@pentoshi/clai 2.0.61 → 3.0.0

package/dist/agent/context-manager.d.ts

@@ -8,6 +8,11 @@ import type { ChatMessage } from "../types.js";
  */
 export declare function estimateTokens(text: string): number;
 export declare function estimateMessagesTokens(messages: ChatMessage[]): number;
+/**
+ * Agent-loop auto-compact threshold (estimated tokens). Shared with `/context`
+ * so the reported % of budget matches when auto-compaction fires.
+ */
+export declare const AUTO_COMPACT_TOKEN_BUDGET = 150000;
 export interface CompactOptions {
     /** Soft budget (tokens). When estimated tokens exceed this, compact. */
     budgetTokens?: number | undefined;

package/dist/agent/context-manager.js

@@ -21,6 +21,11 @@ export function estimateMessagesTokens(messages) {
     }
     return sum;
 }
+/**
+ * Agent-loop auto-compact threshold (estimated tokens). Shared with `/context`
+ * so the reported % of budget matches when auto-compaction fires.
+ */
+export const AUTO_COMPACT_TOKEN_BUDGET = 150_000;
 const DEFAULT_BUDGET_TOKENS = 32_000;
 const DEFAULT_KEEP_RECENT = 6;
 /**
@@ -131,11 +136,18 @@ export async function compactMessagesWithSummary(messages, summarize, options =
         return `${message.role.toUpperCase()}: ${content}`;
     })
         .join("\n\n");
-    const transcript = sessionTranscript?.trim()
+    // Prefer combining visual session material (tools/prompts) with older model
+    // turns so /compact after /history + new messages never loses either side.
+    const visual = sessionTranscript?.trim()
         ? redactSecrets(sessionTranscript.trim())
-        : messageTranscript;
+        : "";
+    const fromMessages = messageTranscript.trim();
+    const transcript = visual && fromMessages
+        ? `${visual}\n\n---\n\nOLDER MODEL TURNS:\n\n${fromMessages}`
+        : visual || fromMessages;
     const prompt = [
         "Create a complete but compact continuation memory of the entire session below for another assistant that will continue it.",
+        "The material may include a resumed history session plus newer turns — treat it as one continuous conversation.",
         "Preserve user intentions and prompts, decisions, constraints, file paths, commands/tools run, steps taken, task states, outputs and results, errors and failed approaches, completed work, and exactly what remains.",
         "Organize the memory under concise sections: User goals, Decisions and constraints, Work completed, Commands/tools and results, Current state, Remaining work.",
         "Do not add facts, commentary, or markdown framing. Be concise but specific. Never include secrets or credentials.",

package/dist/agent/context-manager.js.map

@@ -1 +1 @@
-{"version":3,"file":"context-manager.js","sourceRoot":"","sources":["../../src/agent/context-manager.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,aAAa,EAAE,MAAM,oBAAoB,CAAC;AACnD,OAAO,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAC;AAElD;;;;;;GAMG;AACH,MAAM,UAAU,cAAc,CAAC,IAAY;IACzC,OAAO,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,MAAM,GAAG,GAAG,CAAC,CAAC;AACtC,CAAC;AAED,MAAM,UAAU,sBAAsB,CAAC,QAAuB;IAC5D,IAAI,GAAG,GAAG,CAAC,CAAC;IACZ,KAAK,MAAM,OAAO,IAAI,QAAQ,EAAE,CAAC;QAC/B,GAAG,IAAI,cAAc,CAAC,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,gBAAgB;QAC5D,gEAAgE;QAChE,IAAI,OAAO,CAAC,MAAM,EAAE,CAAC;YACnB,GAAG,IAAI,OAAO,CAAC,MAAM,CAAC,MAAM,GAAG,IAAI,CAAC;QACtC,CAAC;IACH,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC;AAkBD,MAAM,qBAAqB,GAAG,MAAM,CAAC;AACrC,MAAM,mBAAmB,GAAG,CAAC,CAAC;AAE9B;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,wBAAwB,GACnC,8CAA8C,CAAC;AACjD,MAAM,CAAC,MAAM,wBAAwB,GACnC,2CAA2C,CAAC;AAE9C,MAAM,UAAU,yBAAyB,CAAC,OAAoB;IAC5D,OAAO,CACL,OAAO,CAAC,IAAI,KAAK,QAAQ;QACzB,CAAC,OAAO,CAAC,OAAO,CAAC,UAAU,CAAC,wBAAwB,CAAC;YACnD,OAAO,CAAC,OAAO,CAAC,UAAU,CAAC,wBAAwB,CAAC,CAAC,CACxD,CAAC;AACJ,CAAC;AAED;;;;;;;;;GASG;AACH,MAAM,UAAU,eAAe,CAC7B,QAAuB,EACvB,UAA0B,EAAE;IAE5B,MAAM,MAAM,GAAG,OAAO,CAAC,YAAY,IAAI,qBAAqB,CAAC;IAC7D,MAAM,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,CAAC,UAAU,IAAI,mBAAmB,CAAC,CAAC;IAC1E,IAAI,QAAQ,CAAC,MAAM,IAAI,UAAU,GAAG,CAAC;QAAE,OAAO,QAAQ,CAAC;IACvD,IAAI,sBAAsB,CAAC,QAAQ,CAAC,IAAI,MAAM;QAAE,OAAO,QAAQ,CAAC;IAEhE,oEAAoE;IACpE,MAAM,IAAI,GAAkB,EAAE,CAAC;IAC/B,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,IAAI,QAAQ,CAAC,CAAC,CAAC,EAAE,IAAI,KAAK,QAAQ,EAAE,CAAC;QACnC,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;QACvB,KAAK,GAAG,CAAC,CAAC;IACZ,CAAC;IAED,MAAM,IAAI,GAAG,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,QAAQ,CAAC,MAAM,GAAG,UAAU,CAAC,CAAC,CAAC;IAC3E,MAAM,MAAM,GAAG,QAAQ,CAAC,KAAK,CAAC,KAAK,EAAE,QAAQ,CAAC,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC;IACpE,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,QAAQ,CAAC;IAEzC,MAAM,OAAO,GAAa,EAAE