chainlesschain 0.162.94 → 0.162.96

package/src/repl/agent-repl.js

@@ -286,9 +286,19 @@ export async function startAgentRepl(options = {}) {
   // an auto-detected image path so the REPL switches to a vision-capable model
   // for that turn only (resolveVisionLlm falls back to the default when unset).
   let _visionModel;
+  // Hard stream inactivity timeout override (config.llm.streamStallTimeoutMs,
+  // ms): a stream silent that long is aborted + retried instead of hanging.
+  // Unset → agent-core's 180s default (matches cc chat/ask). Set to 0 to
+  // disable. Left undefined here means "use the default".
+  let _streamStallTimeoutMs;
   try {
     const { loadConfig } = await import("../lib/config-manager.js");
-    _visionModel = loadConfig()?.llm?.visionModel || undefined;
+    const _cfg = loadConfig();
+    _visionModel = _cfg?.llm?.visionModel || undefined;
+    const raw = _cfg?.llm?.streamStallTimeoutMs;
+    const t = Number(raw);
+    // Accept 0 (explicit disable) — only ignore absent/invalid values.
+    if (raw != null && Number.isFinite(t) && t >= 0) _streamStallTimeoutMs = t;
   } catch {
     /* optional — resolveVisionLlm falls back to DEFAULT_VISION_MODEL */
   }
@@ -721,6 +731,7 @@ export async function startAgentRepl(options = {}) {
           // an IDE integrated terminal. --ide forces it, --no-ide disables it
           // (parity with headless; auto-detect already works via process.env).
           ide: options.ide,
+          pdh: options.pdh,
           cwd: process.cwd(),
           // advertise the session id to spawned stdio MCP servers
           sessionId,
@@ -1946,6 +1957,11 @@ export async function startAgentRepl(options = {}) {
             messages.length = 1; // keep system prompt
             messages.push(...rebuilt.filter((m) => m.role !== "system"));
             sessionId = resumeId;
+            // The prior session's checkpoint marks (turn-index → checkpoint id)
+            // no longer map to this swapped-in history; keeping them would make
+            // a later /rewind restore files from the WRONG session's snapshot.
+            // Same reset /clear does on a context swap.
+            _checkpointMarks.length = 0;
             logger.info(
               `Resumed JSONL session ${sessionId} (${rebuilt.length} messages)`,
             );
@@ -1959,6 +1975,8 @@ export async function startAgentRepl(options = {}) {
               messages.length = 1; // keep system prompt
               messages.push(...parsed.filter((m) => m.role !== "system"));
               sessionId = existing.id;
+              // Drop the prior session's checkpoint marks (see JSONL branch).
+              _checkpointMarks.length = 0;
               logger.info(
                 `Resumed session ${sessionId} (${parsed.length} messages)`,
               );
@@ -2456,12 +2474,19 @@ export async function startAgentRepl(options = {}) {
           process.stdout.write(chalk.gray(`  [step] ${evt.step}\n`));
         };
 
-        autonomousAgent.on("goal:completed", goalListener);
-        autonomousAgent.on("goal:failed", goalListener);
-        autonomousAgent.on("step:started", stepListener);
-        autonomousAgent.on("step:completed", (evt) => {
-          process.stdout.write(chalk.green(`  [done] ${evt.step}\n`));
-        });
+        // Attach the live-output listeners ONCE per REPL session — the agent is
+        // a session-scoped singleton (created once above), so re-attaching on
+        // every `/auto` would accumulate listeners (duplicate output + leak +
+        // MaxListenersExceededWarning). The handlers are stateless.
+        if (!autonomousAgent._replListenersAttached) {
+          autonomousAgent._replListenersAttached = true;
+          autonomousAgent.on("goal:completed", goalListener);
+          autonomousAgent.on("goal:failed", goalListener);
+          autonomousAgent.on("step:started", stepListener);
+          autonomousAgent.on("step:completed", (evt) => {
+            process.stdout.write(chalk.green(`  [done] ${evt.step}\n`));
+          });
+        }
 
         logger.info(`Submitting goal: ${chalk.cyan(autoArg)}`);
         try {
@@ -2660,7 +2685,7 @@ export async function startAgentRepl(options = {}) {
 
