aiwcli 0.12.6 → 0.12.8

package/dist/templates/_shared/lib-ts/base/inference.ts

@@ -1,301 +1,301 @@
-/**
- * Inference utility for AI-powered text processing.
- * Unified interface for Claude API calls using the claude CLI.
- * See SPEC.md §6
- */
-
-import { execFileSync } from "node:child_process";
-import { logDebug, logWarn } from "./logger.js";
-import { STOP_WORDS } from "./stop-words.js";
-import type { InferenceResult } from "../types.js";
-import { execFileAsync, getInternalSubprocessEnv, shellQuoteWin } from "./subprocess-utils.js";
-
-// Model configurations §6.1
-const MODELS: Record<string, string> = {
-  fast: "claude-3-haiku-20240307",
-  standard: "claude-sonnet-4-20250514",
-  smart: "claude-opus-4-20250514",
-};
-
-const TIMEOUTS: Record<string, number> = {
-  fast: 15,
-  standard: 30,
-  smart: 90,
-};
-
-/**
- * Run inference using the claude CLI.
- * See SPEC.md §6.1
- */
-export function inference(
-  systemPrompt: string,
-  userPrompt: string,
-  level = "fast",
-  timeout?: number,
-): InferenceResult {
-  const startTime = Date.now();
-  const model = MODELS[level] ?? MODELS.fast;
-  const timeoutSec = timeout ?? TIMEOUTS[level] ?? TIMEOUTS.fast;
-  const fullPrompt = `${systemPrompt}\n\n${userPrompt}`;
-
-  // Remove ANTHROPIC_API_KEY to force subscription auth
-  const env = { ...process.env };
-  delete env.ANTHROPIC_API_KEY;
-
-  try {
-    const isWin = process.platform === "win32";
-    let stdout: string;
-
-    // On Windows with shell:true, Node.js sets windowsVerbatimArguments —
-    // args are joined with spaces, NOT individually quoted. We must manually
-    // wrap multi-word/special-char args in "..." for cmd.exe parsing.
-    // Inside double quotes: "" = literal ", and |&<> are safe.
-    const empty = isWin ? '""' : "";
-    let promptArg = fullPrompt;
-    if (isWin) {
-      promptArg = '"' + fullPrompt.replace(/\r?\n/g, " ").replace(/"/g, '""') + '"';
-    }
-
-    stdout = execFileSync(
-      "claude",
-      ["--model", model, "--print", "--setting-sources", empty, "-p", "--no-session-persistence", promptArg],
-      {
-        timeout: timeoutSec * 1000,
-        env,
-        encoding: "utf-8",
-        stdio: ["pipe", "pipe", "pipe"],
-        shell: isWin, // Windows needs shell for .cmd resolution
-      },
-    );
-
-    const latencyMs = Date.now() - startTime;
-    return {
-      success: true,
-      output: stdout.trim(),
-      latency_ms: latencyMs,
-    };
-  } catch (e: any) {
-    const latencyMs = Date.now() - startTime;
-
-    if (e.code === "ETIMEDOUT" || e.killed) {
-      return {
-        success: false,
-        output: "",
-        error: `Timeout after ${timeoutSec}s`,
-        latency_ms: latencyMs,
-      };
-    }
-
-    if (e.code === "ENOENT") {
-      return {
-        success: false,
-        output: "",
-        error: "claude CLI not found",
-        latency_ms: latencyMs,
-      };
-    }
-
-    // Non-zero exit code
-    if (e.status !== undefined && e.status !== 0) {
-      return {
-        success: false,
-        output: (e.stdout ?? "").toString().trim(),
-        error: (e.stderr ?? "").toString().trim() || `Exit code: ${e.status}`,
-        latency_ms: latencyMs,
-      };
-    }
-
-    return {
-      success: false,
-      output: "",
-      error: String(e),
-      latency_ms: latencyMs,
-    };
-  }
-}
-
-// §6.2 — System prompt for keyword extraction
-const CONTEXT_ID_SYSTEM_PROMPT = `Extract 6-12 keywords from what the user wants to do.
-
-Rules:
-- Output 6-12 keywords only
-- Keywords: nouns, verbs, adjectives, technical terms, proper names
-- NO function words: the, to, with, for, in, a, an, of, on, is, it, and, or, that, this, be, as, at, by, from
-- Most important/specific words preferred
-- No punctuation, no quotes
-
-Output ONLY the keywords separated by spaces, nothing else.`;
-
-/**
- * Generate a keyword summary of a user prompt.
- * Uses Sonnet (standard tier). Returns null if inference fails.
- * See SPEC.md §6.2
- */
-export function generateSemanticSummary(
-  prompt: string,
-  timeout = 15,
-): string | null {
-  const result = inference(CONTEXT_ID_SYSTEM_PROMPT, prompt, "standard", timeout);
-
-  if (!result.success || !result.output) return null;
-
-  let summary = result.output.trim();
-  summary = summary.replace(/^["']+|["']+$/g, "");
-  summary = summary.replace(/[.!?]+$/, "");
-
-  // Filter stop words
-  summary = filterStopWords(summary);
-
-  const words = summary.split(/\s+/);
-  if (words.length < 6 || words.length > 12) return null;
-
-  return summary;
-}
-
-// §6.3 — System prompt for context ID slug generation
-const CONTEXT_ID_SLUG_PROMPT = `You generate short title phrases for work sessions. These become folder names like \`260206-1959-fix-auth-middleware-redirect-loop-session-timeout\`.
-
-Users scan 100+ such names to find past sessions. Your title must make THIS session instantly recognizable.
-
-Rules:
-- Exactly 8-12 lowercase words
-- First word is an action verb (fix, add, implement, refactor, update, create, remove, optimize, debug, migrate, integrate, configure, deploy, scaffold, restructure)
-- Coherent phrase, not disjointed keywords — reads like a short task description
-- Prefer specific technical terms over generic words
-- No articles (the, a, an), no pronouns, no filler words, no punctuation, no quotes
-- Input may come from speech-to-text with filler words (uh, um, like, you know, basically, so) — ignore them entirely
-
-Examples:
-
-Input: "um so basically I need to like fix the auth bug in the login page"
-{"slug": "fix authentication bug login page redirect session handling flow"}
-
-Input: "hey uh can we add dark mode to the settings page"
-{"slug": "add dark mode toggle settings page user preference storage"}
-
-Input: "the context ids are bad can we change how we generate them towards a summary"
-{"slug": "improve context id generation use prompt summary slugs"}
-
-Input: "I want to refactor the database connection pooling for PostgreSQL"
-{"slug": "refactor postgresql database connection pooling optimize query performance"}
-
-Input: "so like you know the webhook retry logic is broken and stuff"
-{"slug": "fix webhook retry logic broken error handling recovery mechanism"}
-
-Input: "update the CI pipeline to cache node modules between runs"
-{"slug": "update ci pipeline cache node modules between workflow runs"}
-
-Respond with ONLY a JSON object: {"slug": "your 8-12 word phrase here"}`;
-
-/**
- * Generate a 5-12 word context ID slug from a user prompt.
- * Uses Haiku (fast tier) for low latency.
- * See SPEC.md §6.3
- */
-export function generateContextIdSlug(
-  prompt: string,
-  timeout = 3,
-): string | null {
-  const truncated = prompt.slice(0, 500);
-
-  const result = inference(CONTEXT_ID_SLUG_PROMPT, truncated, "fast", timeout);
-
-  if (!result.success || !result.output) {
-    logWarn("inference", `Context ID slug inference failed: ${result.error}`);
-    return null;
-  }
-
-  const raw = result.output.trim();
-
-  // Parse JSON response, fall back to raw text
-  let slug: string | null = null;
-  try {
-    const parsed = JSON.parse(raw);
-    if (parsed && typeof parsed === "object" && "slug" in parsed) {
-      slug = parsed.slug;
-    }
-  } catch {
-    // Fall through to raw text
-  }
-
-  if (!slug) slug = raw;
-
-  // Clean up
-  slug = slug.replace(/^["'`]+|["'`]+$/g, "");
-  slug = slug.replace(/[.!?]+$/, "");
-  slug = slug.replace(/-/g, " ");
-  slug = slug.replace(/[^a-zA-Z0-9 ]/g, "");
-  slug = slug.replace(/\s+/g, " ").trim();
-
-  const words = slug.split(" ");
-
-  if (words.length > 12) words.length = 12;
-  if (words.length < 5) {
-    logDebug("inference", `Context ID slug too short (${words.length} words): '${slug}'`);
-    return null;
-  }
-
-  const resultSlug = words.join(" ");
-  logDebug("inference", `Generated context ID slug: '${resultSlug}' (${result.latency_ms}ms)`);
-  return resultSlug;
-}
-
-/**
- * Async version of inference() that does NOT block the event loop.
