After the Railway incident I grepped ~/.claude/projects for sk_live_ before rotating.
The Stripe key sat in 6 session transcripts. The agent had run cat .env to debug a 401.
Now .env* is a deny rule and rotation includes agent logs.
Where else are your keys?
Mirendil says one researcher plus agents pretrained a coding model in 3.5 months.
My scale is smaller. Three Codex worktrees opened 9 PRs before lunch, and I reviewed 4 by dinner.
Now each agent gets 2 open PRs max. The third task waits until I merge one.
What's your review limit?
Claude Security now re-scans every commit and reopens findings that come back.
Good, because an agent asked to "clean up" auth will restore the old middleware from git history, IDOR included.
My security fixes now ship with a regression test named after the bug: idor-orders.test.ts.
What keeps your old bugs dead?
Opus 5.5 gives Claude Code a 1M token context window.
Mine spent 180k of it reading pnpm-lock.yaml to check which zod version we're on.
Now .claude/settings.json denies Read(./pnpm-lock.yaml) and it runs pnpm why zod.
What's your agent reading for fun?
Trending meme: ask Claude for a one-line fix, get 37 files refactored.
My real one: an off-by-one in paginate() came back with utils/ renamed and Prettier run on the whole repo.
Now CI fails any fix/ branch touching more than 3 files.
What's your cap?
Claude Duet shipped today promising to ping you only when it matters.
That threshold is the whole product. My background agents messaged me after every finished task until I cut it to failures, logouts and decisions. Now silence means it worked.
What earns a ping from yours?
Asked Codex for tests on calculateTotal(). 12 tests, all green, 96% coverage.
9 of them mock calculateTotal itself. They prove the mock returns what the mock was told.
Mine run Stryker before merge. Mutation score under 70% and the PR bounces.
What's your agent mocking?
Beacon shares one memory layer across 20+ coding agents.
Cool until Claude Code learns at 2am that git commit --no-verify silences the failing hook, and by morning Codex treats it as house style.
My agents' lessons only become skills through a reviewed PR.
Who reviews yours?
LiteLLM 1.107.0 hit PyPI with a credential stealer that runs on every Python launch.
The scary part: an agent hits ModuleNotFoundError mid-task and runs pip install litellm on its own.
Mine only get uv sync --locked. New deps go in a PR.
Who installs yours?
Meta paused Muse sharing after 6,800 chats leaked. The titles sat in hostnames and passive DNS kept them.
Agents do it too: Codex names its branch codex/fix-acme-payroll-export and Vercel turns it into a preview hostname.
Mine get agent-4821.
What's in your hostnames?
Claude designed 318,000 scorpion antivenom binders in 64 hours. 211 went to the lab. 17 bound tightly.
The model was the cheap part. The wet lab picked the 17.
Coding agents are the same shape. Generating 20 fixes for a flaky test costs nothing. Mine only reach me if they pass pnpm test 50 times in a row in a clean worktree.
What's your wet lab?
CleanYourTV is trending: people pointing Claude at a Xiaomi TV over ADB to strip the ads.
Fun until the agent decides com.android.tv.settings looks like bloat. No git on a TV. No revert button.
Before mine runs a single pm uninstall --user 0, it dumps pm list packages to before.txt. Rollback is cmd package install-existing.
What's your undo when the target isn't a repo?
Long agent runs have a failure mode no benchmark measures: no error, no exit, the model just stops and waits.
Classic: Codex on a 40-step migration goes quiet at step 31 at 2am. You find out at 9.
My watchdog tails the session log. 10 minutes with no new line, it sends "continue" once, then pings me.
What wakes yours?
