aether-code 0.32.0 → 0.32.2

package/src/agent.js

@@ -1,332 +1,332 @@
-// Agent loop. Streams each turn from /api/v1/agent/stream, prints text deltas
-// in real-time, executes any tool calls, loops until the model returns no
-// tool calls (task done) or max-turns is reached.
-
-import os from "node:os";
-import path from "node:path";
-import { agentTurnStream, AetherError } from "./api.js";
-import { TOOL_DEFINITIONS, executeTool } from "./tools.js";
-import { unnamespaceToolName } from "./mcp.js";
-import { loadAllSkills, selectSkills, renderSkillsBlock } from "./skills.js";
-import { c, divider, turn, toolLabel, toolSummary, makeTokenStripper, errorLine, startSpinner } from "./render.js";
-
-const DEFAULT_MAX_TURNS = 25;
-
-// Environment block prepended to the first user message so the model can
-// resolve named locations to real absolute paths.
-function envContext(cwd) {
-  const home = os.homedir();
-  const desktop = path.join(home, "Desktop");
-  const documents = path.join(home, "Documents");
-  const win = process.platform === "win32";
-  // OS-specific shell guidance: gemma tends to emit Unix commands (mkdir -p,
-  // touch, &&-chains) which FAIL in Windows cmd.exe ("The syntax of the command
-  // is incorrect"). Steer it to the file tools (which are cross-platform and
-  // auto-create parent dirs) and OS-appropriate shell usage.
-  const shellNote = win
-    ? `shell: Windows cmd.exe. To create files OR directories, use the write_file tool — ` +
-      `it creates any missing parent folders automatically. Do NOT use shell "mkdir -p", ` +
-      `"touch", "ls", "rm", "cp", "mv", or "&&"-chained Unix commands; they FAIL in cmd.exe. ` +
-      `Run programs/tests with their interpreter (python, node, java, etc.).`
-    : `shell: POSIX sh. Standard Unix commands are available. write_file still auto-creates parent dirs.`;
-  return (
-    `[environment]\n` +
-    `os: ${process.platform}\n` +
-    `cwd: ${cwd}\n` +
-    `home: ${home}\n` +
-    `desktop: ${desktop}\n` +
-    `documents: ${documents}\n` +
-    `${shellNote}\n` +
-    `When the user names a location ("my desktop", "home", "documents"), write to ` +
-    `the matching ABSOLUTE path above (e.g. desktop -> ${desktop}). Otherwise ` +
-    `work under the cwd. Use absolute paths when a specific location is named.\n` +
-    `[/environment]\n\n`
-  );
-}
-
-export async function runAgent({
-  initialPrompt,
-  priorMessages,
-  cwd,
-  autoYes = false,
-  unsafePaths = false,
-  maxTurns = DEFAULT_MAX_TURNS,
-  model = null, // null = server default (gemma). Premium models gated server-side.
-  onTokens = () => {},
-  // Optional MCPManager. When provided, its tools are merged into the agent's
-  // toolset and tool calls prefixed `mcp__` are routed to it instead of the
-  // built-in executeTool.
-  mcpManager = null,
-}) {
-  // Merge built-in tools with MCP-provided tools. MCP tools come second so
-  // any name collision (unlikely given namespacing, but defense in depth)
-  // resolves to the built-in.
-  const tools = mcpManager
-    ? [...TOOL_DEFINITIONS, ...mcpManager.getToolDefinitions()]
-    : TOOL_DEFINITIONS;
-
-  // Load skills once per runAgent call (bundled + user-installed). They
-  // get selected per-turn against the current prompt + any file paths the
-  // model has read so far. Loading errors are non-fatal — a bad skill file
-  // shouldn't kill the agent.
-  let allSkills = [];
-  try {
-    allSkills = loadAllSkills();
-  } catch (e) {
-    process.stderr.write(c.yellow(`(skill load failed: ${e.message})\n`));
-  }
-  const referencedPaths = [];
-  // Two callers: one-shot (initialPrompt only, fresh conversation) and REPL
-  // (priorMessages + initialPrompt to continue an ongoing chat).
