@pentoshi/clai 0.13.0 → 1.0.1

package/dist/agent/runner.js

@@ -19,12 +19,25 @@ import { ensureProviderConfigured } from "../commands/providers.js";
 import { rememberThinkingFromText, renderThinkingSummary, } from "../ui/thinking.js";
 import { renderMarkdown, indentAndWrapText } from "../ui/markdown.js";
 import { startThinkingSpinner } from "../ui/spinner.js";
+import { safeCwd } from "../os/cwd.js";
 import { analyzeTask } from "./task-analyzer.js";
 import { LoopGuard } from "./loop-guard.js";
+import { createPlan, loadPlan, savePlan, markTask, } from "../store/plan.js";
+import { renderPlanChecklist, renderPlanSidePane } from "../ui/plan-pane.js";
+/** Render the plan as a right-side pane on wide terminals, else inline. */
+function renderPlanForTerminal(plan) {
+    const cols = process.stdout.columns ?? 0;
+    const side = process.stdout.isTTY
+        ? renderPlanSidePane(plan, cols)
+        : undefined;
+    return side ?? renderPlanChecklist(plan);
+}
 export function createSessionPolicy() {
     return {
         allow: new Set(),
         pentestAuthorized: { value: false },
+        sessionId: `sess-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 8)}`,
+        planApproved: { value: false },
     };
 }
 function tryParseCall(raw) {
@@ -139,6 +152,147 @@ export function parseToolCall(text, options = {}) {
     }
     return undefined;
 }
+// Argument keys that the built-in tools accept. Used to recognize when a
+// model emitted a bare args object (e.g. {"path":"file.pdf"}) — intending a
+// tool call but forgetting the {"name","args"} wrapper and the ```tool fence.
+const TOOL_ARG_KEYS = new Set([
+    "command",
+    "path",
+    "paths",
+    "url",
+    "query",
+    "target",
+    "pattern",
+    "tool",
+    "tools",
+    "files",
+    "content",
+    "calls",
+    "record",
+    "ports",
+    "profile",
+    "id",
+    "lang",
+    "dpi",
+    "psm",
+    "recursive",
+    "oldText",
+    "newText",
+    "expectedReplacements",
+    "goal",
+    "tasks",
+    "taskId",
+    "state",
+    "method",
+    "body",
+    "headers",
+    "maxBytes",
+    "maxResults",
+    "cwd",
+    "name",
+    "concurrency",
+]);
+/**
+ * Strip a single wrapping ```json / ``` fence (if any) and return the inner
+ * text trimmed. Leaves un-fenced text unchanged.
+ */
+function stripLoneFence(text) {
+    const fenced = text
+        .trim()
+        .match(/^```[a-zA-Z]*\s*\n?([\s\S]*?)\n?```$/);
+    return (fenced?.[1] ?? text).trim();
+}
+/**
+ * When a model means to call a tool but emits ONLY a bare JSON object —
+ * either a proper {"name","args"} that the strict matchers missed, or a bare
+ * args object like {"path":"file.pdf"} with the wrapper/fence dropped — this
+ * recognizes it. Returns:
+ *   - { call } when the object is a complete {name, args} tool call, or
+ *   - { argsOnly: true } when it looks like a bare args object (so the caller
+ *     can nudge the model to re-emit a properly named, fenced tool call).
+ * Returns undefined for anything that is plainly a normal prose/JSON answer.
+ */
+export function recognizeBareToolJson(text) {
+    const inner = stripLoneFence(text);
+    // Must be a single JSON object spanning the whole (de-fenced) output.
+    if (!inner.startsWith("{") || !inner.endsWith("}"))
+        return undefined;
+    let parsed;
+    try {
+        parsed = JSON.parse(inner);
+    }
+    catch {
+        return undefined;
+    }
+    if (!parsed || typeof parsed !== "object" || Array.isArray(parsed)) {
+        return undefined;
+    }
+    const obj = parsed;
+    // Complete {name, args} call the earlier matchers didn't catch (e.g. not
+    // anchored to end-of-string). Recover it directly.