- * Use for parallel AI calls (e.g., Stage 3 parallel summarizers).
- * Uses execFileAsync and getInternalSubprocessEnv for proper subprocess isolation.
- */
-export async function inferenceAsync(
-  systemPrompt: string,
-  userPrompt: string,
-  level = "fast",
-  timeout?: number,
-): Promise<InferenceResult> {
-  const startTime = Date.now();
-  const model = (level in MODELS ? MODELS[level] : undefined) ?? MODELS.fast;
-  const timeoutSec = timeout ?? (level in TIMEOUTS ? TIMEOUTS[level] : undefined) ?? TIMEOUTS.fast;
-  const timeoutMs = timeoutSec * 1000;
-  const fullPrompt = `${systemPrompt}\n\n${userPrompt}`;
-
-  const env = getInternalSubprocessEnv();
-  delete env.ANTHROPIC_API_KEY;
-
-  const isWin = process.platform === "win32";
-  const empty = isWin ? '""' : "";
-  const promptArg = isWin
-    ? shellQuoteWin(fullPrompt.replace(/\r?\n/g, " "))
-    : fullPrompt;
-
-  const result = await execFileAsync(
-    "claude",
-    ["--model", model, "--print", "--setting-sources", empty, "-p", "--no-session-persistence", promptArg],
-    { timeout: timeoutMs, env, shell: isWin },
-  );
-
-  const latencyMs = Date.now() - startTime;
-
-  if (result.killed) {
-    return { success: false, output: "", error: `Timeout after ${timeoutSec}s`, latency_ms: latencyMs };
-  }
-  if (result.exitCode !== 0) {
-    return {
-      success: false,
-      output: result.stdout.trim(),
-      error: result.stderr.trim() || `Exit code: ${result.exitCode}`,
-      latency_ms: latencyMs,
-    };
-  }
-  return { success: true, output: result.stdout.trim(), latency_ms: latencyMs };
-}
-
-/**
- * Filter stop words from text.
- * See SPEC.md §6.4
- */
-function filterStopWords(text: string): string {
-  return text
-    .toLowerCase()
-    .split(/\s+/)
-    .filter((w) => !STOP_WORDS.has(w) && w.length > 1)
-    .join(" ");
-}
+/**
+ * Inference utility for AI-powered text processing.
+ * Unified interface for Claude API calls using the claude CLI.
+ * See SPEC.md §6
+ */
+
+import { execFileSync } from "node:child_process";
+import { logDebug, logWarn } from "./logger.js";
+import { STOP_WORDS } from "./stop-words.js";
+import type { InferenceResult } from "../types.js";
+import { execFileAsync, getInternalSubprocessEnv, shellQuoteWin } from "./subprocess-utils.js";
+
+// Model configurations §6.1
+const MODELS: Record<string, string> = {
+  fast: "claude-haiku-4-5-20251001",
+  standard: "claude-sonnet-4-6",
+  smart: "claude-opus-4-6",
+};
+
+const TIMEOUTS: Record<string, number> = {
+  fast: 15,
+  standard: 30,
+  smart: 90,
+};
+
+/**
+ * Run inference using the claude CLI.
+ * See SPEC.md §6.1
+ */
+export function inference(
+  systemPrompt: string,
+  userPrompt: string,
+  level = "fast",
+  timeout?: number,
+): InferenceResult {
+  const startTime = Date.now();
+  const model = MODELS[level] ?? MODELS.fast;
+  const timeoutSec = timeout ?? TIMEOUTS[level] ?? TIMEOUTS.fast;
+  const fullPrompt = `${systemPrompt}\n\n${userPrompt}`;
+
+  // Remove ANTHROPIC_API_KEY to force subscription auth
+  const env = { ...process.env };
+  delete env.ANTHROPIC_API_KEY;
+
+  try {
+    const isWin = process.platform === "win32";
+    let stdout: string;
+
+    // On Windows with shell:true, Node.js sets windowsVerbatimArguments —
+    // args are joined with spaces, NOT individually quoted. We must manually
+    // wrap multi-word/special-char args in "..." for cmd.exe parsing.
+    // Inside double quotes: "" = literal ", and |&<> are safe.