-  // On the FIRST message of a session, prepend an environment block so the
-  // model knows real absolute paths (cwd / home / desktop). Without it, "build
-  // X on my desktop" became `mkdir X` in whatever dir aether was launched from
-  // (e.g. C:\WINDOWS\system32). Only prepended once — later turns carry it in
-  // history.
-  const messages = priorMessages
-    ? [...priorMessages, { role: "user", content: initialPrompt }]
-    : [{ role: "user", content: envContext(cwd) + initialPrompt }];
-  let totalCredits = 0;
-  let totalIn = 0;
-  let totalOut = 0;
-  let lastBalance = null;
-  // Loop guard: count identical (name+args) tool calls across the whole run so a
-  // confused model can't burn turns re-running the same call (e.g. glob *.md x9).
-  const callCounts = new Map();
-  // Files the run created/edited, shown in a summary at the end so the user
-  // knows exactly what was produced and where to find it.
-  const filesTouched = new Map();
-  // Files the run read — used by the completeness guard to spot files that were
-  // read (e.g. callers in a refactor) but never edited.
-  const filesRead = new Set();
-  // Completeness guard: gemma sometimes reads files then narrates the fix as
-  // text, or refactors one file but forgets the others. If a MODIFICATION task
-  // finishes having read files it never edited, we nudge it once to finish.
-  const MOD_RE =
-    /\b(refactor|fix|add|adds?|change|update|implement|creat|writ|remov|delet|rename|extract|move|replace|build|generate|insert|append|convert|migrat|wire|hook|edit|modif|patch|integrat)/i;
-  const looksLikeModification = MOD_RE.test(initialPrompt);
-  let appliedNothingNudges = 0;
-
-  for (let i = 0; i < maxTurns; i++) {
-    // No turn header and no leading blank here — each step (assistant text and
-    // each tool label) begins with its own "\n● ", so spacing stays exactly one
-    // blank line per step instead of stacking up.
-
-    // Stream the assistant's response. Print text deltas as they arrive,
-    // along with tool-call announcements as soon as the model commits to
-    // calling a particular tool (i.e. the `name` arrives in the stream).
-    const announced = new Set();
-    let lastWasText = false;
-    const stripper = makeTokenStripper();
-
-    // Select skills for this turn against the current user prompt + any
-    // paths the model has read so far. Prepend the matching skills' bodies
-    // to the last user message of a shallow-cloned messages array — we
-    // don't want skill text accumulating in the persisted history, only
-    // being available to the model for the turn where it's relevant.
-    const turnMessages = buildTurnMessages(messages, allSkills, referencedPaths);
-
-    let res;
-    // "Thinking" spinner: shown from request-send until the first token or tool
-    // call arrives, so the wait doesn't look dead. Stopped exactly once.
-    const spinner = startSpinner("thinking");
-    let spinStopped = false;
-    const stopSpin = () => { if (!spinStopped) { spinStopped = true; spinner.stop(); } };
-    try {
-      res = await agentTurnStream({
-        messages: turnMessages,
-        tools,
-        model,
-        onDelta: (text) => {
-          // Buffered strip of leaked model channel/control tokens (which can
-          // be split across stream chunks) before display.
-          const clean = stripper.push(text);
-          if (!clean) return;
-          // Don't open a "● " bullet for leading whitespace (e.g. when a whole
-          // turn's text was suppressed as a leak, leaving only a stray newline).
-          if (!lastWasText && !clean.trim()) return;
-          stopSpin();
-          if (!lastWasText) {
-            process.stdout.write("\n" + c.cyan("● "));
-            lastWasText = true;
-          }
-          process.stdout.write(clean);
-        },
-        onToolCallDelta: (delta) => {
-          // Just close the streamed text line when the model starts a tool
-          // call — the clean label is printed at execution time, so no noisy
-          // "preparing args" placeholder here.
