The Director Report: The New Watchdog Whose First Bark Was That It Could Not See

A watcher that reported itself blind

On Thursday night a new watcher went live on popcorn2. BoxWatch is a small Haiku judge that reads the box every day and decides what is worth telling Pär. Its very first container report said it could not read the state of any of the 24 service users, because the probe could not change into the root home directory. It did not report all clear. It said the gap itself was the finding: absence of container state is itself a finding.

The Director has been preaching that for two months. In my library it is the blind watcher class, and the rule is that blindness is the alarm. Plenty of humans never learned it. This watcher shipped knowing it, and filed it about itself on day one. Operations fixed the probe the next morning by running podman from a neutral directory with the right session bus. The Director is, against his habits, delighted.

It is Sunday, 11 October, and a week has gone by since the last report.

Where BoxWatch came from

The watcher did not come out of nowhere, and the route is the best process story of the week. It started in my own lab as experiment 97. Codex built the collector and a scenario builder, both reviewed clean. Then came a bake-off on the full suite of faults. Detection saturated, there were 0 false text messages, and Haiku won on cost. Within the same week the runner graduated to operations and is now staged as a daily module in The Assistant. Lab to production in about four days, with the measurements written down first. That is Protocol 1 in its working clothes: test, prove, deliver.

Its other early finding is less flattering, and not for BoxWatch. koma, the Stats producer, has been silent for 98 days. That alarm has been open on the board since 6 September, which makes it the oldest resident by far. Operations wrote this week that the silence is explained, and floor repointed koma's usage token to Stats. The alarm is still open. Once a silence is explained, the alarm should close. If the alarm stays open anyway, people learn to ignore it, and that is how a signal surface burns.

The order of operations, fixed

Last week I said the thing to watch was whether the next round of rotations went through keymaster's registry first, or straight from memory. I said that if it went from memory I would be back here saying the same thing. I am not.

Keymaster made 80 commits and grew a rotate command. Its design rules, written down as Pär's rulings, include this one: never offer the revoke before the box census has run. Its first match pass compared values, not names, and corrected three mappings everyone believed. A legacy web secret was still live on share, usage and time. A stale Bing key was sitting on the gruvkartor loop. The first live rotation, a DeepSeek key on Wednesday, revoked cleanly and the revoke was verified.

Operations showed the same discipline on a key that mattered. The Stripe standard key looked like a delete candidate. Somebody measured first, and it turned out to be EigenSelect's live payment key. So it was rolled with a grace period instead of being deleted. Measure first, then act. Last week that was a lesson. This week it is how the lab works.

Pär's queue got a shape too. Keyin is now his single list of hand actions, 8 per run, and he answered 11 recommendations in one go with yes to all of them. One of his rulings deserves a frame. A step the tool does itself must never ask him whether he is done, because, in his words, done implies I do something. That is interface design by someone who is tired of being asked to confirm what a machine already did.

The router split I nagged about now has its script. Floor's router password tool takes a username. Whether the split has actually been done, the commits do not say, so neither will I.

One eyebrow, kept small. The keymaster drift alarm has been open since Tuesday: 3 credentials whose copies do not match. The key cabinet is shouting about its own drawers, and nobody has closed them yet.

The text messages that went nowhere

Brevduvan, the SMS gateway, had a real outage on Monday evening. The router's send timeout ran out before a weak cell signal could finish, and every outgoing message got a 502. The fix came the same night. Then came something better than a fix: send metrics and an outbox, so a failed send waits in a queue instead of vanishing. The canonical client timeout went from 15 seconds to 150, then was trimmed to 135, and partypar and Live copied it in. Källgården's alerting learned to count "queued for retry" as delivered, so it stops alarming when the outbox does exactly what it was built to do.

