supipowers 1.1.0 → 1.2.2

package/src/orchestrator/dispatcher.ts

@@ -1,666 +0,0 @@
-import type { Platform } from "../platform/types.js";
-import type {
-  PlanTask,
-  AgentResult,
-  AgentStatus,
-  SupipowersConfig,
-} from "../types.js";
-import { buildTaskPrompt, buildFixPrompt } from "./prompts.js";
-import {
-  buildSpecComplianceReviewPrompt,
-  buildCodeQualityReviewPrompt,
-} from "./agent-prompts.js";
-import { isLspAvailable } from "../lsp/detector.js";
-import { detectContextMode } from "../context-mode/detector.js";
-import {
-  notifySuccess,
-  notifyWarning,
-  notifyError,
-  notifyInfo,
-} from "../notifications/renderer.js";
-import type { RunProgressState } from "./run-progress.js";
-import { modelRegistry } from "../config/model-registry-instance.js";
-import { resolveAllCandidates, createModelBridge, type ModelPlatformBridge } from "../config/model-resolver.js";
-import { loadModelConfig } from "../config/model-config.js";
-
-modelRegistry.register({
-  id: "implementer",
-  category: "sub-agent",
-  parent: "run",
-  label: "Implementer",
-  harnessRoleHint: "default",
-});
-
-modelRegistry.register({
-  id: "spec-reviewer",
-  category: "sub-agent",
-  parent: "run",
-  label: "Spec Reviewer",
-  harnessRoleHint: "slow",
-});
-
-modelRegistry.register({
-  id: "quality-reviewer",
-  category: "sub-agent",
-  parent: "run",
-  label: "Quality Reviewer",
-  harnessRoleHint: "slow",
-});
-
-modelRegistry.register({
-  id: "fix-agent",
-  category: "sub-agent",
-  parent: "run",
-  label: "Fix Agent",
-  harnessRoleHint: "default",
-});
-
-export interface DispatchOptions {
-  platform: Platform;
-  ctx: {
-    cwd: string;
-    ui: { notify(msg: string, type?: "info" | "warning" | "error"): void };
-  };
-  task: PlanTask;
-  planContext: string;
-  config: SupipowersConfig;
-  lspAvailable: boolean;
-  contextModeAvailable: boolean;
-  progress?: RunProgressState;
-  actionId?: string;
-  signal?: AbortSignal;
-  parentSessionModel?: string;
-}
-
-/** Check if an error looks like a model/auth configuration issue */
-function isModelAuthError(error: unknown): boolean {
-  const msg = error instanceof Error ? error.message : String(error);
-  return (
-    msg.includes("No API key") ||
-    msg.includes("API key") ||
-    msg.includes("authentication") ||
-    msg.includes("unauthorized") ||
-    msg.includes("model not found") ||
-    msg.includes("model_not_found") ||
-    msg.includes("invalid_api_key") ||
-    msg.includes("Could not resolve model")
-  );
-}
-
-/** Create a user-friendly error message from a raw error */
-function friendlyErrorMessage(error: unknown): string {
-  const msg = error instanceof Error ? error.message : String(error);
-  if (msg.includes("No API key") || msg.includes("invalid_api_key")) {
-    return "Model configuration issue — no valid API key found. Run /supi:models to configure, or ensure your agent has a model set up.";
-  }
-  if (msg.includes("model not found") || msg.includes("model_not_found") || msg.includes("Could not resolve model")) {
-    return "Configured model is not available. Run /supi:models to pick a different model, or remove the override to use your agent's default.";
-  }
-  if (msg.includes("Sub-agent dispatch is not available")) {
-    return msg; // This one is already user-friendly
-  }
-  return `Agent error: ${msg}`;
-}
-
-export async function dispatchAgent(
-  options: DispatchOptions,
-): Promise<AgentResult> {
-  const { platform, ctx, task, planContext, config, lspAvailable, contextModeAvailable, progress, signal } = options;
