@oracle-agent/oracle 0.14.1 → 0.15.0

package/src/router/proposal.mjs

@@ -1,311 +0,0 @@
-// Oracle Router proposal model.
-//
-// Pure, dependency-free. A "proposal" is an advisory object: a labeled
-// suggestion for what Oracle Control *could* do, carrying no authority of its
-// own. This module MUST NEVER be able to produce an object that authorizes,
-// signs, or broadcasts anything, and MUST NEVER accept or echo back secrets
-// (bearer tokens, private/session keys, passphrases, keystore paths, etc.).
-//
-// Enforcement is defense-in-depth, in this order:
-//   1. Reject: any dangerous field anywhere in the input throws immediately.
-//   2. Strip: the emitted proposal is built from an explicit field allowlist
-//      only (nothing "extra" from the input can leak through even if step 1
-//      had a gap).
-//   3. Pin: every proposal carries fixed, non-overridable advisory markers
-//      (mode: "proposal", authorized: false, signed: false, broadcast: false)
-//      that the caller cannot set to anything else.
-//   4. Freeze: the returned proposal (and its nested plain objects/arrays)
-//      is deep-frozen so downstream code cannot mutate it into something
-//      that looks authorized.
-
-import { classifyRisk } from "./risk-classifier.mjs";
-
-// Case-insensitive key patterns that must never appear anywhere in a
-// proposal request or its emitted output. Deliberately broad: it is safer to
-// over-reject an oddly named benign field than to under-reject a secret.
-export const DANGEROUS_FIELD_PATTERNS = Object.freeze([
-  /priv(ate)?[-_ ]?key/i,
-  /seed[-_ ]?phrase/i,
-  /mnemonic/i,
-  /api[-_ ]?key/i,
-  /api[-_ ]?secret/i,
-  /bearer/i,
-  /session[-_ ]?key/i,
-  /session[-_ ]?secret/i,
-  /session[-_ ]?token/i,
-  /keystore/i,
-  /passphrase/i,
-  /^password$/i,
-  /secret/i,
-  /signature/i,
-  /signed[-_ ]?tx/i,
-  /raw[-_ ]?tx/i,
-  /broadcast/i,
-  /authoriz/i,
-  /^grant$/i,
-  /grant[-_ ]?token/i,
-  /credential/i,
-  /house[-_ ]?wallet/i,
-  /executor[-_ ]?token/i,
-  /access[-_ ]?token/i,
-  /^token$/i,
-  /^sig$/i,
-  /wallet[-_ ]?key/i,
-]);
-
-// Exact key names the router itself emits as PINNED, hardcoded-false/fixed
-// advisory markers (see createProposal below). These key names happen to
-// match the dangerous-field patterns above (e.g. "authorized" contains
-// "authoriz", "canAccessSecrets" contains "secret") precisely BECAUSE they
-// exist to make the proposal's non-authority explicit and readable. They are
-// only ever assigned literal constants by this module, never taken from
-// caller input, so allowing them past the OUTPUT-side scan is safe. This
-// allowlist is NEVER applied when scanning untrusted caller input.
-const PINNED_SAFE_OUTPUT_KEYS = new Set([
-  "mode",
-  "advisoryOnly",
-  "authorized",
-  "signed",
-  "broadcast",
-  "requiresHumanConfirmation",
-  "grantsPermissions",
-  "capabilities",
-  "canClassifyRisk",
-  "canExplain",
-  "canSimulate",
-  "canSign",
-  "canBroadcast",
-  "canAuthorize",
-  "canMintCapability",
-  "canAccessSecrets",
-]);
-
-function isDangerousKey(key, { allowPinnedSafe = false } = {}) {
-  const k = String(key);
-  if (allowPinnedSafe && PINNED_SAFE_OUTPUT_KEYS.has(k)) return false;
-  return DANGEROUS_FIELD_PATTERNS.some((re) => re.test(k));
-}
-
-/**
- * Recursively scan a value for keys matching DANGEROUS_FIELD_PATTERNS.
- * Returns a list of dotted paths where a dangerous key was found. Pure,
- * read-only -- never mutates the input.
- *
- * @param {*} value
- * @param {string} [pathPrefix]
- * @param {object} [opts]
- * @param {boolean} [opts.allowPinnedSafe] when true, exempts this module's own
- *   fixed advisory-marker key names (see PINNED_SAFE_OUTPUT_KEYS). Only ever
- *   pass this when scanning output THIS module constructed itself -- never
- *   when scanning caller-supplied input.
