Asterisk Call Sampler → Telegram QA

• What it is: A small ops utility for sampled QA of call-center calls on Asterisk. The name tg_invoice_listener is misleading (likely a legacy of bun init or a rename) — by the code this is a "call-recording QA sampler → Telegram".
• Type / status / role: cli / ops script · prototype · solo (1 commit by the user).
• Activity period: 2026-04-18 (one day, 1 commit).

• Bun + TypeScript, NO external dependencies — only Bun built-ins: Bun.$ (shell), Bun.file, Bun.env, native fetch/FormData. System sshpass for SSH/SCP over password.
• Data source: Asterisk CDR (asteriskcdrdb MySQL), recording files under /var/spool/asterisk/monitor/YYYY/MM/DD/.
• Delivery: Telegram Bot API (sendAudio, multipart FormData).
• Config via env (SSH host/user/password, TG token/chat, lookback/sample parameters).

• A single commit (2026-04-18): a complete working index.ts script (176 lines).
• Logic: fetchCandidates() (SQL over CDR: ANSWERED, billsec ≥ threshold, recording present, GROUP BY recordingfile, ORDER BY RAND() LIMIT), pickCount() (random 10–15), SCP download into a temp folder, sendAudio() with a formatted caption, tmp cleanup in finally, exit code based on result.

Confirmed by index.ts:

• SSH/SCP automation by password (sshpass -e + strict -o options: host-key check disabled, password-auth only, timeout): access to a remote PBX without keys. → practical ops integration.
• Random sampling for QA (ORDER BY RAND() + pickCount 10–15): a representative sample of recordings instead of all of them. → thoughtful QA approach.
• Zero-dependency on Bun built-ins (Bun.$, Bun.file, native fetch/FormData): no node_modules dependencies, all I/O via the Bun runtime. → a modern, lightweight style.
• Careful resources and errors: temp folder with PID+timestamp, guaranteed cleanup in finally, per-file try/catch, meaningful exit code (fail if nothing was sent). → reliability for cron runs.
• `requireEnv` with an explicit error on a missing variable. → fail-fast config.

• Sessions found: 0. The repo has .claude/ (with a Bun guide in CLAUDE.md) and .cursor/ — AI tooling was used, but no Claude Code transcripts.
• What was done with AI: likely generated/refined with Cursor/Claude (CLAUDE.md is the standard Bun guide).
• AI-workflow patterns: minimal data.

• Working ops utility: Asterisk CDR → SCP → Telegram, ~176 lines, zero deps.
• Configurable sampling, safe cleanup, cron readiness.