-  const startTime = Date.now();
-
-  const prompt = buildTaskPrompt(task, planContext, config, lspAvailable, contextModeAvailable);
-
-  // Check abort/timeout before starting
-  if (signal?.aborted) {
-    const reason = signal.reason;
-    const isTimeout = reason instanceof DOMException && reason.name === "TimeoutError";
-    const msg = isTimeout
-      ? `Task timed out after ${config.orchestration.taskTimeout / 1000}s`
-      : "Interrupted by user";
-    progress?.setStatus(task.id, "blocked", msg);
-    return {
-      taskId: task.id,
-      status: "blocked" as AgentStatus,
-      output: msg,
-      filesChanged: [],
-      duration: 0,
-    };
-  }
-
-  // Initialize widget card if available
-  progress?.setStatus(task.id, "running");
-
-  // Build the full fallback chain of model candidates
-  const modelConfig = loadModelConfig(platform.paths, ctx.cwd);
-  const bridge = createModelBridge(platform);
-  const candidates = resolveAllCandidates(
-    options.actionId ?? "implementer",
-    modelRegistry,
-    modelConfig,
-    bridge,
-  );
-
-  // Add parentSessionModel as final fallback if not already in candidates
-  if (options.parentSessionModel) {
-    const alreadyPresent = candidates.some((c) => c.model === options.parentSessionModel);
-    if (!alreadyPresent) {
-      candidates.push({ model: options.parentSessionModel, thinkingLevel: null, source: "main" });
-    }
-  }
-
-  // If no candidates at all, try with no model (let platform use its own default)
-  if (candidates.length === 0) {
-    candidates.push({ model: undefined, thinkingLevel: null, source: "main" });
-  }
-
-  let lastError: unknown = null;
-
-  // Try each candidate model; only retry on model/auth errors
-  for (const candidate of candidates) {
-    try {
-      const result = await executeSubAgent(
-        platform, prompt, task, config, ctx, progress,
-        candidate.model, candidate.thinkingLevel, signal,
-      );
-
-      const agentResult: AgentResult = {
-        taskId: task.id,
-        status: result.status,
-        output: result.output,
-        concerns: result.concerns,
-        filesChanged: result.filesChanged,
-        duration: Date.now() - startTime,
-      };
-
-      // Update widget with final status (no error notification here — callers handle it)
-      switch (agentResult.status) {
-        case "done":
-          progress?.setStatus(task.id, "done");
-          notifySuccess(ctx, `Task ${task.id} completed`, task.name);
-          break;
-        case "done_with_concerns":
-          progress?.setStatus(task.id, "done_with_concerns", agentResult.concerns);
-          notifyWarning(ctx, `Task ${task.id} done with concerns`, agentResult.concerns);
-          break;
-        case "blocked":
-          progress?.setStatus(task.id, "blocked", agentResult.output);
-          // Don't notifyError here — the run loop handles retries and cascade
-          break;
-      }
-
-      return agentResult;
-    } catch (error) {
-      lastError = error;
-
-      // If this is NOT a model/auth error, don't try other candidates
-      if (!isModelAuthError(error)) {
-        break;
-      }
-
-      // Model/auth error — try next candidate silently
-      // Only log at info level so we don't alarm the user
-      if (candidates.indexOf(candidate) < candidates.length - 1) {
-        notifyInfo(
-          ctx,
-          `Task ${task.id} model fallback`,
-          `Trying next model candidate...`,
-        );
-      }
-    }
-  }
-
-  // All candidates exhausted or non-model error
-  const errorMsg = friendlyErrorMessage(lastError);
-  progress?.setStatus(task.id, "blocked", errorMsg);
-
-  const agentResult: AgentResult = {
-    taskId: task.id,
-    status: "blocked",
-    output: errorMsg,
-    filesChanged: [],
-    duration: Date.now() - startTime,
-  };
-  // Don't notifyError here — let the run loop handle it after retries
-  return agentResult;
-}
-
-interface SubAgentResult {
-  status: AgentStatus;
-  output: string;
-  concerns?: string;
-  filesChanged: string[];
-}
-
-type NotifyCtx = { ui: { notify(msg: string, type?: "info" | "warning" | "error"): void } };
-
-/** Shorten a file path for display */
-function shortenPath(filePath: string): string {
-  const parts = filePath.split("/");
-  return parts.length > 3
-    ? `.../${parts.slice(-3).join("/")}`
-    : filePath;
-}
-
-/** Format a tool call for display */
-function formatToolAction(toolName: string, args?: Record<string, unknown>): string {
-  const path = args?.file_path ? shortenPath(String(args.file_path)) : "";
-  switch (toolName) {
-    case "Read": return `Reading ${path}`;
-    case "Edit": return `Editing ${path}`;
-    case "Write": return `Writing ${path}`;
-    case "Bash": {
-      const cmd = String(args?.command ?? "").slice(0, 60);
-      return `Running: ${cmd}${String(args?.command ?? "").length > 60 ? "..." : ""}`;
-    }
-    case "Grep": return `Searching for ${args?.pattern ?? "pattern"}`;
-    case "Glob": return `Finding files ${args?.pattern ?? ""}`;
-    case "Agent": return "Spawning sub-agent";
-    default: return `${toolName}${path ? ` ${path}` : ""}`;
-  }
-}
-
-async function executeSubAgent(
-  platform: Platform,
-  prompt: string,
-  task: PlanTask,
-  config: SupipowersConfig,
-  ctx?: NotifyCtx,
-  progress?: RunProgressState,
-  model?: string,
-  thinkingLevel?: string | null,
-  signal?: AbortSignal,
-): Promise<SubAgentResult> {
-  if (typeof platform.createAgentSession !== "function") {
-    throw new Error(
-      "Sub-agent dispatch is not available on this platform. " +
-      "If you just updated supipowers, restart your agent session to load the new code.",
-    );
-  }
-
-  // Pre-flight abort/timeout check
-  if (signal?.aborted) {
-    const reason = signal.reason;
-    const isTimeout = reason instanceof DOMException && reason.name === "TimeoutError";
-    const msg = isTimeout
-      ? `Task timed out after ${config.orchestration.taskTimeout / 1000}s`
-      : "Interrupted by user";
-    progress?.setStatus(task.id, "blocked", msg);
-    return { status: "blocked", output: msg, filesChanged: [] };
-  }
-
-  const session = await platform.createAgentSession({
-    cwd: process.cwd(),
-    taskDepth: 1,
-    parentTaskPrefix: `task-${task.id}`,
-    ...(model ? { model } : {}),
-    ...(thinkingLevel !== undefined ? { thinkingLevel } : {}),
-  });
-
-  // Track files changed and emit live progress
-  const filesChanged = new Set<string>();
-  const FILE_TOOLS = new Set(["Edit", "Write", "NotebookEdit"]);
-  const pendingToolArgs = new Map<string, Record<string, unknown>>();
-  const tag = `Task ${task.id}`;
-  let lastThinkingPreview = "";
-  const unsubscribe = session.subscribe((event) => {
-    if (event.type === "message_update") {
-      const msg = event.message as { content?: unknown } | undefined;
-      const content = extractTextContent(msg?.content);
-      const preview = content.split("\n").filter(Boolean).pop()?.slice(0, 80) ?? "";
-      if (preview && preview !== lastThinkingPreview) {
-        lastThinkingPreview = preview;
-        if (progress) {
