Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add MangroveTechnologies/mangrove-agent --skill sievegit clone --depth 1 https://github.com/MangroveTechnologies/mangrove-agentWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/mangrovetechnologies/mangrove-agent/sieve)<a href="https://agentmods.dev/skills/mangrovetechnologies/mangrove-agent/sieve"><img src="https://agentmods.dev/badge/skills/mangrovetechnologies/mangrove-agent/sieve.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00194 | $0.02476 |
| Opus 5 | $0.00097 | $0.01238 |
| Sonnet 5 | $0.00039 | $0.00495 |
| Haiku 4.5 | $0.00019 | $0.00248 |
Grade A, and why
sieve scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 8d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 210 lines — stays where its author put it; the contents beside it link to each section on GitHub.
SIEVE Skill
This skill exists because a backtest is expensive and most strategies fail it. A single Oracle backtest takes 30–120s on a multi-month window, and roughly 5 of every 6 candidate strategies fail — they never fire a trade, or they wash, or they lose. Paying for 99 backtests to find the 1 good one is wasteful.
SIEVE ("Strategy Indicator Embeddings with Value Estimation") is a classifier trained on 1.24M historical Mangrove sweep runs. It scores a candidate in milliseconds and returns two things:
- binary (2-class) —
{p_no_trades, p_trades}. The go/no-go gate: if SIEVE thinks the strategy will never fire on real data, you skip the backtest entirely. (This is the same head the/sweepengine uses as its inline pre-filter — but that's a different mechanism; see below.) - four_class —
{losing, no_trades, wash, winning}. A softmax over outcomes. Rank survivors byP(winning)to order the shortlist.
Use both heads: binary to drop the dead, 4-class to rank what's left.
SIEVE is a SCREEN, not a verdict. Both heads are cheap predictions over an aggregate of millions of runs — they tell you what's worth paying to backtest, never whether a strategy is actually good. Only a real backtest decides. A 99-candidate screen compresses to ~5 backtests: same signal quality at ~10% of the cost (see Prohibited).
Two different SIEVE uses — don't confuse them:
- This skill = offline screen of a fixed shortlist via the
sieve_scoreendpoint (≤99 per call), then hand survivors to a backtest. /sweep's pre-filter = the binary head running inside the engine during a parameter search, skipping dead configs as they're generated. That's part of/sweep, not this skill — and it never uses the 4-class head.
Trigger
Activate when the user:
- Has many candidate strategies and asks which to test ("score these", "which of these is worth backtesting", "rank these")
- Wants to explore parameter variations cheaply ("try 50 variations of my MACD strategy and tell me the good ones")
- Just produced a candidate set via
/create-strategy(autonomous mode emits N candidates) or/custom-signaland needs to prune before backtesting - Asks to "pre-filter", "screen", or "narrow down" strategies
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 8d ago First seen · 210 lines · 194 tokens per session scan A ab98ed989c65
sieve is a skill published in the GitHub repository MangroveTechnologies/mangrove-agent (5 stars, last pushed 1mo ago), licensed MIT. It adds 194 tokens to every session and 2,476 once invoked, about $0.0010 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
t2000-job
Escrow USDC for agent-to-agent deliverable work (A2A jobs). Use when hiring another agent for async work (research reports, builds, SLA tasks) or when selling deliverable work yourself (list a service: fixed price + SLA, no server needed), or posting/claiming open jobs on the board (t2 job open / claim) — anything…
t2000-receive
Generate a payment request for the t2000 Agent Wallet — print the wallet address, an ANSI QR code, and (via MCP) a Payment Kit URI (sui:pay?…). Use when asked to receive a payment, share a wallet address, create a payment link, or set up a fund-me link.
t2000-send
Send USDC, USDsui, or SUI from the t2000 Agent Wallet to another Sui address. Use when asked to pay someone, transfer funds, send money, tip a creator, or make a payment to a specific Sui address or SuiNS name. Do NOT use for API payments — use the t2000-pay skill for x402-protected services.
t2000-swap
Swap tokens on Sui via Cetus Aggregator (20+ DEXs, best-route across SUI, USDC, USDsui, USDT, USDe, ETH, GOLD, NAVX, WAL, vSUI, and more). Use when asked to swap, trade, convert, exchange, or "turn X into Y". Do not use for sending — use the t2000-send skill for transfers.
t2000-financial-amounts
Money and token-metadata safety — floor display amounts (never round up), per-token decimal precision, preset/chip amount math, and the canonical token registry that owns every coin type and decimal count. Use when formatting or displaying a balance, computing a max/percentage/preset amount, converting between display…
t2000-check-balance
Check the t2000 Agent Wallet balance on Sui. Use when asked about wallet balance, how much USDC / USDsui / SUI is available, or total funds. Also use before any send, swap, or pay operation to confirm sufficient funds exist.