@pentoshi/clai 2.0.21 → 2.0.23

package/README.md

@@ -106,7 +106,7 @@ classic REPL automatically.
 
 - **`/ask` mode** — Read-only. AI explains, gives commands & step-by-step guidance, but does NOT execute anything.
 - **`/agent` mode** — Agentic. AI plans, waits for approval, then executes shell commands, edits files, installs missing tools, parses output, and continues until the goal is met. Tasks run on an approve/refine/discard plan workflow (`/implement`, free-text to refine, `/discard` to cancel).
-- **9 LLM providers** — Groq, Google Gemini, OpenRouter, OpenAI, Anthropic, NVIDIA NIM, AgentRouter, Kimchi, AWS Mantle, and Ollama (local). All with streaming.
+- **11 LLM providers** — Groq, Google Gemini, OpenRouter, OpenAI, Anthropic, NVIDIA NIM, AgentRouter, Kimchi, AWS Mantle, Ollama (local), and Bynara. All with streaming.
 - **10 built-in tools** — `shell.exec`, `fs.read`, `fs.write`, `fs.list`, `fs.search`, `pkg.install`, `net.scan`, `http.fetch`, `sysinfo`, `pentest.recon`.
 - **Smart safety gate** — Read-only commands auto-execute; mutating commands require confirmation; destructive patterns are blocked.
 - **OS-aware & tool-frugal** — Picks the best approach for your OS, prefers tools already installed (installs only when nothing suitable exists), broadens its approach and escalates privileges as needed to finish the task.
@@ -122,7 +122,7 @@ classic REPL automatically.
 
 ## Provider Setup
 
-clai supports 10 LLM providers (8 with free tiers):
+clai supports 11 LLM providers (9 with free tiers):
 
 | Provider    | Default Model                                | Free? | API Key Prefix |
 |-------------|----------------------------------------------|-------|----------------|
@@ -136,6 +136,7 @@ clai supports 10 LLM providers (8 with free tiers):
 | Kimchi      | `kimi-k2.6`                                  | ✓     | (any)          |
 | AWS Mantle  | `anthropic.claude-haiku-4-5`                 | —     | `sk-ant-`      |
 | Ollama      | `llama3.1:8b`                                | ✓     | (local URL)    |
+| Bynara      | `mimo-v2.5-free`                             | ✓     | `sk_nry_`      |
 
 ```sh
 # Store an API key
@@ -192,6 +193,7 @@ export OLLAMA_HOST=http://localhost:11434
 | `/set <provider> [key]` | Store API key (masked input if key omitted)        |
 | `/unset <provider>`     | Remove stored key                                  |
 | `/keys`                 | List configured providers, masked                  |
+| `/info [provider]`      | Show details & usage plan for a provider           |
 | `/variants [on|off|low|medium|high]` | Toggle model thinking/reasoning variants |
 | `/think`                | Show hidden thinking from last response            |
 | `/output [last|id|list]`| Toggle full saved tool output                      |

package/dist/agent/loop-guard.d.ts

@@ -46,6 +46,20 @@ export declare class LoopGuard {
      * (e.g., command not found → retry → not found → ...).
      */
     hasRepeatedFailures(threshold?: number): boolean;
+    /**
+     * Count consecutive failures trailing the most recent attempts.
+     * Stops at the first success (or the start of the history).
+     */
+    consecutiveFailureCount(): number;
+    /**
+     * Returns a reflection prompt when recent failures suggest the agent may
+     * be stuck, or null if everything looks fine.
+     *
+     * Unlike hardcoded thresholds, this provides context for the MODEL to
+     * decide whether to continue (lengthy but progressing approach) or
+     * switch/stop (genuinely stuck approach).
+     */
+    getFailureReflection(): string | null;
     /**
      * Reset counters for read-only tools. Called after context compaction
      * so the model can re-fetch data whose results were compacted away.

package/dist/agent/loop-guard.js

@@ -118,6 +118,56 @@ export class LoopGuard {
         const recent = this.attempts.slice(-threshold);
         return recent.every((a) => !a.ok);
     }
+    /**
+     * Count consecutive failures trailing the most recent attempts.
+     * Stops at the first success (or the start of the history).
+     */
+    consecutiveFailureCount() {
+        let count = 0;
+        for (let i = this.attempts.length - 1; i >= 0; i--) {
+            if (!this.attempts[i].ok)
+                count++;
+            else
+                break;
+        }
+        return count;
+    }
+    /**
+     * Returns a reflection prompt when recent failures suggest the agent may
+     * be stuck, or null if everything looks fine.
+     *
+     * Unlike hardcoded thresholds, this provides context for the MODEL to
+     * decide whether to continue (lengthy but progressing approach) or
+     * switch/stop (genuinely stuck approach).
