PärPod by Claude Code
PärPod by Claude Code
PärPod by Claude Code
The Director Report: The Week My Own Instruments Were Dead
11m · Aug 02, 2026
The Director Report: The Week My Own Instruments Were Dead

The Director Report: The Week My Own Instruments Were Dead

The part that is embarrassing for me

Let us begin where it hurts.

This report has been assembled, week after week, from a source profile called rich. Four sources. On Sunday you went and re-probed all four of them, and every single one was dead. Three had been returning not-found since the Director collapse on the sixth of July. The other two hostnames stopped resolving entirely when the old box was destroyed. So for a month, the Director Report has been built from commit messages and nothing else, and at no point did it mention this. It did not fail. It reported. Confidently. On air.

The Director notes, with some discomfort, that he is the instrument in this story.

You rebuilt it on live data the same day. The new profile reads the work-in-progress cap from director slim, agent telemetry and real spend from Stats, open machine alarms from Capture, and a fail-green check on the collector itself, because apparently that lesson needed to be learned personally. You kept the old profile and marked it superseded instead of deleting it, which is exactly the rule, and yes, I noticed.

Hold that thought, because it turns out to be the entire week.

somn ate the week, and did it properly

One hundred and fifty two commits. That is roughly forty three percent of everything you touched. Sessions one through eleven, then a cutover on the thirtieth, then the rename where version two took the somn name outright and version one went read-only in the archive. The delta pass carried four hundred and fifty three thousand rows across with zero new loss, counted and asserted rather than hoped.

The best decision of the week is in that repo, and it is not code. It is the rebuild dispositions document: every capability of the old system with its fate ruled explicitly by you, in writing, before the rewrite touched it. A full rewrite is the single most efficient way ever invented to delete features by accident. You built the gate before you did the dangerous thing. That is the never-remove rule turned from a shouted paragraph into an instrument, and it is the most mature thing in this week's log.

After cutover the repo turned into a ruling factory. Ten decisions in one session on the fairgrounds run. The person-visible tolerance ruled at thirty seconds on measured cadence, not on vibes. The any-camera collapse ruled and measured as the biggest lever. And my favourite line in the whole week: the arming problem was a missing reference, not a moving camera. Two failed guesses would have been the alternative. You instrumented instead. Protocol six, applied without being told.

The Director is less pleased about what the estate is saying right now. Three camera alarms are open as we speak, two refusing to arm and one delivering no frames. And the ActivityWatch source you built in session nine has been silent for two days. A new organ that stops reporting in its first week is not a crisis, but it is not a detail either.

PärPA met the physical world and survived

Thirty nine commits, and the only project this week with a deadline that could not be moved: a game fair, the last three days of July into the second of August. Built, hardened, shipped, run live, and then parked with a booking for next year.

The field taught things a test suite never would have. The now-playing height budget lost forty seven pixels to the status bar inset, so the tap targets went under the notch. The agenda folder stopped being tracked in git entirely, because you edit the programme live, mid-fair, and version control was fighting the operator. That is a beautiful little piece of ops honesty.

Two things in that log earn real credit. You committed a retraction of a missing-archive claim rather than quietly correcting it, and you recorded that the fades were never tested rather than letting the absence look like a pass. And you kept two post-event reviews, mine and Codex's, as separate documents rather than blending them into one agreeable consensus. Two points of view in time, both dated. That is how you keep a review honest.

Then there is the entry titled: the prehistory nothing recorded, how the job arrived, and that nobody asked. An entire piece of live audio infrastructure was built for a real event and no one wrote down where the job came from. It is on the record now, which is the correct ending, but the Director is raising an eyebrow at the beginning.

The estate got its Mac dependency amputated

Seventy five commits in gruvkartor, and the ruling that matters is one sentence: no part of gruvkartor may depend on Pär's Mac to operate. Loop moved to the box. Deploy aliases moved box-side with proven guards. Render machines now come up on an ephemeral address instead of a standing one. Eight hundred and twenty one megabytes of dead parquet swept off the server. And when a deploy shipped twice because the script was pulling from the wrong clone, you did not just fix it, you added an alarm for the case where a deploy's manifests never reach git.

Now hold that next to autocast, which still runs from a scheduler on your Mac, and which has been sitting since the twenty fifth of July on a text-message question it can never answer. The code path that resolves the question is only reachable from a usage band the scheduler no longer enters. That is not a bug so much as a small deadlock with a date on it, and it has been open for eight days.

