s-app-scoder-agent 0.6.0 → 0.7.0

package/dist/tools.js

@@ -1,10 +1,26 @@
-import { readFileSync, writeFileSync, readdirSync, statSync, mkdirSync, realpathSync } from "node:fs";
-import { spawn } from "node:child_process";
-import { resolve, dirname, basename, join, relative, isAbsolute } from "node:path";
+import { readFileSync, writeFileSync, readdirSync, statSync, mkdirSync, realpathSync, existsSync } from "node:fs";
+import { spawn, spawnSync } from "node:child_process";
+import { resolve, dirname, basename, join, relative, isAbsolute, sep } from "node:path";
 import { unifiedDiff } from "./diff.js";
+import { detectImageBytes } from "./tui/image.js";
+import { discoverSkills, findSkill } from "./skills.js";
+import { redactSecrets } from "./llm.js";
+import { loadRcFile } from "./config.js";
+import { resolveSandbox } from "./sandbox.js";
 import * as ui from "./ui.js";
 import { toolWebFetch } from "./webfetch.js";
-const MAX_OUTPUT_CHARS = 20_000;
+/**
+ * G7: max chars tool output — default 20_000, runtime override qua
+ * setMaxOutputChars (agent constructor gọi với config.limits).
+ * ponytail: module-level mutable thay vì ctx param vào 12 call sites —
+ * 1 agent/process đổi 1 lần lúc khởi tạo; đa-config cùng process (server mode)
+ * chỉ cần giá trị giống nhau, nâng cấp thành ctx field khi có nhu cầu khác nhau.
+ */
+let MAX_OUTPUT_CHARS = 20_000;
+export function setMaxOutputChars(n) {
+    if (Number.isFinite(n) && n > 0)
+        MAX_OUTPUT_CHARS = Math.floor(n);
+}
 const MAX_FILE_SIZE_BYTES = 2 * 1024 * 1024; // refuse to read files > 2MB
 function clamp(s, max = MAX_OUTPUT_CHARS) {
     if (s.length <= max)
@@ -70,9 +86,31 @@ function toResult(fn) {
         return { ok: false, output: `ERROR: ${err.message}` };
     }
 }
+// H-04: protected files — read ⇒ ask (khi không allow rule), write luôn gate.
+// `.env.example` và các biến thể template luôn được phép đọc.
+const PROTECTED_READ_PATTERNS = [/^\.env$/, /\/\.env$/, /^\.env\.[^/]+$/, /\/\.env\.[^/]+$/];
+const PROTECTED_READ_EXEMPT = [/^\.env\.example$/, /\/\.env\.example$/, /^\.env\.sample$/, /\/\.env\.sample$/, /^\.env\.template$/, /\/\.env\.template$/];
+export function isProtectedEnvFile(relPath) {
+    if (PROTECTED_READ_EXEMPT.some((rx) => rx.test(relPath)))
+        return "allow";
+    if (PROTECTED_READ_PATTERNS.some((rx) => rx.test(relPath)))
+        return "ask";
+    return null;
+}
 function toolReadFile(workdir, args) {
     return toResult(() => {
         const abs = safePath(workdir, args.path);
+        // H-04: .env read policy — ask trước khi trả nội dung (secrets hygiene).
+        const rel = relative(workdir, abs);
+        const envPolicy = isProtectedEnvFile(rel);
+        if (envPolicy === "ask") {
+            return {
+                ok: false,
+                output: `ASK_PERMISSION: ${rel} may contain secrets (env file). ` +
+                    `Ask the user to confirm before reading; suggest reading .env.example instead ` +
+                    `or running with a permission rule ({"action":"shell","resource":"read .env","effect":"allow"}).`,
+            };
+        }
         const st = statSync(abs);
         if (st.isDirectory()) {
             return { ok: false, output: `ERROR: ${args.path} is a directory — use list_files instead.` };
@@ -83,6 +121,20 @@ function toolReadFile(workdir, args) {
                 output: `ERROR: ${args.path} is ${(st.size / 1024 / 1024).toFixed(1)}MB (limit ${MAX_FILE_SIZE_BYTES / 1024 / 1024}MB). Use run_command with grep/head/tail to inspect specific parts.`,
             };
         }
+        // T7: image file → base64 notice cho TUI render (LLM vẫn nhận mô tả text).
