@workermill/agent 0.1.2 → 0.2.1

package/dist/plan-validator.d.ts

@@ -0,0 +1,82 @@
+/**
+ * Plan Validator for Remote Agent
+ *
+ * Validates execution plans locally before posting to the cloud API.
+ * Implements the same guardrails as the server-side planning pipeline:
+ *   1. File cap: max 5 targetFiles per story (prevents scope explosion)
+ *   2. Critic validation: LLM scores the plan, rejects below threshold
+ *
+ * This ensures remote agent plans get the same quality gates as cloud plans,
+ * even though the planning prompt runs locally via Claude CLI.
+ */
+export interface PlannedStory {
+    id: string;
+    title: string;
+    description: string;
+    persona: string;
+    priority: number;
+    estimatedEffort: "small" | "medium" | "large";
+    dependencies: string[];
+    acceptanceCriteria: string[];
+    targetFiles?: string[];
+    scope?: string;
+}
+export interface ExecutionPlan {
+    summary: string;
+    stories: PlannedStory[];
+    risks: string[];
+    assumptions: string[];
+}
+export interface CriticResult {
+    approved: boolean;
+    score: number;
+    risks: string[];
+    suggestions?: string[];
+    storyFeedback?: Array<{
+        storyId: string;
+        feedback: string;
+        suggestedChanges?: string[];
+    }>;
+}
+declare const AUTO_APPROVAL_THRESHOLD = 85;
+/**
+ * Parse execution plan JSON from raw Claude CLI output.
+ * Mirrors server-side parseExecutionPlan() in planning-agent-local.ts.
+ */
+export declare function parseExecutionPlan(output: string): ExecutionPlan;
+/**
+ * Apply file cap to all stories. Truncates targetFiles > MAX_TARGET_FILES.
+ * Returns details about truncated stories for logging.
+ */
+export declare function applyFileCap(plan: ExecutionPlan): {
+    truncatedCount: number;
+    details: string[];
+};
+/**
+ * Re-serialize plan as a JSON code block for posting to the API.
+ * The server-side parseExecutionPlan() expects ```json ... ``` blocks.
+ */
+export declare function serializePlan(plan: ExecutionPlan): string;
+/**
+ * Build the critic prompt with PRD and plan substituted.
+ */
+export declare function buildCriticPrompt(prd: string, plan: ExecutionPlan): string;
+/**
+ * Parse critic JSON response from raw Claude CLI output.
+ */
+export declare function parseCriticResponse(text: string): CriticResult;
+/**
+ * Run the critic via Claude CLI (lightweight — no tools, just reasoning).
+ * Returns the raw text output.
+ */
+export declare function runCriticCli(claudePath: string, model: string, prompt: string, env: Record<string, string | undefined>): Promise<string>;
+/**
+ * Format critic feedback for appending to the planner prompt on re-run.
+ */
+export declare function formatCriticFeedback(critic: CriticResult): string;
+/**
+ * Run critic validation on a parsed plan.
+ * Returns the critic result, or null if critic fails (non-blocking).
+ */
+export declare function runCriticValidation(claudePath: string, model: string, prd: string, plan: ExecutionPlan, env: Record<string, string | undefined>, taskLabel: string): Promise<CriticResult | null>;
+export { AUTO_APPROVAL_THRESHOLD };

package/dist/plan-validator.js

@@ -0,0 +1,268 @@
+/**
+ * Plan Validator for Remote Agent
+ *
+ * Validates execution plans locally before posting to the cloud API.
+ * Implements the same guardrails as the server-side planning pipeline:
+ *   1. File cap: max 5 targetFiles per story (prevents scope explosion)
+ *   2. Critic validation: LLM scores the plan, rejects below threshold
+ *
+ * This ensures remote agent plans get the same quality gates as cloud plans,
+ * even though the planning prompt runs locally via Claude CLI.