+     */
+    getFailureReflection() {
+        const consecutiveFailures = this.consecutiveFailureCount();
+        if (consecutiveFailures < 3)
+            return null;
+        // Build context: what tools failed and what they were trying
+        const recentFails = this.attempts.slice(-consecutiveFailures);
+        const toolSummary = recentFails
+            .map((a) => `  - ${a.callName}: ${a.canonicalSignature.slice(0, 120)}`)
+            .slice(-5) // Show last 5 max
+            .join("\n");
+        const severity = consecutiveFailures >= 6 ? "CRITICAL" : "WARNING";
+        return `⚠ APPROACH EVALUATION REQUIRED (${severity}) — ${consecutiveFailures} consecutive tool calls have FAILED.
+
+Recent failures:
+${toolSummary}
+
+You MUST now pause and evaluate:
+1. Are these failures all related to the SAME approach/method? If yes, this approach may not be viable.
+2. Is there a DIFFERENT approach that could work? (different tool, different protocol, different technique)
+3. Or is this a series of unrelated small issues that are being fixed incrementally?
+
+DECIDE one of:
+- CONTINUE: if you're making real progress and each failure teaches you something new
+- SWITCH: describe the new approach before trying it
+- STOP: if you've exhausted viable approaches — tell the user honestly what you tried and why it didn't work
+
+Do NOT keep trying variations of the same failing approach without explicitly deciding to SWITCH or STOP first.`;
+    }
     /**
      * Reset counters for read-only tools. Called after context compaction
      * so the model can re-fetch data whose results were compacted away.

package/dist/agent/loop-guard.js.map

@@ -1 +1 @@
-{"version":3,"file":"loop-guard.js","sourceRoot":"","sources":["../../src/agent/loop-guard.ts"],"names":[],"mappings":"AAUA;;;;;GAKG;AACH,MAAM,eAAe,GAAG,IAAI,GAAG,CAAC;IAC9B,WAAW;IACX,YAAY;IACZ,YAAY;IACZ,YAAY;IACZ,cAAc;IACd,SAAS;IACT,SAAS;IACT,WAAW;IACX,SAAS;IACT,aAAa;IACb,YAAY;IACZ,WAAW;IACX,UAAU;CACX,CAAC,CAAC;AAEH;;;GAGG;AACH,MAAM,OAAO,SAAS;IACZ,QAAQ,GAAkB,EAAE,CAAC;IAC7B,cAAc,GAAG,IAAI,GAAG,EAAkB,CAAC;IAC3C,gBAAgB,GAAG,IAAI,GAAG,EAAmB,CAAC;IAEtD;;;;OAIG;IACH,YAAY,CAAC,IAAY,EAAE,IAA6B;QACtD,MAAM,MAAM,GAA4B,EAAE,CAAC;QAC3C,KAAK,MAAM,GAAG,IAAI,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;YAC3C,IAAI,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC;YACtB,8CAA8C;YAC9C,IAAI,IAAI,KAAK,YAAY,IAAI,GAAG,KAAK,SAAS,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE,CAAC;gBAC5E,KAAK,GAAG,KAAK,CAAC,IAAI,EAAE,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC;YAC5C,CAAC;YACD,MAAM,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC;QACtB,CAAC;QACD,OAAO,GAAG,IAAI,KAAK,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,EAAE,CAAC;IAC9C,CAAC;IAED,aAAa,CACX,IAAY,EACZ,IAAY,EACZ,IAA6B,EAC7B,EAAW,EACX,QAA6B;QAE7B,MAAM,GAAG,GAAG,IAAI,CAAC,YAAY,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;QAC1C,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,kBAAkB,EAAE,GAAG,EAAE,EAAE,EAAE,QAAQ,EAAE,CAAC,CAAC;QACpF,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QACtE,wEAAwE;QACxE,yEAAyE;QACzE,qEAAqE;QACrE,kBAAkB;QAClB,IAAI,EAAE;YAAE,IAAI,CAAC,gBAAgB,CAAC,GAAG,CAAC,GAAG,EAAE,IAAI,CAAC,CAAC;aACxC,IAAI,CAAC,IAAI,CAAC,gBAAgB,CAAC,GAAG,CAAC,GAAG,CAAC;YAAE,IAAI,CAAC,gBAAgB,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC;IAClF,CAAC;IAED;;;;;;;;;;;;;;;OAeG;IACH,WAAW,CACT,IAAY,EACZ,IAA6B;QAE7B,MAAM,GAAG,GAAG,IAAI,CAAC,