@livx.cc/agentx 0.99.79 → 0.99.81

package/dist/{Agent-BK-heHrw.d.ts → Agent-RHsiEAiP.d.ts}

@@ -307,6 +307,11 @@ interface RunResult {
         totalTokens: number;
         cacheCreationTokens?: number;
         cacheReadTokens?: number;
+        /** What the context holds after the run: the LAST model call's prompt (a self-looping backend's own figure —
+         *  claude-code sums its calls into promptTokens) — and the window it is measured against (the backend's
+         *  reported one, else `Agent.contextCap`). context % = contextTokens / contextWindow. */
+        contextTokens?: number;
+        contextWindow?: number;
     };
     /** Present iff `finishReason === 'budget'`: the weighted spend vs `maxTokens`, and the last step's prompt
      *  size — a prompt near the limit means the CONVERSATION is too big for one run (compact / new chat). */
@@ -669,8 +674,9 @@ declare class Agent {
      * retained window → halve the window → drop-oldest backstop. Deterministic, no LLM call.
      * `focus` (e.g. `/compact keep the API details`) preserves matching lines from the dropped span verbatim.
      * Returns before/after token estimates so the caller can report honestly (0 freed => say so).
+     * `target` (tokens): a tighter budget than half the window — a host compacting at its own threshold.
      */
-    compactNow(maxMessages?: number, focus?: string): CompactResult;
+    compactNow(maxMessages?: number, focus?: string, target?: number): CompactResult;
     /** Drop inline base64 images from older messages so Bun/JSC can reclaim the heap (see contentBytes). */
     releaseStaleImages(keepRecent?: number): void;
     private runLoop;

package/dist/cli.d.ts

@@ -1,5 +1,5 @@
 #!/usr/bin/env bun
-import { m as RouterState, H as Hooks, n as RunResult, R as ReasoningEffort, A as Agent } from './Agent-BK-heHrw.js';
+import { m as RouterState, H as Hooks, n as RunResult, R as ReasoningEffort, A as Agent } from './Agent-RHsiEAiP.js';
 import { IFilesystem } from '@livx.cc/wcli/core';
 import { M as Message, U as UserQuestion, H as HostBridge, c as ContentPart, o as MessageContent } from './tools-lC3jn60o.js';
 
@@ -331,6 +331,8 @@ declare function jsonResult(res: RunResult, session: SessionData): {
         totalTokens: number;
         cacheCreationTokens?: number;
         cacheReadTokens?: number;
+        contextTokens?: number;
+        contextWindow?: number;
     } | undefined;
     providerFinishReason?: string | undefined;
     ok: boolean;

package/dist/cli.js

@@ -5140,10 +5140,11 @@ var Agent = class _Agent {
    * retained window → halve the window → drop-oldest backstop. Deterministic, no LLM call.
    * `focus` (e.g. `/compact keep the API details`) preserves matching lines from the dropped span verbatim.
    * Returns before/after token estimates so the caller can report honestly (0 freed => say so).
+   * `target` (tokens): a tighter budget than half the window — a host compacting at its own threshold.
    */
-  compactNow(maxMessages = 12, focus) {
+  compactNow(maxMessages = 12, focus, target) {
     const cap = _Agent.contextCap(this.options);
-    const target = Math.round(cap * _Agent.COMPACT_TARGET);
+    target = Math.min(Math.round(cap * _Agent.COMPACT_TARGET), target || Infinity);
     const kr = this.options.keepRecentImages ?? 1;
     const vis2 = this.sendsImages();
     const tokensBefore = estimateTokens(this.transcript, kr, vis2);
@@ -5351,6 +5352,8 @@ var Agent = class _Agent {
         usage.cacheCreationTokens += res.usage.cacheCreationTokens ?? 0;
         usage.cacheReadTokens += res.usage.cacheReadTokens ?? 0;
         const prompt = res.usage.promptTokens ?? 0, cacheRead = res.usage.cacheReadTokens ?? 0;
+        usage.contextTokens = res.usage.contextTokens ?? prompt;
+        usage.contextWindow = res.usage.contextWindow || _Agent.contextCap(o);
         if (prevPrompt < 0) {
           const u = sent.map((m) => m.role).lastIndexOf("user");
           prevPrompt = u > 0 ? estimateTokens(sent.slice(0, u)) : 0;
@@ -20112,6 +20115,22 @@ async function main() {
     }
   });
   if (args.task) {
+    const compactArg = /^\/compact(?:\s+([\s\S]*))?$/.exec(args.task.trim());
+    if (compactArg) {
+      const agent2 = await makeAgent(args, ai, cfg, [], []), store2 = new SessionStore(cwd), session2 = startSession(args, store2, agent2, cwd);
+      const est = (m) => estimateTranscriptTokens(m, agent2.options.keepRecentImages, agent2.options.imageInput?.(agent2.options.model) !== false);
+      const fixed = est(agent2.transcript.filter((m) => m.role === "system"));
+      const r = agent2.compactNow(12, compactArg[1]?.trim() || void 0, Math.round(fixed + (est(agent2.transcript) - fixed) / 2));
+      session2.messages = agent2.transcript;
+      session2.meta.updated = Date.now();
+      store2.save(session2);
+      const obj = { ok: true, finishReason: "stop", text: "", steps: 0, compacted: r, sessionId: session2.meta.id };
+      if (args.outputFormat === "json" || args.outputFormat === "stream-json") process.stdout.write(JSON.stringify(args.outputFormat === "stream-json" ? { type: "result", ...obj } : obj) + "\n");
+      else err(dim(`  compacted ~${(r.tokensBefore / 1e3).toFixed(1)}k \u2192 ~${(r.tokensAfter / 1e3).toFixed(1)}k tok of ${Math.round(r.cap / 1e3)}k
+`));
+      worktreeCleanup?.();
+      process.exit(0);
+    }
     const mounted = await mountMcp(cfg, new McpOAuth({ storePath: join15(cwd, ".agent", "mcp-auth.json") }));
     const agent = await makeAgent(args, ai, cfg, mcpAgentTools(mounted), mounted.map((m) => m.name));
     const store = new SessionStore(cwd);