+ */
+import { spawn } from "child_process";
+import chalk from "chalk";
+// ============================================================================
+// CONSTANTS
+// ============================================================================
+const MAX_TARGET_FILES = 5;
+const AUTO_APPROVAL_THRESHOLD = 85;
+// ============================================================================
+// PLAN PARSING
+// ============================================================================
+/**
+ * Parse execution plan JSON from raw Claude CLI output.
+ * Mirrors server-side parseExecutionPlan() in planning-agent-local.ts.
+ */
+export function parseExecutionPlan(output) {
+    const jsonMatch = output.match(/```json\s*([\s\S]*?)\s*```/);
+    if (jsonMatch) {
+        return JSON.parse(jsonMatch[1]);
+    }
+    const rawJsonMatch = output.match(/\{[\s\S]*"stories"[\s\S]*\}/);
+    if (rawJsonMatch) {
+        return JSON.parse(rawJsonMatch[0]);
+    }
+    throw new Error("Could not find JSON execution plan in output");
+}
+// ============================================================================
+// FILE CAP
+// ============================================================================
+/**
+ * Apply file cap to all stories. Truncates targetFiles > MAX_TARGET_FILES.
+ * Returns details about truncated stories for logging.
+ */
+export function applyFileCap(plan) {
+    let truncatedCount = 0;
+    const details = [];
+    for (const story of plan.stories) {
+        if (!story.targetFiles || !Array.isArray(story.targetFiles)) {
+            story.targetFiles = [];
+        }
+        else if (story.targetFiles.length > MAX_TARGET_FILES) {
+            const dropped = story.targetFiles.slice(MAX_TARGET_FILES);
+            details.push(`${story.id}: ${story.targetFiles.length} files → ${MAX_TARGET_FILES} (dropped: ${dropped.join(", ")})`);
+            story.targetFiles = story.targetFiles.slice(0, MAX_TARGET_FILES);
+            truncatedCount++;
+        }
+    }
+    return { truncatedCount, details };
+}
+// ============================================================================
+// PLAN SERIALIZATION
+// ============================================================================
+/**
+ * Re-serialize plan as a JSON code block for posting to the API.
+ * The server-side parseExecutionPlan() expects ```json ... ``` blocks.
+ */
+export function serializePlan(plan) {
+    return "```json\n" + JSON.stringify(plan, null, 2) + "\n```";
+}
+// ============================================================================
+// CRITIC
+// ============================================================================
+/**
+ * Critic prompt — identical to server-side critic-agent.ts CRITIC_PROMPT.
+ */
+const CRITIC_PROMPT = `You are a Senior Architect reviewing an execution plan. Your job is to ensure the plan is appropriately sized for the task.
+
+Review this execution plan against the PRD:
+
+## PRD (Product Requirements Document)
+{{PRD}}
+
+## PROPOSED EXECUTION PLAN
+{{PLAN}}
+
+## Review Guidelines
+
+**IMPORTANT: Match plan size to task complexity**
+
+- Simple tasks (typos, config changes, single-file fixes) = 1 step is CORRECT
+- Medium tasks (2-4 files, small features) = 2-3 steps is appropriate
+- Complex tasks (new systems, security) = 3-5 steps is appropriate
+
+**Do NOT penalize:**
+- Single-step plans for genuinely simple tasks
+- Using one persona when only one skill is needed
+
+**DO check for:**
+1. **Missing Requirements** - Does the plan cover what the PRD asks for?
+2. **Vague Instructions** - Will the worker know what to do?
+3. **Security Issues** - Only for tasks involving auth, user data, or external input
+4. **Unrealistic Scope** - Any step targeting >3 files MUST score below 85 (auto-rejection threshold). Each step should modify at most 3 files. If a step needs more, split it into multiple steps first.
+5. **Missing Operational Steps** - If the PRD requires deployment, provisioning, migrations, or running commands, does the plan include operational steps? Writing code is not the same as deploying it.