YAAY,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;QAC1C,MAAM,KAAK,GAAG,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QAEhD,IAAI,KAAK,KAAK,CAAC;YAAE,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC;QAEzC,2DAA2D;QAC3D,IAAI,IAAI,CAAC,gBAAgB,CAAC,GAAG,CAAC,GAAG,CAAC,KAAK,KAAK;YAAE,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC;QAEtE,wEAAwE;QACxE,kEAAkE;QAClE,sEAAsE;QACtE,MAAM,OAAO,GACX,IAAI,KAAK,UAAU,IAAI,IAAI,KAAK,cAAc,IAAI,IAAI,KAAK,SAAS,CAAC;QAEvE,oEAAoE;QACpE,0DAA0D;QAC1D,MAAM,SAAS,GAAG,eAAe,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAEpD,IAAI,KAAK,GAAG,SAAS,EAAE,CAAC;YACtB,OAAO;gBACL,KAAK,EAAE,KAAK;gBACZ,MAAM,EAAE,OAAO;oBACb,CAAC,CAAC,GAAG,IAAI,8HAA8H;oBACvI,CAAC,CAAC,GAAG,IAAI,gHAAgH;aAC5H,CAAC;QACJ,CAAC;QAED,qDAAqD;QACrD,OAAO;YACL,KAAK,EAAE,IAAI;YACX,MAAM,EAAE,OAAO;gBACb,CAAC,CAAC,GAAG,IAAI,uBAAuB,KAAK,gJAAgJ;gBACrL,CAAC,CAAC,GAAG,IAAI,uBAAuB,KAAK,0HAA0H;SAClK,CAAC;IACJ,CAAC;IAED,eAAe,CAAC,IAAY,EAAE,IAA6B;QACzD,MAAM,GAAG,GAAG,IAAI,CAAC,YAAY,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;QAC1C,OAAO,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;IAC3C,CAAC;IAED;;;OAGG;IACH,mBAAmB,CAAC,SAAS,GAAG,CAAC;QAC/B,IAAI,IAAI,CAAC,QAAQ,CAAC,MAAM,GAAG,SAAS;YAAE,OAAO,KAAK,CAAC;QACnD,MAAM,MAAM,GAAG,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,SAAS,CAAC,CAAC;QAC/C,OAAO,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IACpC,CAAC;IAED;;;OAGG;IACH,aAAa;QACX,KAAK,MAAM,GAAG,IAAI,CAAC,GAAG,IAAI,CAAC,cAAc,CAAC,IAAI,EAAE,CAAC,EAAE,CAAC;YAClD,MAAM,IAAI,GAAG,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;YACtC,IAAI,eAAe,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC;gBAC9B,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;gBAChC,IAAI,CAAC,gBAAgB,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;YACpC,CAAC;QACH,CAAC;QACD,IAAI,CAAC,QAAQ,GAAG,IAAI,CAAC,QAAQ,CAAC,MAAM,CAClC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,eAAe,CAAC,GAAG,CAAC,CAAC,CAAC,QAAQ,CAAC,CACxC,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,IAAI,aAAa;QACf,OAAO,IAAI,CAAC,QAAQ,CAAC,MAAM,CAAC;IAC9B,CAAC;CACF"}
+{"version":3,"file":"loop-guard.js","sourceRoot":"","sources":["../../src/agent/loop-guard.ts"],"names":[],"mappings":"AAUA;;;;;GAKG;AACH,MAAM,eAAe,GAAG,IAAI,GAAG,CAAC;IAC9B,WAAW;IACX,YAAY;IACZ,YAAY;IACZ,YAAY;IACZ,cAAc;IACd,SAAS;IACT,SAAS;IACT,WAAW;IACX,SAAS;IACT,aAAa;IACb,YAAY;IACZ,WAAW;IACX,UAAU;CACX,CAAC,CAAC;AAEH;;;GAGG;AACH,MAAM,OAAO,SAAS;IACZ,QAAQ,GAAkB,EAAE,CAAC;IAC7B,cAAc,GAAG,IAAI,GAAG,EAAkB,CAAC;IAC3C,gBAAgB,GAAG,IAAI,GAAG,EAAmB,CAAC;IAEtD;;;;OAIG;IACH,YAAY,CAAC,IAAY,EAAE,IAA6B;QACtD,MAAM,MAAM,GAA4B,EAAE,CAAC;QAC3C,KAAK,MAAM,GAAG,IAAI,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC;YAC3C,IAAI,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC;YACtB,8CAA8C;YAC9C,IAAI,IAAI,KAAK,YAAY,IAAI,GAAG,KAAK,SAAS,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE,CAAC;gBAC5E,KAAK,GAAG,KAAK,CAAC,IAAI,EAAE,CAAC,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC;YAC5C,CAAC;YACD,MAAM,CAAC,GAAG,CAAC,GAAG,KAAK,CAAC;QACtB,CAAC;QACD,OAAO,GAAG,IAAI,KAAK,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,EAAE,CAAC;IAC9C,CAAC;IAED,aAAa,CACX,IAAY,EACZ,IAAY,EACZ,IAA6B,EAC7B,EAAW,EACX,QAA6B;QAE7B,MAAM,GAAG,GAAG,IAAI,CAAC,YAAY,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;QAC1C,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,kBAAkB,EAAE,GAAG,EAAE,EAAE,EAAE,QAAQ,EAAE,CAAC,CAAC;QACpF,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