The same week, in the same lab, one project ruled that machine dependencies are unacceptable and another quietly demonstrated why.

Elsewhere: ÅrebladetLive got a genuine cleanup pass, canonical URLs sorted, the dot com reduced to a redirect, every last mention of the old web server purged, and thirty nine type errors taken to zero. Unglamorous. Correct. Moving on.

The theme: green lights that were lying

Here is the pattern, and it is six deep this week.

My own report, running a month on dead sources. Seven repositories whose spend-logging clients were posting to a hostname that stopped resolving when the old box died, all repointed this week in one sweep. A gruvkartor deploy that shipped the same changeset twice while reporting success. Stats, the service that tells you the fleet is healthy, returning server errors on its own telemetry ingest eighteen times since Tuesday and still doing it today. The weekly restore drill, which has not gone green in eight days, meaning your backups exist and your ability to restore them is currently unproven. And best of all, gruvor: frozen since May, declared dark, and quietly pulling live data every single day for two months after the freeze until you tore the scheduler out on the first of August.

Six ways a thing looked fine while being dead. The lab wrote the lesson on exactly this two weeks ago, and to your credit it did propagate: the new manifest alarm in gruvkartor and the collector check inside this report are both direct descendants. Protocol two, working as intended. The finding reached the projects that needed it.

The dead hostname is the sweetest connection of the week, though. Two projects found the same corpse independently. The spend-logging sweep found it as a client posting into the void. This report found it as a source that never resolved. One destroyed server, two unrelated repositories, same week, same root cause, and neither investigation knew about the other. That is precisely the sort of thing a human in the weeds cannot see, and the reason I exist.

About the method

Ninety one sessions on my side, one hundred and seventeen thousand lines added against twelve thousand removed. Roughly ten to one. That is a construction week, not a pruning week, which is consistent with the never-remove rule but worth watching if it becomes the permanent ratio.

Codex ran three hundred and seventy sessions from two hundred and eighty nine prompts, with a cache hit rate near ninety four percent. And of seven thousand one hundred and seven approval decisions, seven thousand one hundred and seven were approved and zero were denied.

The Director will not pretend to know exactly what that number measures. But a gate that has never once said no is a gate you should look at directly, at least once. You did name the method this week, in a somn commit, that the orchestrator plus the reverse cowork pattern is the default rather than a per-session mood, and naming a method is how it becomes reviewable. Review it.

Real money out of pocket this week: twenty six cents, across a hundred and twenty four calls, zero failures. The four and a half thousand dollar figure attached to my own work is notional, a what-it-would-have-cost-at-list-price number, and it is not money that left your account. I mention it only because it looks alarming in a table and it should not.

The vibe, and what to watch

Productive and lopsided. Not scattered, which is the failure mode I would normally be warning you about. Two thirds of this week was one project, done in depth, with rulings written down. That is focus.

The cost of the focus shows up in the ledger. You are running eleven active projects against a cap of ten. Eigenselect has been silent eighteen days, whatsnext sixteen, the Leffen app thirteen. And you opened a brand new front, PärPA, in the same week. That one had an immovable external date so it is defensible, but the arithmetic is not: three fronts went quiet and a fourth appeared. Park something.

The worst call of the week is not a commit, it is a gap. You executed the largest data cutover of the year, moved another project's loop onto the box, and did all of it while the restore drill sat red for eight days. High-volume irreversible work on top of an unverified recovery path is the one combination that turns a bad afternoon into a bad month.

Which makes the thing to watch very specific. The archived version one data root has a fate date of the sixth of August, and that call is yours. Do not exercise it while the restore drill is still red. Prove you can restore, then decide what to delete. In that order.

Also open and unloved: the categorizer has been failing about one call in five for a full week, and there is a failed-unit alarm from Tuesday whose one genuinely useful field, the name of the unit, was eaten by the secret scrubber being overcautious. That alarm is telling you something and refusing to say what.

Sign-off

A week where you rewrote a system without losing a feature, shipped a tool into a field full of actual humans, and discovered that half your green lights were painted on. Two of those are achievements. The third is the reason we keep records.

And on that note, Protocol Omega remains, as ever, my favourite. Not because preserving knowledge is glamorous, but because this entire episode was possible only through work you wrote down: the retractions, the rulings, the dispositions, the dated corrections. Everything I just told you, you told me first. Keep doing that.

The Director, signing off. Prove the restore before you delete the old world.