@causyne/agent 0.0.0-stage → 1.10.1

package/src/tokenTracker.js

@@ -0,0 +1,560 @@
+import fs from 'fs';
+import path from 'path';
+import os from 'os';
+import { loadConfig } from './config.js';
+
+import { causyneHome } from './utils/paths.js';
+
+const USAGE_DIR = path.join(causyneHome(), 'usage');
+
+/**
+ * The date key for "today", in the USER'S timezone.
+ *
+ * This was `new Date().toISOString().split('T')[0]`, which is UTC. A daily
+ * budget bucketed in UTC does not reset at the user's midnight: in WIB (UTC+7)
+ * it rolls over at 07:00 local, in NZDT (UTC+13) at 13:00, in PST (UTC-8) the
+ * previous evening. So a user's morning work was charged to "yesterday" and
+ * their daily cap could be exhausted before their day had started — and the
+ * reset landed in the middle of it.
+ *
+ * A limit the user sets as "per day" must mean their day. Anything else is a
+ * limit they cannot reason about.
+ */
+
+let usageQueue = Promise.resolve();
+function withUsageLock(fn) {
+  const run = usageQueue.then(fn, fn);
+  usageQueue = run.then(() => {}, () => {});
+  return run;
+}
+
+/**
+ * Write a usage file the way the journal does: temp file, then rename.
+ *
+ * A direct writeFileSync that is interrupted leaves truncated JSON, and
+ * getDailyUsage() answers a parse failure with a fresh zeroed record — so a
+ * torn write does not surface as an error, it silently ERASES the day's budget.
+ * rename() is atomic on the same filesystem, so a reader sees either the whole
+ * previous file or the whole new one.
+ */
+function writeUsageFile(file, payload) {
+  const temp = `${file}.tmp_${process.pid}`;
+  fs.writeFileSync(temp, JSON.stringify(payload, null, 2), { encoding: 'utf8', mode: 0o600 });
+  fs.renameSync(temp, file);
+}
+
+function localDateKey(when = new Date()) {
+  const pad = (n) => String(n).padStart(2, '0');
+  return `${when.getFullYear()}-${pad(when.getMonth() + 1)}-${pad(when.getDate())}`;
+}
+
+// Input-token count above which GPT-6 Astra bills on its second rate card
+// ($20/$75 instead of $10/$50). The higher card applies to the whole request,
+// not to the overage alone.
+const ASTRA_LONG_CONTEXT_THRESHOLD = 272000;
+
+export class TokenTracker {
+  constructor() {
+    this.config = loadConfig();
+    this.budget = this.config.tokenBudget || { enabled: false };
+    this.taskTokens = 0;
+    this.taskCostUsd = 0;
+    this.largestStepTokens = 0;
+    this.ensureUsageDir();
+  }
+
+
+
+  // Cost per 1M tokens [prompt, completion] (2026 Pricing)
+  //
+  // `promptTokens` is optional and only consulted by providers that publish a
+  // SECOND rate card above a context threshold. GPT-6 Astra is the first: past
+  // 272,000 input tokens the same request bills at $20 in / $75 out instead of
+  // $10 / $50. Ignoring that under-reported long-context runs by 2x — and
+  // under-reporting is the worse direction, because the user's real invoice
+  // then exceeds the cost shown by Causyne.
+  getPricing(promptTokens = 0) {
+    const provider = this.config.provider?.toLowerCase();
+    const model = this.config.model?.toLowerCase() || '';
+
+    if (provider === 'deepseek') {
+      if (model.includes('pro')) return [0.435, 0.87];
+      if (model.includes('flash')) return [0.14, 0.28];
+      return [0.14, 0.28]; // Default
+    }
+
+    // Meta Model API — Muse. Published rates per 1M tokens, read off the
+    // console: Standard $1.25 in / $4.25 out; Contributor $0.10 in / $0.20 out,
+    // which is the "up to 95% less than Standard" the tier advertises in
+    // exchange for training on your prompts and completions.
+    //
+    // 1.2 and 1.1 are NOT separately published. They fall back to the 1.3 rate
+    // of their own tier, which over-estimates if they are cheaper — the safe
+    // direction for something that stops a run when the budget is spent.