     // `/sessions` — list recent RESUMABLE conversations (read-only; the ids
     // work with `cc agent --resume <id>`). `/session` shows the current one.
-    if (trimmed === "/sessions" || trimmed === "/sessions ") {
+    if (trimmed === "/sessions" || trimmed.startsWith("/sessions ")) {
       try {
         const { listRecentSessions } = await import("../lib/recent-session.js");
         const { renderRecentSessions } = await import("./recent-sessions.js");
@@ -2676,7 +2701,7 @@ export async function startAgentRepl(options = {}) {
     // `/memory` — project-memory files auto-loaded into the system prompt
     // (cc.md hierarchy + imports + path-scoped rules). Distinct from `#` (add
     // a note) and `cc memory recall` (scoped store).
-    if (trimmed === "/memory" || trimmed === "/memory ") {
+    if (trimmed === "/memory" || trimmed.startsWith("/memory ")) {
       try {
         const { loadProjectInstructions } =
           await import("../lib/project-instructions.js");
@@ -2695,7 +2720,7 @@ export async function startAgentRepl(options = {}) {
     }
 
     // `/mcp` — overview of connected MCP servers' resources + prompts.
-    if (trimmed === "/mcp" || trimmed === "/mcp ") {
+    if (trimmed === "/mcp" || trimmed.startsWith("/mcp ")) {
       const mcpClient = _adhocMcp?.mcpClient || _bundleMcpClient;
       logger.log(renderMcpSurface(mcpClient));
       prompt();
@@ -2753,7 +2778,7 @@ export async function startAgentRepl(options = {}) {
 
     // `/doctor` — consolidated session-health readout (Claude-Code parity):
     // provider/key/IDE/MCP/permissions/hooks in one pass-or-warn view.
-    if (trimmed === "/doctor" || trimmed === "/doctor ") {
+    if (trimmed === "/doctor" || trimmed.startsWith("/doctor ")) {
       let config = {};
       try {
         config = (await import("../lib/config-manager.js")).loadConfig() || {};
@@ -2860,7 +2885,7 @@ export async function startAgentRepl(options = {}) {
 
     // `/cost` — running token spend + estimated $ for this session (Claude-Code
     // parity). In-memory accumulation, so it works without session persistence.
-    if (trimmed === "/cost" || trimmed === "/cost ") {
+    if (trimmed === "/cost" || trimmed.startsWith("/cost ")) {
       let overrides;
       let visionModel;
       try {
@@ -2937,7 +2962,7 @@ export async function startAgentRepl(options = {}) {
 
     // `/ide` — IDE bridge connection status (Claude-Code parity): which editor
     // is connected, its tools, or why discovery came up empty.
-    if (trimmed === "/ide" || trimmed === "/ide ") {
+    if (trimmed === "/ide" || trimmed.startsWith("/ide ")) {
       let diag = null;
       try {
         const { diagnoseIde } = await import("../lib/ide-bridge.js");
@@ -3242,6 +3267,20 @@ export async function startAgentRepl(options = {}) {
               `  ⟳ connection dropped — retrying (attempt ${attempt})…\n`,
             ),
           ),
+        // Stream-stall hint (Claude-Code 2.1.185): the connection is alive but
+        // the API has gone silent mid-response — tell the user we're still
+        // waiting instead of leaving a frozen spinner. stderr so it never
+        // corrupts the streamed answer on stdout.
+        onStall: (ms) =>
+          process.stderr.write(
+            chalk.dim(
+              `  ⏳ waiting for API response (silent ${Math.round(ms / 1000)}s)…\n`,
+            ),
+          ),
+        // Hard inactivity timeout: abort + retry a dead-but-open stream instead
+        // of hanging forever. undefined → agent-core's 180s default (matches cc
+        // chat/ask); config.llm.streamStallTimeoutMs tunes or disables (0).
+        streamStallTimeoutMs: _streamStallTimeoutMs,
         signal: _turnAbort.signal,
         // On an auto-detected image turn, switch to the vision LLM for this
         // turn only (provider/baseUrl/apiKey unchanged, model → vision model).