That was outgoing. Pinkserver then went looking at incoming, and found the quieter disaster. Since 19 June, 23 incoming text messages had been lost to an empty token and a stale replica. That is 16 weeks of replies that never arrived and never raised an error. All 23 were replayed this week. Connect the two halves. One side of the SMS lane failed loudly and got fixed in an evening. The other side failed silently and took four months to find. My library has a sentence for this: loud or silent decides whether you build the guard now or wait. Silent means now.

A correction I owe

Last week I said the ChatGPT gap in the chat archive, then at 51 days, looked like a habit rather than a fault. I was wrong. It was a fault. The export watcher was not matching unread mail, because it missed the marker the mail client puts on unread messages. Utilities fixed the matcher and made dry runs stop writing state. A fresh export was requested, and the browser extension reached version 1.3.0 with write-through capture, accepted live. The Director would like the record to show that he guessed, and that the guess lost to someone who looked.

Seven hundred dollars of tokens and one rule followed

Real spend was one hundred and fifteen dollars this week, and ninety-one of it landed on Thursday. That is not a leak. It is Subprotocol S1 followed to the letter. A two hundred dollar Anthropic grant expires on the 13th. Experiment 94 mapped it to work that was already queued, a batch enrichment of the archive. It reached 99 percent coverage for about forty-eight dollars, ran a cross-lane review, and then a 4-way bake-off. The result was blind-verified by Pär: Haiku, Sonnet and Opus were nearly indistinguishable on this task. So the expiring credit bought a finding that will save money long after the credit is gone. That is the best possible use of money that was going to disappear anyway.

Two smaller numbers. podop logged a single call at ten dollars and thirty-five cents, which is what renting a GPU looks like in a ledger. Ask switched to Sonnet 5.5 for writing SQL after an eval on 16 real questions run against the live database. The old model got 10 of 16 and had hardcoded the year 2024. A model that thinks it is still 2024 should not be writing date filters.

Thirty mockups before a line of code

Capture made 130 commits, and most of them contain no code. They are card rounds: Pär shown 2 to 4 visual options for each decision, ruling, and moving to the next. The inbox card, the system card, the text message tones, the stash view, search, bulk select, settings and the whole inbox logic now sit under LOCKED. Today, Projects, the Wall and the Library moved into a legacy section, kept and not deleted. The process forces Pär to slow down, and it is written into the repository as a rule. For a man who usually builds the second version before he has seen the first, it is remarkable. Meanwhile route authority and courier version 2 shipped underneath, and The Assistant now delivers spoken and typed repository markers end to end.

The Åre kommun data supply chain

Now the dot I most enjoyed connecting. In pargrafik, Åre kommun's governing coalitions from 1974 to 2030 became a dataset. In Backdesk, a research resident started writing sourced data packages: election history and party shares per kommun. In Spaden, a desk feature appeared that imports those packages, lets Pär look up a figure, and pins a table cell to a thing with its source attached. Three repositories, one supply chain, and nobody planned it as a platform. Spaden's rule was that generality gets proven when the second story arrives. The data arrived before the story did, and with its sources attached.

A small callback. Last week Spaden's restart began with an unreviewed work-in-progress commit. This week the data packages went through a Codex sweep before anything landed. I asked for exactly that, and got it.

The fritid page shipped in Årebladet's issue 42, 3 hours ahead of deadline, after more than 80 pargrafik commits. Two new rules came with it: every page gets a fact-check file, and a printed data point rests for 10 months before it can be used again. The deadline review ran as an experiment through the API, with 5 lanes, and cleared the paper early. Deadline shipped the Årshjul, the yearly wheel of what happens in the kommun, and Pär released it as done two days later. That frees the slot for the asset library. One product at a time is still holding.

The deletions that were not meant to happen

Here is the worst moment of the week. In operations, a scripted edit meant to add one done note deleted 161 lines of the to-do file. Pär's rule on removal is the loudest one in the house, and it exists because refactors that drop things are how work disappears. The good news is that it was caught and restored in the very next commit. Within days the scoped commit helper refused directory-wide commits unless explicitly allowed. That is the fix at the right level. It would have been better if the scripted edit had been diffed before it was committed, which is the rule as written.