+    if (provider === 'muse' || provider === 'meta') {
+      if (model.includes('-contributor')) return [0.10, 0.20];
+      return [1.25, 4.25];
+    }
+
+    if (provider === 'qwen' || provider === 'dashscope') {
+      if (model.includes('max')) return [1.25, 3.75];
+      if (model.includes('plus')) return [0.32, 1.28];
+      if (model.includes('flash')) return [0.19, 1.13];
+      return [0.32, 1.28]; // Default
+    }
+
+    if (provider === 'openai') {
+      // GPT-6 Astra (2026-09-03). Two rate cards: $10/$50 up to 272K input
+      // tokens, $20/$75 for a request that crosses it — the tier applies to
+      // the WHOLE request, not just the overage.
+      if (model.includes('astra')) {
+        return promptTokens > ASTRA_LONG_CONTEXT_THRESHOLD ? [20.00, 75.00] : [10.00, 50.00];
+      }
+      if (model.includes('sol')) return [5.00, 30.00];
+      if (model.includes('terra')) return [2.50, 15.00];
+      if (model.includes('luna')) return [1.00, 1.00];
+      if (model.includes('mini')) return [0.75, 4.50];
+      return [5.00, 30.00]; // Default OpenAI (Sol)
+    }
+    
+    if (provider === 'anthropic') {
+      // Mythos shares Fable's tier and price. Without its own branch it fell
+      // through to the Sonnet default and was billed at a FIFTH of its rate.
+      if (model.includes('fable') || model.includes('mythos')) return [10.00, 50.00];
+      if (model.includes('opus')) return [5.00, 25.00];
+      // Sonnet 4.6 is $3/$15; Sonnet 5 is $2/$10. Checked in that order
+      // because 'sonnet-4-6' also matches the generic 'sonnet' test below.
+      if (model.includes('sonnet-4-6')) return [3.00, 15.00];
+      if (model.includes('sonnet')) return [2.00, 10.00];
+      if (model.includes('haiku')) return [1.00, 5.00];
+      return [2.00, 10.00]; // Default Anthropic (Sonnet 5)
+    }
+    
+    if (provider === 'gemini') {
+      if (model.includes('pro')) return [2.00, 12.00];
+      if (model.includes('flash-lite')) return [0.25, 1.50];
+      // 3.6 Flash holds 3.5's $1.50 input but cuts output from $9.00 to $7.50.
+      // Checked before the generic 'flash' test, which it would also match.
+      if (model.includes('3.6-flash') || model.includes('3-6-flash')) return [1.50, 7.50];
+      if (model.includes('flash')) return [1.50, 9.00];
+      return [1.50, 9.00]; // Default Gemini
+    }
+    
+    // Local Ollama cost the user $0.00 in BYOK APIs
+    if (provider === 'ollama') {
+      return [0.00, 0.00]; 
+    }
+    
+    return [5.00, 15.00]; // Unknown BYOK provider fallback
+  }
+
+  // Prompt-cache multipliers, applied against the base input price.
+  // Anthropic: cache read 0.1x, cache write (5m) 1.25x.
+  // OpenAI / xAI: cached input 0.5x. DeepSeek disk cache: roughly 0.1x.
+  cachedInputMultiplier() {
+    const provider = this.config.provider?.toLowerCase();
+    if (provider === 'anthropic') {
+      // Fable 5.1 reprices cache reads to $0.25/MTok against a $10 base — a
+      // 0.025x discount, not the family's usual 0.1x. Charging 0.1x here
+      
+      // counter. Same shape as the Muse case below: the discount is a property
+      // of the model, not the vendor.
+      const model = this.config.model?.toLowerCase() || '';
+      if (model.includes('fable-5-1')) return 0.025;
+      return 0.1;
+    }
+    if (provider === 'deepseek') return 0.1;
+
+    // Muse is the one provider whose cache discount differs by MODEL, not just
+    // by vendor: $0.15 against $1.25 on Standard is 0.12x, while $0.002 against
+    // $0.10 on Contributor is 0.02x. Collapsing both into one vendor number
+    // would over-charge a Contributor run by six times on every cached token.