+    const direct = tryParseCall(inner);
+    if (direct)
+        return { call: direct };
+    // Bare args object: every key is a known tool-arg key, and it carries at
+    // least one identifying arg. Don't treat huge/odd objects as tool args.
+    const keys = Object.keys(obj);
+    if (keys.length === 0 || keys.length > 6)
+        return undefined;
+    const allKnown = keys.every((key) => TOOL_ARG_KEYS.has(key));
+    if (allKnown)
+        return { argsOnly: true };
+    return undefined;
+}
+/**
+ * Detect an opened-but-unparseable tool call. This happens when the model's
+ * output is truncated by the token limit mid-JSON: we see the ```tool fence
+ * (or a bare {"name":"...","args" prefix) open, but parseToolCall returns
+ * undefined because the JSON never closed. Without this, the broken block
+ * leaks to the screen as a "final answer" and the requested action (e.g. a
+ * multi-file fs.writeMany scaffold) silently never runs.
+ */
+export function looksLikeTruncatedToolCall(text) {
+    // An opened ```tool fence with no closing fence.
+    const openFence = /```tool\s*\n?/i.test(text);
+    const closeFence = /```tool[\s\S]*?```/i.test(text);
+    if (openFence && !closeFence)
+        return true;
+    // A tool-call JSON object that started but whose braces never balanced.
+    const jsonStart = text.search(/\{\s*"name"\s*:\s*"[A-Za-z][\w.]*"\s*,\s*"args"/);
+    if (jsonStart >= 0) {
+        const slice = text.slice(jsonStart);
+        let depth = 0;
+        let inString = false;
+        let escaped = false;
+        let balanced = false;
+        for (const ch of slice) {
+            if (escaped) {
+                escaped = false;
+                continue;
+            }
+            if (ch === "\\") {
+                escaped = true;
+                continue;
+            }
+            if (ch === '"')
+                inString = !inString;
+            if (inString)
+                continue;
+            if (ch === "{")
+                depth += 1;
+            else if (ch === "}") {
+                depth -= 1;
+                if (depth === 0) {
+                    balanced = true;
+                    break;
+                }
+            }
+        }
+        if (!balanced)
+            return true;
+    }
+    return false;
+}
 /** Extract the text before the tool call block for display purposes */
 function textBeforeToolCall(text) {
     const patterns = [
@@ -177,13 +331,17 @@ function formatToolArgs(call) {
     if (call.name === "fs.writeMany") {
         const files = Array.isArray(call.args.files) ? call.args.files : [];
         const names = files
-            .map((f) => f && typeof f === "object" ? String(f.path ?? "") : "")
+            .map((f) => f && typeof f === "object"
+            ? String(f.path ?? "")
+            : "")
             .filter(Boolean);
         const preview = names.slice(0, 4).join(", ");
         return `${names.length} file(s)${preview ? `: ${preview}${names.length > 4 ? ", …" : ""}` : ""}`;
     }
     if (call.name === "fs.search")
         return String(call.args.pattern ?? "");
+    if (call.name === "image.ocr" || call.name === "pdf.read")
+        return String(call.args.path ?? "");
     if (call.name === "http.fetch" || call.name === "web.fetch")
         return String(call.args.url ?? "");
     if (call.name === "web.search")
@@ -191,7 +349,7 @@ function formatToolArgs(call) {
     if (call.name === "pkg.install")
         return String(call.args.tool ?? "");
     if (call.name === "fs.list")
-        return String(call.args.path ?? process.cwd());
+        return String(call.args.path ?? safeCwd());
     return JSON.stringify(call.args);
 }
 const VOLATILE_SIGNAL_RE = /\b(?:current(?:ly)?|latest|newest|today|now|right now|live|recent|breaking|news|release[sd]?|version|prices?|stocks?|market|rates?|weather|forecast|elections?|results?|rankings?|standings?|stats?|cve|advis(?:ory|ories)|vulnerabilit(?:y|ies))\b/i;
@@ -210,6 +368,12 @@ const BUILD_STACK_RE = /\b(?:react|next(?:\.?js)?|vue|svelte|angular|vite|webpac
 // clearly mean "keep going with what we were doing".