,CAAC;IAC7B,KAAK,MAAM,GAAG,IAAI,MAAM,EAAE,CAAC;QACzB,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YACxB,OAAO,CAAC,IAAI,CAAC,iBAAiB,OAAO,CAAC,GAAG,CAAC,OAAO,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC;QAC7D,CAAC;aAAM,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YACpC,MAAM,IAAI,GAAG,OAAO,CAAC,GAAG,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC;YACvC,IAAI,IAAI;gBAAE,OAAO,CAAC,IAAI,CAAC,gBAAgB,IAAI,EAAE,CAAC,CAAC;QACjD,CAAC;aAAM,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YAC/B,OAAO,CAAC,IAAI,CAAC,kBAAkB,OAAO,CAAC,GAAG,CAAC,OAAO,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC;QAC9D,CAAC;IACH,CAAC;IAED,MAAM,IAAI,GAAgB;QACxB,IAAI,EAAE,QAAQ;QACd,OAAO,EACL,GAAG,wBAAwB,oHAAoH;YAC/I,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC;KACrB,CAAC;IAEF,OAAO,CAAC,GAAG,IAAI,EAAE,IAAI,EAAE,GAAG,IAAI,CAAC,CAAC;AAClC,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,0BAA0B,CAC9C,QAAuB,EACvB,SAA8C,EAC9C,UAA0B,EAAE,EAC5B,iBAAsC;IAEtC,MAAM,MAAM,GAAG,QAAQ,CAAC,MAAM,CAAC;IAC/B,MAAM,YAAY,GAAG,sBAAsB,CAAC,QAAQ,CAAC,CAAC;IACtD,MAAM,QAAQ,GAAG,OAAO,CAAC,YAAY,KAAK,CAAC,CAAC;IAE5C,IAAI,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,CAAC,UAAU,IAAI,mBAAmB,CAAC,CAAC;IACxE,MAAM,KAAK,GAAG,QAAQ,CAAC,CAAC,CAAC,EAAE,IAAI,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACrD,IAAI,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,QAAQ,CAAC,MAAM,GAAG,UAAU,CAAC,CAAC;IAC9D,IAAI,KAAK,GAAG,QAAQ,CAAC,KAAK,CAAC,KAAK,EAAE,SAAS,CAAC,CAAC;IAE7C,yEAAyE;IACzE,qFAAqF;IACrF,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC,IAAI,QAAQ,IAAI,QAAQ,CAAC,MAAM,IAAI,KAAK,GAAG,CAAC,EAAE,CAAC;QACnE,UAAU,GAAG,CAAC,CAAC;QACf,SAAS,GAAG,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC;QAChC,KAAK,GAAG,QAAQ,CAAC,KAAK,CAAC,KAAK,EAAE,SAAS,CAAC,CAAC;IAC3C,CAAC;IAED,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC,IAAI,CAAC,iBAAiB,EAAE,IAAI,EAAE,EAAE,CAAC;QACrD,4DAA4D;QAC5D,OAAO;YACL,QAAQ,EAAE,CAAC,GAAG,QAAQ,CAAC;YACvB,MAAM;YACN,KAAK,EAAE,MAAM;YACb,YAAY;YACZ,WAAW,EAAE,YAAY;YACzB,UAAU,EAAE,KAAK;SAClB,CAAC;IACJ,CAAC;IAED,MAAM,iBAAiB,GAAG,KAAK;SAC5B,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE;QACf,IAAI,OAAO,GAAG,aAAa,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;QAC7C,iEAAiE;QACjE,4DAA4D;QAC5D,IAAI,OAAO,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YACjC,OAAO,GAAG,aAAa,CAAC,OAAO,CAAC,CAAC,OAAO,CAAC;QAC3C,CAAC;QACD,OAAO,GAAG,OAAO,CAAC,IAAI,CAAC,WAAW,EAAE,KAAK,OAAO,EAAE,CAAC;IACrD,CAAC,CAAC;SACD,IAAI,CAAC,MAAM,CAAC,CAAC;IAChB,MAAM,UAAU,GAAG,iBAAiB,EAAE,IAAI,EAAE;QAC1C,CAAC,CAAC,aAAa,CAAC,iBAAiB,CAAC,IAAI,EAAE,CAAC;QACzC,CAAC,CAAC,iBAAiB,CAAC;IACtB,MAAM,MAAM,GAAG;QACb,4HAA4H;QAC5H,sNAAsN;QACtN,+JAA+J;QAC/J,mHAAmH;QACnH,EAAE;QACF,UAAU;KACX,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAEb,uEAAuE;IACvE,mDAAmD;IACnD,wEAAwE;IACxE,qEAAqE;IACrE,MAAM,UAAU,GAAG,aAAa,CAAC,CAAC,MAAM,SAAS,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC;IACnE,MAAM,OAAO,GAAG,aAAa,CAAC,UAAU,CAAC,CAAC,OAAO,CAAC,IAAI,EAAE,CAAC;IACzD,IAAI,CAAC,OAAO;QAAE,MAAM,IAAI,KAAK,CAAC,oDAAoD,CAAC,CAAC;IAEpF,MAAM,IAAI,GAAG,KAAK,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IAC/C,kEAAkE;IAClE,gDAAgD;IAChD,MAAM,IAAI,GAAG,QAAQ,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE;QACjD,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,IAAI,SAAS,CAAC,IAAI,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,CAAC;YAC5D,OAAO,EAAE,GAAG,GAAG,EAAE,OAAO,EAAE,aAAa,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,OAAO,EAAE,CAAC;QACjE,CAAC;QACD,OAAO,GAAG,CAAC;IACb,CAAC,CAAC,CAAC;IACH,MAAM,SAAS,GAAkB;QAC/B,GAAG,IAAI;QACP,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,GAAG,wBAAwB,OAAO,OAAO,EAAE,EAAE;QACxE,GAAG,IAAI;KACR,CAAC;IACF,MAAM,WAAW,GAAG,sBAAsB,CAAC,SAAS,CAAC,CAAC;IACtD,OAAO;QACL,QAAQ,EAAE,SAAS;QACnB,MAAM;QACN,KAAK,EAAE,SAAS,CAAC,MAAM;QACvB,YAAY;QACZ,WAAW;QACX,UAAU,EAAE,IAAI;KACjB,CAAC;AACJ,CAAC;AAED,SAAS,OAAO,CAAC,IAAY,EAAE,QAAgB;IAC7C,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,CAAC;IACjD,IAAI,OAAO,CAAC,MAAM,IAAI,QAAQ;QAAE,OAAO,OAAO,CAAC;IAC/C,OAAO,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,QAAQ,GAAG,CAAC,CAAC,GAAG,CAAC;AAC9C,CAAC"}