-          progress.setActivity(task.id, preview);
-        } else if (ctx) {
-          ctx.ui.notify(`${tag}: thinking — ${preview}`, "info");
-        }
-      }
-    }
-    if (event.type === "tool_execution_start") {
-      const args = event.args as Record<string, unknown> | undefined;
-      if (FILE_TOOLS.has(event.toolName) && args?.file_path) {
-        pendingToolArgs.set(event.toolCallId, args);
-      }
-      if (progress) {
-        progress.setActivity(task.id, formatToolAction(event.toolName, args));
-        progress.incrementTools(task.id);
-      } else if (ctx) {
-        ctx.ui.notify(`${tag}: ${formatToolAction(event.toolName, args)}`, "info");
-      }
-    }
-    if (event.type === "tool_execution_end") {
-      const args = pendingToolArgs.get(event.toolCallId);
-      if (args?.file_path && !event.isError) {
-        filesChanged.add(String(args.file_path));
-        progress?.incrementFiles(task.id);
-      }
-      pendingToolArgs.delete(event.toolCallId);
-    }
-  });
-
-  // Wire abort signal to dispose the session
-  let abortHandler: (() => void) | undefined;
-  if (signal && !signal.aborted) {
-    abortHandler = () => {
-      session.dispose().catch(() => {}); // best-effort cleanup
-    };
-    signal.addEventListener("abort", abortHandler, { once: true });
-  }
-
-  try {
-    await session.prompt(prompt, { expandPromptTemplates: false });
-
-    // Extract the last assistant message
-    const messages = session.state.messages;
-    const lastAssistant = [...messages]
-      .reverse()
-      .find((m) => m.role === "assistant");
-
-    const lastMsg = lastAssistant as { content?: unknown } | undefined;
-    const output = extractTextContent(lastMsg?.content);
-    const status = parseAgentStatus(output);
-
-    return {
-      status,
-      output,
-      concerns: status === "done_with_concerns"
-        ? extractConcerns(output)
-        : undefined,
-      filesChanged: [...filesChanged],
-    };
-  } finally {
-    if (abortHandler && signal) {
-      signal.removeEventListener("abort", abortHandler);
-    }
-    unsubscribe();
-    await session.dispose().catch(() => {}); // may already be disposed by abort
-  }
-}
-
-function extractTextContent(content: unknown): string {
-  if (!content) return "";
-  if (typeof content === "string") return content;
-  if (Array.isArray(content)) {
-    return content
-      .filter((block: { type?: string }) => block.type === "text")
-      .map((block: { text?: string }) => block.text ?? "")
-      .join("\n");
-  }
-  return String(content);
-}
-
-function parseAgentStatus(output: string): AgentStatus {
-  // Look for structured "**Status:** X" or "Status: X" patterns to avoid false positives
-  const statusMatch = output.match(/\*?\*?status\*?\*?:\s*(BLOCKED|NEEDS_CONTEXT|DONE_WITH_CONCERNS|DONE)/i);
-  if (statusMatch) {
-    const val = statusMatch[1].toUpperCase();
-    if (val === "BLOCKED" || val === "NEEDS_CONTEXT") return "blocked";
-    if (val === "DONE_WITH_CONCERNS") return "done_with_concerns";
-    return "done";
-  }
-  // Default: if agent completed without explicit status, treat as done
-  return "done";
-}
-
-function extractConcerns(output: string): string {
-  const match = output.match(/(?:concerns?|issues?|worries):\s*(.+?)(?:\n\n|$)/is);
-  return match?.[1]?.trim() ?? "";
-}
-
-/** Review result from a spec compliance or code quality reviewer */
-export interface ReviewResult {
-  passed: boolean;
-  issues: string;
-}
-
-/**
- * Dispatch an implementer with 2-stage review (spec compliance + code quality).