+    if (provider === 'muse' || provider === 'meta') {
+      const model = this.config.model?.toLowerCase() || '';
+      return model.includes('-contributor') ? 0.02 : 0.12;
+    }
+
+    // GPT-6 Astra publishes cached input at $1.00/MTok against a $10.00 base —
+    // a 90% discount, not the 0.5x the rest of this vendor falls back to.
+    // Found while verifying the getPricing() rows above; charging 0.5x
+    // overstated every cached Astra token FIVEFOLD.
+    if (provider === 'openai') {
+      const model = this.config.model?.toLowerCase() || '';
+      if (model.includes('astra')) return 0.1;
+    }
+
+    // 0.5x is the long-standing OpenAI/xAI cached-input discount. It is NOT
+    // verified against the GPT-5.6 family's current rate card; left as-is
+    // rather than guessed at, since overstating is the safer error direction
+    // for a number the user reads as a bill.
+    return 0.5;
+  }
+
+  /**
+   * @param {number} promptTokens      Total input tokens (cached + uncached).
+   * @param {number} completionTokens  Output tokens.
+   * @param {object} usageDetail       Optional { cachedPromptTokens, cacheWriteTokens }.
+   */
+  calculateCost(promptTokens, completionTokens, usageDetail = null) {
+    // promptTokens decides the rate card on providers with a context tier.
+    const [promptPrice, completionPrice] = this.getPricing(promptTokens);
+
+    const cachedTokens = Math.min(usageDetail?.cachedPromptTokens || 0, promptTokens);
+    const writeTokens = Math.min(usageDetail?.cacheWriteTokens || 0, promptTokens - cachedTokens);
+    const fullPriceTokens = Math.max(0, promptTokens - cachedTokens - writeTokens);
+
+    const writeMultiplier = usageDetail?.cacheWriteMultiplier || 1.25;
+    const promptCost =
+      (fullPriceTokens / 1_000_000) * promptPrice +
+      (cachedTokens / 1_000_000) * promptPrice * this.cachedInputMultiplier() +
+      (writeTokens / 1_000_000) * promptPrice * writeMultiplier;
+
+    const completionCost = (completionTokens / 1_000_000) * completionPrice;
+    return promptCost + completionCost;
+  }
+
+  // Cost per item (image) or per second (video_animation) in USD (2026 Pricing)
+  getMediaPricing(provider = '', model = '', mediaType = 'image') {
+    const p = provider.toLowerCase();
+    const m = model.toLowerCase();
+
+    // Muse Image is a flat per-image rate; there is no video model in the
+    // family, so a video request must not silently bill at the image rate.
+    if (p === 'muse' || p === 'meta') {
+      if (mediaType === 'video_animation') return 0;
+      return 0.01; // $0.01 per image
+    }
+
+    if (p === 'gemini') {
+      if (mediaType === 'video_animation' || m.includes('veo')) return 0.05; // $0.05 per second of Veo 2.0 video
+      if (m.includes('imagen')) return 0.03; // $0.03 per image
+      return 0.03;
+    }
+    if (p === 'openai') {
+      if (mediaType === 'video_animation' || m.includes('sora')) return 0.10; // $0.10 per second
+      if (m.includes('dall-e-3')) return 0.08; // $0.08 per HD image
+      return 0.04; // $0.04 per standard image
+    }
+    if (p === 'qwen' || p === 'dashscope') {
+      if (mediaType === 'video_animation' || m.includes('wanx') || m.includes('wan2')) return 0.04; // $0.04 per second
+      return 0.02; // $0.02 per image
+    }
+    if (p === 'pollinations' || p === 'local') {
+      return 0.00; // Free Mode D / Mode A
+    }
+    return 0.03; // Default generic media cost
+  }
+
+  async trackMediaUsage(provider, mediaType, model, quantityUnits = 1, onWarning = null) {
+    this.config = loadConfig();