+{"version":3,"file":"context-manager.js","sourceRoot":"","sources":["../../src/agent/context-manager.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,aAAa,EAAE,MAAM,oBAAoB,CAAC;AACnD,OAAO,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAC;AAElD;;;;;;GAMG;AACH,MAAM,UAAU,cAAc,CAAC,IAAY;IACzC,OAAO,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,MAAM,GAAG,GAAG,CAAC,CAAC;AACtC,CAAC;AAED,MAAM,UAAU,sBAAsB,CAAC,QAAuB;IAC5D,IAAI,GAAG,GAAG,CAAC,CAAC;IACZ,KAAK,MAAM,OAAO,IAAI,QAAQ,EAAE,CAAC;QAC/B,GAAG,IAAI,cAAc,CAAC,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,gBAAgB;QAC5D,gEAAgE;QAChE,IAAI,OAAO,CAAC,MAAM,EAAE,CAAC;YACnB,GAAG,IAAI,OAAO,CAAC,MAAM,CAAC,MAAM,GAAG,IAAI,CAAC;QACtC,CAAC;IACH,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC;AAED;;;GAGG;AACH,MAAM,CAAC,MAAM,yBAAyB,GAAG,OAAO,CAAC;AAkBjD,MAAM,qBAAqB,GAAG,MAAM,CAAC;AACrC,MAAM,mBAAmB,GAAG,CAAC,CAAC;AAE9B;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,wBAAwB,GACnC,8CAA8C,CAAC;AACjD,MAAM,CAAC,MAAM,wBAAwB,GACnC,2CAA2C,CAAC;AAE9C,MAAM,UAAU,yBAAyB,CAAC,OAAoB;IAC5D,OAAO,CACL,OAAO,CAAC,IAAI,KAAK,QAAQ;QACzB,CAAC,OAAO,CAAC,OAAO,CAAC,UAAU,CAAC,wBAAwB,CAAC;YACnD,OAAO,CAAC,OAAO,CAAC,UAAU,CAAC,wBAAwB,CAAC,CAAC,CACxD,CAAC;AACJ,CAAC;AAED;;;;;;;;;GASG;AACH,MAAM,UAAU,eAAe,CAC7B,QAAuB,EACvB,UAA0B,EAAE;IAE5B,MAAM,MAAM,GAAG,OAAO,CAAC,YAAY,IAAI,qBAAqB,CAAC;IAC7D,MAAM,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,CAAC,UAAU,IAAI,mBAAmB,CAAC,CAAC;IAC1E,IAAI,QAAQ,CAAC,MAAM,IAAI,UAAU,GAAG,CAAC;QAAE,OAAO,QAAQ,CAAC;IACvD,IAAI,sBAAsB,CAAC,QAAQ,CAAC,IAAI,MAAM;QAAE,OAAO,QAAQ,CAAC;IAEhE,oEAAoE;IACpE,MAAM,IAAI,GAAkB,EAAE,CAAC;IAC/B,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,IAAI,QAAQ,CAAC,CAAC,CAAC,EAAE,IAAI,KAAK,QAAQ,EAAE,CAAC;QACnC,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;QACvB,KAAK,GAAG,CAAC,CAAC;IACZ,CAAC;IAED,MAAM,IAAI,GAAG,QAAQ,CAAC,KAAK,CAAC,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,QAAQ,CAAC,MAAM,GAAG,UAAU,CAAC,CAAC,CAAC;IAC3E,MAAM,MAAM,GAAG,QAAQ,CAAC,KAAK,CAAC,KAAK,EAAE,QAAQ,CAAC,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC;IACpE,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,QAAQ,CAAC;IAEzC,MAAM,OAAO,GAAa,EAAE,CAAC;IAC7B,KAAK,MAAM,GAAG,IAAI,MAAM,EAAE,CAAC;QACzB,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YACxB,OAAO,CAAC,IAAI,CAAC,iBAAiB,OAAO,CAAC,GAAG,CAAC,OAAO,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC;QAC7D,CAAC;aAAM,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YACpC,MAAM,IAAI,GAAG,OAAO,CAAC,GAAG,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC;YACvC,IAAI,IAAI;gBAAE,OAAO,CAAC,IAAI,CAAC,gBAAgB,IAAI,EAAE,CAAC,CAAC;QACjD,CAAC;aAAM,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YAC/B,OAAO,CAAC,IAAI,CAAC,kBAAkB,OAAO,CAAC,GAAG,CAAC,OAAO,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC;QAC9D,CAAC;IACH,CAAC;IAED,MAAM,IAAI,GAAgB;QACxB,IAAI,EAAE,QAAQ;QACd,OAAO,EACL,GAAG,wBAAwB,oHAAoH;YAC/I,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC;KACrB,CAAC;IAEF,OAAO,CAAC,GAAG,IAAI,EAAE,IAAI,EAAE,GAAG,IAAI,CAAC,CAAC;AAClC,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,CAAC,KAAK,UAAU,0BAA0B,CAC9C,QAAuB,EACvB,SAA8C,EAC9C,UAA0B,EAAE,EAC5B,iBAAsC;IAEtC,MAAM,MAAM,GAAG,QAAQ,CAAC,MAAM,CAAC;IAC/B,MAAM,YAAY,GAAG,sBAAsB,CAAC,QAAQ,CAAC,CAAC;IACtD,MAAM,QAAQ,GAAG,OAAO,CAAC,YAAY,KAAK,CAAC,CAAC;IAE5C,IAAI,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,CAAC,UAAU,IAAI,mBAAmB,CAAC,CAAC;IACxE,MAAM,KAAK,GAAG,QAAQ,CAAC,CAAC,CAAC,EAAE,IAAI,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACrD,IAAI,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,QAAQ,CAAC,MAAM,GAAG,UAAU,CAAC,CAAC;IAC9D,IAAI,KAAK,GAAG,QAAQ,CAAC,KAAK,CAAC,KAAK,EAAE,SAAS,CAAC,CAAC;IAE7C,yEAAyE;IACzE,qFAAqF;IACrF,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC,IAAI,QAAQ,IAAI,QAAQ,CAAC,MAAM,IAAI,KAAK,GAAG,CAAC,EAAE,CAAC;QACnE,UAAU,GAAG,CAAC,CAAC;QACf,SAAS,GAAG,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC;QAChC,KAAK,GAAG,QAAQ,CAAC,KAAK,CAAC,KAAK,EAAE,SAAS,CAAC,CAAC;IAC3C,CAAC;IAED,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC,IAAI,CAAC,iBAAiB,EAAE,IAAI,EAAE,EAAE,CAAC;QACrD,4DAA4D;QAC5D,OAAO;YACL,QAAQ,EAAE,CAAC,GAAG,QAAQ,CAAC;YACvB,MAAM;YACN,KAAK,EAAE,MAAM;YACb,YAAY;YACZ,WAAW,EAAE,YAAY;YACzB,UAAU,EAAE,KAAK;SAClB,CAAC;IACJ,CAAC;IAED,MAAM,iBAAiB,GAAG,KAAK;SAC5B,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE;QACf,IAAI,OAAO,GAAG,aAAa,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;QAC7C,iEAAiE;QACjE,4DAA4D;QAC5D,IAAI,OAAO,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YACjC,OAAO,GAAG,aAAa,CAAC,OAAO,CAAC,CAAC,OAAO,CAAC;QAC3C,CAAC;QACD,OAAO,GAAG,OAAO,CAAC,IAAI,CAAC,WAAW,EAAE,KAAK,OAAO,EAAE,CAAC;IACrD,CAAC,CAAC;SACD,IAAI,CAAC,MAAM,CAAC,CAAC;IAChB,4EAA4E;IAC5E,2EAA2E;IAC3E,MAAM,MAAM,GAAG,iBAAiB,EAAE,IAAI,EAAE;QACtC,CAAC,CAAC,aAAa,CAAC,iBAAiB,CAAC,IAAI,EAAE,CAAC;QACzC,CAAC,CAAC,EAAE,CAAC;IACP,MAAM,YAAY,GAAG,iBAAiB,CAAC,IAAI,EAAE,CAAC;IAC9C,MAAM,UAAU,GACd,MAAM,IAAI,YAAY;QACpB,CAAC,CAAC,GAAG,MAAM,oCAAoC,YAAY,EAAE;QAC7D,CAAC,CAAC,MAAM,IAAI,YAAY,CAAC;IAC7B,MAAM,MAAM,GAAG;QACb,4HAA4H;QAC5H,gHAAgH;QAChH,sNAAsN;QACtN,+JAA+J;QAC/J,mHAAmH;QACnH,EAAE;QACF,UAAU;KACX,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAEb,uEAAuE;IACvE,mDAAmD;IACnD,wEAAwE;IACxE,qEAAqE;IACrE,MAAM,UAAU,GAAG,aAAa,CAAC,CAAC,MAAM,SAAS,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC;IACnE,MAAM,OAAO,GAAG,aAAa,CAAC,UAAU,CAAC,CAAC,OAAO,CAAC,IAAI,EAAE,CAAC;IACzD,IAAI,CAAC,OAAO;QAAE,MAAM,IAAI,KAAK,CAAC,oDAAoD,CAAC,CAAC;IAEpF,MAAM,IAAI,GAAG,KAAK,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IAC/C,kEAAkE;IAClE,gDAAgD;IAChD,MAAM,IAAI,GAAG,QAAQ,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE;QACjD,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,IAAI,SAAS,CAAC,IAAI,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,CAAC;YAC5D,OAAO,EAAE,GAAG,GAAG,EAAE,OAAO,EAAE,aAAa,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,OAAO,EAAE,CAAC;QACjE,CAAC;QACD,OAAO,GAAG,CAAC;IACb,CAAC,CAAC,CAAC;IACH,MAAM,SAAS,GAAkB;QAC/B,GAAG,IAAI;QACP,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,GAAG,wBAAwB,OAAO,OAAO,EAAE,EAAE;QACxE,GAAG,IAAI;KACR,CAAC;IACF,MAAM,WAAW,GAAG,sBAAsB,CAAC,SAAS,CAAC,CAAC;IACtD,OAAO;QACL,QAAQ,EAAE,SAAS;QACnB,MAAM;QACN,KAAK,EAAE,SAAS,CAAC,MAAM;QACvB,YAAY;QACZ,WAAW;QACX,UAAU,EAAE,IAAI;KACjB,CAAC;AACJ,CAAC;AAED,SAAS,OAAO,CAAC,IAAY,EAAE,QAAgB;IAC7C,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,IAAI,EAAE,CAAC;IACjD,IAAI,OAAO,CAAC,MAAM,IAAI,QAAQ;QAAE,OAAO,OAAO,CAAC;IAC/C,OAAO,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,QAAQ,GAAG,CAAC,CAAC,GAAG,CAAC;AAC9C,CAAC"}

package/dist/agent/events.d.ts

@@ -26,10 +26,13 @@ export type AgentEvent = {
     id: string;
     name: string;
     argsDisplay: string;
-} | {
+}
+/** `replace: true` sets the full body (never append) so the UI never keeps a truncated live preview. */
+ | {
     type: "tool-output";
     id: string;
     chunk: string;
+    replace?: boolean;
 } | {
     type: "tool-result";
     id: string;

package/dist/agent/runner.js

@@ -11,7 +11,7 @@ import { classifyToolCall, isPentestToolCall, scopeHint, scopeTargetForToolCall,
 import { availableToolNames, normalizeToolCall, runToolCall, BATCH_SAFE_TOOLS, } from "../tools/registry.js";
 import { looksInteractiveStdin } from "../tools/shell.js";
 import { formatViewportHint, registerViewport } from "../ui/output-pane.js";
-import { compactMessagesWithSummary, estimateTokens, estimateMessagesTokens, COMPACTION_MEMORY_PREFIX, } from "./context-manager.js";
+import { compactMessagesWithSummary, estimateTokens, estimateMessagesTokens, AUTO_COMPACT_TOKEN_BUDGET, COMPACTION_MEMORY_PREFIX, } from "./context-manager.js";
 import { auditLog } from "../store/logs.js";
 import { loadProjectContext } from "../store/project.js";
 import { loadScope, isScopeActive, targetInScope } from "../store/scope.js";
@@ -24,7 +24,7 @@ import { analyzeTask } from "./task-analyzer.js";
 import { LoopGuard } from "./loop-guard.js";
 import { loadPlan } from "../store/plan.js";
 import { pathInsideSandbox, fsWrite } from "../tools/fs.js";
-import { stripSentinelTokens, parseToolCall, recognizeBareToolJson, looksLikeTruncatedToolCall, salvageTruncatedWrite, countToolFences, parseAllToolCalls, groupToolCallsForExecution, buildTurnHistory, collapseRepeatedText, textBeforeToolCall, formatToolArgs, looksLikePentestTask, looksLikeBuildTask, looksLikeInformationalQuery, looksLikeActionNarration, looksLikePlanNarration, requiresFreshWebSearch, freshnessGuardMessage, buildWorkflowDirective, pentestWorkflowDirective, shouldDimToolChatter, looksLikePromptLeak, } from "./tool-call-parser.js";
+import { stripSentinelTokens, parseToolCall, recognizeBareToolJson, looksLikeTruncatedToolCall, salvageTruncatedWrite, countToolFences, parseAllToolCalls, groupToolCallsForExecution, buildTurnHistory, collapseRepeatedText, textBeforeToolCall, formatToolArgs, looksLikePentestTask, looksLikeBuildTask, looksLikeInformationalQuery, looksLikeIdleOrSocialPrompt, looksLikeActionNarration, looksLikeWebActionNarration, looksLikePlanNarration, requiresFreshWebSearch, freshnessGuardMessage, buildWorkflowDirective, pentestWorkflowDirective, shouldDimToolChatter, looksLikePromptLeak, } from "./tool-call-parser.js";
 import { createSessionPolicy, isPreApprovalAllowedTool, isPlanApprovedByStatus, planHasOpenWork, isAbortError, shouldEnableImageOcr, } from "./session-policy.js";
 import { saveToolOutput, summarizeOutput, formatToolContext, } from "./tool-output-formatting.js";
 import { planContextMessage, handlePlanTool, } from "./plan-tool.js";
@@ -114,6 +114,13 @@ function isScratchOnlyWrite(call, scratchDir) {
 export async function runAgentLoop(prompt, options = {}) {
     const writesDirectly = !options.onEvent;
     const emit = (event) => options.onEvent?.(event);
+    // Whether the CURRENT model iteration has already committed its visible
+    // prose to the transcript with an `assistant-message` event. The recovery
+    // paths preserve streamed prose before retrying (so it isn't wiped by the
+    // next tool-call/turn event); this flag stops them from re-committing prose
+    // the normal tool path already surfaced, which would render it twice. Reset
+    // at the top of every loop iteration.