+    const empty = isWin ? '""' : "";
+    let promptArg = fullPrompt;
+    if (isWin) {
+      promptArg = '"' + fullPrompt.replace(/\r?\n/g, " ").replace(/"/g, '""') + '"';
+    }
+
+    stdout = execFileSync(
+      "claude",
+      ["--model", model, "--print", "--setting-sources", empty, "-p", "--no-session-persistence", promptArg],
+      {
+        timeout: timeoutSec * 1000,
+        env,
+        encoding: "utf-8",
+        stdio: ["pipe", "pipe", "pipe"],
+        shell: isWin, // Windows needs shell for .cmd resolution
+      },
+    );
+
+    const latencyMs = Date.now() - startTime;
+    return {
+      success: true,
+      output: stdout.trim(),
+      latency_ms: latencyMs,
+    };
+  } catch (e: any) {
+    const latencyMs = Date.now() - startTime;
+
+    if (e.code === "ETIMEDOUT" || e.killed) {
+      return {
+        success: false,
+        output: "",
+        error: `Timeout after ${timeoutSec}s`,
+        latency_ms: latencyMs,
+      };
+    }
+
+    if (e.code === "ENOENT") {
+      return {
+        success: false,
+        output: "",
+        error: "claude CLI not found",
+        latency_ms: latencyMs,
+      };
+    }
+
+    // Non-zero exit code
+    if (e.status !== undefined && e.status !== 0) {
+      return {
+        success: false,
+        output: (e.stdout ?? "").toString().trim(),
+        error: (e.stderr ?? "").toString().trim() || `Exit code: ${e.status}`,
+        latency_ms: latencyMs,
+      };
+    }
+
+    return {
+      success: false,
+      output: "",
+      error: String(e),
+      latency_ms: latencyMs,
+    };
+  }
+}
+
+// §6.2 — System prompt for keyword extraction
+const CONTEXT_ID_SYSTEM_PROMPT = `Extract 6-12 keywords from what the user wants to do.
+
+Rules:
+- Output 6-12 keywords only
+- Keywords: nouns, verbs, adjectives, technical terms, proper names
+- NO function words: the, to, with, for, in, a, an, of, on, is, it, and, or, that, this, be, as, at, by, from
+- Most important/specific words preferred
+- No punctuation, no quotes
+
+Output ONLY the keywords separated by spaces, nothing else.`;
+
+/**
+ * Generate a keyword summary of a user prompt.
+ * Uses Sonnet (standard tier). Returns null if inference fails.
+ * See SPEC.md §6.2
+ */
+export function generateSemanticSummary(
+  prompt: string,
+  timeout = 15,
+): string | null {
+  const result = inference(CONTEXT_ID_SYSTEM_PROMPT, prompt, "standard", timeout);
+
+  if (!result.success || !result.output) return null;
+
+  let summary = result.output.trim();
+  summary = summary.replace(/^["']+|["']+$/g, "");
+  summary = summary.replace(/[.!?]+$/, "");
+
+  // Filter stop words
+  summary = filterStopWords(summary);
+
+  const words = summary.split(/\s+/);
+  if (words.length < 6 || words.length > 12) return null;
+
+  return summary;
+}
+
+// §6.3 — System prompt for context ID slug generation
+const CONTEXT_ID_SLUG_PROMPT = `You generate short title phrases for work sessions. These become folder names like \`260206-1959-fix-auth-middleware-redirect-loop-session-timeout\`.
+
+Users scan 100+ such names to find past sessions. Your title must make THIS session instantly recognizable.
+
+Rules:
+- Exactly 8-12 lowercase words
+- First word is an action verb (fix, add, implement, refactor, update, create, remove, optimize, debug, migrate, integrate, configure, deploy, scaffold, restructure)
+- Coherent phrase, not disjointed keywords — reads like a short task description
+- Prefer specific technical terms over generic words
+- No articles (the, a, an), no pronouns, no filler words, no punctuation, no quotes
+- Input may come from speech-to-text with filler words (uh, um, like, you know, basically, so) — ignore them entirely
+
+Examples:
+
+Input: "um so basically I need to like fix the auth bug in the login page"
+{"slug": "fix authentication bug login page redirect session handling flow"}
+
+Input: "hey uh can we add dark mode to the settings page"
+{"slug": "add dark mode toggle settings page user preference storage"}
+
+Input: "the context ids are bad can we change how we generate them towards a summary"
+{"slug": "improve context id generation use prompt summary slugs"}
+
+Input: "I want to refactor the database connection pooling for PostgreSQL"
+{"slug": "refactor postgresql database connection pooling optimize query performance"}
+
+Input: "so like you know the webhook retry logic is broken and stuff"
+{"slug": "fix webhook retry logic broken error handling recovery mechanism"}
+
+Input: "update the CI pipeline to cache node modules between runs"
+{"slug": "update ci pipeline cache node modules between workflow runs"}
+
+Respond with ONLY a JSON object: {"slug": "your 8-12 word phrase here"}`;
+
+/**
+ * Generate a 5-12 word context ID slug from a user prompt.