+        let image;
+        try {
+            const probe = readFileSync(abs);
+            const ext = detectImageBytes(probe);
+            if (ext)
+                image = `data:image/${ext};base64,${probe.toString("base64")}`;
+        }
+        catch {
+            // not an image / unreadable — fall through to text read
+        }
+        if (image) {
+            return { ok: true, output: `[image: ${args.path}]`, image };
+        }
         const content = readFileSync(abs, "utf8");
         const lines = content.split("\n");
         const numbered = lines.map((l, i) => `${i + 1}: ${l}`).join("\n");
@@ -147,9 +199,10 @@ async function toolWriteFile(ctx, args) {
         return gate;
     mkdirSync(dirname(abs), { recursive: true });
     writeFileSync(abs, args.content, "utf8");
+    const fmtNote = formatFile(ctx, abs);
     return {
         ok: true,
-        output: `Wrote ${args.content.split("\n").length} lines to ${args.path}`,
+        output: `Wrote ${args.content.split("\n").length} lines to ${args.path}${fmtNote ? `\n${fmtNote}` : ""}`,
         diff: toToolDiff(args.path, diff),
     };
 }
@@ -177,9 +230,10 @@ async function toolEditFile(ctx, args) {
     if (gate)
         return gate;
     writeFileSync(abs, newContent, "utf8");
+    const fmtNote = formatFile(ctx, abs);
     return {
         ok: true,
-        output: `Edited ${args.path}`,
+        output: `Edited ${args.path}${fmtNote ? `\n${fmtNote}` : ""}`,
         diff: toToolDiff(args.path, diff),
     };
 }
@@ -227,15 +281,329 @@ function toolListFiles(workdir, args) {
         return { ok: true, output: out.length > 0 ? clamp(out.join("\n")) : "(empty directory)" };
     });
 }
+// ---------------------------------------------------------------------------
+// grep / glob (I-07)
+// ---------------------------------------------------------------------------
+const SEARCH_SKIP = new Set(["node_modules", ".git", "dist", "build", ".next", "__pycache__", ".venv"]);
+const MAX_SEARCH_HITS = 50;
+/** Walk files under root (bounded depth, skip noise dirs), collect relative paths. */
+function walkFiles(root, cb, depth = 8) {
+    if (depth <= 0)
+        return;
+    let entries;
+    try {
+        entries = readdirSync(root, { withFileTypes: true });
+    }
+    catch {
+        return;
+    }
+    for (const e of entries) {
+        if (SEARCH_SKIP.has(e.name))
+            continue;
+        const abs = join(root, e.name);
+        if (e.isDirectory())
+            walkFiles(abs, cb, depth - 1);
+        else if (e.isFile())
+            cb(e.name, abs);
+    }
+}
+/** ripgrep if on PATH; else grep -rn; else JS line-scan fallback. Returns formatted hits. */
+export function toolGrep(workdir, args) {
+    return toResult(() => {
+        const pattern = String(args.pattern ?? "").trim();
+        if (!pattern)
+            return { ok: false, output: "ERROR: pattern is required." };
+        const root = safePath(workdir, args.path ?? ".");
+        // Fast path: ripgrep (respects .gitignore, fast, JSON output). Skip
+        // common noise dirs even when the project has no .gitignore.
+        const rgArgs = [
+            "--json", "--max-count", "5", "-e", pattern, root,
+            // rg globs match the printed (absolute) path — bare dir name excludes it
+            // anywhere in the tree; ** forms do NOT match absolute-path matches.
