTH | ~$0.30 | ~$0.30 | ~$0.30 |
| HUNT (all waves) | ~$6-25 | ~$24-120 | ~$2-8 |
| CHAIN | ~$1-3 | ~$1-3 | skipped |
| VERIFY (3 rounds) | ~$3-8 | ~$5-14 (parallel) | ~$1-2 (round 1 only) |
| GRADE | ~$0.30 | ~$0.30 | ~$0.30 |
| REPORT | ~$0.50 | ~$0.50 | skipped |
| Total typical | ~$10-40 | ~$30-140 | ~$3-10 |
Real costs depend on target complexity, scope size, and how many findings survive each verification round.