QACtE,wEAAwE;QACxE,yEAAyE;QACzE,qEAAqE;QACrE,kBAAkB;QAClB,IAAI,EAAE;YAAE,IAAI,CAAC,gBAAgB,CAAC,GAAG,CAAC,GAAG,EAAE,IAAI,CAAC,CAAC;aACxC,IAAI,CAAC,IAAI,CAAC,gBAAgB,CAAC,GAAG,CAAC,GAAG,CAAC;YAAE,IAAI,CAAC,gBAAgB,CAAC,GAAG,CAAC,GAAG,EAAE,KAAK,CAAC,CAAC;IAClF,CAAC;IAED;;;;;;;;;;;;;;;OAeG;IACH,WAAW,CACT,IAAY,EACZ,IAA6B;QAE7B,MAAM,GAAG,GAAG,IAAI,CAAC,YAAY,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;QAC1C,MAAM,KAAK,GAAG,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QAEhD,IAAI,KAAK,KAAK,CAAC;YAAE,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC;QAEzC,2DAA2D;QAC3D,IAAI,IAAI,CAAC,gBAAgB,CAAC,GAAG,CAAC,GAAG,CAAC,KAAK,KAAK;YAAE,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC;QAEtE,wEAAwE;QACxE,kEAAkE;QAClE,sEAAsE;QACtE,MAAM,OAAO,GACX,IAAI,KAAK,UAAU,IAAI,IAAI,KAAK,cAAc,IAAI,IAAI,KAAK,SAAS,CAAC;QAEvE,oEAAoE;QACpE,0DAA0D;QAC1D,MAAM,SAAS,GAAG,eAAe,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAEpD,IAAI,KAAK,GAAG,SAAS,EAAE,CAAC;YACtB,OAAO;gBACL,KAAK,EAAE,KAAK;gBACZ,MAAM,EAAE,OAAO;oBACb,CAAC,CAAC,GAAG,IAAI,8HAA8H;oBACvI,CAAC,CAAC,GAAG,IAAI,gHAAgH;aAC5H,CAAC;QACJ,CAAC;QAED,qDAAqD;QACrD,OAAO;YACL,KAAK,EAAE,IAAI;YACX,MAAM,EAAE,OAAO;gBACb,CAAC,CAAC,GAAG,IAAI,uBAAuB,KAAK,gJAAgJ;gBACrL,CAAC,CAAC,GAAG,IAAI,uBAAuB,KAAK,0HAA0H;SAClK,CAAC;IACJ,CAAC;IAED,eAAe,CAAC,IAAY,EAAE,IAA6B;QACzD,MAAM,GAAG,GAAG,IAAI,CAAC,YAAY,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;QAC1C,OAAO,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;IAC3C,CAAC;IAED;;;OAGG;IACH,mBAAmB,CAAC,SAAS,GAAG,CAAC;QAC/B,IAAI,IAAI,CAAC,QAAQ,CAAC,MAAM,GAAG,SAAS;YAAE,OAAO,KAAK,CAAC;QACnD,MAAM,MAAM,GAAG,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,SAAS,CAAC,CAAC;QAC/C,OAAO,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IACpC,CAAC;IAED;;;OAGG;IACH,uBAAuB;QACrB,IAAI,KAAK,GAAG,CAAC,CAAC;QACd,KAAK,IAAI,CAAC,GAAG,IAAI,CAAC,QAAQ,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YACnD,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAE,CAAC,EAAE;gBAAE,KAAK,EAAE,CAAC;;gBAC9B,MAAM;QACb,CAAC;QACD,OAAO,KAAK,CAAC;IACf,CAAC;IAED;;;;;;;OAOG;IACH,oBAAoB;QAClB,MAAM,mBAAmB,GAAG,IAAI,CAAC,uBAAuB,EAAE,CAAC;QAC3D,IAAI,mBAAmB,GAAG,CAAC;YAAE,OAAO,IAAI,CAAC;QAEzC,6DAA6D;QAC7D,MAAM,WAAW,GAAG,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,mBAAmB,CAAC,CAAC;QAC9D,MAAM,WAAW,GAAG,WAAW;aAC5B,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,CAAC,QAAQ,KAAK,CAAC,CAAC,kBAAkB,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,CAAC;aACtE,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,kBAAkB;aAC5B,IAAI,CAAC,IAAI,CAAC,CAAC;QAEd,MAAM,QAAQ,GAAG,mBAAmB,IAAI,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,SAAS,CAAC;QAEnE,OAAO,mCAAmC,QAAQ,OAAO,mBAAmB;;;EAG9E,WAAW;;;;;;;;;;;;gHAYmG,CAAC;IAC/G,CAAC;IAED;;;OAGG;IACH,aAAa;QACX,KAAK,MAAM,GAAG,IAAI,CAAC,GAAG,IAAI,CAAC,cAAc,CAAC,IAAI,EAAE,CAAC,EAAE,CAAC;YAClD,MAAM,IAAI,GAAG,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;YACtC,IAAI,eAAe,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC;gBAC9B,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;gBAChC,IAAI,CAAC,gBAAgB,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;YACpC,CAAC;QACH,CAAC;QACD,IAAI,CAAC,QAAQ,GAAG,IAAI,CAAC,QAAQ,CAAC,MAAM,CAClC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,eAAe,CAAC,GAAG,CAAC,CAAC,CAAC,QAAQ,CAAC,CACxC,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,IAAI,aAAa;QACf,OAAO,IAAI,CAAC,QAAQ,CAAC,MAAM,CAAC;IAC9B,CAAC;CACF"}