 const CONTINUATION_RE = /^(?:do\s+it|build\s+it|build\s+fully|build\s+it\s+fully|go\s+ahead|continue|proceed|keep\s+going|finish(?:\s+it)?|complete(?:\s+it)?|yes|ok(?:ay)?|make\s+it|run\s+it|next|on\s+your\s+own|build\s+(?:fully\s+)?on\s+your\s+own)\b/i;
 const INCOMPLETE_RE = /\b(?:not\s+complete|incomplete|isn'?t\s+(?:done|complete|working|finished)|doesn'?t\s+work|still\s+(?:broken|missing|failing)|missing\s+(?:files?|parts?)|finish\s+(?:the|it)|complete\s+(?:the|it))\b/i;
+// The synthetic message injected when the user runs /implement to approve a
+// plan ("I approve the plan. Execute it now, task by task…"). It must always
+// count as a build/continuation turn — it contains the word "now", which
+// would otherwise trip the volatile-info freshness guard and divert the run
+// into a pointless web.search instead of executing the plan.
+const PLAN_EXECUTION_RE = /\b(?:approve the plan|execute it (?:now|task by task)|task by task|execute the plan|implement the plan)\b/i;
 /**
  * Decide whether this turn should get a generous step budget because it is
  * a multi-file build, a continuation of one, or a "it's not done yet" nudge.
@@ -221,7 +385,8 @@ export function looksLikeBuildTask(prompt, history) {
     if (BUILD_TASK_RE.test(text) ||
         BUILD_STACK_RE.test(text) ||
         CONTINUATION_RE.test(text) ||
-        INCOMPLETE_RE.test(text)) {
+        INCOMPLETE_RE.test(text) ||
+        PLAN_EXECUTION_RE.test(text)) {
         return true;
     }
     // Inspect recent history: if the conversation was already about building
@@ -245,11 +410,34 @@ export function requiresFreshWebSearch(prompt) {
     if (STATIC_DISAMBIGUATION_RE.test(text) || LOCAL_RUNTIME_RE.test(text)) {
         return false;
     }
+    // Plan-execution and terse continuation turns are never "fetch current
+    // info" turns, even when they contain words like "now". (We intentionally
+    // do NOT exclude on build-stack keywords here — "latest vite version" is a
+    // legitimate version lookup. The runAgentLoop caller additionally gates the
+    // guard on looksLikeBuildTask so a real scaffold turn never searches.)
+    if (PLAN_EXECUTION_RE.test(text) || CONTINUATION_RE.test(text)) {
+        return false;
+    }
     return (VOLATILE_SIGNAL_RE.test(text) ||
         VOLATILE_ROLE_QUERY_RE.test(text) ||
         ROLE_OF_ENTITY_RE.test(text) ||
         EXPLICIT_WEB_LOOKUP_RE.test(text));
 }
+/**
+ * Detect a low-quality "everything in one step" plan task. A single task that
+ * itself enumerates many files/actions (multiple commas, an "and", several
+ * slashes, or an overlong title) means the model lumped the whole build into
+ * one checkbox instead of producing a real ordered checklist.