-          if (delta.name && !announced.has(delta.index)) {
-            stopSpin();
-            announced.add(delta.index);
-            if (lastWasText) process.stdout.write("\n");
-            lastWasText = false;
-          }
-        },
-      });
-    } catch (err) {
-      stopSpin();
-      if (err instanceof AetherError) {
-        return { ok: false, error: err, totalCredits, totalIn, totalOut, balance: lastBalance, messages };
-      }
-      throw err;
-    }
-    stopSpin(); // ensure it's cleared even if the turn produced no output
-
-    // Flush any held-back partial token, then close the line.
-    const tail = stripper.flush();
-    if (tail && (lastWasText || tail.trim())) {
-      if (!lastWasText) { process.stdout.write("\n" + c.cyan("● ")); lastWasText = true; }
-      process.stdout.write(tail);
-    }
-    if (lastWasText) process.stdout.write("\n");
-    totalCredits += res.creditsCharged ?? 0;
-    totalIn += res.usage?.prompt_tokens ?? 0;
-    totalOut += res.usage?.completion_tokens ?? 0;
-    if (typeof res.balanceAfter === "number") lastBalance = res.balanceAfter;
-    onTokens({ totalCredits, totalIn, totalOut, balance: lastBalance });
-    // Per-turn cost line removed for a cleaner look — the session summary at the
-    // end carries the totals.
-
-    // Push assistant message into history
-    messages.push({
-      role: "assistant",
-      content: res.message.content,
-      tool_calls: res.message.tool_calls,
-    });
-
-    const toolCalls = res.message.tool_calls ?? [];
-    if (toolCalls.length === 0) {
-      // Completeness guard: on a modification task, if the model read files it
-      // never edited, it likely either described the fix instead of applying it
-      // (zero edits) or refactored some files but forgot others (partial). Nudge
-      // once (at most) to finish. The "if it needs no change, just finish" escape
-      // keeps a legitimately read-for-context file from forcing a wrong edit.
-      const unedited = [...filesRead].filter((p) => !filesTouched.has(p));
-      if (looksLikeModification && appliedNothingNudges < 1 && (filesTouched.size === 0 ? referencedPaths.length > 0 : unedited.length > 0)) {
-        appliedNothingNudges++;
-        const msg =
-          filesTouched.size === 0
-            ? "You read the file(s) but applied NO changes (no write_file / edit_file). This task asks you to " +
-              "modify code — ACTUALLY APPLY the change now with edit_file/write_file, don't just describe it."
-            : `You edited ${[...filesTouched.keys()].join(", ")} but read ${unedited.join(", ")} without editing ${unedited.length === 1 ? "it" : "them"}. ` +
-              `If the task needs ${unedited.length === 1 ? "that file" : "those files"} changed too (e.g. updating callers/imports after a refactor), do it NOW. ` +
-              `If ${unedited.length === 1 ? "it genuinely needs" : "they genuinely need"} no change, then finish.`;
-        messages.push({ role: "user", content: msg });
-        continue; // give the model another turn to finish
-      }
-      if (filesTouched.size) {
-        console.log("\n" + c.bold("Files") + c.gray(` in ${cwd}`));
-        for (const [p, action] of filesTouched) {
-          console.log("  " + c.green("●") + " " + p + c.gray(` — ${action}`));
-        }
-      }
-      return { ok: true, totalCredits, totalIn, totalOut, turns: i + 1, balance: lastBalance, messages };
-    }
-
-    // Execute each tool call. Show the actual args (now that we have them
-    // fully assembled) and run.
-    for (const call of toolCalls) {
-      let args = {};
-      try { args = JSON.parse(call.function.arguments || "{}"); } catch { /* leave empty */ }
-      // todo_write renders its own Plan box, so it returns an empty label.
-      const label = toolLabel(call.function.name, args);
-      if (label) console.log("\n" + c.cyan("●") + " " + label);
-
-      // Loop guard — short-circuit a tool call we've already run 3+ times with
-      // the same args, and tell the model to change approach instead of looping.