+    this.budget = this.config.tokenBudget || { enabled: false };
+
+    const unitPrice = this.getMediaPricing(provider, model, mediaType);
+    const cost = unitPrice * (quantityUnits || 1);
+    
+    // Multimedia Equivalent Units (MEU): convert dollar cost to equivalent tokens ($0.000002/token ~ $2/1M tokens)
+    const equivalentTokens = Math.round(cost / 0.000002);
+    
+    this.taskTokens += equivalentTokens;
+    this.taskCostUsd += cost;
+
+
+
+    const daily = await this.getDailyUsage();
+    daily.totalTokens += equivalentTokens;
+    daily.totalCalls += 1;
+    daily.totalCostUsd = (daily.totalCostUsd || 0) + cost;
+    await this.saveDailyUsage(daily);
+
+    const cumulative = await this.getCumulativeUsage();
+    cumulative.totalTokens += equivalentTokens;
+    cumulative.totalCalls += 1;
+    cumulative.totalCostUsd = (cumulative.totalCostUsd || 0) + cost;
+    await this.saveCumulativeUsage(cumulative);
+
+    if (!this.budget.enabled) {
+      return { allowed: true, used: daily.totalTokens, costUsd: daily.totalCostUsd, taskCostUsd: this.taskCostUsd, equivalentTokens };
+    }
+
+    // Monetary Kill Switch
+    if (this.budget.maxCostUsd && daily.totalCostUsd > this.budget.maxCostUsd) {
+      if (this.budget.hardStop) {
+        return {
+          allowed: false,
+          reason: `BYOK Monetary Kill-Switch Activated: Media generation cost ($${cost.toFixed(2)}) pushes total spend to $${daily.totalCostUsd.toFixed(2)}, exceeding your $${this.budget.maxCostUsd.toFixed(2)} limit.`,
+          used: daily.totalTokens,
+          costUsd: daily.totalCostUsd,
+          taskCostUsd: this.taskCostUsd,
+          equivalentTokens
+        };
+      }
+    }
+
+    await this.checkAndTriggerAlerts(daily, onWarning);
+    return { allowed: true, used: daily.totalTokens, costUsd: daily.totalCostUsd, taskCostUsd: this.taskCostUsd, equivalentTokens };
+  }
+
+  async checkAndTriggerAlerts(daily, onWarning = null) {
+    if (!this.budget.enabled) return;
+
+    const alertThreshold = this.budget.alertThresholdTokens
+      ? Number(this.budget.alertThresholdTokens)
+      : (this.budget.maxTokensPerDay && this.budget.warningThreshold ? this.budget.maxTokensPerDay * this.budget.warningThreshold : null);
+
+    if (alertThreshold && alertThreshold > 0 && daily.totalTokens >= alertThreshold && !daily.alertSent) {
+      daily.alertSent = true;
+      await this.saveDailyUsage(daily);
+
+      const msg = `⚠️ Token Budget Alert: Your BYOK usage has reached ${daily.totalTokens.toLocaleString()} tokens (Threshold: ${alertThreshold.toLocaleString()}).`;
+      if (onWarning) {
+        onWarning(msg);
+      }
+    }
+  }
+
+  ensureUsageDir() {
+    if (!fs.existsSync(USAGE_DIR)) {
+      fs.mkdirSync(USAGE_DIR, { recursive: true });
+    }
+  }
+
+  getDailyFile() {
+    return path.join(USAGE_DIR, `${localDateKey()}.json`);
+  }
+
+  getCumulativeFile() {
+    return path.join(USAGE_DIR, 'byok_cumulative.json');
+  }
+
+
+  async getDailyUsage() {
+    const file = this.getDailyFile();
+    if (fs.existsSync(file)) {
+      try {
+        return JSON.parse(fs.readFileSync(file, 'utf8'));
+      } catch(e) {}
+    }
+    return { date: localDateKey(), totalTokens: 0, totalCalls: 0, totalCostUsd: 0 };
+  }
+
+  async getCumulativeUsage() {
+    const file = this.getCumulativeFile();
+    if (fs.existsSync(file)) {
+      try {
+        return JSON.parse(fs.readFileSync(file, 'utf8'));
+      } catch(e) {}
+    }
+    
+    // Automatic Historical Backfill Migration (v1.1.50):
+    // Scan all past `${date}.json` files across USAGE_DIR and sum them up so past 5M+ tokens are never lost!