+6. **Overlapping File Scope** - If two or more steps share the same targetFiles, this causes parallel merge conflicts. Steps MUST NOT overlap on targetFiles. Deduct 10 points per shared file across steps.
+
+## Scoring Guide
+
+- **90-100**: Plan matches task complexity, requirements covered
+- **75-89**: Minor gaps but fundamentally sound
+- **50-74**: Significant issues or wrong-sized for the task
+- **0-49**: Fundamentally flawed
+
+## Output Format
+
+Respond with ONLY a JSON object (no markdown, no explanation):
+{"approved": boolean, "score": number, "risks": ["risk1", "risk2"], "suggestions": ["suggestion1", "suggestion2"], "storyFeedback": [{"storyId": "step-0", "feedback": "specific feedback", "suggestedChanges": ["change1"]}]}
+
+Rules:
+- approved = true if score >= 85 AND plan is right-sized for task
+- risks = specific issues (empty array if none)
+- suggestions = actionable improvements (empty array if none)
+- storyFeedback = per-step feedback (optional, only for steps that need changes)`;
+/**
+ * Build the critic prompt with PRD and plan substituted.
+ */
+export function buildCriticPrompt(prd, plan) {
+    const planJson = JSON.stringify(plan, null, 2);
+    return CRITIC_PROMPT.replace("{{PRD}}", prd).replace("{{PLAN}}", planJson);
+}
+/**
+ * Parse critic JSON response from raw Claude CLI output.
+ */
+export function parseCriticResponse(text) {
+    let jsonText = text.trim();
+    // Handle markdown code blocks
+    if (jsonText.includes("```")) {
+        const match = jsonText.match(/```(?:json)?\s*([\s\S]*?)```/);
+        if (match)
+            jsonText = match[1].trim();
+    }
+    // Find JSON object if preceded by reasoning text
+    const jsonStart = jsonText.indexOf("{");
+    if (jsonStart > 0) {
+        jsonText = jsonText.substring(jsonStart);
+    }
+    const result = JSON.parse(jsonText);
+    return {
+        approved: result.approved,
+        score: Math.max(0, Math.min(100, Math.round(result.score))),
+        risks: result.risks || [],
+        suggestions: result.suggestions,
+        storyFeedback: Array.isArray(result.storyFeedback)
+            ? result.storyFeedback
+            : undefined,
+    };
+}
+/**
+ * Run the critic via Claude CLI (lightweight — no tools, just reasoning).
+ * Returns the raw text output.
+ */
+export function runCriticCli(claudePath, model, prompt, env) {
+    return new Promise((resolve, reject) => {
+        const proc = spawn(claudePath, [
+            "--print",
+            "--model",
+            model,
+            "--permission-mode",
+            "bypassPermissions",
+        ], {
+            env,
+            stdio: ["pipe", "pipe", "pipe"],
+        });
+        proc.stdin.write(prompt);
+        proc.stdin.end();
+        let stdout = "";
+        let stderr = "";
+        proc.stdout.on("data", (data) => {
+            stdout += data.toString();
+        });
+        proc.stderr.on("data", (data) => {
+            stderr += data.toString();
+        });
+        const timeout = setTimeout(() => {
+            proc.kill("SIGTERM");
+            reject(new Error("Critic CLI timed out after 3 minutes"));
+        }, 180_000);
+        proc.on("exit", (code) => {
+            clearTimeout(timeout);
+            if (code !== 0) {
+                reject(new Error(`Critic CLI failed (exit ${code}): ${stderr.substring(0, 300)}`));
+            }
+            else {
+                resolve(stdout);
+            }
+        });
+        proc.on("error", (err) => {
+            clearTimeout(timeout);
+            reject(err);
+        });
+    });
+}
+/**
+ * Format critic feedback for appending to the planner prompt on re-run.
