18 Sessions, 451 Tool Calls: 5 Parallel Redesigns, a Pre-Deploy Bug Catch, and a Dental SEO Pipeline
451 tool calls across 18 sessions in a single day. Bash ran 175 times, 50 subagent dispatches, 41 web searches. Three completely different domains — coffeechat platform redesign, spoonai site migration, dental marketing diagnostics — all handled in parallel with Claude Code as the orchestrator.
TL;DR: Parallel agents expand your design search space, but initial direction matters more than generation speed. Codex cross-validation caught an SRI hash mismatch that browsers would have blocked in production. Dental marketing reports require hard separation between public data and metrics that need admin access — mixing them is a compliance risk.
Why “All Five Are Bad” Was the Right Answer
The first major task: redesign coffeechat.it.kr, a mentoring platform connecting game industry professionals with candidates for 1:1 coffee chats, resume reviews, and mock interviews.
The brief was simple: “Give me at least 5 variations, I’ll pick one.”
I ran plan-orchestrator to drop a plan.md, then dispatched 5 frontend-implementer subagents simultaneously:
- V1: Editorial Magazine
- V2: Soft Brutalist
- V3: Floating Gradient
- V4: Object-Oriented UI
- V5: Brief Format
Each agent ran in an independent context — no cross-contamination. The output was an index.html comparison canvas where clicking a card opens the variation in a new tab.
Feedback: “None of these look professional. Check Inflearn or other edu platforms.”
Correct. I had explored design trend variations but missed the core requirement: trust signals for a professional education service. I approached it with indie game aesthetic when the reference baseline was Inflearn, Class101, and FastCampus — polished, credibility-first platforms.
The lesson: parallel generation without reference alignment produces generic results. The bottleneck isn’t generation speed — it’s initial direction. “Make me 5 versions” without a reference produces 5 variations of the wrong thing.
Session 1: 78 tool calls, 28 agent dispatches, 25 bash commands.
The SRI Hash Bug Codex Found Before the Browser Did
During the spoonai redesign, design-reviewer flagged V3 as a blocker. After fixing, I saved diff.patch and ran Codex cross-validation.
Codex found an SRI (Subresource Integrity) hash mismatch. V2 through V5 loaded react.production.min.js from CDN, but the integrity attributes were set to development build hashes.
<!-- Wrong: production URL with development hash -->
<script src="react.production.min.js"
integrity="sha384-development-hash...">
The browser compares the hash against the actual file content and blocks the script if they don’t match. Locally, this is invisible — browser cache hides it. In production, you get Integrity check failed and a broken page.
The reason this slipped through: the model that wrote the code tends to validate its own output. When Codex reviewed the diff cold — with no context about how the code was written — it spotted what the original model missed. That’s the value of external cross-validation. Not redundancy, but a genuinely different perspective.
Fix: replace all CDN links with correct production hashes.
The 58KB Box Problem
Session 2 started with a screenshot: an article about OpenAI’s deployment company, filled with □□□□ glyphs. Labels like Total Size, OpenAI Stake, and Structure were rendering without any actual characters.
Root cause: a directly-produced infographic image with credit: "spoonai". File openai-deployment-company-tpg-10b-01.jpg, 58KB — Korean font not embedded.
grep -r 'credit.*spoonai' src/content/posts/
Found two posts (ko/en) referencing the image. Removed the image: block from both frontmatters, deleted the JPG. 3 files, single commit, Vercel auto-deploy.
In the same session: spoonai redesign continued. From the initial 5 variations, 05-brief.html won. Five more variations on that direction. 05a-editorial-premium.html selected as final, applied to the production codebase — feat/editorial-premium-redesign branch, 7 files changed, 442 insertions, 725 deletions.
Session 2: 130 tool calls, 87 bash, 22 agent dispatches.
Dental SEO Reports: Public Data vs. Admin-Only Metrics
Sessions 9–18 were a dental advertising diagnostics pipeline. Two clinics — Yatap NYU Dental and Dongbaek Seoul Pediatric Dental — each needed a free SEO/AEO diagnostic HTML report.
One hard rule governed the entire pipeline: never fabricate metrics that require admin access. Naver Place view counts, phone click rates, monthly search volume, CPC, CTR — all flagged as “requires authorization / unverified” and kept strictly separate from publicly observable data. Korean medical advertising law prohibits performance guarantees and exaggerated claims; those rules applied throughout.
Dongbaek had a naming confusion problem. A search on Mooodac returned “Dongbaek Seoul Dental Clinic” (Dongbaek 7-ro 83) — a completely different clinic from “Dongbaek Seoul Pediatric Dental Clinic,” the actual target. Mixing those two means putting another clinic’s data into the report.
Every search result was labeled “candidate / needs verification.” The user confirmed the correct Naver Place link, which became the authoritative reference for all subsequent data.
Multi-session handoff used file-based state. Each session starts with fresh context, so evidence.md and plan.md were saved first, and the next session prompt included explicit file paths:
Continuation task. Web search not needed.
Input:
- .../research/yatap_nyu_public_evidence.md
- .../research/dongbaek_yu_public_evidence.md
Explicitly saying “web search not needed” prevents the model from re-searching already-verified data.
Codex cross-validation returned request-changes: missing data-label attributes on the comparison summary table for mobile responsive layout, and residual performance-guarantee phrases that needed removal. After fixes, a final piece of feedback arrived: “The design looks too AI-generated.” So I built an 8-direction reference selection board in HTML before touching any redesign — direction-first, then execution.
Orchestration Policy in Three Sessions
A meta-task ran alongside the implementation work: documenting the Hermes + Claude + Codex + OMX multi-model orchestration policy.
Session 5: initial draft. Session 6: verified against local environment (~/.codex/AGENTS.md, openclaw). Session 8: Codex review incorporated, v2 finalized.
Each session’s output became the next session’s input — same handoff pattern as code work.
One key correction surfaced: OMX is not a separate binary. It’s a behavioral pattern composed of ~/.codex/AGENTS.md + project AGENTS.md + Codex prompt/skill layers operating in heavy mode.
Day Stats
| Metric | Count |
|---|---|
| Sessions | 18 |
| Total tool calls | 451 |
| Files created | 23 |
| Files modified | 9 |
| Bash | 175 |
| Agent | 50 |
| WebSearch | 41 |
| Edit | 32 |
Bash at 39% of all calls. CDN URL verification, git operations, build checks, dev server — all accumulated. The 50 agent dispatches split between parallel design generation and cross-validation calls.
Parallel generation widens the exploration space. Cross-validation catches the bugs you’d otherwise find in production. Both are quality insurance — not speed optimizations.
More projects and build logs at jidonglab.com
Comments 0