@emilia-protocol/openai-agents 0.1.1 → 0.1.2

package/README.md

@@ -47,11 +47,15 @@ const cancelOrder = tool({
 
 const agent = new Agent({ name: 'Ops', tools: [cancelOrder] });
 
-// One gate, configured once. actionFor maps a tool call -> canonical EP action_type.
+// One gate, configured once. actionFor maps a tool call -> the canonical EP
+// action_type the receipt must be bound to. For real safety, bind to the
+// SPECIFIC target the call acts on (not just the tool name) so a receipt for one
+// resource can't authorize another — and fold in the call identity so the same
+// receipt can't be reused across two different tool calls.
 const gate = requireReceiptForOpenAIAgent({
   trustedKeys: [process.env.EMILIA_ISSUER_KEY], // base64url SPKI-DER issuer key(s) you trust
   maxAgeSec: 900,
-  actionFor: (toolName /*, args */) => `openai.tool.${toolName}`,
+  actionFor: (toolName, args) => `openai.tool.${toolName}:${args?.orderId ?? ''}`,
 });
 
 let result = await run(agent, 'Cancel order 4242');
@@ -102,6 +106,9 @@ For every pending tool-approval interruption:
 ## Production note
 
 - **Pin `trustedKeys`** to your real issuer key(s). **Drop `allowInlineKey`** (it only proves integrity, never trust).
+- **Bind to the target, not just the tool.** Make `actionFor` incorporate the specific resource the call touches (and ideally the `callId` / an args hash), so a receipt minted for one action can't be replayed against a different one.
+- A receipt is **consumed only when it actually drives an `approve`** — never on a reject — so a blocked approval stays retryable with a fresh, valid receipt.
+- **Rejections are sanitized:** a reject decision carries only a machine-readable `reason` code, never the signer, the subject, or verifier internals.
 - The **consumed-receipt set is per-process** (replay defense across restarts/instances needs a shared store — see `@emilia-protocol/gate`).
 - This is **necessary, not sufficient**. It composes with — and never substitutes for — the resource owner's own authorization and policy checks. It makes the human approval that OpenAI already asks for into auditable, portable evidence; it does not decide whether the action *should* be allowed.
 

package/index.d.ts

@@ -36,7 +36,13 @@ export interface RequireReceiptForOpenAIAgentOptions {
   allowInlineKey?: boolean;
   /** Reject receipts older than this many seconds. Default 900. */
   maxAgeSec?: number;
-  /** REQUIRED. Map a tool call to the canonical EP action_type the receipt must bind. */
+  /**
+   * REQUIRED. Map a tool call to the canonical EP action_type the receipt must
+   * bind. For per-target binding, incorporate the SPECIFIC resource the call
+   * acts on (e.g. `payment.release:${args.destination}`); to stop a receipt
+   * being reused across distinct calls, fold in the call identity (the
+   * interruption's callId or a stable hash of args).
+   */
   actionFor: (toolName: string, args: unknown) => string;
 }
 

package/index.js

@@ -129,14 +129,26 @@ export function requireReceiptForOpenAIAgent(opts = {}) {
   }
 
   /**
-   * Decide a single interruption against a single receipt. Pure: no side effects
-   * EXCEPT marking a receipt_id consumed when (and only when) it is the valid
-   * receipt that earns an approve.
+   * Decide a single interruption against a single receipt.
+   *
+   * The decision NEVER carries the raw verifier output — only an allowlisted,
+   * sanitized shape ({decision, action, toolName, callId, reason} plus, on the
+   * approve path only, receipt_id + the accountable subject). A rejection never
+   * leaks the signer, the verifier's `detail`, or any other library internals.
+   *
+   * Consume-on-approve: when `decide` returns an `approve`, it marks the
+   * receipt_id consumed (replay defense) — and ONLY then. A reject (including a
+   * replayed or invalid receipt) never consumes anything, so a blocked approval
+   * stays retryable with a fresh, valid receipt. Replay is still checked here,
+   * before any approval, so a receipt can satisfy at most one interruption.
    */
   function decide(interruption, receipt) {
     const toolName = interruptionToolName(interruption);
     const callId = interruptionCallId(interruption);
     const args = interruptionArgs(interruption);
+    // Bind the receipt to THIS tool call. actionFor SHOULD incorporate the
+    // specific call identity (callId or a hash of args) so a single receipt
+    // cannot be reused across distinct tool calls — see the README note.
     const action = toolName != null ? actionFor(toolName, args) : null;
 