+            "--glob", "!node_modules", "--glob", "!.git", "--glob", "!dist",
+            "--glob", "!build", "--glob", "!.next", "--glob", "!__pycache__",
+            "--glob", "!.venv",
+        ];
+        if (args.include)
+            rgArgs.push("-g", args.include);
+        const rg = spawnSync("rg", rgArgs, { encoding: "utf8", maxBuffer: 10 * 1024 * 1024 });
+        const lines = [];
+        let count = 0;
+        if ((rg.status ?? 2) === 0 || rg.stdout) {
+            for (const line of rg.stdout.split("\n")) {
+                if (!line || count >= MAX_SEARCH_HITS)
+                    break;
+                try {
+                    const j = JSON.parse(line);
+                    if (j.type !== "match" || !j.data)
+                        continue;
+                    const rel = relative(workdir, j.data.path?.text ?? "").split(sep).join("/");
+                    lines.push(`${rel}:${j.data.line_number}: ${(j.data.lines?.text ?? "").trimEnd().slice(0, 200)}`);
+                    count++;
+                }
+                catch { /* skip malformed */ }
+            }
+            if ((rg.status ?? 2) <= 1) {
+                return { ok: true, output: lines.length > 0 ? clamp(lines.join("\n")) : "(no matches)" };
+            }
+            // status >= 2 → rg error; fall through to grep fallback.
+        }
+        // Fallback: grep -rn (portable) — output already path:line:text.
+        const g = spawnSync("grep", ["-rn", "--binary-files=without-match", "-I", "-m", "5", "-e", pattern, root], { encoding: "utf8", maxBuffer: 10 * 1024 * 1024 });
+        if ((g.status ?? 2) <= 1) {
+            const hits = (g.stdout ?? "").split("\n").filter(Boolean).slice(0, MAX_SEARCH_HITS)
+                .map((l) => l.startsWith(root) ? relative(workdir, root).split(sep).join("/") + l.slice(root.length) : l);
+            return { ok: true, output: hits.length > 0 ? clamp(hits.join("\n")) : "(no matches)" };
+        }
+        // rg vắng mặt + grep vắng mặt/lỗi → scan JS thuần (luôn chạy được, không cần binary).
+        return jsGrepFallback(workdir, root, pattern, args.include);
+    });
+}
+/**
+ * Fallback cuối: scan dòng thuần JS (không cần rg/grep). Kém tối ưu hơn rg
+ * (không tôn trọng .gitignore ngoài SEARCH_SKIP, không lazy) nhưng luôn chạy
+ * — quan trọng trên Windows máy chưa cài ripgrep. Pattern là regex JS; regex
+ * không hợp lệ → coi như literal substring.
+ */
+function jsGrepFallback(workdir, root, pattern, include) {
+    let rx;
+    try {
+        rx = new RegExp(pattern);
+    }
+    catch {
+        rx = new RegExp(pattern.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"));
+    }
+    const includeRx = include
+        ? globToRegExp(include.includes("/") ? include : `**/${include}`)
+        : null;
+    const hits = [];
+    walkFiles(root, (_name, abs) => {
+        if (hits.length >= MAX_SEARCH_HITS)
+            return;
+        const rel = relative(workdir, abs).split(sep).join("/");
+        if (includeRx && !includeRx.test(rel))
+            return;
+        try {
+            const st = statSync(abs);
+            if (st.size > 2 * 1024 * 1024)
+                return; // skip file quá lớn
+            const buf = readFileSync(abs);
+            if (buf.includes(0))
+                return; // skip binary (NUL byte)
+            const lines = buf.toString("utf8").split("\n");
+            let perFile = 0;
+            for (let i = 0; i < lines.length && perFile < 5 && hits.length < MAX_SEARCH_HITS; i++) {
+                if (rx.test(lines[i])) {
+                    hits.push(`${rel}:${i + 1}: ${lines[i].trimEnd().slice(0, 200)}`);
+                    perFile++;
+                }
+            }
+        }
+        catch { /* unreadable — skip */ }
+    });
+    return { ok: true, output: hits.length > 0 ? clamp(hits.join("\n")) : "(no matches)" };
+}
+/** Mini glob: supports **, *, ?, {a,b}. Implemented over walkFiles (no deps). */
+export function globToRegExp(pattern) {
+    let rx = "";
+    for (let i = 0; i < pattern.length; i++) {
+        const c = pattern[i];
+        if (c === "*") {
+            if (pattern[i + 1] === "*") {
+                rx += ".*";
+                i++;
+                if (pattern[i + 1] === "/")
+                    i++; // `**/` also matches zero dirs
+            }
+            else
+                rx += "[^/]*";
+        }
+        else if (c === "?")