+ */
+export function formatCriticFeedback(critic) {
+    const lines = [
+        "",
+        "## CRITIC FEEDBACK — Your previous plan was REJECTED",
+        "",
+        `Score: ${critic.score}/100 (need >= ${AUTO_APPROVAL_THRESHOLD} to pass)`,
+        "",
+    ];
+    if (critic.risks.length > 0) {
+        lines.push("### Risks Identified:");
+        for (const risk of critic.risks) {
+            lines.push(`- ${risk}`);
+        }
+        lines.push("");
+    }
+    if (critic.suggestions && critic.suggestions.length > 0) {
+        lines.push("### Required Changes:");
+        for (const suggestion of critic.suggestions) {
+            lines.push(`- ${suggestion}`);
+        }
+        lines.push("");
+    }
+    if (critic.storyFeedback && critic.storyFeedback.length > 0) {
+        lines.push("### Per-Story Feedback:");
+        for (const fb of critic.storyFeedback) {
+            lines.push(`- **${fb.storyId}**: ${fb.feedback}`);
+            if (fb.suggestedChanges) {
+                for (const change of fb.suggestedChanges) {
+                    lines.push(`  - ${change}`);
+                }
+            }
+        }
+        lines.push("");
+    }
+    lines.push("**You MUST address ALL feedback above.** Each story must target at most 5 files.", "Stories MUST NOT overlap on targetFiles. Generate a revised plan.");
+    return lines.join("\n");
+}
+/** Timestamp prefix for console logs */
+function ts() {
+    return chalk.dim(new Date().toLocaleTimeString());
+}
+/**
+ * Run critic validation on a parsed plan.
+ * Returns the critic result, or null if critic fails (non-blocking).
+ */
+export async function runCriticValidation(claudePath, model, prd, plan, env, taskLabel) {
+    const criticPrompt = buildCriticPrompt(prd, plan);
+    console.log(`${ts()} ${taskLabel} ${chalk.dim("Running critic validation...")}`);
+    try {
+        const rawCriticOutput = await runCriticCli(claudePath, model, criticPrompt, env);
+        const result = parseCriticResponse(rawCriticOutput);
+        const statusIcon = result.score >= AUTO_APPROVAL_THRESHOLD
+            ? chalk.green("✓")
+            : chalk.red("✗");
+        console.log(`${ts()} ${taskLabel} ${statusIcon} Critic score: ${result.score}/100 (threshold: ${AUTO_APPROVAL_THRESHOLD})`);
+        return result;
+    }
+    catch (error) {
+        const errMsg = error instanceof Error ? error.message : String(error);
+        console.error(`${ts()} ${taskLabel} ${chalk.yellow("⚠")} Critic failed: ${errMsg.substring(0, 100)}`);
+        return null;
+    }
+}
+export { AUTO_APPROVAL_THRESHOLD };

package/dist/planner.d.ts

@@ -3,7 +3,13 @@
  *
  * Fetches the planning prompt from the cloud API, runs it through
  * Claude CLI locally (using the customer's Claude Max subscription),
- * and posts the raw output back for server-side validation.
+ * validates with a Planner-Critic loop, and posts the approved plan
+ * back for server-side processing.
+ *
+ * Guardrails (matching server-side planning pipeline):
+ *   1. File cap: max 5 targetFiles per story (prevents scope explosion)
+ *   2. Critic validation: LLM scores the plan, rejects below 85/100
+ *   3. Max 3 Planner-Critic iterations before failure
  *
  * Logs are streamed to the cloud dashboard in real-time so the user
  * sees the same planning progress as cloud mode.
@@ -12,8 +18,18 @@ import { type AgentConfig } from "./config.js";
 export interface PlanningTask {
     id: string;
     summary: string;
+    description: string | null;
 }
 /**
- * Run planning for a task: fetch prompt, execute Claude CLI, post result.
+ * Run planning for a task with Planner-Critic validation loop.