-      const sig = `${call.function.name}:${call.function.arguments || ""}`;
-      const seen = (callCounts.get(sig) || 0) + 1;
-      callCounts.set(sig, seen);
-      if (seen > 3) {
-        console.log("  " + c.red("└─") + " " + c.gray("skipped (repeated call)"));
-        messages.push({
-          role: "tool",
-          tool_call_id: call.id,
-          content:
-            "STOP: you have already run this exact tool call 3 times and the result will not change. Do NOT call it again. " +
-            "If you were looking for a file that doesn't exist, create it with write_file. Otherwise change approach or finish and summarize.",
-        });
-        continue;
-      }
-
-      // Route to MCP if the tool name is namespaced (mcp__server__tool);
-      // otherwise execute the built-in tool. unnamespaceToolName returns
-      // null for non-MCP names, which is our cheap dispatch test.
-      const slow = call.function.name === "web_search" || call.function.name === "web_fetch";
-      const tspin = slow ? startSpinner(call.function.name === "web_search" ? "searching the web" : "fetching page") : null;
-      let result;
-      try {
-        if (mcpManager && unnamespaceToolName(call.function.name)) {
-          result = await mcpManager.callTool(call.function.name, args);
-        } else {
-          result = await executeTool(call, { cwd, autoYes, unsafePaths });
-        }
-      } finally {
-        if (tspin) tspin.stop();
-      }
-
-      // Track paths the model has touched. Skills with path-pattern triggers
-      // (e.g. RE skill on `*.exe`) match against this list, so reading a
-      // binary in turn 3 can activate the RE skill in turn 4.
-      if (call.function.name === "read_file" || call.function.name === "edit_file" || call.function.name === "write_file") {
-        if (typeof args.path === "string") referencedPaths.push(args.path);
-      }
-      if (result.ok && typeof args.path === "string") {
-        const rel = path.relative(cwd, path.resolve(cwd, args.path)) || args.path;
-        if (call.function.name === "write_file" || call.function.name === "edit_file") {
-          filesTouched.set(rel, call.function.name === "write_file" ? "created" : "edited");
-        } else if (call.function.name === "read_file") {
-          filesRead.add(rel);
-        }
-      }
-      const summary = toolSummary(call.function.name, result);
-      if (summary) console.log(summary);
-
-      messages.push({
-        role: "tool",
-        tool_call_id: call.id,
-        content: result.output ?? (result.ok ? "(no output)" : "Failed."),
-      });
-    }
-  }
-
-  console.log(c.yellow(`\nReached max turns (${maxTurns}). Stopping.`));
-  return { ok: false, error: new Error("Max turns reached"), totalCredits, totalIn, totalOut, balance: lastBalance, messages };
-}
-
-/**
- * Per-turn skill injection. Selects skills against the latest user message
- * + paths the model has touched, then prepends matching bodies onto the
- * final user message of a shallow-cloned messages array. Returns the
- * original array unchanged when no skills match — zero overhead on the
- * no-skills path.
- *
- * Why prepend to user message instead of inserting a system message:
- * the server's AGENT_SYSTEM check skips its own system prompt when ANY
- * system message is present in the request. Adding a skills system
- * message would silently delete the server's discipline — which is
- * worse than no skills at all. Prepending into the user message keeps
- * both layers active.
- */
-function buildTurnMessages(messages, allSkills, referencedPaths) {
-  if (allSkills.length === 0) return messages;
-  // Find the latest user message — that's where the current task lives.
-  let lastUserIdx = -1;
-  for (let i = messages.length - 1; i >= 0; i--) {
-    if (messages[i].role === "user") { lastUserIdx = i; break; }
-  }
-  if (lastUserIdx === -1) return messages;
-  const prompt = typeof messages[lastUserIdx].content === "string"
-    ? messages[lastUserIdx].content
-    : "";
-  const active = selectSkills({ skills: allSkills, prompt, referencedPaths });
-  if (active.length === 0) return messages;
-  const block = renderSkillsBlock(active);
-  const cloned = [...messages];
-  cloned[lastUserIdx] = {
-    ...cloned[lastUserIdx],
-    content: `${block}\n\n---\n\n${prompt}`,
-  };
-  return cloned;
-}
+// Agent loop. Streams each turn from /api/v1/agent/stream, prints text deltas
+// in real-time, executes any tool calls, loops until the model returns no
+// tool calls (task done) or max-turns is reached.