+ */
+export function isLumpedSingleTask(taskTitles) {
+    if (taskTitles.length !== 1)
+        return false;
+    const only = taskTitles[0];
+    return ((only.match(/,/g)?.length ?? 0) >= 2 ||
+        /\band\b/i.test(only) ||
+        (only.match(/\//g)?.length ?? 0) >= 2 ||
+        only.length > 90);
+}
 function freshnessGuardMessage(now = new Date()) {
     return (`Freshness guard for this turn: the latest user prompt appears to ask for current, volatile, or externally verifiable information. The present moment is ${currentDateTimeContext(now)}. ` +
         "Before answering, call web.search FIRST with a concise query derived from the user prompt. " +
@@ -378,12 +566,161 @@ async function confirmToolExecution(call, autoConfirm, session) {
         default: true,
     });
 }
+/** Build the system-context block describing the session's active plan. */
+function planContextMessage(plan, approved) {
+    const lines = [];
+    lines.push(`ACTIVE PLAN for this session (goal: ${plan.goal}, status: ${plan.status}):`);
+    if (plan.detail.trim())
+        lines.push(plan.detail.trim());
+    lines.push("Tasks:");
+    plan.tasks.forEach((t, i) => {
+        lines.push(`  ${i + 1}. [${t.id}] (${t.state}) ${t.title}`);
+    });
+    if (approved) {
+        lines.push("The user APPROVED this plan. Execute it task by task NOW: before starting a task call " +
+            'task.update with {"taskId":"<id>","state":"in_progress"}, do the work with real tool calls, ' +
+            'then call task.update {"taskId":"<id>","state":"done"} (or "failed"/"skipped" with a note). ' +
+            "Actually run installs and start servers — never claim something ran without a successful tool call. " +
+            "When all tasks are done, verify and give a final summary.");
+    }
+    else {
+        lines.push("This plan is NOT yet approved. If the user is refining it, update it with plan.create again. " +
+            "Do NOT execute tasks until the user runs /implement.");
+    }
+    return lines.join("\n");
+}
+/**
+ * Handle plan.create / task.update inline. These are session-scoped and
+ * persisted via the plan store so the user can view the plan (Ctrl+P) and
+ * the agent keeps it in context across the whole session.
+ */
+async function handlePlanTool(call, session, ctx) {
+    void ctx;
+    if (call.name === "plan.create") {
+        const goal = typeof call.args.goal === "string" ? call.args.goal : "";
+        const detail = typeof call.args.detail === "string" ? call.args.detail : "";
+        const kind = typeof call.args.kind === "string" ? call.args.kind : "general";
+        const rawTasks = Array.isArray(call.args.tasks) ? call.args.tasks : [];
+        const taskTitles = rawTasks
+            .map((t) => (typeof t === "string" ? t : ""))
+            .filter(Boolean);
+        if (!goal || taskTitles.length === 0) {
+            return {
+                handled: true,
+                ok: false,
+                display: chalk.red("  ✗ plan.create needs a non-empty goal and at least one task title\n"),
+                modelNote: "plan.create failed: provide a string goal and a non-empty tasks array of step titles.",
+            };
+        }
+        // Reject a low-quality "everything in one step" plan. A single task that
+        // itself enumerates many files/actions (commas, "and", slashes) is a sign
+        // the model lumped the whole build into one checkbox — split it so the
+        // user gets a real, trackable checklist and the executor works step by step.
+        if (isLumpedSingleTask(taskTitles)) {
+            return {
+                handled: true,
+                ok: false,
+                display: chalk.red("  ✗ plan.create: that single task lumps the whole build into one step\n"),
+                modelNote: "plan.create rejected: you put everything into ONE task. Break it into 3-8 SEPARATE, " +
+                    "ordered tasks — each a distinct action, e.g. 'scaffold package.json + vite config', " +
+                    "'create index.html + entry (main.jsx)', 'build App + Post components', 'add posts data + styles', " +
+                    "'install deps and run dev server to verify'. Call plan.create again with that tasks array.",
+            };
+        }
+        const plan = createPlan({
+            sessionId: session.sessionId,
+            goal,
+            detail,
+            taskTitles,
+            kind,
+        });
+        await savePlan(plan).catch(() => undefined);
+        // A freshly (re)created plan resets approval — the user must /implement.
