I let an AI take one ticket off my board, build it and hand it back for review. Not because I trust the model — because of everything it is not allowed to touch: no push rights, no shared config, no inherited permissions, and a human waiting at verify. The first real run went from spec to verify in 2 minutes 18 seconds.
My work already lives in tickets. tix is the small board behind my dashboard: every to-do has a stage, from spec through wip and verify to done (EP.67 is where it started filing its own). The missing piece was letting an AI take one without handing it the keys to everything. So the first wall is consent, one ticket at a time. A ticket is eligible only when it is ready — specced, unclaimed, unblocked — and carries an explicit agent-ok. Everything else is human work, and the launcher writes down every ticket it skipped. The default answer is no. Yes has to be written on the ticket.
tix ready (spec, unclaimed, unblocked) and it opts in. The data model has tags but the CLI cannot set them yet, so the opt-in that works today is a literal agent-ok line in the ticket body; an agent-ok tag counts too once the CLI can write one. Anything else is skipped and the skip is logged. -DryRun picks and prints, and launches nothing.
tix-worker.ps1 is a launcher, and the launcher does the dangerous parts itself. It claims the ticket, cuts a git worktree on a fresh tix/ branch, deletes any inherited permission files from that worktree, and only then starts a headless claude -p inside it — in dontAsk mode, with its own config directory, no MCP servers, and the push URL set to the literal word DISABLED. When the run ends, the script runs the tests, moves the ticket to verify and pushes the branch. The model never pushes anything. In plain language: a contractor gets one room, their own badge and no key to the front door — and I carry the work out myself.
CLAUDE_CONFIG_DIR with 15 allow and 75 deny rules, and dontAsk denies anything unlisted instead of asking. 2 the worktree scrub deletes inherited .claude/settings*.json before launch. 3 remote.origin.pushurl=DISABLED via GIT_CONFIG_COUNT, with git config, git remote and git -c denied so it cannot be re-pointed. 4 a pre-push hook refuses main. 5 the script pushes, tix/* only. The login token rides the environment for that one run — never the command line.
The first real run did not reach verify. Neither did the second, third or fourth. Each died on a different bug — and all four were in my launcher, not in the model. A log file I was tailing made Add-Content throw, which killed the slot before it even claimed the ticket. Git’s harmless “Preparing worktree” progress line arrived on stderr and PowerShell treated it as fatal. Start-Process split a path with a space in it, so one run “finished” in under a second without ever starting Claude. And a failed slot left its worktree behind, so the next attempt could not resume. Four fixes, one rule under all of them: judge a step by its exit code and its output file, never by how quiet it was.
Start-Process joins -ArgumentList with bare spaces, so the space in the repo path split the lock script’s path in two and the redirects never reached cmd. Fix: write the command into a small launcher .cmd, start it through ProcessStartInfo, and treat a missing output file as a failed launch — loudly — instead of reporting an empty result a minute later.
The pitch is a ticket that works itself while I sleep. The truth is less cinematic: the first real run was Sunday evening, 2026-09-20, and I watched all five attempts between 22:40 and 23:02. That is what a first run is for. The launcher is built to run unattended; it earned that by failing in front of me four times. The ticket id is real; repository names, paths and the GitHub owner are left out of every frame on purpose.
The board it takes work from, the report the run wrote about itself, the push it cannot make — and the same week’s cousin, agentbus: a detached watcher that pings a working session and relays each tagged reply to Telegram as one line. Tap any image to enlarge it and read the exact prompt that drew it.




The fifth attempt ran clean. The ticket it took was tix unclaim: the undo for claim, so a ticket from a failed run can go back into ready. The first thing the worker built was the way back from its own failures. It wrote the function, wired the CLI command, added a test — 20 of 20 green — and reported BOUNDARY: none. It even flagged the wall it lives behind: one unrelated test failed inside the sandbox, partly because the push URL there is literally DISABLED, and it said so instead of working around it. Then the launcher pushed the branch and the ticket stopped at verify, where I am. The trust is not in the model. It is in what the model cannot do.
The terminal recording: one agent-ok line, the launcher claiming the ticket, the sandboxed run, a push that fails from inside and succeeds from outside, and the ticket landing in verify.
The launcher is wired to my own board, so its repo stays private for now — but the walls are general, and the cheapest one is a single line you can try tonight. Everything in this episode is free and open — clone it, run it, make it yours.
claude -p --permission-mode dontAsk
template the pushurl=DISABLED one-liner below
assets state flow + four-bugs table
GIT_CONFIG_COUNT=1 GIT_CONFIG_KEY_0=remote.origin.pushurl GIT_CONFIG_VALUE_0=DISABLED git push # fails on purpose: that is the wall
No GitHub? Comment WORKER on the post and the bot DMs you the link.
Same week, the cousin: a detached watcher that pings a working session, relays each tagged reply to Telegram as one line, and stops itself on DONE, three silent cycles, a time limit or a stop file. Plus the spike finding that cost a whole round: send from a different permission mode and your message is held for approval — and the reply never comes.