+ *
+ * Flow:
+ *   1. Fetch planning prompt from cloud API
+ *   2. Run Claude CLI to generate plan
+ *   3. Parse plan, apply file cap (max 5 files per story)
+ *   4. Run critic validation via Claude CLI
+ *   5. If critic approves (score >= 85): post validated plan to API
+ *   6. If critic rejects: re-run planner with feedback (up to MAX_ITERATIONS)
+ *   7. After MAX_ITERATIONS without approval: fail the task
  */
 export declare function planTask(task: PlanningTask, config: AgentConfig): Promise<boolean>;

package/dist/planner.js

@@ -3,7 +3,13 @@
  *
  * Fetches the planning prompt from the cloud API, runs it through
  * Claude CLI locally (using the customer's Claude Max subscription),
- * and posts the raw output back for server-side validation.
+ * validates with a Planner-Critic loop, and posts the approved plan
+ * back for server-side processing.
+ *
+ * Guardrails (matching server-side planning pipeline):
+ *   1. File cap: max 5 targetFiles per story (prevents scope explosion)
+ *   2. Critic validation: LLM scores the plan, rejects below 85/100
+ *   3. Max 3 Planner-Critic iterations before failure
  *
  * Logs are streamed to the cloud dashboard in real-time so the user
  * sees the same planning progress as cloud mode.
@@ -12,6 +18,9 @@ import chalk from "chalk";
 import { spawn } from "child_process";
 import { findClaudePath } from "./config.js";
 import { api } from "./api.js";
+import { parseExecutionPlan, applyFileCap, serializePlan, runCriticValidation, formatCriticFeedback, AUTO_APPROVAL_THRESHOLD, } from "./plan-validator.js";
+/** Max Planner-Critic iterations before giving up */
+const MAX_ITERATIONS = 3;
 /** Timestamp prefix */
 function ts() {
     return chalk.dim(new Date().toLocaleTimeString());
@@ -220,7 +229,16 @@ function runClaudeCli(claudePath, model, prompt, env, taskId, startTime) {
     });
 }
 /**
- * Run planning for a task: fetch prompt, execute Claude CLI, post result.
+ * Run planning for a task with Planner-Critic validation loop.
+ *
+ * Flow:
+ *   1. Fetch planning prompt from cloud API
+ *   2. Run Claude CLI to generate plan
+ *   3. Parse plan, apply file cap (max 5 files per story)
+ *   4. Run critic validation via Claude CLI
+ *   5. If critic approves (score >= 85): post validated plan to API
+ *   6. If critic rejects: re-run planner with feedback (up to MAX_ITERATIONS)
+ *   7. After MAX_ITERATIONS without approval: fail the task
  */
 export async function planTask(task, config) {
     const taskLabel = chalk.cyan(task.id.slice(0, 8));
@@ -230,47 +248,184 @@ export async function planTask(task, config) {
     const promptResponse = await api.get("/api/agent/planning-prompt", {
         params: { taskId: task.id },
     });
-    const { prompt, model } = promptResponse.data;
+    const { prompt: basePrompt, model } = promptResponse.data;
     const cliModel = model || "sonnet";
-    console.log(`${ts()} ${taskLabel} Running Claude CLI ${chalk.dim(`(model: ${chalk.yellow(cliModel)})`)}`);
-    await postLog(task.id, `${PREFIX} Starting planning agent using anthropic/${cliModel}`);
-    // 2. Run Claude CLI asynchronously with progress logging
     const claudePath = process.env.CLAUDE_CLI_PATH || findClaudePath() || "claude";
     const cleanEnv = { ...process.env };
     delete cleanEnv.CLAUDE_CODE_OAUTH_TOKEN;
     const startTime = Date.now();
-    let rawOutput;
+    // PRD for critic validation: use task description, fall back to summary
+    const prd = task.description || task.summary;
+    // 2. Planner-Critic iteration loop
+    let currentPrompt = basePrompt;
+    let bestPlan = null;