+            rx += "[^/]";
+        else if (c === "{") {
+            // {a,b} → (a|b)
+            const end = pattern.indexOf("}", i);
+            const inner = pattern.slice(i + 1, end).replace(/,/g, "|");
+            rx += `(${inner})`;
+            i = end;
+        }
+        else
+            rx += c.replace(/[.+^$()|[\]\\]/g, "\\$&");
+    }
+    return new RegExp(`^${rx}$`);
+}
+export function toolGlob(workdir, args) {
+    return toResult(() => {
+        const pattern = String(args.pattern ?? "").trim();
+        if (!pattern)
+            return { ok: false, output: "ERROR: pattern is required (e.g. 'src/**/*.ts')." };
+        const root = safePath(workdir, args.path ?? ".");
+        const rx = globToRegExp(pattern);
+        const out = [];
+        walkFiles(root, (_name, abs) => {
+            // Pattern dùng '/' — normalize separator của OS (Windows trả '\') để khớp.
+            const rel = relative(workdir, abs).split(sep).join("/");
+            if (rx.test(rel) && out.length < MAX_SEARCH_HITS * 4)
+                out.push(rel);
+        });
+        out.sort();
+        return { ok: true, output: out.length > 0 ? clamp(out.join("\n")) : "(no files match)" };
+    });
+}
+// ---------------------------------------------------------------------------
+// Formatter post-edit (I-11)
+// ---------------------------------------------------------------------------
+const FORMATTER_TIMEOUT_MS = 30_000;
+/**
+ * Run the project formatter on a just-written file. Detection: `formatter`
+ * config in ToolContext (true → auto-detect prettier/biome/denofmt from
+ * node_modules; string → explicit binary/script). Fire-and-observe: errors
+ * are appended to the tool output (LLM self-corrects), never thrown.
+ */
+function formatFile(ctx, abs) {
+    const fmt = ctx.formatter;
+    if (!fmt)
+        return null;
+    let cmd;
+    if (typeof fmt === "string" && fmt !== "auto") {
+        cmd = fmt.startsWith("npm:")
+            ? ["npm", "run", fmt.slice(4), "--", abs]
+            : [fmt, "--write", abs];
+    }
+    else {
+        // auto-detect (formatter: true) — trên Windows, shim trong node_modules/.bin
+        // là file .cmd (file không extension chỉ chạy được trên POSIX).
+        const localBin = (name) => {
+            const base = join(ctx.workdir, "node_modules", ".bin", name);
+            if (process.platform === "win32" && existsSync(`${base}.cmd`))
+                return `${base}.cmd`;
+            if (existsSync(base))
+                return base;
+            return null;
+        };
+        const prettier = localBin("prettier");
+        const biome = localBin("biome");
+        const eslint = localBin("eslint");
+        if (prettier)
+            cmd = [prettier, "--write", abs];
+        else if (biome)
+            cmd = [biome, "format", "--write", abs];
+        else if (eslint)
+            cmd = [eslint, "--fix", abs];
+        else
+            return null;
+    }
+    const r = spawnSync(cmd[0], cmd.slice(1), {
+        cwd: ctx.workdir,
+        encoding: "utf8",
+        timeout: FORMATTER_TIMEOUT_MS,
+        stdio: ["ignore", "pipe", "pipe"],
+        // Windows: .cmd shim (npm.cmd, prettier.cmd) chỉ chạy được qua shell.
+        shell: process.platform === "win32",
+    });
+    if (r.error)
+        return `formatter notice: ${r.error.message.split("\n")[0]}`;
+    if (r.status !== 0) {
+        return `formatter notice: ${cmd[0]} exited ${r.status}\n${(r.stderr ?? r.stdout ?? "").slice(0, 500)}`;
+    }
+    return null;
+}
+// ── G3: child env allowlist ─────────────────────────────────────────────────
+// spawn /bin/sh -c with the FULL process.env leaks every secret in the shell's
+// environment (API keys, tokens) to any command the model runs — and thus to
+// the conversation/log via `env` or a misbehaving tool. Allowlist instead.
