add TG skill

2026-05-11 06:11:15 +00:00 · 2026-03-17 12:59:05 +01:00
parent d7bbb7587e
commit 761d5e0790
30 changed files with 4869 additions and 42 deletions
--- a/.claude/skills/tdlib-telegram/SKILL.md
+++ b/.claude/skills/tdlib-telegram/SKILL.md
@@ -0,0 +1,301 @@
+---
+name: tdlib-telegram
+description: >
+  Reference guide for building Telegram integrations with TDLib (tdl/node).
+  Covers rate limits, FLOOD_WAIT handling, file size constraints, bot vs user account
+  differences, and safe code patterns. Use this skill whenever writing or modifying code
+  that calls Telegram APIs via TDLib, the Bot API, or any Telegram client library — including
+  sending messages, downloading/uploading files, scanning channels, forwarding messages,
+  managing subscriptions, or handling notifications. Also use when debugging 429 errors,
+  FLOOD_WAIT, or silent message drops.
+---
+
+# TDLib / Telegram Development Guide
+
+This skill provides the rate limits, constraints, and patterns you need to write correct
+Telegram integrations. The limits below come from official Telegram documentation and
+well-established community findings (Telegram does not publish exact numbers for all limits).
+
+## Telegram Rate Limits
+
+These are approximate safe boundaries. Telegram's actual limits are dynamic and depend on
+account age, history, and request type. The correct strategy is to respect these as guidelines
+and always handle FLOOD_WAIT errors gracefully.
+
+### Bot Accounts
+
+| Operation | Limit | Notes |
+|-----------|-------|-------|
+| Messages to same chat | ~1 msg/sec | Bursts OK, sustained exceeds limit |
+| Messages in a group | 20 msgs/min | Hard limit per group chat |
+| Bulk notifications (different users) | ~30 msgs/sec | Global across all chats |
+| Message edits in a group | ~20 edits/min | Community-observed |
+| API requests (global) | ~30 req/sec | All request types combined |
+| Paid broadcasts | up to 1000 msgs/sec | Requires Telegram Stars balance |
+
+### User Accounts (TDLib)
+
+| Operation | Limit | Notes |
+|-----------|-------|-------|
+| API requests (global) | ~30 req/sec | All request types combined |
+| Messages in a group | ~20 msgs/min | Same as bot |
+| Channel history reads | No published limit | But pagination + delay is essential |
+| Joining groups | Very strict | FLOOD_WAIT often 30-300+ seconds |
+
+### File Size Limits
+
+| Context | Upload | Download |
+|---------|--------|----------|
+| Bot API (standard) | 50 MB | 20 MB |
+| Bot API (local server) | 2,000 MB | 2,000 MB |
+| User account (TDLib) | 2 GB | 2 GB |
+| Premium user (TDLib) | 4 GB | 4 GB |
+
+### Message & Content Limits
+
+| Item | Limit |
+|------|-------|
+| Message text length | 4,096 chars |
+| Media caption | 1,024 chars (4,096 premium) |
+| Album / media group | 10 items max |
+| Forwarded messages per request | `forwarded_message_count_max` (TDLib option) |
+| Inline keyboard buttons | 100 entities |
+| Formatting entities per message | 100 |
+| Scheduled messages per chat | 100 |
+| Bot commands | 100 max |
+
+### Forum & Group Limits
+
+| Item | Limit |
+|------|-------|
+| Topics per group | 1,000,000 |
+| Topic title | 128 chars |
+| Group members | 200,000 |
+| Admins per group | 50 |
+| Bots per group | 20 |
+| Pinned topics | 5 |
+
+## FLOOD_WAIT — How It Works
+
+When you exceed rate limits, Telegram returns a `FLOOD_WAIT_X` error (or HTTP 429 with
+`retry_after`). This is a **mandatory pause** — the value `X` is the number of seconds you
+must wait before ANY request will succeed. It blocks the entire client, not just the
+operation that triggered it.
+
+### The Right Way to Handle It
+
+```typescript
+// Extract the wait duration from the error
+function extractFloodWaitSeconds(err: unknown): number | null {
+  const message = err instanceof Error ? err.message : String(err);
+
+  // Pattern 1: FLOOD_WAIT_30
+  const flood = message.match(/FLOOD_WAIT_(\d+)/i);
+  if (flood) return parseInt(flood[1], 10);
+
+  // Pattern 2: "retry after 30"
+  const retry = message.match(/retry after (\d+)/i);
+  if (retry) return parseInt(retry[1], 10);
+
+  // Pattern 3: HTTP 429 without explicit seconds
+  if (String((err as any)?.code) === "429") return 30;
+
+  return null; // Not a rate limit error
+}
+
+// Wrap any TDLib call with automatic retry
+async function withFloodWait<T>(fn: () => Promise<T>, maxRetries = 5): Promise<T> {
+  for (let attempt = 0; attempt <= maxRetries; attempt++) {
+    try {
+      return await fn();
+    } catch (err) {
+      const wait = extractFloodWaitSeconds(err);
+      if (wait === null || attempt >= maxRetries) throw err;
+
+      // Add 1-5s jitter to prevent thundering herd
+      const jitter = 1000 + Math.random() * 4000;
+      await sleep(wait * 1000 + jitter);
+    }
+  }
+  throw new Error("Unreachable");
+}
+```
+
+### Key Rules
+
+- **Always respect the wait duration.** Never retry before `retry_after` expires.