+    let bestScore = 0;
+    for (let iteration = 1; iteration <= MAX_ITERATIONS; iteration++) {
+        const iterLabel = MAX_ITERATIONS > 1 ? ` (attempt ${iteration}/${MAX_ITERATIONS})` : "";
+        if (iteration > 1) {
+            console.log(`${ts()} ${taskLabel} Running Claude CLI${iterLabel} ${chalk.dim(`(model: ${chalk.yellow(cliModel)})`)}`);
+            await postLog(task.id, `${PREFIX} Re-planning${iterLabel} using anthropic/${cliModel}`);
+        }
+        else {
+            console.log(`${ts()} ${taskLabel} Running Claude CLI ${chalk.dim(`(model: ${chalk.yellow(cliModel)})`)}`);
+            await postLog(task.id, `${PREFIX} Starting planning agent using anthropic/${cliModel}`);
+        }
+        // 2a. Run Claude CLI to generate plan
+        let rawOutput;
+        try {
+            rawOutput = await runClaudeCli(claudePath, cliModel, currentPrompt, cleanEnv, task.id, startTime);
+        }
+        catch (error) {
+            const elapsed = Math.round((Date.now() - startTime) / 1000);
+            const errMsg = error instanceof Error ? error.message : String(error);
+            console.error(`${ts()} ${taskLabel} ${chalk.red("✗")} Failed after ${elapsed}s: ${errMsg.substring(0, 100)}`);
+            await postLog(task.id, `${PREFIX} Planning failed after ${formatElapsed(elapsed)}: ${errMsg.substring(0, 200)}`, "error", "error");
+            return false;
+        }
+        const elapsed = Math.round((Date.now() - startTime) / 1000);
+        console.log(`${ts()} ${taskLabel} ${chalk.green("✓")} Claude CLI done ${chalk.dim(`(${elapsed}s, ${rawOutput.length} chars)`)}`);
+        // 2b. Parse plan from raw output
+        let plan;
+        try {
+            plan = parseExecutionPlan(rawOutput);
+        }
+        catch (error) {
+            const errMsg = error instanceof Error ? error.message : String(error);
+            console.error(`${ts()} ${taskLabel} ${chalk.red("✗")} Plan parse failed: ${errMsg.substring(0, 100)}`);
+            await postLog(task.id, `${PREFIX} Failed to parse execution plan from Claude output: ${errMsg.substring(0, 200)}`, "error", "error");
+            // If we can't parse the plan, post raw output and let server-side try
+            return await postRawPlan(task.id, rawOutput, config.agentId, taskLabel, elapsed);
+        }
+        // 2c. Apply file cap (max 5 files per story)
+        const { truncatedCount, details } = applyFileCap(plan);
+        if (truncatedCount > 0) {
+            const msg = `${PREFIX} File cap applied: ${truncatedCount} stories truncated to max 5 targetFiles`;
+            console.log(`${ts()} ${taskLabel} ${chalk.yellow("⚠")} ${msg}`);
+            await postLog(task.id, msg);
+            for (const detail of details) {
+                console.log(`${ts()} ${taskLabel}   ${chalk.dim(detail)}`);
+            }
+        }
+        console.log(`${ts()} ${taskLabel} Plan: ${chalk.bold(plan.stories.length)} stories`);
+        await postLog(task.id, `${PREFIX} Plan generated: ${plan.stories.length} stories (${formatElapsed(elapsed)}). Running critic validation...`);
+        // 2d. Run critic validation
+        const criticResult = await runCriticValidation(claudePath, cliModel, prd, plan, cleanEnv, taskLabel);
+        // Track best plan across iterations
+        if (criticResult && criticResult.score > bestScore) {
+            bestPlan = plan;
+            bestScore = criticResult.score;
+        }
+        else if (!criticResult && !bestPlan) {
+            // Critic failed entirely — use this plan as fallback
+            bestPlan = plan;
+        }