- */
-export function findDangerousFields(value, pathPrefix = "", opts = {}) {
-  const hits = [];
-  if (Array.isArray(value)) {
-    value.forEach((item, i) => {
-      hits.push(...findDangerousFields(item, `${pathPrefix}[${i}]`, opts));
-    });
-    return hits;
-  }
-  if (value && typeof value === "object") {
-    for (const [key, val] of Object.entries(value)) {
-      const path = pathPrefix ? `${pathPrefix}.${key}` : key;
-      if (isDangerousKey(key, opts)) hits.push(path);
-      hits.push(...findDangerousFields(val, path, opts));
-    }
-    return hits;
-  }
-  return hits;
-}
-
-/** Throws if `value` contains any dangerous field anywhere, at any depth.
- *  Pass { allowPinnedSafe: true } ONLY for output this module itself built. */
-export function assertNoDangerousFields(value, label = "proposal input", opts = {}) {
-  const hits = findDangerousFields(value, "", opts);
-  if (hits.length) {
-    throw new Error(
-      `${label} contains disallowed secret/signing/broadcast field(s): ${hits.join(", ")}`
-    );
-  }
-}
-
-/** Deep-clones plain objects/arrays with any dangerous keys removed. Defense
- *  in depth only -- createProposal() rejects rather than relying on this to
- *  silently clean up the input. */
-export function stripDangerousFields(value) {
-  if (Array.isArray(value)) return value.map(stripDangerousFields);
-  if (value && typeof value === "object") {
-    const out = {};
-    for (const [key, val] of Object.entries(value)) {
-      if (isDangerousKey(key)) continue;
-      out[key] = stripDangerousFields(val);
-    }
-    return out;
-  }
-  return value;
-}
-
-function deepFreeze(value) {
-  if (Array.isArray(value)) {
-    value.forEach(deepFreeze);
-    return Object.freeze(value);
-  }
-  if (value && typeof value === "object") {
-    for (const v of Object.values(value)) deepFreeze(v);
-    return Object.freeze(value);
-  }
-  return value;
-}
-
-// Fields the router is allowed to read off a proposal request. Anything not
-// in this list is dropped when the proposal is built, regardless of what the
-// caller sent -- this is what makes step 2 (Strip) effective even if a
-// dangerous field used a name the regex list hasn't seen yet.
-const ALLOWED_INPUT_FIELDS = Object.freeze([
-  "kind",
-  "target",
-  "notionalUsd",
-  "chainId",
-  "venue",
-  "destinationKnown",
-  "unlimitedApproval",
-  "crossChain",
-  "note",
-]);
-
-function pickAllowed(input) {
-  const out = {};
-  for (const key of ALLOWED_INPUT_FIELDS) {
-    if (input && Object.prototype.hasOwnProperty.call(input, key) && input[key] !== undefined) {
-      out[key] = input[key];
-    }
-  }
-  return out;
-}
-
-let proposalCounter = 0;
-function nextProposalId(nowMs) {
-  proposalCounter += 1;
-  return `proposal-${nowMs}-${proposalCounter}`;
-}
-
-/**
- * Build an advisory-only proposal object. Never authorizes, signs, or
- * broadcasts anything; the returned object cannot be mistaken for one that
- * does because its authority markers are pinned to false/"proposal" and it is
- * deep-frozen.
- *
- * @param {object} input
- * @param {string} input.kind            see risk-classifier.mjs for recognized kinds
- * @param {string} [input.target]        human-readable description of the target (e.g. "HL BTC-PERP")
- * @param {number} [input.notionalUsd]
- * @param {number} [input.chainId]
- * @param {string} [input.venue]
- * @param {boolean} [input.destinationKnown]
- * @param {boolean} [input.unlimitedApproval]
- * @param {boolean} [input.crossChain]
- * @param {string} [input.note]          free-text rationale; itself scanned for dangerous fields if an object is passed by mistake
- * @param {number} [nowMs]               injectable clock for deterministic tests
- * @returns {object} frozen proposal
- */
-export function createProposal(input = {}, nowMs = Date.now()) {
-  if (input === null || typeof input !== "object" || Array.isArray(input)) {
-    throw new Error("proposal input must be a plain object");
-  }
-
-  // Step 1: reject. Scan the FULL raw input (not just the allowlisted
-  // subset) so a caller cannot smuggle a secret in under an allowed key name
-  // collision, nor bury one in an unlisted field and have it silently
-  // dropped without anyone noticing the attempt.
-  assertNoDangerousFields(input, "proposal input");
-
-  const kind = String(input.kind || "").trim().toLowerCase();
-  if (!kind) throw new Error("proposal input: kind is required");
-
-  // Step 2: strip / allowlist. Only known-safe fields are read from input.
-  const safeInput = pickAllowed(input);
-  safeInput.kind = kind;
-
-  const risk = classifyRisk({
-    kind,
-    notionalUsd: safeInput.notionalUsd,
-    destinationKnown: safeInput.destinationKnown,
-    unlimitedApproval: safeInput.unlimitedApproval,
-    crossChain: safeInput.crossChain,
-  });
-
-  const proposal = {
-    id: nextProposalId(Number(nowMs) || Date.now()),
-    createdAt: new Date(Number(nowMs) || Date.now()).toISOString(),
-
-    kind: safeInput.kind,
-    target: safeInput.target != null ? String(safeInput.target) : null,
-    notionalUsd: safeInput.notionalUsd != null ? Number(safeInput.notionalUsd) : null,
-    chainId: safeInput.chainId != null ? Number(safeInput.chainId) : null,
-    venue: safeInput.venue != null ? String(safeInput.venue) : null,
-    note: safeInput.note != null ? String(safeInput.note) : null,
-
-    risk: {
-      tier: risk.tier,
-      score: risk.score,
-      reasons: risk.reasons,
-    },
-
-    // Step 3: pinned advisory markers. These are NOT derived from input and
-    // cannot be overridden by any input field -- they are literal constants.
-    mode: "proposal",
-    advisoryOnly: true,
-    authorized: false,
-    signed: false,
-    broadcast: false,
-    requiresHumanConfirmation: true,
-    grantsPermissions: false,
-    capabilities: Object.freeze({
-      canClassifyRisk: true,
-      canExplain: true,
-      canSimulate: kind === "simulate",
-      canSign: false,
-      canBroadcast: false,
-      canAuthorize: false,
-      canMintCapability: false,
-      canAccessSecrets: false,
-    }),
-  };
-
-  // Belt-and-suspenders: re-verify the object we are about to hand out
-  // contains no dangerous keys before freezing, in case a future edit to
-  // ALLOWED_INPUT_FIELDS accidentally lets one through. allowPinnedSafe is
-  // required here because this module's own fixed advisory markers (e.g.
-  // "authorized", "broadcast", "canAccessSecrets") intentionally reuse
-  // wording from the dangerous-pattern list to stay self-documenting; they
-  // are hardcoded constants above, never derived from input.
-  assertNoDangerousFields(proposal, "proposal output", { allowPinnedSafe: true });
-
-  return deepFreeze(proposal);
-}
-
-/**
- * Defense-in-depth guard for anything downstream that receives a proposal
- * object from the router and wants to double check, before acting on it in
- * any way, that it really is advisory-only. Throws if the object does not
- * look like a well-formed, unauthorized proposal.
- */
-export function assertAdvisoryProposal(proposal) {
-  if (!proposal || typeof proposal !== "object") {
-    throw new Error("proposal required");
-  }
-  if (proposal.mode !== "proposal") throw new Error("proposal: mode must be \"proposal\"");
-  if (proposal.authorized !== false) throw new Error("proposal: authorized must be false");
-  if (proposal.signed !== false) throw new Error("proposal: signed must be false");
-  if (proposal.broadcast !== false) throw new Error("proposal: broadcast must be false");
-  if (proposal.advisoryOnly !== true) throw new Error("proposal: advisoryOnly must be true");
-  if (proposal.grantsPermissions !== false) {
-    throw new Error("proposal: grantsPermissions must be false");
-  }
-  const caps = proposal.capabilities || {};
-  if (caps.canSign || caps.canBroadcast || caps.canAuthorize || caps.canMintCapability || caps.canAccessSecrets) {
-    throw new Error("proposal: capabilities must not grant sign/broadcast/authorize/mint/secret access");
-  }
-  assertNoDangerousFields(proposal, "proposal", { allowPinnedSafe: true });
-  return true;
-}

package/src/router/risk-classifier.mjs

@@ -1,119 +0,0 @@
-// Oracle Router risk classifier.