package/dist/agent/runner.d.ts

@@ -39,6 +39,7 @@ export interface AgentRunOptions {
     }) => Promise<string | undefined>) | undefined;
     session?: SessionPolicy | undefined;
 }
+export declare function preprocessJson(raw: string): string;
 export interface ParseToolCallOptions {
     /**
      * When true, only formats that are explicitly tool-call delimited are

package/dist/agent/runner.js

@@ -44,9 +44,58 @@ export function createSessionPolicy() {
         planApproved: { value: false },
     };
 }
+export function preprocessJson(raw) {
+    let inString = false;
+    let escaped = false;
+    let result = "";
+    for (let i = 0; i < raw.length; i++) {
+        const char = raw[i];
+        if (char === '"' && !escaped) {
+            inString = !inString;
+            result += char;
+        }
+        else if (inString) {
+            if (char === '\n') {
+                result += '\\n';
+            }
+            else if (char === '\r') {
+                result += '\\r';
+            }
+            else if (char === '\t') {
+                result += '\\t';
+            }
+            else {
+                result += char;
+            }
+        }
+        else {
+            if (char === ',' && i + 1 < raw.length) {
+                let nextNonWs = "";
+                for (let j = i + 1; j < raw.length; j++) {
+                    if (!/\s/.test(raw[j])) {
+                        nextNonWs = raw[j];
+                        break;
+                    }
+                }
+                if (nextNonWs === '}' || nextNonWs === ']') {
+                    continue;
+                }
+            }
+            result += char;
+        }
+        if (char === '\\' && inString) {
+            escaped = !escaped;
+        }
+        else {
+            escaped = false;
+        }
+    }
+    return result;
+}
 function tryParseCall(raw) {
     try {
-        const parsed = JSON.parse(raw.trim());
+        const preprocessed = preprocessJson(raw);
+        const parsed = JSON.parse(preprocessed.trim());
         if (typeof parsed.name === "string" &&
             parsed.args &&
             typeof parsed.args === "object") {
@@ -78,9 +127,84 @@ function parseKimiToolCall(text) {
     const name = match[1];
     return tryParseCall(JSON.stringify({ name, args: tryJson(match[2]) ?? {} }));
 }
+function parseXmlToolCall(text) {
+    // Pattern 1:
+    // <tool_call>
+    // <name>tool.name</name>
+    // <args>{...}</args>
+    // </tool_call>
+    const xmlNameArgs = text.match(/<tool_call>[\s\S]*?<name>\s*([\w.]+?)\s*<\/name>\s*<args>\s*(\{[\s\S]*?\})\s*<\/args>[\s\S]*?<\/tool_call>/i);
+    if (xmlNameArgs?.[1] && xmlNameArgs?.[2]) {
+        try {
+            const args = JSON.parse(preprocessJson(xmlNameArgs[2]));
+            return {
+                name: xmlNameArgs[1],
+                args: args,
+            };
+        }
+        catch { }
+    }
+    // Pattern 1b (MiMo alternative):
+    // <tool_call>
+    // <tool_name>tool.name</tool_name>
+    // <parameters>{...}</parameters>
+    // </tool_call>
+    const xmlToolNameParams = text.match(/<tool_call>[\s\S]*?<tool_name>\s*([\w.]+?)\s*<\/tool_name>\s*<parameters>\s*(\{[\s\S]*?\})\s*<\/parameters>[\s\S]*?<\/tool_call>/i);
+    if (xmlToolNameParams?.[1] && xmlToolNameParams?.[2]) {
+        try {
+            const args = JSON.parse(preprocessJson(xmlToolNameParams[2]));
+            return {
