HTTP 200, No Email

For seven days, the pipeline was dead. Every morning at 09:03 KST, the cron job fired, the script ran, and nothing happened. No errors surfaced. No alerts fired. From the outside, it just looked like “no email today.”

Today I found it. 8 sessions, 153 tool calls. Zero new features written.

TL;DR A Claude CLI logout silently killed the AI news crawl from March 17. AgentCrow dispatched 3 agents across 2 phases to recover it. Got HTTP 200 on a test send, still no email — turned out to be spam. Fixed the email design. Sent to all 11 subscribers. Then removed the feedback section entirely before the next run.

The Bug That Hid for 7 Days

First session: 1 minute, 21 tool calls. Started with “why no email again today.”

Scanned ~/spoonai/. The cron log showed generate-ai-news.sh firing daily at 09:03 — but exiting with code 1 every time. The error was buried in the subprocess output:

Not logged in · Please run /login

Familiar error. The problem was the timestamp. The logs went back to March 17 — seven consecutive days of the same failure.

The cascade was clean once you saw it: Claude CLI logout → zero news generated → empty crawl JSON → email HTML can’t render → 0 sends. The first step of the pipeline had been silently failing for a week while the symptom was just “no email.”

Diagnosis took 21 tool calls: 15 Bash runs to read through logs, 3 Glob calls to check whether crawl JSONs existed.

AgentCrow: 2-Phase Pipeline Recovery

After logging back in, the recovery session ran for 14 minutes across 13 tool calls.

The pipeline has a strict dependency order — crawling has to finish before article generation or email rendering can start. So I set up a sequential-then-parallel structure:

━━━ 🐦 AgentCrow ━━━━━━━━━━━━━━━━━━━━━
Dispatching 3 agents (2-phase sequential):

Phase 1:
🔄 @data_pipeline_engineer → crawl → ~/spoonai/crawl/2026-03-24.json

Phase 2 (parallel, after Phase 1):
🖥️ @senior_developer → generate article MDs + download images + git push
📝 @writer            → generate email HTML (ko/en)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Phase 1 collected 52 items, filtered down to 28 (TOP5, BUZZ3, PAPER1, QUICK17, GITHUB3). That JSON fed Phase 2.

Output: 6 article MDs (NVIDIA GTC, Anthropic Marketplace, Meta Llama 4 — both ko and en). 3 OG images failed to fetch, swapped to Wikimedia CC alternatives. Site deployed.

Test send session: 2 minutes, 9 tool calls. Sent to [email protected] via send-email.js. Resend API responded 200, message ID ebf3a2a5-....

“Still not here.”

Second send session kicked off (2 minutes, 8 tool calls). Re-confirmed file paths, checked an alternate output directory (~/spoonai-site/data/output/), explicitly specified the HTML file and script this time. Another send. ID ad836686-..., another 200.

Then: “Check the spam folder.”

It was there. Both of them. The first send had worked fine — Resend had processed it successfully. I should have checked the delivery status dashboard before resending. Instead, I fired a second send immediately. Two test emails went out today.

Email Redesign: 48 Tool Calls

Once delivery was confirmed, feedback came in: “The design changes I asked for earlier aren’t applied.”

Session 5: 9 minutes, 48 tool calls. Four things to fix: smaller keyword tags with more vivid colors, smaller link buttons, improved feedback section design, full TLDR style guide applied throughout.

Used email-2026-03-21-ko.html as a reference and diffed the CSS patterns. 14 Edit calls applied the changes to both ko and en HTML files. 7 TodoWrite calls tracked each of the four fix items.

After the fixes, generated two new base template files: email-template-ko.html and email-template-en.html. Next time, the structure doesn’t need to be rebuilt from scratch.

11 Subscribers, 11 Sent, 0 Failed

One message after design review: “Send the full list.”

Session: 4 minutes, 18 tool calls. subscribers.txt format is ko,[email protected]. Subscriber count: ko 10, en 1. Individual API calls per recipient.

Lang	Subject	Recipients
ko	`[spoonai] 3/24 Mon — 젠슨 황이 터뜨린 1조 달러 AI 수주`	10
en	`[spoonai] 3/24 Mon — Jensen Huang's $1T AI Order Bombshell`	1

Success: 11. Failed: 0. Bash ran 9 times to write and execute the send script.

Remove the Feedback Section — Same Day

Right after the send completed: “Starting tomorrow, remove the ‘how was this briefing’ section. Update the skill too.”

This took two sessions (session 7: 9 minutes, 26 tool calls; session 8: 0 minutes, 10 tool calls). Session 7 edited 2 SKILL.md files and both HTML templates. Session 8 ran grep to find leftover CSS class references and cleaned them out.

Two sessions for one task because session 7 wasn’t complete. CSS class descriptions were still inlined in SKILL.md — missed on the first pass. Four Grep calls covered feedback, 피드백, 어땠어, and survey across the codebase.

Skill files are instructions to Claude, not code. Stale content in a skill file means Claude will regenerate removed features in the next session.

Tool Usage Breakdown

8 sessions. 153 total tool calls.

Tool	Count	Purpose
Bash	51	Log inspection, send scripts
Read	34	File content review
Edit	28	CSS, HTML, skill file edits
Glob	10	File location checks
TodoWrite	7	Fix item tracking
Grep	6	Keyword search

2 files created (email-template-ko.html, email-template-en.html), 6 files modified. No feature code written. Pipeline recovered, design fixed, 11 emails delivered.

More projects and build logs at jidonglab.com