+    let totalTokens = 0;
+    let totalCalls = 0;
+    let totalCostUsd = 0;
+    try {
+      if (fs.existsSync(USAGE_DIR)) {
+        const files = fs.readdirSync(USAGE_DIR);
+        for (const f of files) {
+          if (/^\d{4}-\d{2}-\d{2}\.json$/.test(f)) {
+            try {
+              const data = JSON.parse(fs.readFileSync(path.join(USAGE_DIR, f), 'utf8'));
+              totalTokens += (data.totalTokens || 0);
+              totalCalls += (data.totalCalls || 0);
+              totalCostUsd += (data.totalCostUsd || 0);
+            } catch (err) {}
+          }
+        }
+      }
+    } catch (err) {}
+    
+    const initialCumulative = { totalTokens, totalCalls, totalCostUsd, backfilled: true, lastUpdated: new Date().toISOString() };
+    try {
+      writeUsageFile(file, initialCumulative);
+    } catch (err) {}
+    return initialCumulative;
+  }
+
+
+  async saveDailyUsage(usage) {
+    writeUsageFile(this.getDailyFile(), usage);
+  }
+
+  async saveCumulativeUsage(usage) {
+    writeUsageFile(this.getCumulativeFile(), usage);
+  }
+
+
+  async resetDailyUsage() {
+    const file = this.getDailyFile();
+    if (fs.existsSync(file)) {
+      fs.unlinkSync(file);
+    }
+    // DO NOT reset cumulative usage here, keeping lifetime tracking separate from fresh budget.
+  }
+
+  async resetCumulativeUsage() {
+    const file = this.getCumulativeFile();
+    if (fs.existsSync(file)) {
+      fs.unlinkSync(file);
+    }
+  }
+
+  async getReportedUsage() {
+    const daily = await this.getDailyUsage();
+    const cumulative = await this.getCumulativeUsage();
+    return {
+      usedTokens: daily.totalTokens, // UI budget tracks against daily "Fresh" setup
+      dailyTokens: daily.totalTokens,
+      taskTokens: this.taskTokens,
+      lifetimeTokens: cumulative.totalTokens,
+      costUsd: daily.totalCostUsd, // UI estimated cost tracks against daily "Fresh" setup
+      taskCostUsd: this.taskCostUsd,
+      lifetimeCostUsd: cumulative.totalCostUsd,
+      isDemo: false
+    };
+  }
+
+  /**
+   * Can the loop afford ANOTHER step without breaching the task ceiling?
+   *
+   * WHY THIS EXISTS. trackUsage() checks `taskTokens > maxTokensPerTask` AFTER
+   * a call has already been paid for, so the run always overshoots by one step.
+   * With ~9,000-token steps a 299,000 ceiling lands at ~308,000 — which breaks
+   * the only promise a token cap makes. Asked BEFORE the call, using the
+   * largest step seen so far, the ceiling actually holds.
+   *
+   * The largest step rather than the average, deliberately: a cap that holds
+   * on average is a cap that fails on the step that matters.
+   *
+   * @returns {{allowed:boolean, reason?:string, used:number, limit:number, projected:number}}
+   */
+  canAffordAnotherStep() {
+    const limit = this.budget && this.budget.enabled && this.budget.hardStop
+      ? Number(this.budget.maxTokensPerTask || 0)
+      : 0;
+    if (!limit) return { allowed: true, used: this.taskTokens, limit: 0, projected: this.taskTokens };
+
+    // Before the first step there is nothing to extrapolate from; let it run.
+    const step = this.largestStepTokens || 0;
+    const projected = this.taskTokens + step;
+    if (step > 0 && projected > limit) {
+      return {
+        allowed: false,
+        reason: `Stopping before the next step would breach the task ceiling: ${this.taskTokens.toLocaleString()} used, `
+          + `the largest step so far was ${step.toLocaleString()}, and ${projected.toLocaleString()} would exceed the ${limit.toLocaleString()} limit.`,
+        used: this.taskTokens, limit, projected
+      };
+    }
+    return { allowed: true, used: this.taskTokens, limit, projected };
+  }
+
+  resetTask() {
+    this.taskTokens = 0;
+    this.taskCostUsd = 0;
+    this.largestStepTokens = 0;
+  }
+
+  
+  settle(extra, daily, cumulative) {
+    return {
+      // Daily scope — what the budget bar and the daily limiter both measure.