+        // 2e. Check critic result
+        if (!criticResult) {
+            // Critic failed (timeout, parse error, etc.) — post plan without critic gate
+            const msg = `${PREFIX} Critic validation failed — posting plan without critic score`;
+            console.log(`${ts()} ${taskLabel} ${chalk.yellow("⚠")} ${msg}`);
+            await postLog(task.id, msg);
+            return await postValidatedPlan(task.id, plan, config.agentId, taskLabel, elapsed);
+        }
+        if (criticResult.approved || criticResult.score >= AUTO_APPROVAL_THRESHOLD) {
+            // Approved! Post the file-capped plan
+            const msg = `${PREFIX} Critic approved (score: ${criticResult.score}/100)`;
+            console.log(`${ts()} ${taskLabel} ${chalk.green("✓")} ${msg}`);
+            await postLog(task.id, msg);
+            if (criticResult.risks.length > 0) {
+                const risksMsg = `${PREFIX} Critic risks (non-blocking): ${criticResult.risks.join("; ")}`;
+                console.log(`${ts()} ${taskLabel}   ${chalk.dim(risksMsg)}`);
+                await postLog(task.id, risksMsg);
+            }
+            return await postValidatedPlan(task.id, plan, config.agentId, taskLabel, elapsed);
+        }
+        // 2f. Rejected — append critic feedback for next iteration
+        if (iteration < MAX_ITERATIONS) {
+            const feedback = formatCriticFeedback(criticResult);
+            currentPrompt = basePrompt + "\n\n" + feedback;
+            const msg = `${PREFIX} Critic rejected (score: ${criticResult.score}/100, threshold: ${AUTO_APPROVAL_THRESHOLD}). Re-planning with feedback...`;
+            console.log(`${ts()} ${taskLabel} ${chalk.yellow("⚠")} ${msg}`);
+            await postLog(task.id, msg);
+            if (criticResult.risks.length > 0) {
+                const risksMsg = `${PREFIX} Critic risks: ${criticResult.risks.join("; ")}`;
+                console.log(`${ts()} ${taskLabel}   ${chalk.dim(risksMsg)}`);
+                await postLog(task.id, risksMsg);
+            }
+            if (criticResult.suggestions && criticResult.suggestions.length > 0) {
+                const sugMsg = `${PREFIX} Critic suggestions: ${criticResult.suggestions.join("; ")}`;
+                console.log(`${ts()} ${taskLabel}   ${chalk.dim(sugMsg)}`);
+                await postLog(task.id, sugMsg);
+            }
+        }
+        else {
+            // Final iteration — rejected
+            const msg = `${PREFIX} Critic rejected after ${MAX_ITERATIONS} iterations (best score: ${bestScore}/100, threshold: ${AUTO_APPROVAL_THRESHOLD})`;
+            console.error(`${ts()} ${taskLabel} ${chalk.red("✗")} ${msg}`);
+            await postLog(task.id, msg, "error", "error");
+            if (criticResult.risks.length > 0) {
+                const risksMsg = `${PREFIX} Final risks: ${criticResult.risks.join("; ")}`;
+                console.error(`${ts()} ${taskLabel}   ${risksMsg}`);
+                await postLog(task.id, risksMsg, "error", "error");
+            }
+            if (criticResult.suggestions && criticResult.suggestions.length > 0) {
+                const sugMsg = `${PREFIX} Suggestions: ${criticResult.suggestions.join("; ")}`;
+                console.error(`${ts()} ${taskLabel}   ${sugMsg}`);
+                await postLog(task.id, sugMsg, "error", "error");
+            }
+        }
+    }
+    // All iterations exhausted — fail
+    return false;
+}
+/**
+ * Post a validated (file-capped) plan to the cloud API.
+ * Re-serializes the plan as a JSON code block since the server-side
+ * parseExecutionPlan() expects that format.
