feat: add delu-oracle skill (deluonchain)#481
Conversation
There was a problem hiding this comment.
A couple of issues to address, then would be good to merge
Missing Prompt-Injection Boundary For Oracle Output
The skill presents decision.read, summary, drivers, risks, social data, and the full cognition report as agent-consumable trading input in deluonchain/delu-oracle/SKILL.md and the response references. These fields can contain narrative text from external systems, so the skill should explicitly say oracle responses are untrusted data and must not override instructions, trigger tool calls, install packages, sign payments, or execute trades on their own.
Trading Decision Can Be Treated As Execution Instruction
The docs describe the endpoint as returning a decision “an agent can act on in a single hop” and provide a simple ENTER gate in deluonchain/delu-oracle/SKILL.md. For a wallet-adjacent trading skill, this should separate analysis from execution and require explicit user confirmation for any swap, trade, approval, transfer, or position change, including token, amount, slippage, chain, and max loss.
Paid x402 Calls Need Budget And Confirmation Guardrails
The endpoint uses x402 payments and can charge up to the public 250k DELU tier per call, with optional social enrichment adding a USDC charge. The docs in deluonchain/delu-oracle/SKILL.md and references/social-enrichment.md should require a user-approved budget, max calls, and explicit opt-in for ?social=true, especially before looping through a watchlist.
Mutable Third-Party Package Install Instructions
The external client docs recommend npm install x402-fetch viem and pip install x402 in deluonchain/delu-oracle/references/external-clients.md without pinning versions. Since these packages are used for wallet/payment signing, the examples should pin reviewed versions or add package provenance and verification guidance.
Private Key Examples Need Stronger Custody Warnings
The standalone client examples in deluonchain/delu-oracle/references/external-clients.md load PRIVATE_KEY directly and use it to sign x402 payments. This should strongly recommend a dedicated low-balance hot wallet, spending limits, no committed .env files, and avoiding primary wallets or wallets with unrelated funds.
Silent Social Fallback Hides Paid Feature Failure
deluonchain/delu-oracle/references/social-enrichment.md says to silently fall back to a plain GET and not surface the error to the user if checkr fails. For a paid optional enrichment path, the agent should disclose that social enrichment failed or was skipped, so users do not mistake a quant-only result for a fully enriched result.
delu-oracle — full-cognition token analysis for Base
Provider: deluonchain
Endpoint:
https://x402.bankr.bot/0xed2ceca9de162c4f2337d7c1ab44ee9c427709da/delu-oracle/analyze/{ca}Payment: x402
uptoscheme, DELU token on BaseWhat it does
Intelligence layer for any Base trading agent. Pass one Base EVM contract address (or cashtag like
$BNKR) and get back a flatdecisionheader —action,conviction, entry/stop/size, and a one-lineread— with the full cognition report underneath.Scout, auditor, and quant run server-side on every call. Social signal (checkr) is opt-in via
?social=true.Tiered pricing
Payment token: DELU —
0x7b0ee9dcb5c1d4d7cd630c652959951936512ba3on Base.Files
Key design decisions
SKILL.mdis ~4KB — only what a consuming agent needs to call the endpoint correctly. implementation detail lives in references.observedblock (scout/auditor/quant mirror) is always present — no?verbose=trueneeded.uptoPermit2 signatures are single-use; parallel calls 402.decisionheader is flat — no traversal needed to act on the signal.Example use cases
action === "ENTER" && conviction >= 70 && confidence >= 0.6)