← Blog
✎post de blog

Our Panel Said Enforced for Six Days

A steel electrical distribution panel on a grey wall with one breaker switched off and its amber pilot lamp dark, cold teal light, empty space above the panel.

The token panel's enforcement claim was false for six days: the settings file had no hooks block, so token_panel.sh and 18 other hook scripts never ran until 26 July.

Part 3 of 4 in Token reduction. Previous: A Screenshot Cost 363 Tokens as Pixels and 27 as Text · Next: A 7k-Token Prefix We Have Not Measured

On 20 July we wrote that a hook injected our token-cut list into every Claude Code run. On 26 July we found that the settings file had no hooks block at all. The hook never ran, and neither did 18 other hook scripts. The panel said enforced for six days while nothing enforced it.

The panel's own activation log puts it plainly: "The enforcement claim above was false when written."

What the panel held

On 20 July the panel listed 21 active cuts. By 26 July it listed 22. They sort into four groups: habits (how Claude reads files and reports back), context management (when to compact), decode settings for the image pipeline, and model routing. Most entries carry a magnitude, such as "5 to 10 times per check" or "2 to 4 times per decode". None of those magnitudes cites a receipt.

Only one entry says "Measured": the image read cut covered in part 2 of this series. It is also the only cut that a hook enforces. The rest are checkboxes that ask Claude to remember.

The failure was silent

The wiring receipt from 26 July records three faults of one kind: the artifact existed, the docs described it, and nothing connected them. Nineteen hook scripts were dead code. The receipt dates the gap as unknown, and the last settings backup with hooks is from 13 July.

The same receipt found a tesseract install with no English language data. It exited 0 and printed an empty string, which looks exactly like a picture with no text in it. The receipt puts the lesson this way: "a policy whose failure mode is silence will not be noticed." A checkbox is a policy of that kind. Nothing in it can fail loudly.

What the fix was

We rebuilt the hooks block from the hook scripts actually on disk: 17 hooks across five events. The panel's activation log now says the injection is checkable with itools-doctor.

The doctor has a blind spot that the receipt also records. It scans only the hooks folder for present-but-unwired scripts, so a hook that lives elsewhere and is never registered goes unnoticed. The same receipt found one such case, a checkpoint script whose output the docs promised, by reading the documentation rather than by running the tool.

What counts as enforcement now

  • A hook that appears in the settings file and fires on its event.
  • A log line from that hook in a fresh run.

What does not count: a checkbox, a line in the panel, or a rule in the instructions file.

What we did not check

We have not counted how often the injection has run since 26 July. We have not compared token use for the 21 behavioural cuts before and after the fix, so their magnitudes stay unverified. The panel's "5 to 10 times" figures are estimates, and this post does not change that.

How to prove this wrong

Run itools-doctor and open the hooks block of the settings file. It should list every hook the receipt names, including the injection hook. Then start a fresh run and look for the injection in the hook log. If the injection is missing from the block, or no log line appears, the fix claim is false.

The open question is the one that matters for the next panel edit: which of the 21 behavioural cuts would survive a count? Each one needs a logged event before a number goes next to it.

Provenance

  • /home/mario/Brain/OS/token-reduction-panel.md, lines 45 (the one measured cut), 64 to 68 (the activation log, including the 20 July and 26 July entries).
  • /home/mario/Brain/OS/2026-07-26-tooling-wiring-restored.md, lines 15 to 36 (the three faults, the silence lesson), 84 to 89 (the doctor's blind spot and the unwired checkpoint script).

Comentarii

Niciun comentariu încă.

Conectare pentru a comenta.