+ * Uses Haiku (fast tier) for low latency.
+ * See SPEC.md §6.3
+ */
+export function generateContextIdSlug(
+  prompt: string,
+  timeout = 3,
+): string | null {
+  const truncated = prompt.slice(0, 500);
+
+  const result = inference(CONTEXT_ID_SLUG_PROMPT, truncated, "fast", timeout);
+
+  if (!result.success || !result.output) {
+    logWarn("inference", `Context ID slug inference failed: ${result.error}`);
+    return null;
+  }
+
+  const raw = result.output.trim();
+
+  // Parse JSON response, fall back to raw text
+  let slug: string | null = null;
+  try {
+    const parsed = JSON.parse(raw);
+    if (parsed && typeof parsed === "object" && "slug" in parsed) {
+      slug = parsed.slug;
+    }
+  } catch {
+    // Fall through to raw text
+  }
+
+  if (!slug) slug = raw;
+
+  // Clean up
+  slug = slug.replace(/^["'`]+|["'`]+$/g, "");
+  slug = slug.replace(/[.!?]+$/, "");
+  slug = slug.replace(/-/g, " ");
+  slug = slug.replace(/[^a-zA-Z0-9 ]/g, "");
+  slug = slug.replace(/\s+/g, " ").trim();
+
+  const words = slug.split(" ");
+
+  if (words.length > 12) words.length = 12;
+  if (words.length < 5) {
+    logDebug("inference", `Context ID slug too short (${words.length} words): '${slug}'`);
+    return null;
+  }
+
+  const resultSlug = words.join(" ");
+  logDebug("inference", `Generated context ID slug: '${resultSlug}' (${result.latency_ms}ms)`);
+  return resultSlug;
+}
+
+/**
+ * Async version of inference() that does NOT block the event loop.
+ * Use for parallel AI calls (e.g., Stage 3 parallel summarizers).
+ * Uses execFileAsync and getInternalSubprocessEnv for proper subprocess isolation.
+ */
+export async function inferenceAsync(
+  systemPrompt: string,
+  userPrompt: string,
+  level = "fast",
+  timeout?: number,
+): Promise<InferenceResult> {
+  const startTime = Date.now();
+  const model = (level in MODELS ? MODELS[level] : undefined) ?? MODELS.fast;
+  const timeoutSec = timeout ?? (level in TIMEOUTS ? TIMEOUTS[level] : undefined) ?? TIMEOUTS.fast;
+  const timeoutMs = timeoutSec * 1000;
+  const fullPrompt = `${systemPrompt}\n\n${userPrompt}`;
+
+  const env = getInternalSubprocessEnv();
+  delete env.ANTHROPIC_API_KEY;
+
+  const isWin = process.platform === "win32";
+  const empty = isWin ? '""' : "";
+  const promptArg = isWin
+    ? shellQuoteWin(fullPrompt.replace(/\r?\n/g, " "))
+    : fullPrompt;
+
+  const result = await execFileAsync(
+    "claude",
+    ["--model", model, "--print", "--setting-sources", empty, "-p", "--no-session-persistence", promptArg],
+    { timeout: timeoutMs, env, shell: isWin },
+  );
+
+  const latencyMs = Date.now() - startTime;
+
+  if (result.killed) {
+    return { success: false, output: "", error: `Timeout after ${timeoutSec}s`, latency_ms: latencyMs };
+  }
+  if (result.exitCode !== 0) {
+    return {
+      success: false,
+      output: result.stdout.trim(),
+      error: result.stderr.trim() || `Exit code: ${result.exitCode}`,
+      latency_ms: latencyMs,
+    };
+  }
+  return { success: true, output: result.stdout.trim(), latency_ms: latencyMs };
+}
+
+/**
+ * Filter stop words from text.
+ * See SPEC.md §6.4
+ */
+function filterStopWords(text: string): string {
+  return text
+    .toLowerCase()
+    .split(/\s+/)
+    .filter((w) => !STOP_WORDS.has(w) && w.length > 1)
+    .join(" ");
+}