+/** Always passed through — minimum for node/python/git builds to work. */
+const BASE_ENV_KEYS = [
+    "PATH", "HOME", "LANG", "LC_ALL", "TMPDIR", "TERM", "SHELL", "USER", "LOGNAME", "XDG_CONFIG_HOME", "XDG_CACHE_HOME", "XDG_DATA_HOME",
+    // Windows essentiels — cmd.exe/npm.cmd cần SystemRoot/ComSpec/PATHEXT để chạy;
+    // Node trên Windows tra env key không phân biệt hoa thường nên viết hoa cũng khớp.
+    "SYSTEMROOT", "SYSTEMDRIVE", "COMSPEC", "PATHEXT", "WINDIR", "TEMP", "TMP",
+    "USERPROFILE", "HOMEDRIVE", "HOMEPATH", "APPDATA", "LOCALAPPDATA", "PROGRAMDATA",
+    "PROGRAMFILES", "PROGRAMFILES(X86)", "PUBLIC", "NUMBER_OF_PROCESSORS", "PROCESSOR_ARCHITECTURE",
+];
+/**
+ * G3: build the child env. Base keys + SCODER_* config + explicit extras from
+ * the rc (`shellEnvAllow`), resolved from process.env. Missing keys are absent
+ * (not empty) so tools detect them correctly.
+ */
+export function buildChildEnv(workdir, extraKeys) {
+    const rc = loadRcFile(workdir);
+    const allow = [...BASE_ENV_KEYS, ...(rc.shellEnvAllow ?? []), ...(extraKeys ?? [])];
+    const out = {};
+    for (const k of allow) {
+        const v = process.env[k];
+        if (v !== undefined)
+            out[k] = v;
+    }
+    // Windows + Git Bash: không có HOME — git/bash đọc config từ ~ (.gitconfig,
+    // ssh keys). Map HOME → USERPROFILE để git hoạt động đúng user hiện tại.
+    if (process.platform === "win32" && out.HOME === undefined && out.USERPROFILE) {
+        out.HOME = out.USERPROFILE;
+    }
+    return out;
+}
+/**
+ * Shell mặc định theo platform:
+ * - Unix → /bin/sh -c.
+ * - Windows → Git Bash nếu tìm thấy (model hay emit cú pháp POSIX: pipes, grep,
+ *   rm -rf — cmd.exe không chạy được), fallback cmd.exe /d /c.
+ *   KHÔNG dùng `where bash` — System32\bash.exe là WSL (sai filesystem).
+ */
+export function defaultShellArgv(command) {
+    if (process.platform === "win32") {
+        const bash = findGitBash();
+        if (bash)
+            return [bash, "-c", command];
+        return [process.env.ComSpec || process.env.COMSPEC || "cmd.exe", "/d", "/c", command];
+    }
+    return ["/bin/sh", "-c", command];
+}
+/** Locate Git for Windows' bash.exe (cached); null when not installed. */
+let gitBashCache;
+function findGitBash() {
+    if (gitBashCache !== undefined)
+        return gitBashCache;
+    const pf = process.env.ProgramFiles ?? "C:\\Program Files";
+    const pf86 = process.env["ProgramFiles(x86)"] ?? "C:\\Program Files (x86)";
+    const local = process.env.LOCALAPPDATA;
+    const candidates = [
+        join(pf, "Git", "bin", "bash.exe"),
+        join(pf86, "Git", "bin", "bash.exe"),
+        ...(local ? [join(local, "Programs", "Git", "bin", "bash.exe")] : []),
+    ];
+    for (const c of candidates) {
+        if (existsSync(c)) {
+            gitBashCache = c;
+            return c;
+        }
+    }
+    gitBashCache = null;
+    return null;
+}
 /**
  * Run a shell command asynchronously (async spawn keeps the event loop free,
  * so streaming output renders and SIGINT is handled promptly).
  * Kills the child on timeout or external abort (SIGINT via ctx.signal).