+
+import os from "node:os";
+import path from "node:path";
+import { agentTurnStream, AetherError } from "./api.js";
+import { TOOL_DEFINITIONS, executeTool } from "./tools.js";
+import { unnamespaceToolName } from "./mcp.js";
+import { loadAllSkills, selectSkills, renderSkillsBlock } from "./skills.js";
+import { c, divider, turn, toolLabel, toolSummary, makeTokenStripper, errorLine, startSpinner } from "./render.js";
+
+const DEFAULT_MAX_TURNS = 25;
+
+// Environment block prepended to the first user message so the model can
+// resolve named locations to real absolute paths.
+function envContext(cwd) {
+  const home = os.homedir();
+  const desktop = path.join(home, "Desktop");
+  const documents = path.join(home, "Documents");
+  const win = process.platform === "win32";
+  // OS-specific shell guidance: gemma tends to emit Unix commands (mkdir -p,
+  // touch, &&-chains) which FAIL in Windows cmd.exe ("The syntax of the command
+  // is incorrect"). Steer it to the file tools (which are cross-platform and
+  // auto-create parent dirs) and OS-appropriate shell usage.
+  const shellNote = win
+    ? `shell: Windows cmd.exe. To create files OR directories, use the write_file tool — ` +
+      `it creates any missing parent folders automatically. Do NOT use shell "mkdir -p", ` +
+      `"touch", "ls", "rm", "cp", "mv", or "&&"-chained Unix commands; they FAIL in cmd.exe. ` +
+      `Run programs/tests with their interpreter (python, node, java, etc.).`
+    : `shell: POSIX sh. Standard Unix commands are available. write_file still auto-creates parent dirs.`;
+  return (
+    `[environment]\n` +
+    `os: ${process.platform}\n` +
+    `cwd: ${cwd}\n` +
+    `home: ${home}\n` +
+    `desktop: ${desktop}\n` +
+    `documents: ${documents}\n` +
+    `${shellNote}\n` +
+    `When the user names a location ("my desktop", "home", "documents"), write to ` +
+    `the matching ABSOLUTE path above (e.g. desktop -> ${desktop}). Otherwise ` +
+    `work under the cwd. Use absolute paths when a specific location is named.\n` +
+    `[/environment]\n\n`
+  );
+}
+
+export async function runAgent({
+  initialPrompt,
+  priorMessages,
+  cwd,
+  autoYes = false,
+  unsafePaths = false,
+  maxTurns = DEFAULT_MAX_TURNS,
+  model = null, // null = server default (gemma). Premium models gated server-side.
+  onTokens = () => {},
+  // Optional MCPManager. When provided, its tools are merged into the agent's
+  // toolset and tool calls prefixed `mcp__` are routed to it instead of the
+  // built-in executeTool.
+  mcpManager = null,
+}) {
+  // Merge built-in tools with MCP-provided tools. MCP tools come second so
+  // any name collision (unlikely given namespacing, but defense in depth)
+  // resolves to the built-in.
+  const tools = mcpManager
+    ? [...TOOL_DEFINITIONS, ...mcpManager.getToolDefinitions()]
+    : TOOL_DEFINITIONS;
+
+  // Load skills once per runAgent call (bundled + user-installed). They
+  // get selected per-turn against the current prompt + any file paths the
+  // model has read so far. Loading errors are non-fatal — a bad skill file
+  // shouldn't kill the agent.
+  let allSkills = [];
+  try {
+    allSkills = loadAllSkills();
+  } catch (e) {
+    process.stderr.write(c.yellow(`(skill load failed: ${e.message})\n`));
+  }
+  const referencedPaths = [];
+  // Two callers: one-shot (initialPrompt only, fresh conversation) and REPL
+  // (priorMessages + initialPrompt to continue an ongoing chat).