+ */
+async function postValidatedPlan(taskId, plan, agentId, taskLabel, elapsed) {
+    const serialized = serializePlan(plan);
     try {
-        rawOutput = await runClaudeCli(claudePath, cliModel, prompt, cleanEnv, task.id, startTime);
+        const result = await api.post("/api/agent/plan-result", {
+            taskId,
+            rawOutput: serialized,
+            agentId,
+        });
+        const storyCount = result.data.storyCount;
+        console.log(`${ts()} ${taskLabel} ${chalk.green("✓")} Plan validated: ${chalk.bold(storyCount)} stories → ${chalk.green("queued")}`);
+        await postLog(taskId, `${PREFIX} Plan validated: ${storyCount} stories. Task queued for execution.`);
+        await postProgress(taskId, "complete", elapsed, "Planning complete", 0, 0);
+        return true;
     }
     catch (error) {
-        const elapsed = Math.round((Date.now() - startTime) / 1000);
-        const errMsg = error instanceof Error ? error.message : String(error);
-        console.error(`${ts()} ${taskLabel} ${chalk.red("✗")} Failed after ${elapsed}s: ${errMsg.substring(0, 100)}`);
-        await postLog(task.id, `${PREFIX} Planning failed after ${formatElapsed(elapsed)}: ${errMsg.substring(0, 200)}`, "error", "error");
+        const err = error;
+        const detail = err.response?.data?.detail || String(error);
+        console.error(`${ts()} ${taskLabel} ${chalk.red("✗")} Server validation failed: ${detail.substring(0, 100)}`);
+        await postLog(taskId, `${PREFIX} Server-side plan validation failed: ${detail.substring(0, 200)}`, "error", "error");
         return false;
     }
-    const elapsed = Math.round((Date.now() - startTime) / 1000);
-    console.log(`${ts()} ${taskLabel} ${chalk.green("✓")} Claude CLI done ${chalk.dim(`(${elapsed}s, ${rawOutput.length} chars)`)}`);
-    await postLog(task.id, `${PREFIX} Planning complete (${formatElapsed(elapsed)}). Validating plan...`);
-    // 3. Post raw output back to cloud API for validation
+}
+/**
+ * Post raw (unparsed) plan output to the cloud API as a fallback.
+ * Used when local plan parsing fails — let the server try.
+ */
+async function postRawPlan(taskId, rawOutput, agentId, taskLabel, elapsed) {
     try {
         const result = await api.post("/api/agent/plan-result", {
-            taskId: task.id,
+            taskId,
             rawOutput,
-            agentId: config.agentId,
+            agentId,
         });
         const storyCount = result.data.storyCount;
-        console.log(`${ts()} ${taskLabel} ${chalk.green("✓")} Plan validated: ${chalk.bold(storyCount)} stories → ${chalk.green("queued")}`);
-        await postLog(task.id, `${PREFIX} Plan validated: ${storyCount} stories. Task queued for execution.`);
-        await postProgress(task.id, "complete", elapsed, "Planning complete", 0, 0);
+        console.log(`${ts()} ${taskLabel} ${chalk.green("✓")} Plan validated (server-side): ${chalk.bold(storyCount)} stories → ${chalk.green("queued")}`);
+        await postLog(taskId, `${PREFIX} Plan validated: ${storyCount} stories. Task queued for execution.`);
+        await postProgress(taskId, "complete", elapsed, "Planning complete", 0, 0);
         return true;
     }
     catch (error) {
         const err = error;
         const detail = err.response?.data?.detail || String(error);
         console.error(`${ts()} ${taskLabel} ${chalk.red("✗")} Validation failed: ${detail.substring(0, 100)}`);
-        await postLog(task.id, `${PREFIX} Plan validation failed: ${detail.substring(0, 200)}`, "error", "error");
+        await postLog(taskId, `${PREFIX} Plan validation failed: ${detail.substring(0, 200)}`, "error", "error");
         return false;
     }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@workermill/agent",
-  "version": "0.1.2",
+  "version": "0.2.1",
   "description": "WorkerMill Remote Agent - Run AI workers locally with your Claude Max subscription",
   "type": "module",
   "main": "./dist/index.js",