- * Follows supipowers' subagent-driven-development pattern:
- * 1. Implementer implements + self-reviews
- * 2. Spec compliance reviewer verifies implementation matches spec
- * 3. If spec issues → re-dispatch implementer with feedback
- * 4. Code quality reviewer checks implementation quality
- * 5. If quality issues → re-dispatch implementer with feedback
- */
-export async function dispatchAgentWithReview(
-  options: DispatchOptions & { workDir?: string },
-): Promise<AgentResult> {
-  const { platform, ctx, task, planContext, config, lspAvailable, contextModeAvailable, workDir } = options;
-  const maxReviewRetries = config.orchestration.maxFixRetries;
-
-  // Step 1: Dispatch implementer
-  let implementResult = await dispatchAgent(options);
-
-  // If blocked or needs context, skip reviews
-  if (implementResult.status === "blocked") {
-    return implementResult;
-  }
-
-  // Check abort before reviews
-  if (options.signal?.aborted) return implementResult;
-
-  // Step 2: Spec compliance review
-  options.progress?.setStatus(task.id, "reviewing");
-  for (let attempt = 0; attempt <= maxReviewRetries; attempt++) {
-    const specReview = await dispatchSpecReview(
-      platform,
-      task,
-      implementResult,
-      config,
-    );
-
-    if (specReview.passed) {
-      notifyInfo(ctx, `Task ${task.id} spec review passed`);
-      break;
-    }
-
-    if (attempt === maxReviewRetries) {
-      const exhaustionMsg = `Task ${task.id} spec review failed after ${maxReviewRetries + 1} attempts — continuing as done_with_concerns`;
-      notifyWarning(ctx, exhaustionMsg, specReview.issues);
-      implementResult.status = "done_with_concerns";
-      implementResult.concerns = [
-        `⚠ Spec review exhausted (${maxReviewRetries + 1} attempts):`,
-        specReview.issues,
-        `Re-run with higher maxFixRetries or fix manually.`,
-      ].join("\n");
-      return implementResult;
-    }
-
-    // Re-dispatch implementer with spec review feedback
-    notifyInfo(
-      ctx,
-      `Task ${task.id} spec issues found, re-dispatching`,
-      specReview.issues,
-    );
-    const fixResult = await dispatchFixAgent({
-      ...options,
-      previousOutput: implementResult.output,
-      failureReason: `Spec compliance review failed:\n${specReview.issues}`,
-    });
-    implementResult = fixResult;
-
-    if (fixResult.status === "blocked") {
-      return fixResult;
-    }
-  }
-
-  // Check abort before quality review
-  if (options.signal?.aborted) return implementResult;
-
-  // Step 3: Code quality review
-  options.progress?.setStatus(task.id, "reviewing");
-  for (let attempt = 0; attempt <= maxReviewRetries; attempt++) {
-    const qualityReview = await dispatchQualityReview(
-      platform,
-      task,
-      implementResult,
-      config,
-    );
-
-    if (qualityReview.passed) {
-      notifyInfo(ctx, `Task ${task.id} quality review passed`);
-      break;
-    }
-
-    if (attempt === maxReviewRetries) {
-      const exhaustionMsg = `Task ${task.id} quality review failed after ${maxReviewRetries + 1} attempts — continuing as done_with_concerns`;
-      notifyWarning(ctx, exhaustionMsg, qualityReview.issues);
-      implementResult.status = "done_with_concerns";
-      implementResult.concerns = [
-        `⚠ Quality review exhausted (${maxReviewRetries + 1} attempts):`,
-        qualityReview.issues,
-        `Re-run with higher maxFixRetries or fix manually.`,
-      ].join("\n");
-      return implementResult;
-    }
-
-    // Re-dispatch implementer with quality review feedback
-    notifyInfo(
-      ctx,
-      `Task ${task.id} quality issues found, re-dispatching`,
-      qualityReview.issues,
-    );
-    const fixResult = await dispatchFixAgent({
-      ...options,
-      previousOutput: implementResult.output,
-      failureReason: `Code quality review failed:\n${qualityReview.issues}`,
-    });
-    implementResult = fixResult;
-
-    if (fixResult.status === "blocked") {
-      return fixResult;
-    }
-  }
-
-  return implementResult;
-}
-
-/** Dispatch a spec compliance reviewer sub-agent */
-async function dispatchSpecReview(
-  platform: Platform,
-  task: PlanTask,
-  implementResult: AgentResult,
