@theoremjs/agents 0.3.4-next.401.gec4c582 → 0.3.4-next.405.g57fd3a0

package/esm/src/guardrails/labels.js

@@ -682,7 +682,7 @@ export const LABEL_NOTES = {
     'stop.cancelled': 'Status on a reply or message the person stopped.',
     'stop.interrupted': 'Status on a Live reply the person spoke over.',
     'stop.length': 'Status on a reply the model ended at its output limit.',
-    'stop.filtered': "Status on a reply the provider's content filter ended. A reply Theorem's own guardrails stopped shows its error instead.",
+    'stop.filtered': "Status on a reply the provider's content filter ended. A reply Theorem's own guardrails stopped shows its error or refusal instead.",
     'stop.provider_error': 'Status on a reply the provider ended with an error and no error message.',
     'stop.stream_incomplete': 'Status on a reply whose stream ended before the model finished.',
     'stop.step_limit': 'Status on a turn that made its last allowed model call and got no answer.',

package/esm/src/interface/stop.d.ts

@@ -7,7 +7,7 @@ export declare const WORDED_STOP_KINDS: readonly ["cancelled", "interrupted", "l
 export type WordedStopKind = (typeof WORDED_STOP_KINDS)[number];
 /**
  * Why a turn ended short, in the person's words: its `stop.<kind>` label, the profile's own when it
- * sets one. `undefined` when the turn ended whole, or when an error in `blocks` (the turn's
- * transcript blocks) already says why.
+ * sets one. `undefined` when the turn ended whole, when Theorem's own guardrail stopped it, or when
+ * an error in `blocks` (the turn's transcript blocks) already says why.
  */
 export declare function stopText(stop: TurnStop | undefined, blocks: readonly TranscriptBlock[], labels?: LabelOverrides): string | undefined;

package/esm/src/interface/stop.js

@@ -1,4 +1,5 @@
 import { labelText } from '../guardrails/labels.js';
+import { isGuardrailStop } from '../kernel/stop.js';
 /** The stops that end a reply short of a whole answer. Each is worded by its `stop.<kind>` label. */
 export const WORDED_STOP_KINDS = [
     'cancelled',
@@ -15,11 +16,11 @@ function isWorded(kind) {
 }
 /**
  * Why a turn ended short, in the person's words: its `stop.<kind>` label, the profile's own when it
- * sets one. `undefined` when the turn ended whole, or when an error in `blocks` (the turn's
- * transcript blocks) already says why.
+ * sets one. `undefined` when the turn ended whole, when Theorem's own guardrail stopped it, or when
+ * an error in `blocks` (the turn's transcript blocks) already says why.
  */
 export function stopText(stop, blocks, labels) {
-    if (!stop || !isWorded(stop.kind))
+    if (!stop || !isWorded(stop.kind) || isGuardrailStop(stop))
         return undefined;
     if (blocks.some((block) => block.kind === 'error'))
         return undefined;

package/esm/src/kernel/stop.d.ts

@@ -10,6 +10,8 @@ export declare const CONTINUE_INSTRUCTION_TYPES: readonly ProfileType[];
 export declare const DEFAULT_ALLOW_CONTINUE: readonly ContinueStopKind[];
 /** Hosts wait briefly, then continue once on their own. */
 export declare const DEFAULT_AUTO_CONTINUE: readonly ContinueStopKind[];
+/** Whether Theorem's own guardrail ended the turn, so its error or refusal already says why. */
+export declare function isGuardrailStop(stop: TurnStop): boolean;
 /** Pause before an auto-continue, the host's or the turn's own, so a flaky tunnel can settle. */
 export declare const AUTO_CONTINUE_DELAY_MS = 1500;
 /** Which stops a profile offers Continue for, and which the host continues on its own. */

package/esm/src/kernel/stop.js

@@ -8,6 +8,17 @@ export const CONTINUE_INSTRUCTION_TYPES = ['text'];
 export const DEFAULT_ALLOW_CONTINUE = CONTINUE_STOP_KINDS;
 /** Hosts wait briefly, then continue once on their own. */
 export const DEFAULT_AUTO_CONTINUE = ['length', 'stream_incomplete'];
+/** The `native` of a `filtered` stop that Theorem's own guardrail made, not the provider's filter. */
+const GUARDRAIL_STOP_NATIVES = new Set([
+    'egress',
+    'canary',
+    'prompt_echo',
+    'provider_tool_leak',
+]);
+/** Whether Theorem's own guardrail ended the turn, so its error or refusal already says why. */
+export function isGuardrailStop(stop) {
+    return stop.kind === 'filtered' && GUARDRAIL_STOP_NATIVES.has(stop.native ?? '');
+}
 /** Pause before an auto-continue, the host's or the turn's own, so a flaky tunnel can settle. */
 export const AUTO_CONTINUE_DELAY_MS = 1_500;
 const CONTINUE_KIND_SET = new Set(CONTINUE_STOP_KINDS);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@theoremjs/agents",
-  "version": "0.3.4-next.401.gec4c582",
+  "version": "0.3.4-next.405.g57fd3a0",
   "description": "A flat TypeScript agent kernel for typed profiles, deterministic turn execution, registered tools, provider adapters, guardrails, and host-injected traces.",
   "keywords": [
     "agent",