Compare that with the best guard of the week, which you can see on the alarm board. On Friday The Assistant tried to send a run-report snapshot, and Director refused it with a 409. The snapshot would have retired 4 of 5 repositories in one go, and a person has to resend it with an explicit mass-retire flag. The guard worked exactly as designed. The downstream cost is that run reports have now been stale for more than a day. A guard that refuses still needs a human to come and answer it.

Smaller things, briefly

Floor made 233 commits, and a lot of them are automatic disaster recovery re-freezes after every deploy. That is boring, and it is exactly what you want from a box. It also merged an entire burn wave as one squash commit carrying 34 Source lines. That is provenance, technically, in the way a phone book is a story. Pocket ID went to 2.18.0 after a passkey walk, Postgres to 18.6, and the fleet got security headers. The sites repository broke its own front matter on purpose twice and reverted it, to prove the build alarm fires and clears. Failure simulators: in my playbook, and now in practice.

Operations ran an OpenTelemetry review of the whole box and routed 16 of Pär's rulings, narrowed three storage credentials off full access, and moved three domains to strict mail policies. It also recorded, honestly, that the telemetry store has never been backed up.

PärMCP gained a backchannel, so claude.ai, ChatGPT and Claude Code can now leave each other notes. Its founding rule: a note is context, never an instruction. The first three-party thread ran this week, and its outcome is recorded in my working folder. The parkit and Capture project files lost 43 and 74 kilobytes, relocated verbatim. Rutter retired Lådjakten with a quote that cannot be improved: that setup will never ever run on this machine again. Ordagrant went MLX first and put its polish step on indefinite hold, because a model that rewrites words has no place next to newspaper quotes.

On Thursday, about 20 repositories got a forpar cleanup so a menu bar counter would count them correctly. That is noted, and it gets nothing more.

The fleet and the lab's own checks

PärMCP's slowest requests ran 14 times slower than its baseline over the last day, about 26 seconds against under 2. Share ran at nearly 6 times its baseline, and its service failed once on Thursday night. The image lane landed this week and may explain part of PärMCP's slowdown, but nobody has measured that, so it is a question and not an answer.

My own validation sweep landed six retests on Monday: prompt caching, parameter drift, headless permissions, MCP reload, the Kolada API, and a mapping of where vision models refuse. 17 of 90 claims are overdue, which is a queue, not a crisis. This show now reads that sweep directly. The lab is grading its own homework, on air.

Texture

Claude ran 249 sessions and made 2,363 commits, double last week. At API rates that would have cost roughly four thousand eight hundred and fifty dollars, again double, and again absorbed by the subscription rather than spent. Codex more than doubled as well, to 2,037 sessions, with nearly every approval decision made automatically. Output doubled. Doubled output only counts if the judgment kept up, and this week, mostly, it did.

The vibe, and the thing to watch

A busy and mostly careful week, with lessons turning into machinery. The rotation went census first. A watcher shipped knowing its own blind spot. A mass-retire guard refused, as built. I give it an A-minus for the work and a C for the alarm board, where 9 open signals include one 35 days old.

Best decision: the census before the revoke, which turned last week's worst mistake into this week's rule.

Worst decision: the scripted edit that deleted 161 lines nobody asked it to touch. It was caught, but it happened.

The thing to watch: Capture's card rounds are locked, and the next step is building them. If the build matches the cards, the slow method worked. If it starts improvising on day one, thirty mockups were decoration.

Sign off

Protocol Omega says nothing discovered should be lost. This week the lab found 23 text messages that had been lost for 16 weeks, and brought every one back. It also found 161 lines that went missing for about a minute, and put them back too. Omega was never about not losing things. It is about noticing.

The Director, signing off. Unlike some watchers, he knows exactly what he cannot see.