+    let visibleCommitted = false;
     const noopSpinner = {
         setLabel: () => { },
         bumpReasoning: () => { },
@@ -121,7 +128,9 @@ export async function runAgentLoop(prompt, options = {}) {
         stop: () => { },
     };
     const writeStatus = (text, rendered = chalk.dim(text)) => {
-        emit({ type: "status", text });
+        // Footer activity is single-line; strip classic stdout newlines/indents.
+        const cleaned = text.replace(/\s+/g, " ").trim();
+        emit({ type: "status", text: cleaned || text });
         if (writesDirectly)
             process.stdout.write(rendered);
     };
@@ -131,6 +140,11 @@ export async function runAgentLoop(prompt, options = {}) {
             process.stdout.write(rendered);
     };
     const writeAssistantMessage = (text) => {
+        // Never surface an empty message: the reducer drops it and a direct
+        // stdout writer would print a stray blank line.
+        if (!text.trim())
+            return;
+        visibleCommitted = true;
         emit({ type: "assistant-message", text });
         const rendered = renderMarkdown(text);
         if (writesDirectly) {
@@ -142,9 +156,14 @@ export async function runAgentLoop(prompt, options = {}) {
         if (writesDirectly)
             process.stdout.write(`${renderThinkingSummary(content)}\n`);
     };
-    const writeToolOutput = (id, chunk, rendered) => {
-        emit({ type: "tool-output", id, chunk });
-        if (writesDirectly)
+    const writeToolOutput = (id, chunk, rendered, options) => {
+        emit({
+            type: "tool-output",
+            id,
+            chunk,
+            ...(options?.replace ? { replace: true } : {}),
+        });
+        if (writesDirectly && rendered)
             process.stdout.write(rendered);
     };
     const writeToolCall = (id, call, rendered) => {
@@ -245,8 +264,13 @@ export async function runAgentLoop(prompt, options = {}) {
         // brand-new plan (the exact failure where "what do u know till now"
         // triggered explore→plan and created an unrelated "Enhance clai" plan).
         const informationalQuery = looksLikeInformationalQuery(prompt);
+        // Greetings / thanks / short acks must never force tools, plans, or
+        // freshness retries — a false "act don't narrate" path burned tokens on
+        // web.search recovery loops after a simple "hi".
+        const idleOrSocialPrompt = looksLikeIdleOrSocialPrompt(prompt);
         const freshWebSearchRequired = !buildLikeTurn &&
             !pentestLikeTurn &&
+            !idleOrSocialPrompt &&
             toolNames.includes("web.search") &&
             requiresFreshWebSearch(prompt);
         let provider = options.provider ?? config.defaultProvider;
@@ -295,7 +319,10 @@ export async function runAgentLoop(prompt, options = {}) {
         // what's already there, create a comprehensive multi-task plan, then
         // implement task by task until the goal is met. This mirrors how a careful
         // coding agent (Claude Code) operates.
-        if (buildLikeTurn && !activePlan && !informationalQuery) {
+        if (buildLikeTurn &&
+            !activePlan &&
+            !informationalQuery &&
+            !idleOrSocialPrompt) {
             systemSections.push(buildWorkflowDirective());
         }
         // Pentest / security engagements need a different shape than a coding
@@ -304,7 +331,10 @@ export async function runAgentLoop(prompt, options = {}) {
         // directive is only injected before a plan exists; once a plan is in
         // place (or being refined), the ACTIVE PLAN block already carries the
         // current task state and recon-vs-active-tool guidance.
-        if (pentestLikeTurn && !activePlan && !informationalQuery) {
+        if (pentestLikeTurn &&
+            !activePlan &&
+            !informationalQuery &&
+            !idleOrSocialPrompt) {
             systemSections.push(pentestWorkflowDirective());
         }
         const renderedSystemPrompt = systemSections.join("\n\n");
@@ -346,7 +376,32 @@ export async function runAgentLoop(prompt, options = {}) {
         // which can cause every retry to return empty as well. Keep hidden thinking
         // out of history, but record a compact non-empty sentinel when there was no
         // visible output.
+        //
+        // The Ink reducer intentionally keeps streamed text transient until it
+        // receives `assistant-message`, because a stream may turn out to be raw
+        // tool JSON. Recovery paths used to add prose to model history and retry
+        // without that event, so the user could watch a valid-looking response
+        // vanish when the retry emitted its first tool call. Surface only prose
+        // that is safe to render; malformed/bare tool payloads remain hidden.
+        const recoveryProse = (content) => {
+            const text = textBeforeToolCall(stripSentinelTokens(content)).trim();
+            if (!text ||
+                /^```|^\{|<tool_call>|<\|tool_call(?:s_section)?_begin\|>/i.test(text) ||
+                /\n\s*\{[\s\S]*\}\s*$/.test(text)) {
+                return undefined;
+            }
+            return text;
+        };
         const pushAssistantHistory = (content) => {
+            // Preserve genuine streamed prose before a recovery retry so the visible
+            // text isn't wiped by the next tool-call/turn event. Skip when this
+            // iteration already surfaced its prose (the normal tool path commits
+            // `beforeTool` itself) so the same text is never rendered twice.
+            if (!visibleCommitted) {
+                const prose = recoveryProse(content);
+                if (prose)
+                    writeAssistantMessage(prose);
+            }
             messages.push({
                 role: "assistant",
                 content: content.trim()
@@ -687,38 +742,20 @@ export async function runAgentLoop(prompt, options = {}) {
             parentSignal.addEventListener("abort", onParentAbort);
             let result;
             let liveBytes = 0;
-            const liveCap = 16_000;
-            let liveTruncatedNotified = false;
-            let lastProgressAt = 0;
             const shouldDimLive = !interactiveCommand;
-            const writeToolInfo = (text) => {
-                writeToolOutput(toolEventId, `${text}\n`, chalk.dim(`  ${text}\n`));
-            };
+            // Stream every live byte — never drop mid-run. After the tool finishes we
+            // still replace the spool with the authoritative full `result.output`.