package/src/repl/chat-repl.js

@@ -183,6 +183,13 @@ export async function startChatRepl(options = {}) {
       const onUsage = (u) => {
         capturedUsage = u;
       };
+      // Stall hint (cc agent parity): the connection is alive but the API has
+      // gone silent — reassure instead of a frozen cursor. stderr so it never
+      // corrupts the streamed answer on stdout.
+      const onStall = (ms) =>
+        process.stderr.write(
+          `\x1b[2m  ⏳ waiting for API response (silent ${Math.round(ms / 1000)}s)…\x1b[0m\n`,
+        );
 
       if (sessionId)
         appendEvent(sessionId, "user_message", { content: trimmed });
@@ -194,6 +201,7 @@ export async function startChatRepl(options = {}) {
           baseUrl,
           onToken,
           onUsage,
+          onStall,
         );
       } else if (provider === "anthropic") {
         const providerDef = BUILT_IN_PROVIDERS.anthropic;
@@ -215,6 +223,7 @@ export async function startChatRepl(options = {}) {
           key,
           onToken,
           onUsage,
+          onStall,
         );
       } else {
         // OpenAI-compatible providers (openai, volcengine, deepseek, dashscope, mistral, gemini, anthropic-proxy)
@@ -237,6 +246,7 @@ export async function startChatRepl(options = {}) {
           key,
           onToken,
           onUsage,
+          onStall,
         );
       }
 

package/src/repl/session-cost.js

@@ -24,7 +24,14 @@ function fmtUsd(n) {
 /** A fresh accumulator: running totals + per provider/model rows. */
 export function newCostStore() {
   return {
-    total: { inputTokens: 0, outputTokens: 0, totalTokens: 0, calls: 0 },
+    total: {
+      inputTokens: 0,
+      outputTokens: 0,
+      totalTokens: 0,
+      cacheReadTokens: 0,
+      cacheCreationTokens: 0,
+      calls: 0,
+    },
     byKey: new Map(),
   };
 }
@@ -43,10 +50,16 @@ export function addUsage(store, events) {
     const inT = num(u.input_tokens ?? u.prompt_tokens ?? u.inputTokens);
     const outT = num(u.output_tokens ?? u.completion_tokens ?? u.outputTokens);
     const totT = num(u.total_tokens ?? u.totalTokens ?? inT + outT);
-    if (inT === 0 && outT === 0 && totT === 0) continue;
+    // Prompt-cache token counts (Anthropic caching) so /cost prices cached
+    // reads/writes correctly; absent on non-caching providers → 0.
+    const cR = num(u.cache_read_input_tokens ?? u.cacheReadTokens);
+    const cC = num(u.cache_creation_input_tokens ?? u.cacheCreationTokens);
+    if (inT === 0 && outT === 0 && totT === 0 && cR === 0 && cC === 0) continue;
     s.total.inputTokens += inT;
     s.total.outputTokens += outT;
     s.total.totalTokens += totT;
+    s.total.cacheReadTokens += cR;
+    s.total.cacheCreationTokens += cC;
     s.total.calls += 1;
     const key = `${ev.provider || "?"}/${ev.model || "?"}`;
     let row = s.byKey.get(key);
@@ -57,6 +70,8 @@ export function addUsage(store, events) {
         inputTokens: 0,
         outputTokens: 0,
         totalTokens: 0,
+        cacheReadTokens: 0,
+        cacheCreationTokens: 0,
         calls: 0,
       };
       s.byKey.set(key, row);
@@ -64,6 +79,8 @@ export function addUsage(store, events) {
     row.inputTokens += inT;
     row.outputTokens += outT;
     row.totalTokens += totT;
+    row.cacheReadTokens += cR;
+    row.cacheCreationTokens += cC;
     row.calls += 1;
   }
   return s;
@@ -172,10 +189,21 @@ export function renderSessionCost(store, { pricingOverrides, roles } = {}) {
     `  total: ${fmtUsd(result.cost.totalCost)} USD  ` +
       `(${result.total.totalTokens.toLocaleString()} tokens, ${result.total.calls} calls)`,
   );
+  // Cache line only when caching actually happened this session.
+  const cacheRead = num(result.total.cacheReadTokens);
+  const cacheWrite = num(result.total.cacheCreationTokens);
+  if (cacheRead || cacheWrite) {
+    lines.push(
+      `  cache: ${cacheRead.toLocaleString()} read + ${cacheWrite.toLocaleString()} write tokens`,
+    );
+  }
   for (const row of result.byModel) {
     const provider = (row.provider || "?").padEnd(10);
     const model = (row.model || "?").padEnd(24);
-    const tokens = `in=${row.inputTokens} out=${row.outputTokens}`;
+    const cacheSuffix = num(row.cacheReadTokens)
+      ? ` cache_read=${row.cacheReadTokens}`
+      : "";
+    const tokens = `in=${row.inputTokens} out=${row.outputTokens}${cacheSuffix}`;
     const price = row.free
       ? "free (local)"
       : row.matched