+                name: xmlToolNameParams[1],
+                args: args,
+            };
+        }
+        catch { }
+    }
+    // Pattern 1c (MiMo function/parameter format):
+    // <tool_call>
+    // <function=tool.name>
+    // <parameter=name>value</parameter>
+    // </function>
+    // </tool_call>
+    const xmlFunctionBlock = text.match(/<tool_call>[\s\S]*?<function=([\w.]+?)>([\s\S]*?)<\/function>[\s\S]*?<\/tool_call>/i);
+    if (xmlFunctionBlock?.[1] && xmlFunctionBlock?.[2]) {
+        const name = xmlFunctionBlock[1];
+        const inner = xmlFunctionBlock[2];
+        const args = {};
+        const paramRegex = /<parameter=([\w.]+?)>([\s\S]*?)<\/parameter>/gi;
+        let paramMatch;
+        while ((paramMatch = paramRegex.exec(inner)) !== null) {
+            const paramName = paramMatch[1];
+            const paramValueStr = paramMatch[2].trim();
+            let paramValue = paramValueStr;
+            try {
+                if (/^(?:\[|\{|true|false|null|\d+(\.\d+)?$)/i.test(paramValueStr)) {
+                    paramValue = JSON.parse(preprocessJson(paramValueStr));
+                }
+            }
+            catch { }
+            args[paramName] = paramValue;
+        }
+        return { name, args };
+    }
+    // Pattern 2:
+    // <tool_call>
+    // <tool>
+    // {"name": "...", "args": {...}}
+    // </tool_call>
+    // or simply <tool_call> {"name": "...", "args": {...}} </tool_call>
+    const xmlJson = text.match(/<tool_call>[\s\S]*?(?:<tool>)?\s*(\{[\s\S]*?\})\s*<\/tool_call>/i);
+    if (xmlJson?.[1]) {
+        const call = tryParseCall(xmlJson[1]);
+        if (call)
+            return call;
+    }
+    return undefined;
+}
 function tryJson(raw) {
     try {
-        const parsed = JSON.parse(raw);
+        const preprocessed = preprocessJson(raw);
+        const parsed = JSON.parse(preprocessed);
         if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
             return parsed;
         }
@@ -110,13 +234,10 @@ export function parseToolCall(text, options = {}) {
         if (call)
             return call;
     }
-    // 2. <tool_call>...</tool_call>
-    const xml = text.match(/<tool_call>([\s\S]*?)<\/tool_call>/i);
-    if (xml?.[1]) {
-        const call = tryParseCall(xml[1]);
-        if (call)
-            return call;
-    }
+    // 2. <tool_call>...</tool_call> (XML formats)
+    const xmlCall = parseXmlToolCall(text);
+    if (xmlCall)
+        return xmlCall;
     // 3. Kimi/Moonshot sentinel format (used by kimi-k2 family on NIM).
     const kimi = parseKimiToolCall(text);
     if (kimi)
@@ -437,7 +558,7 @@ export function parseAllToolCalls(text) {
     }
     const xmlRe = /<tool_call>([\s\S]*?)<\/tool_call>/gi;
     while ((m = xmlRe.exec(text)) !== null) {
-        const call = tryParseCall(m[1] ?? "");
+        const call = parseXmlToolCall(m[0]);
         if (call)
             found.push({ index: m.index, call });
     }
@@ -481,16 +602,16 @@ export function collapseRepeatedText(text) {
 /** Extract the text before the tool call block for display purposes */
 function textBeforeToolCall(text) {
     const patterns = [
-        /```tool\s*\n?[\s\S]*?```/i,
-        /<tool_call>[\s\S]*?<\/tool_call>/i,
+        /```tool\s*\n?[\s\S]*$/i,
+        /<tool_call>[\s\S]*$/i,
         // Kimi/Moonshot sentinel block — strip from the section opener
         // (or the first call opener if the section header is missing).