-//
-// Pure, dependency-free, deterministic. No I/O, no network, no chain reads,
-// no imports. The router uses this to LABEL a proposed action's risk tier for
-// advisory purposes only -- the label is context for a human or for Oracle
-// Control, never an authorization decision. Oracle Control independently
-// validates and gates every real action; nothing here grants or implies
-// permission to sign, broadcast, or spend.
-
-export const RISK_TIERS = Object.freeze(["low", "medium", "high", "critical"]);
-
-// Kinds the router treats as pure read/advisory work (no custody exposure).
-export const READ_ONLY_KINDS = Object.freeze(["read", "quote", "explain", "simulate"]);
-
-// Kinds that describe a draft custody-adjacent action. These are still just
-// labels on a proposal -- classifying something as "swap" does not create a
-// swap, sign anything, or move funds.
-export const CUSTODY_ADJACENT_KINDS = Object.freeze([
-  "approval",
-  "transfer",
-  "swap",
-  "order",
-  "bridge",
-]);
-
-function normalizeKind(kind) {
-  const k = String(kind || "").trim().toLowerCase();
-  return k || "unknown";
-}
-
-// Coarse USD notional bands. Unknown/invalid notional is NOT treated as "low"
-// -- callers must not be able to escape risk scoring by omitting a number.
-function notionalBand(notionalUsd) {
-  if (notionalUsd == null) return null;
-  const n = Number(notionalUsd);
-  if (!Number.isFinite(n) || n < 0) return null;
-  if (n > 100_000) return 3;
-  if (n > 10_000) return 2;
-  if (n > 1_000) return 1;
-  return 0;
-}
-
-function tierFromScore(score) {
-  if (score >= 7) return "critical";
-  if (score >= 4) return "high";
-  if (score >= 2) return "medium";
-  return "low";
-}
-
-/**
- * Classify the risk tier of a proposed action. Pure function: same input
- * always produces the same output, and calling it has zero side effects.
- *
- * @param {object} [input]
- * @param {string} [input.kind]               "read" | "quote" | "explain" | "simulate" |
- *                                             "approval" | "transfer" | "swap" | "order" |
- *                                             "bridge" | any other string (treated conservatively)
- * @param {number} [input.notionalUsd]         approximate USD notional, if known
- * @param {boolean} [input.destinationKnown]   true if destination is a reviewed/allowlisted venue
- * @param {boolean} [input.unlimitedApproval]  true if the action requests an unbounded/MaxUint256-style approval
- * @param {boolean} [input.crossChain]         true if the action spans chains (bridge-like)
- * @returns {{ tier: "low"|"medium"|"high"|"critical", score: number, reasons: readonly string[] }}
- */
-export function classifyRisk(input = {}) {
-  const kind = normalizeKind(input?.kind);
-  const reasons = [];
-  let score = 0;
-
-  const isReadOnly = READ_ONLY_KINDS.includes(kind);
-  const isCustodyAdjacent = CUSTODY_ADJACENT_KINDS.includes(kind);
-
-  if (isReadOnly) {
-    reasons.push(`kind "${kind}" is read/advisory-only`);
-  } else if (isCustodyAdjacent) {
-    score += 2;
-    reasons.push(`kind "${kind}" is custody-adjacent (+2)`);
-  } else {
-    score += 3;
-    reasons.push(`kind "${kind}" is unrecognized; treated conservatively (+3)`);
-  }
-
-  const band = notionalBand(input?.notionalUsd);
-  if (band != null && band > 0) {
-    score += band;
-    reasons.push(`notional band ${band} (+${band})`);
-  } else if (band == null && !isReadOnly) {
-    score += 1;
-    reasons.push("notional unknown for a non-read action (+1)");
-  }
-
-  if (input?.unlimitedApproval) {
-    score += 3;
-    reasons.push("unlimited/unbounded approval requested (+3)");
-  }
-
-  if (input?.destinationKnown === false) {
-    score += 2;
-    reasons.push("destination not a reviewed/allowlisted venue (+2)");
-  }
-
-  if (input?.crossChain) {
-    score += 1;
-    reasons.push("cross-chain action (+1)");
-  }
-
-  return Object.freeze({
-    tier: tierFromScore(score),
-    score,
-    reasons: Object.freeze(reasons),
-  });
-}
-
-export function isReadOnlyKind(kind) {
-  return READ_ONLY_KINDS.includes(normalizeKind(kind));
-}
-
-export function isCustodyAdjacentKind(kind) {
-  return CUSTODY_ADJACENT_KINDS.includes(normalizeKind(kind));
-}