+        session.planApproved.value = false;
+        const checklist = renderPlanForTerminal(plan);
+        const display = chalk.cyan("  ● planning\n") +
+            checklist +
+            "\n" +
+            chalk.dim("  ✦ plan created — press Ctrl+P to view it, or type /implement to approve and run it\n");
+        return {
+            handled: true,
+            ok: true,
+            display,
+            modelNote: `Plan saved with ${plan.tasks.length} task(s). STOP here and wait. ` +
+                "Do NOT start executing tasks until the user approves with /implement. " +
+                "When approved you will receive a message telling you to begin; then work task by task, " +
+                "calling task.update to mark each in_progress before and done after you finish it.",
+        };
+    }
+    // task.update
+    const plan = await loadPlan(session.sessionId).catch(() => undefined);
+    if (!plan) {
+        return {
+            handled: true,
+            ok: false,
+            display: chalk.red("  ✗ task.update: no active plan — call plan.create first\n"),
+            modelNote: "task.update failed: there is no active plan. Call plan.create first.",
+        };
+    }
+    const taskId = typeof call.args.taskId === "string" ? call.args.taskId : "";
+    const stateRaw = typeof call.args.state === "string" ? call.args.state : "";
+    const note = typeof call.args.note === "string" ? call.args.note : undefined;
+    const validStates = [
+        "pending",
+        "in_progress",
+        "done",
+        "failed",
+        "skipped",
+    ];
+    if (!validStates.includes(stateRaw)) {
+        return {
+            handled: true,
+            ok: false,
+            display: chalk.red(`  ✗ task.update: state must be one of ${validStates.join(", ")}\n`),
+            modelNote: `task.update failed: state must be one of ${validStates.join(", ")}.`,
+        };
+    }
+    const ok = markTask(plan, taskId, stateRaw, note);
+    if (!ok) {
+        const ids = plan.tasks.map((t) => t.id).join(", ");
+        return {
+            handled: true,
+            ok: false,
+            display: chalk.red(`  ✗ task.update: unknown taskId "${taskId}" (have: ${ids})\n`),
+            modelNote: `task.update failed: unknown taskId. Valid ids: ${ids}.`,
+        };
+    }
+    if (plan.status === "draft" || plan.status === "approved") {
+        plan.status = "in_progress";
+    }
+    const allDone = plan.tasks.every((t) => t.state === "done" || t.state === "skipped" || t.state === "failed");
+    if (allDone)
+        plan.status = "completed";
+    await savePlan(plan).catch(() => undefined);
+    const checklist = renderPlanForTerminal(plan);
+    return {
+        handled: true,
+        ok: true,
+        display: checklist + "\n",
+        modelNote: allDone
+            ? "Task updated. ALL tasks are now finished. Verify the result and give your final summary."
+            : "Task updated. Continue with the next pending task.",
+    };
+}
 export async function runAgentLoop(prompt, options = {}) {
     const config = getConfig();
     const maxSteps = options.maxSteps ?? 30;
     const projectContext = await loadProjectContext();
     const toolNames = availableToolNames();
-    const freshWebSearchRequired = toolNames.includes("web.search") && requiresFreshWebSearch(prompt);
+    // Build / scaffold / continuation turns must NEVER be diverted into a
+    // web.search for "current info". The /implement directive ("Execute it
+    // now…") and prompts like "create a react app" contain words such as
+    // "now"/"latest" that trip the volatile-info regex; without this guard the
+    // agent burns its turn searching the date instead of writing files.
+    const buildLikeTurn = looksLikeBuildTask(prompt, options.history);
+    const freshWebSearchRequired = !buildLikeTurn &&
+        toolNames.includes("web.search") &&
+        requiresFreshWebSearch(prompt);
     const systemSections = [renderAgentSystemPrompt(toolNames.join(", "))];
     if (projectContext) {
         systemSections.push(`Project context from .clai/context.md:\n${projectContext}`);
@@ -391,17 +728,39 @@ export async function runAgentLoop(prompt, options = {}) {
     if (freshWebSearchRequired) {
         systemSections.push(freshnessGuardMessage());
     }
-    const fullSystemPrompt = systemSections.join("\n\n");
-    const messages = [
-        { role: "system", content: fullSystemPrompt },
-        ...(options.history ?? []),
-        { role: "user", content: prompt },
-    ];
     let provider = options.provider ?? config.defaultProvider;
     await ensureProviderConfigured(provider);
     let model = options.model ?? config.defaultModel;
     let lastAnswer = "";
     const session = options.session ?? createSessionPolicy();
+    // ── Active plan context ────────────────────────────────────────────
+    // If this session already has a plan, inject it so the model keeps it in
+    // context. When the user has approved it (via /implement) we instruct the
+    // agent to execute task by task; otherwise the agent should refine/wait.