     const base = { action, toolName, callId };
@@ -155,14 +167,26 @@ export function requireReceiptForOpenAIAgent(opts = {}) {
       action, // binds the receipt's claim.action_type to THIS tool call
     });
     if (!v.ok) {
+      // Sanitized: only the reason code crosses the boundary — never v.signer,
+      // v.subject, or v.detail.
       return { decision: 'reject', ...base, reason: v.reason || 'invalid_receipt' };
     }
 
     // Replay defense: a receipt_id may earn an approval only once per process.
-    if (v.receipt_id && consumed.has(v.receipt_id)) {
+    // Belt-and-suspenders: also refuse if the SAME receipt has already been used
+    // for a DIFFERENT call (compound key), so one receipt can't satisfy two
+    // distinct interruptions even if actionFor collides.
+    const consumeKey = v.receipt_id;
+    const callKey = v.receipt_id && callId != null ? `${v.receipt_id}#${callId}` : null;
+    if (v.receipt_id && (consumed.has(consumeKey) || (callKey && consumed.has(callKey)))) {
       return { decision: 'reject', ...base, reason: 'receipt_replayed', receipt_id: v.receipt_id };
     }
-    if (v.receipt_id) consumed.add(v.receipt_id);
+    // Consume ONLY on the approve path, below — record both the receipt_id and
+    // the call-scoped key so a later call with the same receipt is refused.
+    if (v.receipt_id) {
+      consumed.add(consumeKey);
+      if (callKey) consumed.add(callKey);
+    }
 
     return {
       decision: 'approve',
@@ -222,13 +246,41 @@ export function requireReceiptForOpenAIAgent(opts = {}) {
     const rejected = [];
     const decisions = [];
 
+    // Undo a consumption recorded by decide() when the approval is not actually
+    // driven (no approve() to call, or approve() threw) — keeps it retryable.
+    const unconsume = (d) => {
+      if (!d.receipt_id) return;
+      consumed.delete(d.receipt_id);
+      if (d.callId != null) consumed.delete(`${d.receipt_id}#${d.callId}`);
+    };
+
     for (const interruption of interruptions) {
       const receipt = lookupReceipt(receipts, interruption);
       const d = decide(interruption, receipt);
       decisions.push(d);
       if (d.decision === 'approve') {
-        approved.push(interruption);
-        if (state && typeof state.approve === 'function') state.approve(interruption);
+        // The receipt was consumed inside decide() ONLY because this is an
+        // approve. Drive the SDK's approval now; if approve() is missing or
+        // throws, roll the consumption back so the (un-driven) approval stays
+        // retryable — a receipt is spent only when it actually drives an approve.
+        if (state && typeof state.approve === 'function') {
+          try {
+            state.approve(interruption);
+            approved.push(interruption);
+          } catch (err) {
+            unconsume(d);
+            d.decision = 'reject';
+            d.reason = 'approve_failed';
+            rejected.push(interruption);
+            if (state && typeof state.reject === 'function') {
+              state.reject(interruption, { message: `EMILIA: approve_failed` });
+            }
+          }
+        } else {
+          // No approve() to drive -> do not spend the receipt.
+          unconsume(d);
+          approved.push(interruption);
+        }
       } else {
         rejected.push(interruption);
         if (state && typeof state.reject === 'function') {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@emilia-protocol/openai-agents",
-  "version": "0.1.1",
+  "version": "0.1.2",
   "description": "Turn the OpenAI Agents SDK human-in-the-loop tool-approval step into a verifiable, offline-checkable EMILIA authorization receipt. OpenAI pauses an agent and asks for approval; EMILIA makes that approval portable, tamper-evident evidence — a named human accountably authorized this exact tool call. Composes WITH OpenAI's approval primitive (needsApproval / interruptions / state.approve|reject); does not replace it. Reference implementation, experimental.",
   "type": "module",
   "main": "index.js",