@genesislcap/ai-assistant 14.451.3 → 14.451.4

package/src/main/main.template.ts

@@ -28,7 +28,11 @@ import { ANIMATION_DEFS } from './main.types';
 
 function unknownToolPayload(tc: ChatToolCall): string {
   if (!isChatToolCallUnknown(tc)) return '';
-  const lines = [`${tc.name} — tool does not exist`];
+  // A stale tool was real earlier in this agent's lifetime but isn't available
+  // in the current state — distinct from a hallucinated name that never existed.
+  const lines = [
+    tc.stale ? `${tc.name} — no longer available at this step` : `${tc.name} — tool does not exist`,
+  ];
   if (tc.availableTools.length > 0) {
     lines.push('Available tools:');
     tc.availableTools.forEach((t) => lines.push(`  • ${t}`));

package/src/state/debug-event-log.ts

@@ -45,6 +45,7 @@ export type MetaEventType =
   | 'turn.retry'
   | 'turn.error'
   | 'tool.failed'
+  | 'tool.unresolved'
   // Routing / providers
   | 'agent.handoff'
   | 'agent.pinned'
@@ -95,6 +96,7 @@ export const META_EVENT_IMPORTANCE: Record<MetaEventType, MetaEventImportance> =
   'turn.start': 'normal',
   'turn.end': 'normal',
   'turn.retry': 'normal',
+  'tool.unresolved': 'normal',
   'agent.handoff': 'normal',
   'agent.pinned': 'normal',
   'agent.unpinned': 'normal',
@@ -130,7 +132,7 @@ export interface MetaEvent {
  * allowed to float above this cap rather than lose a failure signal; in normal
  * use the frequent `low`/`normal` events keep it near the cap. Entries are cheap.
  */
-const DEFAULT_MAX_META_EVENTS = 400;
+const DEFAULT_MAX_META_EVENTS = 800;
 
 interface MetaEventBuffer {
   events: MetaEvent[];
@@ -186,7 +188,8 @@ export function recordMetaEvent(
  * - `exception`              — an uncaught error escaped the tool loop (catch-all).
  * - `malformed-function-call`— the provider returned an unparseable tool call.
  * - `empty-response`         — the model returned no content and no tool calls.
- * - `unknown-tool-limit`     — the model repeatedly called tools that don't exist.
+ * - `unknown-tool-limit`     — the model repeatedly called tools it couldn't dispatch,
+ *                              whether hallucinated or stale (real earlier, retired now).
  * - `max-iterations`         — the tool loop hit its iteration cap.
  */
 export type TurnFailureReason =
@@ -242,7 +245,7 @@ export const DEBUG_LOG_README: readonly string[] = [
   "kind:'turn'.`agentSnapshot` — the active agent's own view of its internal state, captured at that turn. An agent opts into this by exposing a `getDebugSnapshot()` that returns JSON-serializable per-state info; stateful/flow agents wire it automatically, so you can watch a flow advance turn-by-turn (e.g. current step, cursor, collected fields, pending changes). Absent for agents that don't expose one.",
   "kind:'event' — a meta/lifecycle event. `type` names it (see below); `detail` carries structured data. `detail.placement` is the emitting UI instance: 'bubble' (collapsed), 'panel' (popped-out), or 'standalone'.",
   "Each 'event' also has an `importance`: 'high' (failures/limits — turn.error, tool.failed, file.read-failed, suggestions.failed, context.threshold-crossed), 'normal' (session flow — connects, turns, retries, handoffs, agent/provider changes, interactions), or 'low' (skippable UI/bookkeeping noise — panel.toggled, attachment.added, driver.wired/unwired, context.updated). To skim, ignore importance:'low'; to triage a failure, filter to importance:'high' then read the nearby messages and turns. A 'high' turn.error is often preceded by one or more 'normal' turn.retry events for the same reason — read them together to see how many attempts were made before bailing. 'message' and 'turn' entries carry no importance — they are the substance, always read them.",
-  'Event types: assistant.connected/disconnected (mount + placement + whether the session was created or restored), assistant.popout/popin (window placement), driver.created/wired/unwired (which driver is live and why it stops/starts responding across a popout), state.changed (idle↔loading), turn.start/turn.end (turn boundary; turn.end carries durationMs), turn.retry (a recoverable in-turn retry — detail.reason plus attempt/maxAttempts; for malformed calls also finishMessage), turn.error (a turn failed or hit a guardrail — detail.reason is one of exception/malformed-function-call/empty-response/unknown-tool-limit/max-iterations, plus reason-specific diagnostics: attempts, finishMessage, unknownTools + availableTools, iterations + limit, or name + message for exceptions), tool.failed (a tool threw), agent.handoff (routing; from=null is the initial activation), agent.pinned/unpinned (forced routing), provider.selected (model/provider for the upcoming turns), interaction.requested/resolved (blocking user widgets — explain quiet gaps), context.updated/threshold-crossed (token + cost), panel.toggled, attachment.added, file.read-failed, suggestions.failed.',
+  'Event types: assistant.connected/disconnected (mount + placement + whether the session was created or restored), assistant.popout/popin (window placement), driver.created/wired/unwired (which driver is live and why it stops/starts responding across a popout), state.changed (idle↔loading), turn.start/turn.end (turn boundary; turn.end carries durationMs), turn.retry (a recoverable in-turn retry — detail.reason plus attempt/maxAttempts; for malformed calls also finishMessage), turn.error (a turn failed or hit a guardrail — detail.reason is one of exception/malformed-function-call/empty-response/unknown-tool-limit/max-iterations, plus reason-specific diagnostics: attempts, finishMessage, unknownTools (split into staleTools — real earlier this activation but retired by the current state or hidden behind an open exclusive fold — and hallucinatedTools — never advertised) + availableTools, iterations + limit, or name + message for exceptions), tool.failed (a tool threw), tool.unresolved (the model called a tool that could not be dispatched — detail.kind is folded/fold-hidden/stale/unknown, plus tool + agent and, for the counted kinds, the consecutive streak; the recurring lead-up to an unknown-tool-limit turn.error), agent.handoff (routing; from=null is the initial activation), agent.pinned/unpinned (forced routing), provider.selected (model/provider for the upcoming turns), interaction.requested/resolved (blocking user widgets — explain quiet gaps), context.updated/threshold-crossed (token + cost), panel.toggled, attachment.added, file.read-failed, suggestions.failed.',
   "`meta` holds context captured at export time: agentSummary (full agent configs), context (active model, token usage, session cost), activeDebugSnapshot (the active agent's `getDebugSnapshot()` taken fresh at export — reflects state NOW, which may have advanced beyond the last turn's agentSnapshot), debug (optional host-supplied debug state), host, and the export timestamp.",
   'To debug a failure: find the last turn.error or tool.failed, then read upward for the user message, the turn(s), and the agent/provider/state events that led into it.',
 ];