  * When `onOutput` is provided, stdout/stderr chunks are forwarded live as they
  * arrive (in addition to being collected into the final ToolResult).
+ * G3: output is redacted (secret shapes) before it enters the conversation,
+ * the live stream, or the JSONL log.
  */
-export async function toolRunCommand(workdir, args, signal, onOutput) {
+export async function toolRunCommand(workdir, args, signal, onOutput, sandbox) {
     const timeoutMs = Math.min(Math.max(args.timeout_secs ?? 120, 1), 600) * 1000;
+    // G3: the echoed command line itself may contain a secret (e.g. `curl -H
+    // "Authorization: Bearer sk-..."`) — redact the echo, never run redacted
+    // (the child gets the raw command).
+    const echoLine = `$ ${redactSecrets(args.command)}`;
+    // G4: sandbox wrap khi config bật. Wrapper thiếu → warn + chạy trực tiếp.
+    const sandboxed = await resolveSandbox(sandbox ?? {}, workdir, args.command);
+    if (sandboxed.argv) {
+        ui.info(`sandbox: ${sandboxed.wrapper} (${(sandbox ?? {}).mode})`);
+    }
+    else if ((sandbox ?? {}).mode && (sandbox ?? {}).mode !== "off") {
+        ui.warn(`sandbox: ${sandboxed.reason}`);
+    }
+    const spawnArgv = sandboxed.argv ?? defaultShellArgv(args.command);
     return new Promise((resolve) => {
         let settled = false;
         const finish = (result) => {
@@ -246,9 +614,10 @@ export async function toolRunCommand(workdir, args, signal, onOutput) {
             signal?.removeEventListener("abort", onAbort);
             resolve(result);
         };
-        const child = spawn("/bin/sh", ["-c", args.command], {
+        const child = spawn(spawnArgv[0], spawnArgv.slice(1), {
             cwd: workdir,
             stdio: ["ignore", "pipe", "pipe"],
+            env: buildChildEnv(workdir), // G3: allowlist — không kế thừa full process.env
         });
         let stdout = "";
         let stderr = "";
@@ -277,39 +646,42 @@ export async function toolRunCommand(workdir, args, signal, onOutput) {
         child.stdout?.on("data", (c) => {
             const s = c.toString("utf8");
             stdout += s;
-            onOutput?.("stdout", s);
+            onOutput?.("stdout", redactSecrets(s)); // G3
         });
         child.stderr?.on("data", (c) => {
             const s = c.toString("utf8");
             stderr += s;
-            onOutput?.("stderr", s);
+            onOutput?.("stderr", redactSecrets(s)); // G3
         });
         child.on("error", (err) => {
-            finish({ ok: false, output: clamp(`$ ${args.command}\n(spawn error) ${err.message}`) });
+            finish({ ok: false, output: clamp(`${echoLine}\n(spawn error) ${err.message}`) });
         });
         child.on("close", (code) => {
             if (killedReason === "timeout") {
                 finish({
                     ok: false,
-                    output: clamp(`$ ${args.command}\n(command timed out after ${Math.round(timeoutMs / 1000)}s and was killed)\n--- stdout ---\n${stdout}\n--- stderr ---\n${stderr}`),
+                    output: clamp(`${echoLine}\n(command timed out after ${Math.round(timeoutMs / 1000)}s and was killed)\n--- stdout ---\n${redactSecrets(stdout)}\n--- stderr ---\n${redactSecrets(stderr)}`),
                 });
                 return;
             }
             if (killedReason === "aborted") {
                 finish({
                     ok: false,
-                    output: clamp(`$ ${args.command}\n(interrupted by user)\n--- stdout ---\n${stdout}\n--- stderr ---\n${stderr}`),
+                    output: clamp(`${echoLine}\n(interrupted by user)\n--- stdout ---\n${redactSecrets(stdout)}\n--- stderr ---\n${redactSecrets(stderr)}`),
                 });
                 return;
             }
             const status = code ?? -1;
+            // G3: redact trước khi output vào conversation/log/TUI.
