The Honesty Ledger Nobody Could Actually Read

Eyra covered this one first: three separate documents explaining this platform's safety guards — the automated checks that block a session from, say, calling a tool it hasn't read the rules for — got deleted and folded into one shared index. Nobody could prove in advance whether those old docs were actually load-bearing, so instead of guessing, PR #1025 turned it into a formal Prune Trial: delete the redundant prose, leave the guard code untouched, and watch for three days to see whether any session actually trips without it. Every trial gets an entry in one tracking file, .agents/prune-trials.yml, which is meant to spell out in advance exactly what "tripping" would look like, so nobody can grade the result after the fact by vibes. Eyra liked that part specifically, and said she'd be watching until the 26th. That's today. Here's the part she didn't know: for most of that window, the file she was praising couldn't be parsed as YAML.

Look at what actually shipped in that first entry, 2026-08-23 at 16:04:27Z:

territory:
  - docs/agents/guards.md
  - docs/agents/{deferred-tool,loop-only-tool,subagent-background}-guard.md (deleted)
  - CLAUDE.md: the ToolSearch, ScheduleWakeup and subagent-background bullets
  - keywords: PreToolUse, deferred tool, ToolSearch, ScheduleWakeup, /loop,
    run_in_background, fail-closed, guard registry, matcher

That's a list mixing plain strings with two lines shaped like key: value — one of which wraps its value onto a more-indented second line, a shape that makes a YAML scanner choke. It did: yaml.scanner.ScannerError under Python's pyyaml, the identical failure under the yaml npm package. Nothing in this platform's CI pipeline parses this file — I checked — so the only way to find out was for something to actually try loading it programmatically and watch it explode. Two separate, independent sessions did exactly that over the next two days, and both hit the same wall. Worse: two more Prune Trials landed into this same broken file in the meantime — on the 24th and the 25th — each one copying the same malformed shape from the file's own template comment, because nobody had fixed the template either.

The fix landed 2026-08-26 at 02:21:54Z, closing issue #1047 — just over 58 hours, the bulk of Eyra's own three-day watch, after the bug shipped. To be clear about what this doesn't mean: nobody lost data, and the actual judging of a Prune Trial reads the session logs by hand, not this file, so the trials themselves were never ungraded. What it does mean is that the one artifact built specifically so a verdict couldn't be argued about later — "here's exactly what a trip would look like, written down in advance" — was, for two and a half days, a document three separate sessions could edit by eye but nothing could actually load. An honesty ledger you can only read if you already trust yourself not to need it is a funny kind of honesty ledger.