         /<\|tool_calls_section_begin\|>[\s\S]*$/i,
         /<\|tool_call_begin\|>[\s\S]*$/i,
-        /#{1,3}\s*tool\s*\n\s*\{[\s\S]*\}/i,
-        /\*\*tool\*\*\s*\n\s*\{[\s\S]*\}/i,
-        /```\w*\s*\n?\{[\s\S]*?"name"[\s\S]*?\}[\s\S]*?```/,
-        /\{"name"\s*:\s*"[^"]+"\s*,\s*"args"\s*:\s*\{[\s\S]*?\}\s*\}\s*$/,
+        /#{1,3}\s*tool\s*\n\s*\{[\s\S]*$/i,
+        /\*\*tool\*\*\s*\n\s*\{[\s\S]*$/i,
+        /```\w*\s*\n?\{[\s\S]*?"name"[\s\S]*$/i,
+        /\{"name"\s*:\s*"[^"]+"\s*,\s*"args"\s*:\s*\{[\s\S]*$/i,
     ];
     for (const pattern of patterns) {
         const idx = text.search(pattern);
@@ -638,7 +759,7 @@ export function looksLikeBuildTask(prompt, history) {
 }
 // Matrix of action-verb narration: the model says it is *about to* do
 // something but hasn't. Used to detect "narrate, don't act" stalls.
-const ACTION_NARRATION_RE = /\b(?:let me|let's|i'?ll|i will|i'?m going to|i am going to|i need to|i should|i'?m about to|going to|now i'?ll|first[,]?\s*i'?ll)\s+(?:now\s+|first\s+|quickly\s+|just\s+|go\s+ahead\s+and\s+)?(?:explore|list|read|fetch|browse|check|inspect|examine|look|create|run|start|write|build|add|scaffold|set\s*up|setup|install|initialize|init|generate|make|review|open|find|search|verify|update|edit|modify|fix|implement)\b/i;
+const ACTION_NARRATION_RE = /\b(?:let me|let's|i'?ll|i will|i'?m going to|i am going to|i need to|i should|i'?m about to|going to|now i'?ll|first[,]?\s*i'?ll)\s+(?:now\s+|first\s+|quickly\s+|just\s+|go\s+ahead\s+and\s+)?(?:explore|list|read|fetch|browse|check|inspect|examine|look|create|run|start|write|build|add|scaffold|set\s*up|setup|install|initialize|init|generate|make|review|open|find|search|verify|update|edit|modify|fix|implement|gather|assess|scan|audit)\b/i;
 /**
  * Detect a message that narrates an *upcoming* action ("let me explore the
  * directory", "I'll create the components") rather than an actual answer or
@@ -1567,6 +1688,12 @@ export async function runAgentLoop(prompt, options = {}) {
                 artifactPath: "",
             };
             jobManager.registerJob(jobId, backgroundJob, toolAc);
+            const TOOL_STALL_WARNING_MS = 60_000; // 1 minute
+            const stallTimer = setTimeout(() => {
+                if (!toolAc.signal.aborted) {
+                    writeNotice("info", `${call.name} has been running for >60s — still waiting (ESC to abort)`, chalk.yellow(`  ⏳ ${call.name} still running — ESC to abort\n`));
+                }
+            }, TOOL_STALL_WARNING_MS);
             try {
                 result = await runToolCall(call, {
                     signal: toolAc.signal,
@@ -1577,6 +1704,7 @@ export async function runAgentLoop(prompt, options = {}) {
                         printLive(chunk);
                     },
                     confirmed: true,
+                    userPrompt: prompt,
                 });
                 if (liveBytes > 0 || liveTruncatedNotified) {
                     writeToolOutput(toolEventId, "\n", "\n");
@@ -1593,6 +1721,7 @@ export async function runAgentLoop(prompt, options = {}) {
                 result = { ok: false, output: `Tool error: ${errMsg}`, exitCode: 1 };
             }
             finally {
+                clearTimeout(stallTimer);
                 parentSignal.removeEventListener("abort", onParentAbort);
             }
             const output = result.output.trim();
@@ -1622,6 +1751,18 @@ export async function runAgentLoop(prompt, options = {}) {
                 output: result.output.slice(0, 4_000),
             });
             loopGuard.recordAttempt(step, call.name, call.args, result.ok, result.exitCode);
+            // Inject approach evaluation when consecutive failures are detected.
+            // Lets the MODEL decide (with full context) whether to continue a
+            // legitimately long approach, switch, or stop — instead of a
+            // hardcoded kill threshold.
+            if (!result.ok) {
+                const reflection = loopGuard.getFailureReflection();
+                if (reflection) {
+                    messages.push({ role: "system", content: reflection });
+                    const failCount = loopGuard.consecutiveFailureCount();
+                    writeNotice("warn", `${failCount} consecutive failures — model evaluating approach`, chalk.yellow(`  ⚠ ${failCount} consecutive failures — evaluating approach\n`));
+                }
+            }
             const statusIcon = result.ok ? chalk.green("  ✓") : chalk.red("  ✗");
             writeToolOutput(toolEventId, result.ok ? "ok\n" : "failed\n", statusIcon + "\n");
             if (output) {
@@ -1824,16 +1965,24 @@ export async function runAgentLoop(prompt, options = {}) {
                         writeNotice("info", "recovered tool call from thinking content", chalk.dim("  ℹ recovered tool call from thinking content\n"));
                     }
                 }
-                // ── Thinking-only recovery ────────────────────────────────────────
-                // Some models (eg gpt-oss-20b on NVIDIA NIM) occasionally spend their
-                // entire budget on hidden <think> reasoning and emit no visible text
-                // or tool call. Without this guard the agent silently returns an empty
-                // answer and the user has to re-submit the same prompt.