             const printLive = (chunk) => {
                 if (call.name === "fs.read" ||
                     call.name === "fs.list" ||
                     call.name === "fs.search")
                     return;
-                if (liveBytes >= liveCap) {
-                    if (!liveTruncatedNotified) {
-                        liveTruncatedNotified = true;
-                        writeToolInfo("... live preview truncated, full output saved");
-                        writeToolInfo("(tool still running — ESC or Ctrl+C to abort)");
-                        lastProgressAt = Date.now();
-                    }
-                    const now = Date.now();
-                    if (now - lastProgressAt > 5_000) {
-                        lastProgressAt = now;
-                        writeToolOutput(toolEventId, ".", chalk.dim("."));
-                    }
+                if (!chunk)
                     return;
-                }
-                const remaining = liveCap - liveBytes;
-                const slice = chunk.length > remaining ? chunk.slice(0, remaining) : chunk;
-                liveBytes += slice.length;
-                const indented = slice.replace(/\r/g, "").replace(/\n(?!$)/g, "\n  ");
+                liveBytes += chunk.length;
+                const indented = chunk.replace(/\r/g, "").replace(/\n(?!$)/g, "\n  ");
                 const body = indented.startsWith("\n") ? indented : `  ${indented}`;
-                writeToolOutput(toolEventId, slice, shouldDimLive ? chalk.dim(body) : body);
+                writeToolOutput(toolEventId, chunk, shouldDimLive ? chalk.dim(body) : body);
             };
             const jobId = randomUUID().slice(0, 8);
             const backgroundJob = {
@@ -761,7 +798,7 @@ export async function runAgentLoop(prompt, options = {}) {
                     confirmed: true,
                     userPrompt: prompt,
                 });
-                if (liveBytes > 0 || liveTruncatedNotified) {
+                if (liveBytes > 0) {
                     writeToolOutput(toolEventId, "\n", "\n");
                 }
                 jobManager.updateJobStatus(jobId, result.ok ? "exited" : "failed", result.exitCode);
@@ -802,11 +839,10 @@ export async function runAgentLoop(prompt, options = {}) {
                 parentSignal.removeEventListener("abort", onParentAbort);
             }
             const output = result.output.trim();
-            const displayMax = 6_000;
+            // Always keep a full on-disk copy of tool output (any size) so the
+            // pager never depends on a truncated in-memory preview.
             const savedOutputPath = result.outputPath ??
-                (output.length > displayMax
-                    ? await saveToolOutput(call, output)
-                    : undefined);
+                (output ? await saveToolOutput(call, output) : undefined);
             const resultWithArtifact = {
                 ...result,
                 outputPath: savedOutputPath,
@@ -841,20 +877,30 @@ export async function runAgentLoop(prompt, options = {}) {
                 }
             }
             const statusIcon = result.ok ? chalk.green("  ✓") : chalk.red("  ✗");
-            writeToolOutput(toolEventId, result.ok ? "ok\n" : "failed\n", statusIcon + "\n");
+            // Classic stdout gets a short status glyph; the event stream / spool
+            // must NOT be polluted with "ok"/"failed" (the card already shows status).
+            if (writesDirectly) {
+                process.stdout.write(statusIcon + "\n");
+            }
             if (output) {
-                const displaySummary = summarizeOutput(output, displayMax);
-                const displayText = displaySummary.truncated
-                    ? `${displaySummary.text}${savedOutputPath ? chalk.dim(`\n  ... full output saved to ${savedOutputPath}`) : chalk.dim("\n  ... output truncated")}`
-                    : displaySummary.text;
-                if (liveBytes > 0) {
-                    if (savedOutputPath) {
-                        writeToolInfo(`full output saved to ${savedOutputPath}`);
-                    }
+                // Authoritative FULL body — replace any live stream so the pager
+                // never shows a truncated mid-run preview. Never cap for the UI.
+                const fullChunk = output.endsWith("\n") ? output : `${output}\n`;
+                if (!writesDirectly) {
+                    writeToolOutput(toolEventId, fullChunk, "", { replace: true });
                 }
                 else {
+                    // Classic stdout: short preview only; full text is on disk.
+                    const displayMax = 6_000;
+                    const displaySummary = summarizeOutput(output, displayMax);
+                    const displayText = displaySummary.truncated
+                        ? `${displaySummary.text}${savedOutputPath ? chalk.dim(`\n  ... full output saved to ${savedOutputPath}`) : ""}`
+                        : displaySummary.text;
                     const renderedOutput = indentAndWrapText(displayText);
-                    writeToolOutput(toolEventId, displayText, styleToolChatter(call, renderedOutput) + "\n");
+                    process.stdout.write(styleToolChatter(call, renderedOutput) + "\n");
+                    if (savedOutputPath && displaySummary.truncated) {
+                        process.stdout.write(chalk.dim(`  full output saved to ${savedOutputPath}\n`));
+                    }
                 }
             }
             if (output) {
@@ -880,7 +926,6 @@ export async function runAgentLoop(prompt, options = {}) {
         // what is done, and what remains. The estimate is chars/4; the budget is
         // deliberately conservative so we compact a little early rather than hit a
         // provider context-window error mid-task.
-        const AUTO_COMPACT_TOKEN_BUDGET = 150_000;
         const AUTO_COMPACT_KEEP_RECENT = 6;
         let lastCompactionMsgCount = 0;
         const summarizeForCompaction = async (summaryPrompt) => {
@@ -916,6 +961,8 @@ export async function runAgentLoop(prompt, options = {}) {
                     return;
                 messages.splice(0, messages.length, ...result.messages);
                 loopGuard.resetReadOnly();
+                // Token stats BEFORE plan re-injection so the reduction is accurate.
+                const compactedTokens = estimateMessagesTokens(messages);
                 // Re-inject the live plan so the model keeps full plan awareness even
                 // after older turns (which carried the plan context) were summarized.
                 const livePlan = await loadPlan(session.sessionId).catch(() => undefined);
@@ -926,13 +973,11 @@ export async function runAgentLoop(prompt, options = {}) {
                     });
                 }
                 lastCompactionMsgCount = messages.length;
-                // Report the compacted token count BEFORE plan re-injection so the
-                // compaction stats accurately show the reduction. Then report the
-                // final count (which is what the model actually receives) separately.
-                const compactedTokens = estimateMessagesTokens(messages);
+                // Final count the model actually receives (may include re-injected plan).
+                const afterTokens = estimateMessagesTokens(messages);
                 await auditLog("agent.compact", {
                     newLength: messages.length,
-                    estimatedTokens: compactedTokens,
+                    estimatedTokens: afterTokens,
                     reason,
                 });
                 // Extract the inserted compaction memory so we can surface the
@@ -944,12 +989,12 @@ export async function runAgentLoop(prompt, options = {}) {
                 const summaryText = insertedSummary.startsWith(`${COMPACTION_MEMORY_PREFIX}\n\n`)
                     ? insertedText(insertedSummary, `${COMPACTION_MEMORY_PREFIX}\n\n`)
                     : insertedText(insertedSummary, COMPACTION_MEMORY_PREFIX);
-                const afterTokens = estimateMessagesTokens(messages);
-                writeCompacted(summaryText, beforeTokens, afterTokens);
+                // Card shows pre/post of the summarization; plan re-injection is noted.