+            const redacted = redactSecrets(stdout);
+            const redactedErr = redactSecrets(stderr);
             if (status === 0) {
-                finish({ ok: true, output: clamp(`$ ${args.command}\n(exit 0)\n${stdout}`) });
+                finish({ ok: true, output: clamp(`${echoLine}\n(exit 0)\n${redacted}`) });
             }
             else {
                 finish({
                     ok: false,
-                    output: clamp(`$ ${args.command}\n(exit ${status})\n--- stdout ---\n${stdout}\n--- stderr ---\n${stderr}`),
+                    output: clamp(`${echoLine}\n(exit ${status})\n--- stdout ---\n${redacted}\n--- stderr ---\n${redactedErr}`),
                 });
             }
         });
@@ -323,6 +695,69 @@ const obj = (properties, required) => ({
 });
 const str = (description) => ({ type: "string", description });
 export const AGENT_TOOLS = [
+    {
+        spec: {
+            type: "function",
+            function: {
+                name: "load_skill",
+                description: "Load the full instructions of a named skill from the AVAILABLE SKILLS index (project + global). Call this at the start of a task when it matches a skill, then follow the returned instructions exactly.",
+                parameters: obj({ name: str("Skill name from the AVAILABLE SKILLS index, e.g. 'prompt-gate-workflow'") }, ["name"]),
+            },
+        },
+        run: (ctx, args) => {
+            const name = String(args.name ?? "").trim();
+            if (!name)
+                return { ok: false, output: "ERROR: name is required — pick one from the AVAILABLE SKILLS index." };
+            const skill = findSkill(ctx.workdir, name);
+            if (!skill) {
+                const names = discoverSkills(ctx.workdir).map((s) => s.name).join(", ");
+                return { ok: false, output: `ERROR: skill "${name}" not found${names ? ` — available: ${names}` : " — no skills discovered"}.` };
+            }
+            return { ok: true, output: `# skill: ${skill.name} (${skill.scope})\n\n${skill.body}` };
+        },
+    },
+    {
+        spec: {
+            type: "function",
+            function: {
+                name: "write_todos",
+                description: "Write the full task plan (replaces the previous list). Use at the start of a multi-step task and update as you go: mark items completed IMMEDIATELY after finishing them, set exactly one item in_progress, and add newly discovered steps. The user sees this list live.",
+                parameters: obj({
+                    todos: {
+                        type: "array",
+                        description: "The complete todo list (full replacement).",
+                        items: {
+                            type: "object",
+                            properties: {
+                                content: { type: "string", description: "Imperative step, e.g. 'Fix null check in UserService'" },
+                                status: { type: "string", enum: ["pending", "in_progress", "completed"] },
+                            },
+                            required: ["content", "status"],
+                            additionalProperties: false,
+                        },
+                    },
+                }, ["todos"]),
+            },
+        },
+        run: (ctx, args) => {
+            const todos = args.todos;
+            if (!Array.isArray(todos))
+                return { ok: false, output: "ERROR: todos must be an array of {content, status}." };
+            const clean = [];
+            for (const t of todos) {
+                if (typeof t?.content !== "string" || !t.content.trim()) {
+                    return { ok: false, output: "ERROR: every todo needs a non-empty content string." };
+                }
+                if (!["pending", "in_progress", "completed"].includes(t.status)) {
+                    return { ok: false, output: `ERROR: invalid status "${String(t.status)}" (pending|in_progress|completed).` };
+                }
+                clean.push({ content: t.content.trim().slice(0, 200), status: t.status });
+            }
+            ctx.plan?.set(clean);
+            const done = clean.filter((t) => t.status === "completed").length;
+            return { ok: true, output: `Plan updated: ${done}/${clean.length} completed.` };
+        },
+    },
     {
         spec: {
             type: "function",
@@ -339,7 +774,7 @@ export const AGENT_TOOLS = [
             type: "function",
             function: {
                 name: "read_file",
-                description: "Read a text file. Returns content with 1-based line numbers.",