+    const activePlan = await loadPlan(session.sessionId).catch(() => undefined);
+    if (activePlan) {
+        systemSections.push(planContextMessage(activePlan, session.planApproved.value));
+    }
+    const fullSystemPrompt = systemSections.join("\n\n");
+    const userMessage = { role: "user", content: prompt };
+    if (options.images && options.images.length > 0) {
+        userMessage.images = options.images;
+    }
+    const messages = [
+        { role: "system", content: fullSystemPrompt },
+        ...(options.history ?? []),
+        userMessage,
+    ];
+    const recoveryUserMessage = (content) => {
+        const message = { role: "user", content };
+        if (options.images && options.images.length > 0) {
+            // Some OpenAI-compatible gateways/models attend most strongly to the
+            // latest user turn. Keep the image attached on recovery nudges so a
+            // thinking-only retry does not degrade into OCR/tool guessing.
+            message.images = options.images;
+        }
+        return message;
+    };
     // Track recent tool calls to detect models stuck in a loop calling the
     // same tool with the same arguments over and over (e.g. pentest.recon
     // called 3× on the same target without summarizing).
@@ -409,6 +768,13 @@ export async function runAgentLoop(prompt, options = {}) {
     // Track consecutive thinking-only responses so we can nudge the model
     // to actually act instead of silently returning an empty answer.
     let emptyVisibleRetries = 0;
+    // Track tool calls truncated by the token limit so we can ask the model
+    // to retry in smaller pieces instead of leaking broken JSON as an answer.
+    let truncatedToolRetries = 0;
+    // Track bare-args JSON tool calls (missing the {name,args} wrapper / fence)
+    // so we can nudge the model to re-emit a proper fenced call a few times
+    // before giving up, instead of leaking the JSON as a final answer.
+    let bareToolJsonRetries = 0;
     // For volatile live-info prompts, make one corrective pass if a model
     // ignores the freshness guard and tries to answer from stale memory.
     let sawFreshWebSearch = false;
@@ -428,7 +794,7 @@ export async function runAgentLoop(prompt, options = {}) {
     // like a build/scaffold or a continuation of one.
     const analysis = analyzeTask(prompt);
     const hasHistory = (options.history?.length ?? 0) > 0;
-    const buildLike = looksLikeBuildTask(prompt, options.history);
+    const buildLike = buildLikeTurn;
     let stepBudget = analysis.complexity === "simple"
         ? 15
         : analysis.complexity === "standard"
@@ -472,9 +838,11 @@ export async function runAgentLoop(prompt, options = {}) {
                 temperature: 0.2,
                 // Reasoning models can spend a lot on hidden thinking; give
                 // them headroom so the visible answer / tool call isn't
-                // truncated to silence. Keep the no-thinking default lean so
-                // fast models like kimi-k2.6 respond instantly.
-                maxTokens: config.thinking?.enabled ? 8_192 : 4_096,
+                // truncated to silence. The non-thinking budget must be large
+                // enough for a multi-file fs.writeMany payload — a truncated
+                // tool-call JSON fails to parse and used to leak a broken
+                // ```tool block to the screen with no files written.