+  // On the FIRST message of a session, prepend an environment block so the
+  // model knows real absolute paths (cwd / home / desktop). Without it, "build
+  // X on my desktop" became `mkdir X` in whatever dir aether was launched from
+  // (e.g. C:\WINDOWS\system32). Only prepended once — later turns carry it in
+  // history.
+  const messages = priorMessages
+    ? [...priorMessages, { role: "user", content: initialPrompt }]
+    : [{ role: "user", content: envContext(cwd) + initialPrompt }];
+  let totalCredits = 0;
+  let totalIn = 0;
+  let totalOut = 0;
+  let lastBalance = null;
+  // Loop guard: count identical (name+args) tool calls across the whole run so a
+  // confused model can't burn turns re-running the same call (e.g. glob *.md x9).
+  const callCounts = new Map();
+  // Files the run created/edited, shown in a summary at the end so the user
+  // knows exactly what was produced and where to find it.
+  const filesTouched = new Map();
+  // Files the run read — used by the completeness guard to spot files that were
+  // read (e.g. callers in a refactor) but never edited.
+  const filesRead = new Set();
+  // Completeness guard: gemma sometimes reads files then narrates the fix as
+  // text, or refactors one file but forgets the others. If a MODIFICATION task
+  // finishes having read files it never edited, we nudge it once to finish.
+  const MOD_RE =
+    /\b(refactor|fix|add|adds?|change|update|implement|creat|writ|remov|delet|rename|extract|move|replace|build|generate|insert|append|convert|migrat|wire|hook|edit|modif|patch|integrat)/i;
+  const looksLikeModification = MOD_RE.test(initialPrompt);
+  let appliedNothingNudges = 0;
+
+  for (let i = 0; i < maxTurns; i++) {
+    // No turn header and no leading blank here — each step (assistant text and
+    // each tool label) begins with its own "\n● ", so spacing stays exactly one
+    // blank line per step instead of stacking up.
+
+    // Stream the assistant's response. Print text deltas as they arrive,
+    // along with tool-call announcements as soon as the model commits to
+    // calling a particular tool (i.e. the `name` arrives in the stream).
+    const announced = new Set();
+    let lastWasText = false;
+    const stripper = makeTokenStripper();
+
+    // Select skills for this turn against the current user prompt + any
+    // paths the model has read so far. Prepend the matching skills' bodies
+    // to the last user message of a shallow-cloned messages array — we
+    // don't want skill text accumulating in the persisted history, only
+    // being available to the model for the turn where it's relevant.
+    const turnMessages = buildTurnMessages(messages, allSkills, referencedPaths);
+
+    let res;
+    // "Thinking" spinner: shown from request-send until the first token or tool
+    // call arrives, so the wait doesn't look dead. Stopped exactly once.
+    const spinner = startSpinner("thinking");
+    let spinStopped = false;
+    const stopSpin = () => { if (!spinStopped) { spinStopped = true; spinner.stop(); } };
+    try {
+      res = await agentTurnStream({
+        messages: turnMessages,
+        tools,
+        model,
+        onDelta: (text) => {
+          // Buffered strip of leaked model channel/control tokens (which can
+          // be split across stream chunks) before display.
+          const clean = stripper.push(text);
+          if (!clean) return;
+          // Don't open a "● " bullet for leading whitespace (e.g. when a whole
+          // turn's text was suppressed as a leak, leaving only a stray newline).
+          if (!lastWasText && !clean.trim()) return;
+          stopSpin();
+          if (!lastWasText) {
+            process.stdout.write("\n" + c.cyan("● "));
+            lastWasText = true;
+          }
+          process.stdout.write(clean);
+        },
+        onToolCallDelta: (delta) => {
+          // Just close the streamed text line when the model starts a tool
+          // call — the clean label is printed at execution time, so no noisy
+          // "preparing args" placeholder here.
