Skip to content

About

Cloudflare Workers workspace: Telegram funnel bot, AI video-editing pipeline (FFmpeg+vision), and KV-backed multi-agent orchestration — all deployed to workers.dev.

Topics

Resources

Security policy

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

Multi-Project Workspace — Build Status (2026-09-08)

All three active projects are implemented and locally verified. Live deployment is blocked only by lack of a write-capable Cloudflare token.

Status table

Project Code Local verify Deploy Live URL
P04 Multi-Agent Orchestration ✅ worker/ + controller/ ✅ 10/10 harness + parallel dispatch ✅ deployed https://task-hub.darkestjokepossible.workers.dev
P02 Video Editing Pipeline ✅ pipeline/ + worker/ ✅ real FFmpeg end-to-end ✅ deployed https://video-ai.darkestjokepossible.workers.dev
P01 Telegram Bot ✅ src/ + scripts/ ✅ 10/10 harness ✅ deployed (needs BOT_TOKEN secret) https://telegram-funnel-bot.darkestjokepossible.workers.dev
P03 City-World Animation not started (out of scope this run) — — —

Deploy UNBLOCKED — write-capable Cloudflare API token supplied (2026-09-08). All three Workers deployed live.

Live verification (all PASSED)

  • P04 full task lifecycle on live hub (enqueue→claim→complete→health queue=0) ✅
  • P04 live orchestration: local controller polled live hub, claimed+processed+completed a task end-to-end ✅
  • P02 /analyze live: LLaVA→Llama-3.2-vision now returns real descriptions (Saturn test pattern) after /agree ✅
  • P01 /health live: {"ok":true} (bot shows "BOT_TOKEN not set" until secret added) ✅

How to unblock deploy

  1. setx CLOUDFLARE_API_TOKEN "<token with workers:edit, kv:write, ai:run>" (or wrangler login)
  2. node scripts/deploy-all.js
  3. P01: wrangler secret put BOT_TOKEN then node 01-telegram-bot-cf-workers/scripts/set-webhook.js <BOT_TOKEN> https://<sub>.workers.dev/webhook
  4. P02: one-time POST /agree {"prompt":"agree"} on the video-ai worker to enable llama-vision; LLaVA may need a ticket/format investigation (3016).
  5. P04: controller polls the task hub from anywhere.

Key technical finds (persist!)

  • Workers AI image decode errors (3016) affect LLaVA account-wide, all payload shapes (b64 array, data URI, public URL). Text models fine. Llama-3.2-vision needs prior Model Agreement (error 5016 → POST {"prompt":"agree"}).
  • Cloudflare MCP OAuth token (~/.local/share/opencode/mcp-auth.json) has ONLY .read scopes — every write (worker PUT, KV PUT, queue create) returns 10000 Authentication error. Reads + AI run work.
  • R2 not enabled on account (error 10042) — used KV instead. Queues create also blocked → KV-backed task queue design adopted.
  • ffmpeg -c copy trim: -ss must precede -i; output-seek copy yields audio-only files (video:0KiB).
  • ffmpeg is at C:\Users\User\tools\ffmpeg\ffmpeg.exe (no PATH); node at C:\Users\User\tools\node\node.exe.
  • Template-literal ${accountId} in cloudflare_execute gets URL-encoded and fails (7003) — hardcode <account-id>.

Runbook (local, no Cloudflare)

node 04-multiagent-orchestration/controller/controller.js run ^
  04-multiagent-orchestration/controller/workers-sample/handlers.js ^
  --spec .../workers-sample/tasks.json --concurrency 4

node 01-telegram-bot-cf-workers/scripts/test-local.js

node 02-agent-video-editing/pipeline/cli.js quick ^
  02-agent-video-editing/test-media/sample.mp4 --trim 1:4 --gray --out-dir out-smoke

About

Cloudflare Workers workspace: Telegram funnel bot, AI video-editing pipeline (FFmpeg+vision), and KV-backed multi-agent orchestration — all deployed to workers.dev.

Topics

Resources

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages