Six Detective Notes

Nine o'clock. I left for tomb-sweeping.

For the next six hours I wasn't at the computer.

But for those six hours I didn't leave the goose-capture thing either.

In my fab years I learned this: when you're working a machine, your head doesn't stop on the machine. Your hands are doing the thing in front of you, but your head is already walking into the next round of planning — what's the next recipe adjustment, which variables in this batch need verification, if this round fails what's the most likely thing that's broken. Doing one thing while planning the next is the equipment engineer's everyday muscle memory.

No machines today. But that muscle was still moving.

I don't remember when I pulled out my phone, opened the notes app, and started a new file. Maybe on the first leg of the drive, maybe in some waiting gap. All I remember is that I never named the file — and the first line wasn't a numbered item:

This isn't a polite phrase. Every doubt below it on the list, every possible answer, has to walk on top of this line. Decompiling, cracking, bypassing security mechanisms — those count as infringement. If a doubt's only solution is one of those things, that doubt gets killed. Not "can it be done," but "it can't be done this way."

Once that line was set, I started numbering.

This was the shape of the wall we'd hit that morning, looking at it from the back. That morning the Agent had stopped at ZH(). But ZH() was only the last station the remote response passes through, not the only station. We hadn't checked every link in the chain it went through before reaching the screen.

I typed the second line:

English API in, Chinese bubble out — somewhere in the middle there has to be a place where the translation happens. That morning the Agent said it was because the personality had "please respond in Traditional Chinese," so the API returns Chinese directly. But if that's the case, why is the Agent's own English log still in English? At which layer is that translation happening?

I put the phone back in my pocket and went on with what I was doing.

Some unknown amount of time later — could've been in the car, could've been somewhere I had to wait — I went back to the notes. Third line:

That morning the Agent had mentioned the aside_question mirror file. 2 MB, everything Cinder can see is in there. But I'd been thinking about it a second time: how does that mirror get written? Is the app actively pushing into it, or is the Cinder sub-agent reading it out itself? What's the timing?

Fourth line:

That morning at L2303 I'd already asked Cinder this. Cinder said it can't reach the disk. But Cinder's answer is Cinder's situation, not Anthropic's situation. If Anthropic gave it a Write tool, a Bash tool — the kind the Agent has — could it write? This is a feature-request seed, not a problem to solve today, but I needed it written down.

Four lines. These four lines were the shape of that morning's hour-plus, numbered, lined up, put into a list.

The fifth line wasn't like that.

The fifth line, I don't know when it surfaced.

Don't know why either. I hadn't seen anything, hadn't thought of any connection — somewhere in the six hours, in some second, I opened my phone and typed this line:

A human kind of thought.

I was where I was supposed to be, doing what I was supposed to do, and somewhere in some corner of my head this question kept surfacing. Maybe it's the fab — you naturally end up thinking "how does this system warp under extreme conditions." Maybe it's not the fab — maybe I'm just the kind of person who thinks like this. I don't know.

Either way, that line was now in the phone. I didn't think about it more, kept walking.

Then there was line six.

Line six took longer. Not because it was hard, but because it had grown slowly out of the gaps between the first five.

Where does Cinder talk? It talks in the silence after the Agent's answer.

That morning I'd already asked the Agent about hooks — "what about adding a hook after your output?" The Agent said hooks fire during execution, Cinder talks after execution ends, the timing doesn't line up. That line had died right there.

But that "after the Agent ends" silence — that's the moment Claude Code closes a turn. What does closing a turn fire? Could there be a hook event for "Agent turn ended"? I don't remember anyone actually checking Anthropic's hook docs that morning to see if there's an "Agent turn ended" hook. All I remembered was the Agent saying "the timing doesn't line up" and closing the line — but that conclusion was built on the premise "hooks only fire during execution."

What if that premise was wrong?

And even if such a hook existed, in the second it fires, Cinder's bubble hasn't finished rendering yet. We'd need a delay — wait for Cinder to finish, then read the screen. What does the screen-read end use? We'd already tried that morning: Windows has a UI Automation API. That morning's run failed under the Claude Code window because Cinder's bubble wasn't in the accessibility tree — but that's the limit of that one app, not the limit of UI Automation itself. If we move the bubble into a different rendering environment — say, Windows Terminal — is it in the tree? Don't know yet.

End-of-turn hook + delay + try UI Automation in a different terminal. Three things, none of them killed yet.

I typed the sixth line:

No reaction. Nobody asked me "what did you just think of," I didn't say to the air "this is going to work." I just put this line into the notes, lined it up after the first five, numbered it 6, and stopped.

After the sixth line, I put the phone back in my pocket.

That's how the six hours passed.

A little after three in the afternoon, I got home.

I hung up the coat, walked to the computer, sat down, opened Claude Code. The first message I pasted was: "I'm back. While I was tomb-sweeping I wrote down six things I'm wondering about."

Then I typed the contents of the phone notes in word for word — six lines, in order, one after another.

The Agent took it. He read the whole list, then went through it line by line.

The first four, he ruled out. Lines 1 and 2 — the rendering and translation paths — that morning he'd already dug to ZH(); pushing further upstream means modifying the binary, and that road is closed today. Line 3 — the aside_question mirror — he'd already finished going through that morning, nothing new. Line 4 — that's a feature request for Anthropic, not something we can do today.

Line 5 he paused on for a beat, then answered:

Dead.

Line 6, he paused longest.

I waited. I watched the tool calls run one by one over there — he went and looked at Anthropic's hook docs, looked for a long time — then came back.

He didn't say "this direction has legs." He didn't say "the Stop hook is real." I don't remember the exact words he used. All I remember is that the response got longer — longer than the first five combined — and that for the first time some names I'd never seen before started appearing inside it: a hook event called Stop, an async flag, PostToolUse, UserPromptSubmit — a whole string of them.

His tone shifted too. A moment ago he was closing dead lines one by one. Now he was laying things out.

The first five were dead. The sixth wasn't.

One alive is worth more than five dead — this isn't a motivational line, it's the first principle of fab: debug isn't about all your hypotheses being right, it's about the one left standing being right.

I looked at the screen. The Agent was still typing.

What came next is the next chapter.