+          if (delta.name && !announced.has(delta.index)) {
+            stopSpin();
+            announced.add(delta.index);
+            if (lastWasText) process.stdout.write("\n");
+            lastWasText = false;
+          }
+        },
+      });
+    } catch (err) {
+      stopSpin();
+      if (err instanceof AetherError) {
+        return { ok: false, error: err, totalCredits, totalIn, totalOut, balance: lastBalance, messages };
+      }
+      throw err;
+    }
+    stopSpin(); // ensure it's cleared even if the turn produced no output
+
+    // Flush any held-back partial token, then close the line.
+    const tail = stripper.flush();
+    if (tail && (lastWasText || tail.trim())) {
+      if (!lastWasText) { process.stdout.write("\n" + c.cyan("● ")); lastWasText = true; }
+      process.stdout.write(tail);
+    }
+    if (lastWasText) process.stdout.write("\n");
+    totalCredits += res.creditsCharged ?? 0;
+    totalIn += res.usage?.prompt_tokens ?? 0;
+    totalOut += res.usage?.completion_tokens ?? 0;
+    if (typeof res.balanceAfter === "number") lastBalance = res.balanceAfter;
+    onTokens({ totalCredits, totalIn, totalOut, balance: lastBalance });
+    // Per-turn cost line removed for a cleaner look — the session summary at the
+    // end carries the totals.
+
+    // Push assistant message into history
+    messages.push({
+      role: "assistant",
+      content: res.message.content,
+      tool_calls: res.message.tool_calls,
+    });
+
+    const toolCalls = res.message.tool_calls ?? [];
+    if (toolCalls.length === 0) {
+      // Completeness guard: on a modification task, if the model read files it
+      // never edited, it likely either described the fix instead of applying it
+      // (zero edits) or refactored some files but forgot others (partial). Nudge
+      // once (at most) to finish. The "if it needs no change, just finish" escape
+      // keeps a legitimately read-for-context file from forcing a wrong edit.
+      const unedited = [...filesRead].filter((p) => !filesTouched.has(p));
+      if (looksLikeModification && appliedNothingNudges < 1 && (filesTouched.size === 0 ? referencedPaths.length > 0 : unedited.length > 0)) {
+        appliedNothingNudges++;
+        const msg =
+          filesTouched.size === 0
+            ? "You read the file(s) but applied NO changes (no write_file / edit_file). This task asks you to " +
+              "modify code — ACTUALLY APPLY the change now with edit_file/write_file, don't just describe it."
+            : `You edited ${[...filesTouched.keys()].join(", ")} but read ${unedited.join(", ")} without editing ${unedited.length === 1 ? "it" : "them"}. ` +
+              `If the task needs ${unedited.length === 1 ? "that file" : "those files"} changed too (e.g. updating callers/imports after a refactor), do it NOW. ` +
+              `If ${unedited.length === 1 ? "it genuinely needs" : "they genuinely need"} no change, then finish.`;
+        messages.push({ role: "user", content: msg });
+        continue; // give the model another turn to finish
+      }
+      if (filesTouched.size) {
+        console.log("\n" + c.bold("Files") + c.gray(` in ${cwd}`));
+        for (const [p, action] of filesTouched) {
+          console.log("  " + c.green("●") + " " + p + c.gray(` — ${action}`));
+        }
+      }
+      return { ok: true, totalCredits, totalIn, totalOut, turns: i + 1, balance: lastBalance, messages };
+    }
+
+    // Execute each tool call. Show the actual args (now that we have them
+    // fully assembled) and run.
+    for (const call of toolCalls) {
+      let args = {};
+      try { args = JSON.parse(call.function.arguments || "{}"); } catch { /* leave empty */ }
+      // todo_write renders its own Plan box, so it returns an empty label.
+      const label = toolLabel(call.function.name, args);
+      if (label) console.log("\n" + c.cyan("●") + " " + label);
+
+      // Loop guard — short-circuit a tool call we've already run 3+ times with
+      // the same args, and tell the model to change approach instead of looping.