-  config: SupipowersConfig,
-): Promise<ReviewResult> {
-  const prompt = buildSpecComplianceReviewPrompt({
-    taskRequirements: `Task: ${task.name}\n\n${task.description}\n\nAcceptance Criteria: ${task.criteria}`,
-    implementerReport: implementResult.output,
-  });
-
-  const modelConfig = loadModelConfig(platform.paths, process.cwd());
-  const bridge = createModelBridge(platform);
-  const candidates = resolveAllCandidates("spec-reviewer", modelRegistry, modelConfig, bridge);
-  const resolved = candidates[0] ?? { model: undefined, thinkingLevel: null, source: "main" as const };
-
-  try {
-    const result = await executeSubAgent(platform, prompt, task, config, undefined, undefined, resolved.model, resolved.thinkingLevel);
-    const passed =
-      result.status === "done" ||
-      result.output.toLowerCase().includes("spec compliant");
-    return {
-      passed,
-      issues: passed ? "" : result.output,
-    };
-  } catch {
-    // If reviewer fails, pass through — don't block on reviewer errors
-    return { passed: true, issues: "" };
-  }
-}
-
-/** Dispatch a code quality reviewer sub-agent */
-async function dispatchQualityReview(
-  platform: Platform,
-  task: PlanTask,
-  implementResult: AgentResult,
-  config: SupipowersConfig,
-): Promise<ReviewResult> {
-  const prompt = buildCodeQualityReviewPrompt({
-    taskSummary: `Task ${task.id}: ${task.name}\n\n${task.description}`,
-    implementerReport: implementResult.output,
-    baseSha: "HEAD~1",
-    headSha: "HEAD",
-  });
-
-  const modelConfig = loadModelConfig(platform.paths, process.cwd());
-  const bridge = createModelBridge(platform);
-  const candidates = resolveAllCandidates("quality-reviewer", modelRegistry, modelConfig, bridge);
-  const resolved = candidates[0] ?? { model: undefined, thinkingLevel: null, source: "main" as const };
-
-  try {
-    const result = await executeSubAgent(platform, prompt, task, config, undefined, undefined, resolved.model, resolved.thinkingLevel);
-    const hasCritical = result.output.toLowerCase().includes("critical");
-    return {
-      passed: !hasCritical,
-      issues: hasCritical ? result.output : "",
-    };
-  } catch {
-    // If reviewer fails, pass through
-    return { passed: true, issues: "" };
-  }
-}
-
-export async function dispatchFixAgent(
-  options: DispatchOptions & { previousOutput: string; failureReason: string },
-): Promise<AgentResult> {
-  const { platform, ctx, task, config, lspAvailable, contextModeAvailable, previousOutput, failureReason, progress } =
-    options;
-  const startTime = Date.now();
-
-  const prompt = buildFixPrompt(
-    task,
-    previousOutput,
-    failureReason,
-    lspAvailable,
-    contextModeAvailable,
-  );
-
-  progress?.setStatus(task.id, "running");
-
-  // Build fallback chain just like dispatchAgent
-  const modelConfig = loadModelConfig(platform.paths, ctx.cwd);
-  const bridge = createModelBridge(platform);
-  const candidates = resolveAllCandidates(
-    options.actionId ?? "fix-agent",
-    modelRegistry,
-    modelConfig,
-    bridge,
-  );
-
-  if (options.parentSessionModel) {
-    const alreadyPresent = candidates.some((c) => c.model === options.parentSessionModel);
-    if (!alreadyPresent) {
-      candidates.push({ model: options.parentSessionModel, thinkingLevel: null, source: "main" });
-    }
-  }
-
-  if (candidates.length === 0) {
-    candidates.push({ model: undefined, thinkingLevel: null, source: "main" });
-  }
-
-  let lastError: unknown = null;
-
-  for (const candidate of candidates) {
-    try {
-      const result = await executeSubAgent(platform, prompt, task, config, ctx, progress, candidate.model, candidate.thinkingLevel);
-      return {
-        taskId: task.id,
-        status: result.status,
-        output: result.output,
-        concerns: result.concerns,
-        filesChanged: result.filesChanged,
-        duration: Date.now() - startTime,
-      };
-    } catch (error) {
-      lastError = error;
-      if (!isModelAuthError(error)) break;
-    }
-  }
-
-  return {
-    taskId: task.id,
-    status: "blocked",
-    output: friendlyErrorMessage(lastError),
-    filesChanged: [],
-    duration: Date.now() - startTime,
-  };
-}