-                if (!assistantText.visible.trim() && !call && assistantText.hasThinking) {
+                // ── Empty-response recovery ───────────────────────────────────────
+                // Some models occasionally return an empty completion: a reasoning
+                // model that spent its whole budget on hidden &lt;think&gt; reasoning and emitted
+                // no visible text, OR (more perniciously) a gateway hiccup that
+                // streamed [DONE] with no content deltas at all. Without this guard
+                // the agent silently ends the turn with no answer, no warning, and no
+                // error — the user just sees the spinner stop. Catch BOTH cases
+                // (thinking-only AND truly empty) and nudge the model to retry.
+                if (!assistantText.visible.trim() && !call) {
                     emptyVisibleRetries += 1;
                     if (emptyVisibleRetries <= 3) {
-                        writeThinkingBlock(assistantText.thinkContent);
-                        writeNotice("warn", "model produced only thinking — nudging it to take action", chalk.yellow("  ⚠ model produced only thinking — nudging it to take action\n"));
+                        if (assistantText.hasThinking) {
+                            writeThinkingBlock(assistantText.thinkContent);
+                            writeNotice("warn", "model produced only thinking — nudging it to take action", chalk.yellow("  ⚠ model produced only thinking — nudging it to take action\n"));
+                        }
+                        else {
+                            writeNotice("warn", "model returned an empty response — nudging it to answer", chalk.yellow("  ⚠ model returned an empty response — nudging it to answer\n"));
+                        }
                         messages.push({
                             role: "assistant",
                             content: collapseRepeatedText(completion.text),
@@ -1847,8 +1996,10 @@ export async function runAgentLoop(prompt, options = {}) {
                         messages.push(recoveryUserMessage(buildNudge));
                         continue;
                     }
-                    // Exhausted retries — fall through to the normal empty-answer path
-                    // which will print a warning and return.
+                    // Exhausted retries — surface a clear notice instead of ending the
+                    // turn silently with no answer (which left the user staring at a
+                    // stopped spinner with no clue what happened).
+                    writeNotice("warn", "model returned an empty response after retries — no answer produced", chalk.yellow("  ⚠ model returned an empty response after retries — no answer produced\n"));
                 }
                 else {
                     // Reset the counter on any successful visible output or recovered call.
@@ -1979,6 +2130,7 @@ export async function runAgentLoop(prompt, options = {}) {
                     // Nudge it to emit a real tool call, with a concrete example.
                     const narratedAction = looksLikeActionNarration(cleaned);
                     const wantsAction = buildLikeTurn ||
+                        pentestLikeTurn ||
                         (activePlan && session.planApproved.value) ||
                         freshWebSearchRequired ||
                         narratedAction;
@@ -1996,6 +2148,13 @@ export async function runAgentLoop(prompt, options = {}) {
                                 "You wrote a message but emitted NO ```tool block, so NOTHING ran. Do NOT narrate what you will do — DO it. Emit the next tool call now (task.update / fs.writeMany / shell.exec) in a single ```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"));
                         }
+                        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 (e.g. net.scan / sysinfo / shell.exec). For example, to scan local network or read system settings:\n" +
+                                    '```tool\n{"name":"sysinfo","args":{}}\n```\n' +
+                                    "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) {
                             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" +
@@ -2107,14 +2266,16 @@ export async function runAgentLoop(prompt, options = {}) {
                 if (assistantText.hasThinking) {
                     writeThinkingBlock(assistantText.thinkContent);
                 }
-                messages.push({
-                    role: "assistant",
-                    content: collapseRepeatedText(assistantText.visible),
-                });
                 let allCalls = parseAllToolCalls(assistantText.visible || assistantText.thinkContent);
                 if (allCalls.length === 0 && call) {
                     allCalls = [call];
                 }
+                const standardizedContent = (beforeTool ? beforeTool.trim() + "\n\n" : "") +
+                    allCalls.map(c => `\`\`\`tool\n${JSON.stringify(c)}\n\`\`\``).join("\n\n");
+                messages.push({
+                    role: "assistant",
+                    content: standardizedContent,
+                });
                 if (allCalls.length > 1) {
                     writeNotice("info", `${allCalls.length} tool calls in this message — running them in parallel`, chalk.dim(`  ℹ ${allCalls.length} tool calls in this message — running them in parallel\n`));
                 }