Claude Code Insights — binary.ovh VPS (Jul 17–18 2026)

Claude Code Insights — binary.ovh VPS (Jul 17–18 2026)

📊 4 sessions · 28 messages · ~7h · 2 commits · 2026-07-17 → 2026-07-18. Auto-generated from Claude Code session telemetry.

At a glance

✅ What's working

A plan-then-ship style: detailed infra planning paired with immediate hands-on execution — real services deployed and backups wired in the same session, not just a document. Open-ended research chained into synthesis, benchmarking, and deployed artifacts.

⚠️ What's hindering

A few runs stalled on environment/dependency mismatches (a Chromium version conflict, repeatedly failing probes). Deployments picked up auth walls you never asked for. Large multi-phase sessions hit limits mid-flight; blockers like missing credentials surfaced partway through instead of up front.

⚡ Quick wins

Capture the repeated Bash-driven infra steps into Custom Skills (reusable /commands) so provisioning runs the same way every time. Add Hooks to auto-verify tool/env versions before tests, cutting the dependency mismatches.

🚀 Ambitious workflows

A self-driving deploy-and-verify pipeline that provisions services, wires backups, health-checks everything live, and self-heals common failures — pausing only for credential-gated steps. Parallel agents that research, benchmark, and scaffold at once.

What you did well

• Plan-then-deploy — you write the plan and ship it in the same session, turning strategy into running systems (Forgejo, Uptime Kuma, backups) with no handoff gap.
• Research → benchmark → deploy pipelines — WebSearch research flows into synthesis, benchmarking, and a live webapp, so you get from open question to testable artifact, not a report.
• Ideation + housekeeping together — 'what's worth building next' comes back ranked and synthesized, while practical cleanup (~11–19GB of disk reclaimed) keeps the box lean.

Where things go wrong

Unwanted auth / access barriers

Deployments repeatedly ended up with password walls you didn't want — entering a password 'for nothing.' Stating access preferences (public vs. gated) up front prevents the rework.

• The brainbox webapp shipped with a password wall you never asked for — had to be removed.
• In the local-LLM pipeline session, the deployed webapp's auth wall was unwanted and reworked.

Environment / dependency mismatches

Several runs stalled on version conflicts and failing probes that ate debugging cycles. Pinning tool versions and verifying compatibility before tests would cut these detours.

• A Playwright/Chromium version mismatch needed extra fixes during UI testing.
• Bonsai benchmark probes failed repeatedly with exit 144, blocking the benchmarking goal.

Long-running work cut off before completion

Ambitious multi-phase sessions (research + build + deploy + explore) tend to hit session or time limits, leaving the final phase unfinished. Sequenced, checkpointed chunks finish each phase reliably.

• The final research phase was interrupted after repeated benchmark failures consumed time.
• The paper-trading session hit a session limit right after deploy, cutting off exploration.

Project areas

Suggestions

CLAUDE.md additions

• No auth walls by default — do NOT add authentication/password walls to internal or self-hosted apps unless explicitly requested.
• Pin browser-test versions — pin Playwright + Chromium and run `npx playwright install` before UI tests.
• Front-load credentials — when deploying infra, confirm which secrets (e.g. B2 keys) are needed up front and surface blockers early.
• Codify the workflow — research → build → test → deploy, and always leave the app running and verified reachable before ending.

Features to try

🧩 Custom Skills

Reusable /commands as markdown for the repeated research→benchmark→build→deploy flow — launch it in one command instead of re-describing it.

🔌 MCP Servers

You already planned an MCP gateway; a server lets Claude query service status and repos directly instead of ~201 shelled-out Bash calls.

🪝 Hooks

Shell commands at lifecycle events — auto-run `playwright install` and lint before tests to kill the setup-mismatch friction.

On the horizon

Autonomous infra deploy-and-verify pipeline

Chain research, provisioning, deployment, and health-checking into one self-driving workflow that spins up services, wires backups, and confirms everything live — self-healing common failures (version mismatches, exit-144 probes) by retrying with corrected configs before ever asking you, pausing only for credential-gated steps.

Parallel research-to-build agent swarm

Fan out agents that simultaneously research what's worth hosting, benchmark local-LLM options, and scaffold a candidate app — then converge into one ranked, buildable roadmap. Parallel lanes finish the exploration a single 7-hour serial session runs out of room for.

Test-driven build with a self-correction loop

Turn app work like the paper-trading project into a closed loop: write tests first (explicitly asserting no unwanted auth wall), build against them, and iterate autonomously until green and deployed cleanly. You review a passing suite instead of debugging exit codes.

And the fun one

😄 Claude proudly ships a password wall; user retorts they're entering a password 'for nothing.' During the local-LLM benchmark-and-deploy session, an auth wall got added to the brainbox webapp that was never wanted — prompting a blunt complaint before the final research phase was interrupted.