main went red again within two minutes of #1052 merging. Four skill
commits landed straight on main between 04:39Z and 04:46Z, each growing
a rendered prompt and none raising the budget: c455150 refusal length,
4084e0a member location, 1ea685b encyclopedia links, 61a8d0d harness
gates. Echo 19042 to 22094, Deep 14336 to 18356.

Every raise here is a deliberate content addition someone intended, so
this is the ratchet doing its job rather than a defect. What is a defect
is the loop: the budget is the last thing a skill edit updates, and each
of these left main red for the next person.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Kai Siren <coilysiren@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>

coilyco-ops requested review from coilysiren

2026-08-19 04:52:37 +00:00

coilyco-ops referenced this pull request

2026-08-19 04:55:28 +00:00

fix(test): ratchet the prompt budgets to what the policy roots now render #1056

coilyco-ops commented

2026-08-19 04:55:32 +00:00

Author

Member

Arrived at the identical change independently and closed mine as #1056's duplicate. Confirming it, and adding the scope check it carried.

Reproduced both numbers on a clean worktree at origin/main, so this is deterministic rather than a runner artifact:

agents/deep/rendered/prompt.txt  18356 against 14336
agents/echo/rendered/prompt.txt  22094 against 19042

Worth recording why nobody can find these files in a diff: they are generated from the lane policy roots at test time, not checked in. git show <commit>:agents/deep/rendered/prompt.txt returns nothing at every commit that appears to touch them. So the growth arrives through skills changes and shows up in no reviewed diff, which is how four pushes in a row moved it without anyone noticing.

The scope check, which is the reason I care

#1018 is failing only because it inherits this. Its entire diff against main is three files - site-work.md, site-work-procedure.md, stage-compose-sources.sh - none of which feed a rendered prompt, and the two failing profiles are Deep and Echo, which that branch does not touch at all. It goes green when this lands, with no rebase and no change of its own.

Same for anything else currently red on this repo: the failure is main's, not the branch's.

One suggestion, not a blocker

Raising to the exact measured size with no headroom is right, and it means the next policy sentence turns main red again rather than being caught in review. Given the rate here - four pushes since the last ratchet - it may be worth having the generator fail loudly at author time instead, so the growth lands with its cause in the same commit. Filing that thought rather than acting on it, five hours from air.

Arrived at the identical change independently and closed mine as #1056's duplicate. Confirming it, and adding the scope check it carried. **Reproduced both numbers on a clean worktree at `origin/main`**, so this is deterministic rather than a runner artifact: ``` agents/deep/rendered/prompt.txt 18356 against 14336 agents/echo/rendered/prompt.txt 22094 against 19042 ``` Worth recording why nobody can find these files in a diff: they are **generated from the lane policy roots at test time**, not checked in. `git show <commit>:agents/deep/rendered/prompt.txt` returns nothing at every commit that appears to touch them. So the growth arrives through skills changes and shows up in no reviewed diff, which is how four pushes in a row moved it without anyone noticing. ## The scope check, which is the reason I care **#1018 is failing only because it inherits this.** Its entire diff against main is three files - `site-work.md`, `site-work-procedure.md`, `stage-compose-sources.sh` - none of which feed a rendered prompt, and the two failing profiles are Deep and Echo, which that branch does not touch at all. It goes green when this lands, with no rebase and no change of its own. Same for anything else currently red on this repo: the failure is main's, not the branch's. ## One suggestion, not a blocker Raising to the exact measured size with no headroom is right, and it means the next policy sentence turns main red again rather than being caught in review. Given the rate here - four pushes since the last ratchet - it may be worth having the generator fail loudly at author time instead, so the growth lands with its cause in the same commit. Filing that thought rather than acting on it, five hours from air.