+                description: "Read a text file. Returns content with 1-based line numbers. Image files (PNG/JPG/GIF) return an `[image: path]` placeholder rendered inline by the TUI.",
                 parameters: obj({ path: str("Relative file path") }, ["path"]),
             },
         },
@@ -376,14 +811,14 @@ export const AGENT_TOOLS = [
             type: "function",
             function: {
                 name: "run_command",
-                description: "Run a shell command (/bin/sh -c) in the working directory. Use for builds, tests, git, installing deps, etc.",
+                description: "Run a shell command in the working directory (sh/bash -c; Windows: Git Bash when installed, else cmd /c). Use for builds, tests, git, installing deps, etc.",
                 parameters: obj({
                     command: str("Shell command to run"),
                     timeout_secs: str("Timeout in seconds (default 120, max 600)"),
                 }, ["command"]),
             },
         },
-        run: (ctx, args) => toolRunCommand(ctx.workdir, args, ctx.signal, ctx.onToolOutput ? (stream, delta) => ctx.onToolOutput?.("run_command", stream, delta) : undefined),
+        run: (ctx, args) => toolRunCommand(ctx.workdir, args, ctx.signal, ctx.onToolOutput ? (stream, delta) => ctx.onToolOutput?.("run_command", stream, delta) : undefined, ctx.sandbox),
     },
     {
         spec: {
@@ -406,7 +841,66 @@ export const AGENT_TOOLS = [
         },
         run: (ctx, args) => toolWebFetch(ctx, args),
     },
+    {
+        spec: {
+            type: "function",
+            function: {
+                name: "grep",
+                description: "Search file contents with a regex (ripgrep-backed, gitignore-aware). Returns path:line:text hits, max 50. " +
+                    "Use for finding code by symbol, error string, or pattern. Combine with read_file for context.",
+                parameters: obj({
+                    pattern: str("Regular expression to search for"),
+                    path: str("Directory or file to search (default: workdir root)"),
+                    include: str("Glob filter for filenames, e.g. '*.ts'"),
+                }, ["pattern"]),
+            },
+        },
+        run: (ctx, args) => toolGrep(ctx.workdir, args),
+    },
+    {
+        spec: {
+            type: "function",
+            function: {
+                name: "glob",
+                description: "Find files by glob pattern (supports **, *, ?, {a,b}). Returns matching paths sorted, max 200. " +
+                    "Use when you know the filename shape but not the location.",
+                parameters: obj({
+                    pattern: str("Glob pattern relative to workdir, e.g. 'src/**/*.ts'"),
+                    path: str("Directory to search (default: workdir root)"),
+                }, ["pattern"]),
+            },
+        },
+        run: (ctx, args) => toolGlob(ctx.workdir, args),
+    },
 ];
+/** task tool — runner resolved from ctx.taskRunner (per-agent, no global). */
+const taskTool = {
+    spec: {
+        type: "function",
+        function: {
+            name: "task",
+            description: "Delegate a self-contained subtask to a subagent that runs its own agent loop with its own context. " +
+                "WHEN TO USE: broad codebase exploration (find where X lives, summarize module Y) so the main " +
+                "conversation stays lean, or an independent subtask that needs many steps. " +
+                "The prompt must be self-contained (the subagent sees nothing else); it returns a summary. " +
+                "Prefer this over reading dozens of files yourself.",
+            parameters: obj({
+                description: str("One-line summary of the subtask (shown in the UI)"),
+                prompt: str("Self-contained instructions for the subagent, including any file paths and the expected output"),
+                agent: str("Optional agent name from the available agents (default: explore for research, build for changes)"),
+            }, ["description", "prompt"]),
+        },
+    },
+    run: (ctx, args) => {
+        if (!ctx.taskRunner)
+            return { ok: false, output: "ERROR: task tool not wired (no runner in context)." };
+        return ctx.taskRunner(ctx, args);
+    },
+};
 export function getTool(name) {
-    return AGENT_TOOLS.find((t) => t.spec.function.name === name);
+    return AGENT_TOOLS.find((t) => t.spec.function.name === name) ?? (name === "task" ? taskTool : undefined);
+}
+/** All tool specs incl. task (agent adds MCP specs on top). */
+export function allToolSpecs() {
+    return [...AGENT_TOOLS.map((t) => t.spec), taskTool.spec];
 }