+                maxTokens: config.thinking?.enabled ? 16_384 : 8_192,
                 signal: options.signal,
                 thinking: config.thinking,
             }, (token) => {
@@ -526,12 +894,10 @@ export async function runAgentLoop(prompt, options = {}) {
                 process.stdout.write(`${renderThinkingSummary(assistantText.thinkContent)}\n`);
                 process.stdout.write(chalk.yellow("  ⚠ model produced only thinking — nudging it to take action\n"));
                 messages.push({ role: "assistant", content: completion.text });
-                messages.push({
-                    role: "user",
-                    content: "You only produced internal reasoning with no visible answer or tool call. " +
-                        "You MUST either call a tool using the ```tool format or provide your final answer. " +
-                        "Do NOT just think — take action NOW.",
-                });
+                messages.push(recoveryUserMessage("You only produced internal reasoning with no visible answer or tool call. " +
+                    "You MUST either call a tool using the ```tool format or provide your final answer. " +
+                    "If images are attached, inspect them directly for visual details (text, colors, layout, spacing, style) instead of using OCR unless explicitly needed. " +
+                    "Do NOT just think — take action NOW."));
                 continue;
             }
             // Exhausted retries — fall through to the normal empty-answer path
@@ -541,10 +907,42 @@ export async function runAgentLoop(prompt, options = {}) {
             // Reset the counter on any successful visible output.
             emptyVisibleRetries = 0;
         }
-        const call = parseToolCall(assistantText.visible, {
+        let call = parseToolCall(assistantText.visible, {
             strict: getConfig().parserStrict,
         });
+        // Recovery: the model meant to call a tool but emitted a bare JSON object
+        // with no ```tool fence — either a complete {name,args} the strict
+        // matchers missed (recover it directly), or just an args object like
+        // {"path":"file.pdf"} with the wrapper dropped (nudge a retry below so
+        // the requested action runs instead of the JSON leaking as the answer).
+        let bareArgsOnly = false;
+        let recoveredFromBareJson = false;
         if (!call) {
+            const bare = recognizeBareToolJson(assistantText.visible);
+            if (bare?.call) {
+                call = bare.call;
+                recoveredFromBareJson = true;
+                process.stdout.write(chalk.dim("  ℹ recovered an unfenced tool call from bare JSON\n"));
+            }
+            else if (bare?.argsOnly) {
+                bareArgsOnly = true;
+            }
+        }
+        if (!call) {
+            if (bareArgsOnly) {
+                bareToolJsonRetries += 1;
+                if (bareToolJsonRetries <= 3) {
+                    process.stdout.write(chalk.yellow("  ⚠ tool call missing its name/fence — asking the model to re-emit a proper ```tool block\n"));
+                    messages.push({ role: "assistant", content: assistantText.visible });
+                    messages.push(recoveryUserMessage("Your previous message was a bare JSON args object with no tool name and no ```tool fence, so NOTHING ran. " +
+                        "Reply with ONLY a fenced ```tool block of the form " +
+                        '`{"name": "<tool>", "args": { ... }}`. For example, to read a PDF:\n' +
+                        '```tool\n{"name":"pdf.read","args":{"path":"/abs/file.pdf"}}\n```\n' +
+                        "Choose the correct tool name for the task and include those args."));
+                    continue;
+                }
+                // Exhausted retries — fall through to the normal answer path.
+            }
             // Detect the case where the model emitted sentinel-style tool-call
             // markers but the body was malformed or truncated. Printing those
             // raw tokens looks like a crash to the user — instead, ask the
@@ -552,15 +950,33 @@ export async function runAgentLoop(prompt, options = {}) {
             if (/<\|tool_call(?:s_section)?_begin\|>|<\|tool_call_argument_begin\|>/i.test(assistantText.visible)) {
                 process.stdout.write(chalk.yellow("  ⚠ tool call was malformed or cut off — asking the model to retry in JSON form\n"));
                 messages.push({ role: "assistant", content: assistantText.visible });
-                messages.push({
-                    role: "user",
-                    content: "Your previous tool call was malformed or truncated. " +
-                        "Reply with ONLY a fenced ```tool block containing valid JSON " +
-                        'of the form `{"name": "<tool>", "args": { ... }}`. ' +
-                        "Do not use <|tool_call_begin|> markers.",
-                });
+                messages.push(recoveryUserMessage("Your previous tool call was malformed or truncated. " +
+                    "Reply with ONLY a fenced ```tool block containing valid JSON " +
+                    'of the form `{"name": "<tool>", "args": { ... }}`. ' +
+                    "Do not use <|tool_call_begin|> markers."));
                 continue;
             }
+            // Detect a tool call that opened but was cut off by the token limit
+            // (most common with a large multi-file fs.writeMany). Retrying with a
+            // nudge to split the work is far better than rendering broken JSON as
+            // a final answer and leaving the project half-created.