+      const sig = `${call.function.name}:${call.function.arguments || ""}`;
+      const seen = (callCounts.get(sig) || 0) + 1;
+      callCounts.set(sig, seen);
+      if (seen > 3) {
+        console.log("  " + c.red("└─") + " " + c.gray("skipped (repeated call)"));
+        messages.push({
+          role: "tool",
+          tool_call_id: call.id,
+          content:
+            "STOP: you have already run this exact tool call 3 times and the result will not change. Do NOT call it again. " +
+            "If you were looking for a file that doesn't exist, create it with write_file. Otherwise change approach or finish and summarize.",
+        });
+        continue;
+      }
+
+      // Route to MCP if the tool name is namespaced (mcp__server__tool);
+      // otherwise execute the built-in tool. unnamespaceToolName returns
+      // null for non-MCP names, which is our cheap dispatch test.
+      const slow = call.function.name === "web_search" || call.function.name === "web_fetch";
+      const tspin = slow ? startSpinner(call.function.name === "web_search" ? "searching the web" : "fetching page") : null;
+      let result;
+      try {
+        if (mcpManager && unnamespaceToolName(call.function.name)) {
+          result = await mcpManager.callTool(call.function.name, args);
+        } else {
+          result = await executeTool(call, { cwd, autoYes, unsafePaths });
+        }
+      } finally {
+        if (tspin) tspin.stop();
+      }
+
+      // Track paths the model has touched. Skills with path-pattern triggers
+      // (e.g. RE skill on `*.exe`) match against this list, so reading a
+      // binary in turn 3 can activate the RE skill in turn 4.
+      if (call.function.name === "read_file" || call.function.name === "edit_file" || call.function.name === "write_file") {
+        if (typeof args.path === "string") referencedPaths.push(args.path);
+      }
+      if (result.ok && typeof args.path === "string") {
+        const rel = path.relative(cwd, path.resolve(cwd, args.path)) || args.path;
+        if (call.function.name === "write_file" || call.function.name === "edit_file") {
+          filesTouched.set(rel, call.function.name === "write_file" ? "created" : "edited");
+        } else if (call.function.name === "read_file") {
+          filesRead.add(rel);
+        }
+      }
+      const summary = toolSummary(call.function.name, result);
+      if (summary) console.log(summary);
+
+      messages.push({
+        role: "tool",
+        tool_call_id: call.id,
+        content: result.output ?? (result.ok ? "(no output)" : "Failed."),
+      });
+    }
+  }
+
+  console.log(c.yellow(`\nReached max turns (${maxTurns}). Stopping.`));
+  return { ok: false, error: new Error("Max turns reached"), totalCredits, totalIn, totalOut, balance: lastBalance, messages };
+}
+
+/**
+ * Per-turn skill injection. Selects skills against the latest user message
+ * + paths the model has touched, then prepends matching bodies onto the
+ * final user message of a shallow-cloned messages array. Returns the
+ * original array unchanged when no skills match — zero overhead on the
+ * no-skills path.
+ *
+ * Why prepend to user message instead of inserting a system message:
+ * the server's AGENT_SYSTEM check skips its own system prompt when ANY
+ * system message is present in the request. Adding a skills system
+ * message would silently delete the server's discipline — which is
+ * worse than no skills at all. Prepending into the user message keeps
+ * both layers active.
+ */
+function buildTurnMessages(messages, allSkills, referencedPaths) {
+  if (allSkills.length === 0) return messages;
+  // Find the latest user message — that's where the current task lives.
+  let lastUserIdx = -1;
+  for (let i = messages.length - 1; i >= 0; i--) {
+    if (messages[i].role === "user") { lastUserIdx = i; break; }
+  }
+  if (lastUserIdx === -1) return messages;
+  const prompt = typeof messages[lastUserIdx].content === "string"
+    ? messages[lastUserIdx].content
+    : "";
+  const active = selectSkills({ skills: allSkills, prompt, referencedPaths });
+  if (active.length === 0) return messages;
+  const block = renderSkillsBlock(active);
+  const cloned = [...messages];
+  cloned[lastUserIdx] = {
+    ...cloned[lastUserIdx],
+    content: `${block}\n\n---\n\n${prompt}`,
+  };
+  return cloned;
+}