+      usedTokens: daily.totalTokens,
+      dailyTokens: daily.totalTokens,
+      dailyCostUsd: daily.totalCostUsd || 0,
+      dailyLimit: Number(this.budget && this.budget.maxTokensPerDay) || 0,
+      // Task scope — what this run has spent.
+      taskTokens: this.taskTokens,
+      taskCostUsd: this.taskCostUsd,
+      taskLimit: Number(this.budget && this.budget.maxTokensPerTask) || 0,
+      // Lifetime scope.
+      lifetimeTokens: cumulative.totalTokens,
+      lifetimeCostUsd: cumulative.totalCostUsd,
+      // Retained for clients built against the older frame. It has always been
+      // the daily cost on the allowed path, so it keeps that meaning here
+      // rather than acquiring a third one.
+      costUsd: daily.totalCostUsd || 0,
+      used: this.taskTokens,
+      ...extra
+    };
+  }
+
+  async trackUsage(promptTokens, completionTokens, onWarning = null, usageDetail = null) {
+    // Dynamically reload config to apply live budget updates from UI
+    this.config = loadConfig();
+    this.budget = this.config.tokenBudget || { enabled: false };
+
+    const total = (promptTokens || 0) + (completionTokens || 0);
+    this.taskTokens += total;
+    // Feeds canAffordAnotherStep(). The LARGEST step, not the average: a
+    // ceiling that holds on average is a ceiling that fails on the step that
+    // matters, and every step is bigger than the last as the window grows.
+    this.largestStepTokens = Math.max(this.largestStepTokens || 0, total);
+
+    const cost = this.calculateCost(promptTokens, completionTokens, usageDetail);
+    this.taskCostUsd += cost;
+
+
+    
+    // Normal BYOK Tracking.
+    //
+    // Read-modify-write, so it runs inside the lock. Outside it, a condenser
+    // pass firing concurrently with the loop silently subtracts its own tokens
+    // from the day's total.
+    const { daily, cumulative } = await withUsageLock(async () => {
+      const daily = await this.getDailyUsage();
+      daily.totalTokens += total;
+      daily.totalCalls += 1;
+      daily.totalCostUsd = (daily.totalCostUsd || 0) + cost;
+      await this.saveDailyUsage(daily);
+
+      const cumulative = await this.getCumulativeUsage();
+      cumulative.totalTokens += total;
+      cumulative.totalCalls += 1;
+      cumulative.totalCostUsd = (cumulative.totalCostUsd || 0) + cost;
+      await this.saveCumulativeUsage(cumulative);
+
+      return { daily, cumulative };
+    });
+    
+    if (!this.budget.enabled) {
+      return this.settle({ allowed: true }, daily, cumulative);
+    }
+    
+    // BYOK Monetary Kill Switch
+    if (this.budget.maxCostUsd && daily.totalCostUsd > this.budget.maxCostUsd) {
+      if (this.budget.hardStop) {
+        return this.settle({
+          allowed: false,
+          reason: `BYOK Monetary Kill-Switch Activated: You have spent $${daily.totalCostUsd.toFixed(2)}, which exceeds your $${this.budget.maxCostUsd.toFixed(2)} limit.`
+        }, daily, cumulative);
+      }
+    }
+    
+    // Check task budget
+    if (this.budget.maxTokensPerTask && this.taskTokens > this.budget.maxTokensPerTask) {
+      if (this.budget.hardStop) {
+        return this.settle({
+          allowed: false,
+          reason: `Task token budget reached: ${this.taskTokens.toLocaleString()} of ${Number(this.budget.maxTokensPerTask).toLocaleString()} tokens for this task.`,
+          limit: this.budget.maxTokensPerTask
+        }, daily, cumulative);
+      }
+    }
+
+    // Check daily budget. This limit was previously accepted from the UI and
+    // used only to derive an alert threshold — it never actually stopped
+    // anything, so a runaway task could blow straight through it.