+            if (looksLikeTruncatedToolCall(assistantText.visible)) {
+                truncatedToolRetries += 1;
+                if (truncatedToolRetries <= 3) {
+                    process.stdout.write(chalk.yellow("  ⚠ tool call was cut off (output too long) — asking the model to retry in smaller pieces\n"));
+                    messages.push({ role: "assistant", content: assistantText.visible });
+                    messages.push({
+                        role: "user",
+                        content: "Your previous tool call was cut off before it finished — the JSON was incomplete, so NOTHING ran. " +
+                            "Retry now with a COMPLETE, valid ```tool block. " +
+                            "If it was a large fs.writeMany, split it into SMALLER batches (3-5 files per call, and keep each file's content concise) " +
+                            "so the whole JSON fits in one response. Do NOT claim any file was written until a tool call actually succeeds.",
+                    });
+                    continue;
+                }
+                // Exhausted retries — fall through so we don't loop forever, but the
+                // user at least sees the (broken) output and the stop notice.
+            }
             // Normal final-answer path: strip any stray sentinel tokens that
             // somehow leaked into prose so the answer renders cleanly.
             const cleaned = stripSentinelTokens(assistantText.visible);
@@ -612,8 +1028,12 @@ export async function runAgentLoop(prompt, options = {}) {
         if (loopCheck.reason) {
             process.stdout.write(chalk.dim(`  ℹ ${loopCheck.reason}\n`));
         }
-        // Print only non-thinking text before the tool call.
-        const beforeTool = textBeforeToolCall(assistantText.visible);
+        // Print only non-thinking text before the tool call. When the call was
+        // recovered from a bare JSON object (the whole message WAS the call),
+        // there is no prose to show — skip it so we don't echo the raw JSON.
+        const beforeTool = recoveredFromBareJson
+            ? ""
+            : textBeforeToolCall(assistantText.visible);
         if (beforeTool) {
             process.stdout.write(renderMarkdown(beforeTool) + "\n");
         }
@@ -621,6 +1041,25 @@ export async function runAgentLoop(prompt, options = {}) {
             process.stdout.write(`${renderThinkingSummary(assistantText.thinkContent)}\n`);
         }
         messages.push({ role: "assistant", content: assistantText.visible });
+        // ── Plan / task tools (session-scoped, handled inline) ─────────────
+        // These don't go through the generic registry because they need the
+        // session id and mutate the live plan that the user can view (Ctrl+P).
+        if (call.name === "plan.create" || call.name === "task.update") {
+            const planResult = await handlePlanTool(call, session, {
+                loopGuard,
+                step,
+            });
+            if (planResult.handled) {
+                productiveSteps += 1;
+                loopGuard.recordAttempt(step, call.name, call.args, planResult.ok, 0);
+                process.stdout.write(planResult.display);
+                messages.push({
+                    role: "tool",
+                    content: `Tool ${call.name} result (ok=${planResult.ok}):\n${planResult.modelNote}`,
+                });
+                continue;
+            }
+        }
         const scope = await loadScope();
         const decision = classifyToolCall(call, { scope });
         await auditLog("tool.classified", {
@@ -809,7 +1248,9 @@ export async function runAgentLoop(prompt, options = {}) {
                 ? String(call.args.command ?? "").split(/\s+/)[0]
                 : call.name === "net.scan"
                     ? "nmap"
-                    : undefined;
+                    : call.name === "image.ocr"
+                        ? "tesseract"
+                        : undefined;
             if (cmdName) {
                 process.stdout.write(chalk.yellow(`  ⚠ ${cmdName} not found — asking model to install and retry\n`));
                 messages.push({