+- **Add jitter.** Without it, multiple clients retry simultaneously and trigger another FLOOD_WAIT.
+- **Non-rate-limit errors should fail fast.** Only retry on FLOOD_WAIT, not on other errors.
+- **Don't artificially throttle below ~1 req/sec.** Telegram's own guidance (via grammY docs)
+  is to send requests as fast as you need and handle 429 errors. Fixed low-frequency throttling
+  wastes throughput without preventing floods.
+
+## Code Patterns
+
+### Pattern: Sequential Send Queue
+
+When sending notifications to multiple users, use a sequential queue with a per-message delay.
+Never fire concurrent sends — you will hit the 30 msg/sec global limit instantly.
+
+```typescript
+let sendQueue: Promise<void> = Promise.resolve();
+
+function queueSend(chatId: bigint, text: string): void {
+  sendQueue = sendQueue
+    .then(() => withFloodWait(() => sendTextMessage(chatId, text)))
+    .then(() => sleep(50)) // ~20 msgs/sec, well under 30 limit
+    .catch((err) => log.error({ err, chatId }, "Send failed"));
+}
+```
+
+### Pattern: Paginated Scanning with Delay
+
+When reading channel history or enumerating topics, always add a delay between pages:
+
+```typescript
+while (hasMorePages) {
+  const result = await invokeWithTimeout(client, { _: "getChatHistory", ... });
+  processMessages(result.messages);
+
+  if (result.messages.length < limit) break;
+
+  await sleep(1000); // 1 second between pages — prevents FLOOD_WAIT on large channels
+}
+```
+
+### Pattern: Event Listener Before Action
+
+When waiting for TDLib async events (upload confirmation, download completion), always
+attach the event listener BEFORE starting the operation. If you attach after, fast
+operations can complete before the listener exists, causing the promise to hang forever.
+
+```typescript
+// CORRECT: listener first, then action
+client.on("update", handleUpdate);
+const tempMsg = await client.invoke({ _: "sendMessage", ... });
+tempMsgId = tempMsg.id; // handler now knows which message to match
+
+// WRONG: action first, then listener — race condition!
+const tempMsg = await client.invoke({ _: "sendMessage", ... });
+client.on("update", handleUpdate); // may miss updateMessageSendSucceeded
+```
+
+### Pattern: Download/Upload Timeouts
+
+Scale timeouts with file size. TDLib downloads/uploads are asynchronous — without a timeout,
+a stalled transfer hangs the entire pipeline.
+
+```typescript
+const timeoutMs = Math.max(
+  10 * 60_000,                    // minimum 10 minutes
+  (fileSizeMB / 1024) * 10 * 60_000  // 10 minutes per GB
+);
+```
+
+### Pattern: TDLib Client Lifecycle
+
+Always close TDLib clients in a `finally` block. Unclosed clients leak memory and file
+descriptors, and can leave TDLib's internal database locked.
+
+```typescript
+const client = await createTdlibClient(account);
+try {
+  // ... use client ...
+} finally {
+  await closeTdlibClient(client);
+}
+```
+
+## Anti-Patterns
+
+### Never: Concurrent TDLib Sends Without Queue
+
+```typescript
+// BAD: fires all sends concurrently — will trigger FLOOD_WAIT immediately
+await Promise.all(users.map((u) => sendTextMessage(u.chatId, msg)));
+
+// GOOD: sequential with delay
+for (const user of users) {
+  await withFloodWait(() => sendTextMessage(user.chatId, msg));
+  await sleep(50);
+}
+```
+
+### Never: Bare client.invoke() Without Retry
+
+Every `client.invoke()` call can return FLOOD_WAIT at any time. Bare calls will crash
+on rate limits instead of retrying.