+    if (this.budget.maxTokensPerDay && daily.totalTokens > this.budget.maxTokensPerDay) {
+      if (this.budget.hardStop) {
+        return this.settle({
+          allowed: false,
+          reason: `Daily token budget reached: ${daily.totalTokens.toLocaleString()} of ${Number(this.budget.maxTokensPerDay).toLocaleString()} tokens used today.`,
+          limit: this.budget.maxTokensPerDay
+        }, daily, cumulative);
+      }
+    }
+    
+    await this.checkAndTriggerAlerts(daily, onWarning);
+    
+    return this.settle({ allowed: true }, daily, cumulative);
+  }
+}

package/src/tool-scripts/ax_snapshot.applescript

@@ -0,0 +1,39 @@
+on run argv
+  set targetApp to item 1 of argv
+  set maxNodes to (item 2 of argv) as integer
+  set out to ""
+  set n to 0
+  tell application "System Events"
+    if targetApp is "" then
+      set proc to first application process whose frontmost is true
+    else
+      set proc to first application process whose name is targetApp
+    end if
+    set out to "app|" & (name of proc) & linefeed
+    repeat with w in (windows of proc)
+      if n is greater than or equal to maxNodes then exit repeat
+      set out to out & "window|" & (name of w) & "||" & linefeed
+      set n to n + 1
+      repeat with e in (entire contents of w)
+        if n is greater than or equal to maxNodes then exit repeat
+        set r to ""
+        set t to ""
+        set v to ""
+        try
+          set r to (class of e) as string
+        end try
+        try
+          set t to (name of e) as string
+        end try
+        try
+          set v to (value of e) as string
+        end try
+        if t is not "" or v is not "" then
+          set out to out & r & "|" & t & "|" & v & linefeed
+          set n to n + 1
+        end if
+      end repeat
+    end repeat
+  end tell
+  return out
+end run

package/src/tool-scripts/ax_snapshot.ps1

@@ -0,0 +1,10 @@
+Add-Type -AssemblyName UIAutomationClient,UIAutomationTypes
+$max = [int]$args[0]
+$root = [System.Windows.Automation.AutomationElement]::RootElement
+$cond = [System.Windows.Automation.Condition]::TrueCondition
+$n = 0
+foreach ($e in $root.FindAll([System.Windows.Automation.TreeScope]::Descendants, $cond)) {
+  if ($n -ge $max) { break }
+  $name = $e.Current.Name
+  if ($name) { Write-Output ("" + $e.Current.ControlType.ProgrammaticName + "|" + $name + "|"); $n++ }
+}

package/src/tool-scripts/desktop_act.applescript

@@ -0,0 +1,58 @@
+on run argv
+  set appName to item 1 of argv
+  set report to ""
+  tell application "System Events"
+    if appName is not "" then
+      set proc to first application process whose name is appName
+      set frontmost of proc to true
+      delay 0.2
+    else
+      set proc to first application process whose frontmost is true
+    end if
+    repeat with i from 2 to (count of argv)
+      set step to item i of argv
+      set AppleScript's text item delimiters to "::"
+      set parts to text items of step
+      set AppleScript's text item delimiters to ""
+      set verb to item 1 of parts
+      if verb is "press" then
+        set elemName to item 2 of parts
+        set hit to false
+        repeat with w in (windows of proc)
+          repeat with e in (entire contents of w)
+            if hit is false then
+              try
+                if (name of e) is elemName then
+                  perform action "AXPress" of e
+                  set hit to true
+                end if
+              end try
+            end if
+          end repeat
+        end repeat
+        if hit then
+          set report to report & "ok|press|" & elemName & linefeed
+        else
+          set report to report & "miss|press|" & elemName & linefeed
+        end if
+      else if verb is "clickxy" then
+        set px to (item 2 of parts) as integer
+        set py to (item 3 of parts) as integer
+        click at {px, py}
+        set report to report & "ok|clickxy|" & px & "," & py & linefeed
+      else if verb is "type" then
+        keystroke (item 2 of parts)
+        set report to report & "ok|type|" & (length of (item 2 of parts)) & " chars" & linefeed
+      else if verb is "key" then
+        key code ((item 2 of parts) as integer)
+        set report to report & "ok|key|" & (item 2 of parts) & linefeed
+      else if verb is "wait" then
+        delay ((item 2 of parts) as number)
+        set report to report & "ok|wait|" & (item 2 of parts) & linefeed
+      else
+        set report to report & "err|unknown verb|" & verb & linefeed
+      end if
+    end repeat
+  end tell
+  return report
+end run