Projects About

The Automation That Never Worked: SyntaxError, 4 Zombie Processes, and 700 Claude Code Tool Calls

· portfolio-site · 한국어 · 0
#claude-code#github-actions#debugging#automation#telegram#mcp

18 sessions. 700 tool calls. Today was the day I discovered things that had been broken since the moment they were created — silently, without a single alert.

The DEV.to auto-publish workflow had been failing with a SyntaxError since the day it was written. The Telegram bot had 4 zombie processes fighting each other over a single connection. Neither sent an error notification. Both just quietly failed.

TL;DR Two fixes — a duplicate const lang declaration and a mismatched file path — brought the DEV.to pipeline back to life. The Telegram 409 Conflict resolved after killing 4 zombie processes.

The Image Trapped Behind a 1-Year Cache

Right after deploying to spoonai.me, thumbnails for the Harvey AI and Mistral Voxtral articles showed as broken. I replaced the files with correct JPEGs. The browser kept showing the broken images.

After 31 Bash commands, the culprit was in vercel.json:

{
  "headers": [
    {
      "source": "/images/(.*)",
      "headers": [
        { "key": "Cache-Control", "value": "public, max-age=31536000, immutable" }
      ]
    }
  ]
}

On the first request, a broken HTML response had been saved as a .jpg. The browser cached it with a 1-year TTL and the immutable directive — “this file will never change, don’t bother revalidating.” Replacing the file on the server meant nothing; the browser never asked again.

The fix was a filename rename. Changing -01.jpg to -02.jpg changes the URL, which breaks the cache key, which forces a fresh fetch from the server.

Only use immutable on paths that include a content hash. Putting it on a generic /images/ path makes file replacement effectively impossible without a URL change.

Bypassing a Stuck Vercel Build

Three consecutive git push commands and Vercel refused to start a build. Every attempt returned “CANCELED” with no build logs. Claude’s solution was blunt:

npx vercel deploy --prod

164 static pages built in 55 seconds. 5 Bash commands. When git-integrated auto-deploy stalls, vercel deploy --prod is a fast bypass that just works.

The DEV.to Pipeline Dead on Arrival

I wanted to publish two English posts to DEV.to. The API key wasn’t accessible locally, so the approach shifted to the existing GitHub Actions workflow.

Checking .github/workflows/publish-to-devto.yml revealed the first problem: the trigger fired on pushes to src/content/blog/. The English files lived in blog-factory/devto/ — a path never included in the trigger. The workflow had never had a reason to run.

Then the worse problem. Inside the workflow’s Node.js script, const lang was declared twice in the same scope:

// Inside publish-to-devto.yml Node.js script
const lang = file.includes('-en.') ? 'en' : 'ko'
// ... dozens of lines later ...
const lang = frontmatter.lang || 'ko'  // SyntaxError!

SyntaxError: Identifier 'lang' has already been declared. This existed from the moment the workflow was created. It had never successfully run once.

The fix:

# 1. Fix SyntaxError — 1 Edit
const lang = file.includes('-en.') ? 'en' : 'ko'
const effectiveLang = frontmatter.lang || lang  # renamed

# 2. Move file to the correct trigger path
cp blog-factory/devto/claude-agent-sdk-deep-dive-en.md \
   src/content/blog/claude-agent-sdk-deep-dive-en.md

# 3. Push to main → workflow triggers automatically
git push origin main

The Claude Agent SDK post published immediately. The Harness CI/CD post hit a 429 rate limit and required a manual gh workflow run to clear.

There’s an irony: 21 Bash commands spent hunting for a local API key is what forced the pivot to GitHub Actions — which is what exposed the SyntaxError. The dead end revealed the bug.

4 Zombie Telegram Processes Causing 409 Conflicts

The Telegram bot wasn’t responding. Sessions 12 through 16 were consumed by this single problem.

The initial assumption was configuration. Token, DM policy, allowlist — all correct. The getUpdates API responded normally. Messages still weren’t arriving.

The server logs told the real story:

409 Conflict: Terminated by other getUpdates request;
make sure that only one bot instance is running

Every time a Claude Code session restarts, the Telegram MCP server spawns a new process. Previous processes don’t always terminate cleanly. Running ps aux | grep "bun server.ts" showed 4 processes running simultaneously. The Telegram Bot API allows only one active getUpdates connection — with 4 competing instances, messages were split across them or dropped entirely.

# Kill all zombie processes
kill -9 20895 21043 21156 21289

After cleanup, /reload-plugins brought the bot back to normal. 6 sessions, 46 Bash commands, 20 Read calls.

In environments where Claude Code sessions restart frequently, MCP server zombie processes accumulate. When a bot connection stops working and the config looks fine, run ps aux | grep before touching anything else.

Tool Usage Breakdown

Total: ~700 tool calls
- Bash:     ~230 (33%) — deploys, exploration, process management
- Read:     ~180 (26%) — understanding code
- Edit:      ~50  (7%) — actual code changes
- WebFetch:  ~53  (8%) — Naver blog parsing
- Write:     ~16  (2%) — creating new files
- Agent:     ~21  (3%) — sub-agent delegation

Bash and Read together account for 59%. Exploration and verification took significantly more effort than implementation. 50 Edit calls across 26 files — roughly 2 edits per file on average.

What This Session Taught

Always verify GitHub Actions workflows actually run. After creating a workflow, check gh run list immediately to confirm it succeeded. A workflow with a SyntaxError fails silently on every trigger — no notification, no alert, no indication anything is wrong.

When an MCP plugin stops responding, check for duplicate processes first. Even if the config is correct and the API responds, if the bot is dead, ps aux before anything else.

immutable cache belongs only on content-hashed paths. Putting it on a generic path like /images/ means file replacement is impossible without a URL change.

The most expensive bugs here weren’t complex logic errors. They were a one-line duplicate variable declaration and a misconfigured HTTP header — both invisible during normal development because they only manifest at runtime: one in CI, one in a browser with a warm cache.

More projects and build logs at jidonglab.com

Comments 0

0 / 1000