+
+```typescript
+// BAD: crashes on FLOOD_WAIT
+await client.invoke({ _: "sendMessage", ... });
+
+// GOOD: retries automatically
+await withFloodWait(() => client.invoke({ _: "sendMessage", ... }));
+```
+
+### Never: Retry Without Respecting retry_after
+
+```typescript
+// BAD: fixed 1-second retry ignores Telegram's wait requirement
+catch (err) { await sleep(1000); retry(); }
+
+// GOOD: extract and respect the actual wait time
+catch (err) {
+  const wait = extractFloodWaitSeconds(err);
+  if (wait !== null) await sleep(wait * 1000 + jitter);
+  else throw err;
+}
+```
+
+### Never: Ignore FLOOD_WAIT in Bots
+
+Bot accounts get the same FLOOD_WAIT as user accounts. The bot API's 429 response
+blocks ALL operations for the specified duration — not just the chat that triggered it.
+A single unhandled flood in a notification loop can make the entire bot unresponsive.
+
+## Bot vs User Account Differences
+
+| Capability | Bot | User (TDLib) |
+|-----------|-----|-------------|
+| Read channel history | No (unless admin) | Yes |
+| Send to users who haven't started bot | No | N/A |
+| Join groups via invite link | No (must be added) | Yes |
+| Forward messages (send_copy) | Yes | Yes |
+| File upload limit | 50 MB (standard API) | 2 GB |
+| File download limit | 20 MB (standard API) | 2 GB |
+| Auth method | Bot token | Phone + SMS code |
+| Rate limit profile | Same FLOOD_WAIT | Same FLOOD_WAIT |
+
+## TDLib-Specific Notes
+
+### BigInt Chat IDs
+
+TDLib uses numeric chat IDs. Supergroups and channels use negative IDs (e.g., `-1001234567890`).
+When passing to `client.invoke()`, convert with `Number(chatId)` — TDLib's JSON interface
+doesn't handle BigInt. Be aware that very large IDs may lose precision with `Number()`,
+though current Telegram IDs are within safe integer range.
+
+### TDLib Options (Runtime Queryable)
+
+These are read-only values you can query at runtime via `getOption`:
+- `message_text_length_max` — max message text length
+- `message_caption_length_max` — max caption length
+- `forwarded_message_count_max` — max forwards per request
+
+### Session State
+
+TDLib persists session state to disk. Each account needs its own state directory.
+Running two clients on the same state directory simultaneously will corrupt the database.
+Use separate directories per account, and separate volumes in Docker for worker vs bot.
+
+## Docker Considerations
+
+- **prebuilt-tdlib**: The `prebuilt-tdlib` npm package provides platform-specific TDLib
+  binaries. Container base image must match (e.g., `node:20-bookworm-slim` for Debian x64).
+- **Volumes**: Mount persistent volumes for TDLib state directories — losing state forces
+  full re-authentication.
+- **Graceful shutdown**: Wait for active operations to finish before closing DB connections.
+  TDLib operations in flight will fail if the database pool is closed underneath them.
+- **Health checks**: TDLib services don't expose HTTP — use database connectivity as the
+  health signal instead.
--- a/.claude/skills/tdlib-telegram/evals/evals.json
+++ b/.claude/skills/tdlib-telegram/evals/evals.json
@@ -0,0 +1,23 @@
+{
+  "skill_name": "tdlib-telegram",
+  "evals": [
+    {
+      "id": 1,
+      "prompt": "Add a new bot command /broadcast that sends a text message to ALL users who have a TelegramLink in the database. The admin triggers it from the web app. Add it to the bot's command handler and create an API endpoint that triggers it.",
+      "expected_output": "Code that uses a sequential send queue with withFloodWait wrapping each sendTextMessage call, a delay between sends (~50ms), and does NOT use Promise.all or concurrent sends. Should handle errors per-user without stopping the broadcast.",
+      "files": []
+    },
+    {
+      "id": 2,
+      "prompt": "The worker keeps crashing with 'FLOOD_WAIT_35' errors when scanning a source channel that has about 10,000 messages. It happens during the getChannelMessages pagination loop. How do I fix this?",
+      "expected_output": "Diagnosis that the apiDelayMs between pages may be too low or the retry logic isn't wrapping the pagination calls. Should recommend ensuring all getChatHistory/searchChatMessages calls go through withFloodWait/invokeWithTimeout, and that sleep(config.apiDelayMs) exists between pages. Should NOT suggest reducing scan frequency as the primary fix.",
+      "files": []
+    },
+    {
+      "id": 3,
+      "prompt": "Write a utility function for the worker that downloads a file from a source Telegram channel message and re-uploads it to the destination channel. It needs to handle files up to 2GB and work reliably in Docker.",
+      "expected_output": "Code that: (1) wraps download in withFloodWait with size-scaled timeout, (2) attaches upload event listener BEFORE calling sendMessage, (3) uses temp directory with cleanup in finally block, (4) handles the 2GB Telegram limit correctly, (5) uses try/finally for client cleanup if applicable.",
+      "files": []
+    }
+  ]
+}