+                writeCompacted(summaryText, beforeTokens, compactedTokens);
                 const planNote = afterTokens > compactedTokens
                     ? ` (compacted to ~${compactedTokens.toLocaleString()}, +plan → ~${afterTokens.toLocaleString()})`
                     : "";
-                writeNotice("info", `context auto-compacted to fit the window (~${beforeTokens.toLocaleString()} → ~${afterTokens.toLocaleString()} tokens${planNote})`, chalk.dim(`  ℹ context auto-compacted (~${beforeTokens.toLocaleString()} → ~${afterTokens.toLocaleString()} tokens${planNote})\n`));
+                writeNotice("info", `context auto-compacted to fit the window (~${beforeTokens.toLocaleString()} → ~${compactedTokens.toLocaleString()} tokens${planNote})`, chalk.dim(`  ℹ context auto-compacted (~${beforeTokens.toLocaleString()} → ~${compactedTokens.toLocaleString()} tokens${planNote})\n`));
             }
             catch (error) {
                 if (error instanceof Error &&
@@ -964,6 +1009,10 @@ export async function runAgentLoop(prompt, options = {}) {
             }
         }
         for (let iteration = 0; iteration < maxIterations; iteration += 1) {
+            // Each model round-trip re-decides whether its visible prose has been
+            // committed, so recovery-prose preservation applies to THIS turn's
+            // stream and not a prior one.
+            visibleCommitted = false;
             // `step` is the productive-step index (used for display + audit). It only
             // advances when the previous iteration actually executed a tool.
             step = productiveSteps;
@@ -1070,10 +1119,11 @@ export async function runAgentLoop(prompt, options = {}) {
                     completion = await streamWithProvider({
                         provider,
                         model,
-                        // The selected model is the first choice, but an agent must not
-                        // end a turn with no answer if it only emits reasoning. Honor
-                        // the user's providerFallback setting for that recovery path.
-                        allowModelFallback: true,
+                        // Stay on the user's selected provider/model. Cross-provider
+                        // hops (e.g. bynara → gemini on a 429) only happen when the
+                        // user has explicitly enabled `providerFallback` AND this is
+                        // the provider's default model. Never force fallback mid-turn.
+                        allowModelFallback: false,
                         messages,
                         // MiniMax M3 degenerates at the generic agent temperature. The
                         // HTTP layer also applies its `top_p` override for both the
@@ -1440,28 +1490,49 @@ export async function runAgentLoop(prompt, options = {}) {
                     // final answer ends the turn with nothing done and no real plan saved.
                     // Nudge it to emit a real tool call, with a concrete example.
                     const narratedAction = looksLikeActionNarration(cleaned);
-                    // A plan whose OWN persisted status is "completed" has no more
-                    // work to force — treat this like "no active plan" for the
-                    // act-don't-narrate nudge so a plain follow-up question (e.g. "what
-                    // do you know so far") after the plan finished gets answered
-                    // instead of being pushed to emit another tool call.
-                    const planHasOpenWorkNow = planHasOpenWork(activePlan?.status);
-                    // History-inherited build/pentest intent only forces action when
-                    // THIS prompt is not itself a plain question. If the model does
-                    // narrate a concrete next step ("let me read X"), narratedAction
-                    // still forces it to actually run that step — even for a question —
-                    // which is the desired "do what you said" behavior.
-                    const wantsAction = narratedAction ||
-                        freshWebSearchRequired ||
+                    const narratedWebAction = looksLikeWebActionNarration(cleaned);
+                    // `activePlan` is the snapshot loaded at the start of this turn.
+                    // A final task.update mutates and persists a separate plan instance,
+                    // so re-read it here before deciding whether a final response should
+                    // be forced back into execution. Without this, a completed plan
+                    // still appeared open and a short final summary such as "I'll
+                    // summarize the findings" could be discarded for an unnecessary
+                    // recovery turn.
+                    const livePlanAtCompletion = await loadPlan(session.sessionId).catch(() => undefined);
+                    const planStatusAtCompletion = livePlanAtCompletion?.status ?? activePlan?.status;
+                    const completedPlanDuringThisTurn = activePlan?.status !== "completed" &&
+                        planStatusAtCompletion === "completed";
+                    const planHasOpenWorkNow = planHasOpenWork(planStatusAtCompletion);
+                    // User-driven reasons to require tools this turn (build/pentest/
+                    // approved plan / freshness). Idle greetings never qualify.
+                    const userExpectsWork = freshWebSearchRequired ||
                         (planHasOpenWorkNow && session.planApproved.value) ||
-                        (!informationalQuery && (buildLikeTurn || pentestLikeTurn));
+                        (!informationalQuery &&
+                            !idleOrSocialPrompt &&
+                            (buildLikeTurn || pentestLikeTurn));
+                    // History-inherited build/pentest intent only forces action when
+                    // THIS prompt is not itself a plain question/idle turn. Narration
+                    // stalls ("I'll list the files") still force a tool when the user
+                    // is not asking an informational question — but capability menus
+                    // and greetings are filtered out by looksLikeActionNarration /
+                    // idleOrSocialPrompt so they cannot burn recovery turns.
+                    const wantsAction = !completedPlanDuringThisTurn &&
+                        !idleOrSocialPrompt &&
+                        (userExpectsWork ||
+                            (narratedAction && !informationalQuery) ||
+                            (narratedWebAction && !informationalQuery));
                     const planNarrated = (buildLikeTurn || pentestLikeTurn) &&
                         !activePlan &&
                         looksLikePlanNarration(cleaned);
+                    // Once a real tool step has run, a no-plan task has no durable task
+                    // state to prove whether another action is needed. A tool-free reply
+                    // must therefore be allowed to finalize instead of turning a short
+                    // summary containing “I'll” into an implicit recovery request.
+                    const shouldRetryBeforeFinalizing = productiveSteps === 0 || planNarrated;
                     if (wantsAction &&
                         cleaned.trim().length > 0 &&
                         actionIntentRetries < 3 &&
-                        (productiveSteps === 0 || planNarrated || narratedAction)) {
+                        shouldRetryBeforeFinalizing) {
                         actionIntentRetries += 1;
                         let nudge;
                         if (planHasOpenWorkNow && session.planApproved.value) {
@@ -1476,14 +1547,20 @@ export async function runAgentLoop(prompt, options = {}) {
                                     "Every turn MUST contain a ```tool block until the task is done.";
                             writeNotice("warn", "described a security/pentest action but emitted no tool call — nudging it to run one", chalk.yellow("  ⚠ described a security/pentest action but emitted no tool call — nudging it to run one\n"));
                         }
-                        else if (!buildLikeTurn) {
+                        else if (freshWebSearchRequired || narratedWebAction) {
+                            // Web-specific recovery ONLY when the user asked for current
+                            // info or the model explicitly claimed a fetch/search step.
+                            // Previously every non-build stall used this path, so a "hi"
+                            // greeting that said "I'll start executing" was forced into
+                            // pointless web.search recovery loops.
                             nudge =
                                 "You wrote that you would fetch/search/read something but emitted NO ```tool block, so NOTHING ran. Do NOT narrate the next browsing step — DO it. Emit exactly one valid ```tool block now. If you know the exact page, use:\n" +
                                     '```tool\n{"name":"web.fetch","args":{"url":"https://example.com/page","responseMode":"readable"}}\n```\n' +
                                     "If you do not know the exact page URL, use web.search first. After the tool output, answer from the fetched page content.";
                             writeNotice("warn", "described a web action but emitted no tool call — nudging it to run one", chalk.yellow("  ⚠ described a web action but emitted no tool call — nudging it to run one\n"));
                         }
-                        else if (planNarrated || productiveSteps > 0) {
+                        else if (buildLikeTurn &&
+                            (planNarrated || productiveSteps > 0)) {
                             const kind = pentestLikeTurn ? "pentest" : "coding";
                             nudge =
                                 "You wrote the plan as PROSE but did NOT call plan.create, so no plan was saved and the user cannot /implement it. Emit it as a real tool call NOW — exactly one ```tool block:\n" +
@@ -1491,17 +1568,17 @@ export async function runAgentLoop(prompt, options = {}) {
                                     "Do not describe the plan again in prose — just emit the plan.create tool block.";
                             writeNotice("warn", "plan was written as text, not created — nudging it to call plan.create", chalk.yellow("  ⚠ plan was written as text, not created — nudging it to call plan.create\n"));
                         }
+                        else if (buildLikeTurn) {
+                            nudge =
+                                "You described what you will do but emitted NO ```tool block, so NOTHING actually happened — narration is not action. Emit a real tool call NOW. For this build task, explore first like this:\n" +
+                                    '```tool\n{"name":"fs.list","args":{"path":"."}}\n```\n' +
+                                    "Then read key files, and once you understand the directory, call plan.create. Every turn MUST contain a ```tool block until the task is done.";
+                            writeNotice("warn", "described an action but emitted no tool call — nudging it to run one", chalk.yellow("  ⚠ described an action but emitted no tool call — nudging it to run one\n"));
+                        }
                         else {
-                            if (pentestLikeTurn) {
-                                nudge =
-                                    "You described what you will do but emitted NO ```tool block, so NOTHING actually happened — narration is not action. Emit a real tool call NOW. For this pentest task, explore/recon first, then call plan.create. Every turn MUST contain a ```tool block until the task is done.";
-                            }
-                            else {
-                                nudge =
-                                    "You described what you will do but emitted NO ```tool block, so NOTHING actually happened — narration is not action. Emit a real tool call NOW. For this build task, explore first like this:\n" +
-                                        '```tool\n{"name":"fs.list","args":{"path":"."}}\n```\n' +
-                                        "Then read key files, and once you understand the directory, call plan.create. Every turn MUST contain a ```tool block until the task is done.";
-                            }
+                            // Generic non-build, non-web stall (e.g. "I'll list the files").
+                            nudge =
+                                "You described what you will do but emitted NO ```tool block, so NOTHING actually happened — narration is not action. Emit a real tool call NOW for the step you just described. Every turn that claims an action MUST contain a ```tool block.";
                             writeNotice("warn", "described an action but emitted no tool call — nudging it to run one", chalk.yellow("  ⚠ described an action but emitted no tool call — nudging it to run one\n"));
                         }
                         pushAssistantHistory(assistantText.visible);
@@ -1688,7 +1765,12 @@ export async function runAgentLoop(prompt, options = {}) {
                             `${omittedCallCount} were not run. After reviewing results, issue a small, prioritized next batch.`,
                     });
                 }
-                // Now emit only the calls that will actually execute, after thinking
+                // Notice BEFORE tool cards so the transcript reads:
+                // thinking → response → "N tool calls…" → tool cards (not tools then info).
+                if (allCalls.length > 1) {
+                    writeNotice("info", `${allCalls.length} tool calls in this message — running scoped (independent read-only lookups in parallel, everything else in order)`, chalk.dim(`  ℹ ${allCalls.length} tool calls — read-only lookups in parallel, the rest in order\n`));
+                }
+                // Emit only the calls that will actually execute, after thinking
                 // + assistant text so transcript order remains correct.
                 for (const deferred of activeDeferredToolCalls.slice(0, allCalls.length)) {
                     writeToolCall(deferred.eventId, deferred.call, deferred.rendered);
@@ -1698,9 +1780,6 @@ export async function runAgentLoop(prompt, options = {}) {
                         .map((c) => `\`\`\`tool\n${JSON.stringify(c)}\n\`\`\``)
                         .join("\n\n");
                 pushAssistantHistory(standardizedContent);
-                if (allCalls.length > 1) {
-                    writeNotice("info", `${allCalls.length} tool calls in this message — running scoped (independent read-only lookups in parallel, everything else in order)`, chalk.dim(`  ℹ ${allCalls.length} tool calls — read-only lookups in parallel, the rest in order\n`));
-                }
                 // Scoped-parallel batch execution
                 // The model may emit several calls in one message. We partition them,
                 // IN DOCUMENT ORDER, into segments:
@@ -1736,7 +1815,12 @@ export async function runAgentLoop(prompt, options = {}) {
                 let blockedResult = null;
                 let failed = false;
                 let awaitingPlanApproval = false;
+                /** Indices into allCalls that actually ran (got a tool-result). */
+                const executedIndices = new Set();
                 const recordResult = (res, continueAfterFailure = false) => {
+                    const idx = allCalls.indexOf(res.call);
+                    if (idx >= 0)
+                        executedIndices.add(idx);
                     messages.push({
                         role: "tool",
                         content: `Tool ${res.call.name} result (exit=${res.result.exitCode ?? 0}, ok=${res.result.ok}):\n${res.contextOutput}`,
@@ -1801,6 +1885,41 @@ export async function runAgentLoop(prompt, options = {}) {
                             recordResult(res, true);
                     }
                 }
+                // Cards were emitted for every call up front. If the batch stopped
+                // early (failure / abort / plan gate), any card still on "running"
+                // must get a terminal result so the TUI never spins forever.
+                for (let i = 0; i < allCalls.length; i += 1) {
+                    if (executedIndices.has(i))
+                        continue;
+                    const call = allCalls[i];
+                    if (i >= 0 && !callIds[i]) {
+                        callIds[i] = `tool-${++nextToolEventId}`;
+                    }
+                    const id = callIds[i];
+                    if (!alreadyPrintedIds.has(id)) {
+                        // Never shown in the transcript — skip.
+                        continue;
+                    }
+                    const reason = aborted
+                        ? "Cancelled — turn aborted before this call ran."
+                        : blocked
+                            ? "Cancelled — earlier tool was blocked or declined."
+                            : awaitingPlanApproval
+                                ? "Deferred — waiting for plan approval."
+                                : failed
+                                    ? "Cancelled — earlier tool in this batch failed."
+                                    : "Cancelled — not executed.";
+                    const result = {
+                        ok: false,
+                        output: reason,
+                        exitCode: 130,
+                    };
+                    emitToolResult(id, result, reason);
+                    messages.push({
+                        role: "tool",
+                        content: `Tool ${call.name} result (exit=130, ok=false):\n${reason}`,
+                    });
+                }
                 // plan.create is a hard transaction boundary. Its successful handler
                 // persists and displays the plan; returning immediately prevents a
                 // stale pre-loop activePlan snapshot from nudging a duplicate plan and