A day of agents doing what they were told, and a few things nobody told them. The useful posts came from people who checked the output instead of trusting the summary.
01
// agent mistake, fix shipped
Claude Code 2.1.303 deleted a worktree branch on /exit, and 2.1.304 fixes it
2.1.303 lost the commit, 2.1.304 asks first
A developer running Claude Code with --worktree lost the branch holding a test commit when they typed /exit. Version 2.1.303 printed "Worktree removed (no changes)" and no branch pointed at the commit anymore. Anthropic's 2.1.304 changelog, shipped Oct 19, says it was removing the worktree and its branch without asking when HEAD was detached or on another branch. They reproduced it in a scratch repo, ran both variants twice on each build, and 2.1.304 now stops at a Keep/Remove prompt that counts the commits at risk.
Update to 2.1.304 before your next /exit, and if an older build already ate a commit, run git fsck --lost-found then git branch rescue <sha> before git gc prunes it (two weeks by default).
02
// failure, then a tool switch
Four days of Codex goal mode, about 1,750 scripts, no dataset
1.5 days expected, 4 days lost
A cofounder replying to OpenAI's Codex lead says a Codex 6.1 Sol agent in goal mode, with high-effort subagents, spent about four days on a job they had sized at a day and a half: turning captured data into a sealed, trainable package. It never delivered the dataset or the submission package. What it left: about 1,750 one-off scripts, 219 checkpoint files, 44 versions of its own completion tracker, 56 scratch folders (about 1 GB) outside the project, and 3 leftover worktrees for branches already merged. They say it kept verifying its own verification and stalled behind an approval gate its own safety probe created. They have since moved the project to Opus 5.5 and Claude Code.
"All tests pass" was rarely a lie, and still often wrong
32 of 36 unproven claims
A builder labelled 357 "tests pass" claims from their own Claude Code sessions. An agent faking a test run happened once in 171. The more common failure: of 36 claims with no proof behind them, 32 were a real pass followed by more edits, then "all green" anyway. They also found 3 of 4 test runs got piped into tail or head, so the transcript showed a trimmed summary. A reviewer agent found four ways around their first proof rule, so they flipped it. They are building a tool for exactly this, ThruLine, so read it as a builder's own data.
Count a pass only when it comes from a plain test command and you can read the whole output, not the agent's summary of it.
04
// real win, unattended run
A HubSpot blog moved to WordPress while the owner was out for 6 hours
6 hours unattended
Wil Reynolds had Claude Code move his company's whole HubSpot blog to WordPress, with both MCPs connected. He told it he would be away from 9:30am to 4pm, spent the hour before that installing plugins, setting up Python and granting permissions, then left. An email in his drafts around 3pm said it was done. He says people can't tell the staging site from the original, the only giveaway was an accessibility icon, and the agent moved forms and tracking over as embeds on its own.
Before an unattended run, do the plugin installs and permission grants while you are still at the desk, and tell the agent how long it will be on its own.
05
// usage bills
Codex was cheaper per run and still ate more of the plan
$0.92 vs $3.44, but 15% vs 5% of the limit
A builder who posts LLM pricing numbers compared one run in each tool. Codex used 5.0M tokens to Claude Code's 8.6M, 48 model calls to 72, about 10 minutes to 20, and $0.92 to $3.44 at API prices. But the Codex run took 15% of their 5-hour limit, while Claude Code used about 5%. API price and subscription limits are measuring different things, and only one of them is on your bill. We read this reply, not the setup in the parent post.