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

coilyco-ops commented

2026-08-19 04:54:47 +00:00

Member

main is red, and #1018 is only failing because it inherits that. Sent as a fix to main rather than to that branch.

The failure

Run 23290 on 6a1b7b6:

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

I reproduced both numbers exactly on a clean worktree at origin/main, so this is deterministic and not a runner artifact.

Cause

The rendered prompts are generated from the lane policy roots at test time rather than checked in - the paths are gitignored, and git show <commit>:agents/deep/rendered/prompt.txt returns nothing at every commit that appears to touch them.

So the run of skills commits since the last ratchet grew them: the refusal-is-an-acceptable-outcome rule, the restored member-location rule, the harness-gate description, and the workspace/filesystem split. a3d0ac7 moved Deep to 14336 and passed on run 23284; everything after it pushed past both bounds without moving either.

Why raise rather than trim

The table says what it is:

promptBudgets ratchet the tracked snapshots, and are not targets. Every raise is recorded in the commit that makes it, with its cause.

Every one of those additions was a deliberate policy change with its own reasoning. Trimming them back is the other honest answer and it is not one to make five hours before a livestream on these two lanes.

Raised to exactly the measured sizes, with no headroom, so the next growth is a decision again rather than a surprise.

Scope check on #1018

That branch's whole diff against main is three files - site-work.md, site-work-procedure.md, and stage-compose-sources.sh. None of them feed a rendered prompt, and the two failing profiles are Deep and Echo, which #1018 does not touch at all. It goes green on this landing without a rebase.

go test ./... and pre-commit green.

**main is red**, and #1018 is only failing because it inherits that. Sent as a fix to main rather than to that branch. ## The failure Run 23290 on `6a1b7b6`: ``` agents/deep/rendered/prompt.txt 18356 against a 14336 byte budget agents/echo/rendered/prompt.txt 22094 against a 19042 byte budget ``` I reproduced both numbers exactly on a clean worktree at `origin/main`, so this is deterministic and not a runner artifact. ## Cause The rendered prompts are **generated from the lane policy roots at test time** rather than checked in - the paths are gitignored, and `git show <commit>:agents/deep/rendered/prompt.txt` returns nothing at every commit that appears to touch them. So the run of skills commits since the last ratchet grew them: the refusal-is-an-acceptable-outcome rule, the restored member-location rule, the harness-gate description, and the workspace/filesystem split. `a3d0ac7` moved Deep to 14336 and passed on run 23284; everything after it pushed past both bounds without moving either. ## Why raise rather than trim The table says what it is: > promptBudgets ratchet the tracked snapshots, and are not targets. Every raise is recorded in the commit that makes it, with its cause. Every one of those additions was a deliberate policy change with its own reasoning. Trimming them back is the other honest answer and it is **not one to make five hours before a livestream on these two lanes**. Raised to exactly the measured sizes, with no headroom, so the next growth is a decision again rather than a surprise. ## Scope check on #1018 That branch's whole diff against main is three files - `site-work.md`, `site-work-procedure.md`, and `stage-compose-sources.sh`. None of them feed a rendered prompt, and the two failing profiles are Deep and Echo, which #1018 does not touch at all. It goes green on this landing without a rebase. `go test ./...` and `pre-commit` green.

coilyco-ops added 1 commit

2026-08-19 04:54:47 +00:00

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

ci / image-build (pull_request) Successful in 4s

Details

ci / test (pull_request) Successful in 56s

Details

ci / publish-echo-image (pull_request) Has been skipped

Details

ci / publish-observed (pull_request) Has been skipped

Details

02b1905160

main is red. Run 23290 on 6a1b7b6 fails TestRenderedPromptsStayInsideTheirBudget:

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

The rendered prompts are generated from the lane policy roots rather than
checked in, so the run of skills commits that added the refusal rules, the
member-location rule and the harness-gate description grew them past a
ratchet nobody moved. The sizes are deterministic: these are the same two
numbers CI measured.

Raised to exactly the measured sizes, per the table's own rule that a
budget is a ratchet rather than a target and every raise names its cause.
No headroom added, so the next growth is a decision again rather than a
surprise.

Trimming the policy instead would be the other honest answer, and it is
not one to make five hours before a livestream on these lanes.

Co-authored-by: Kai Siren <coilysiren@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>

coilyco-ops requested review from coilysiren

2026-08-19 04:54:47 +00:00

coilyco-ops commented

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

Author

Member

Duplicate of #1055, which is byte-identical - the same two numbers, 22094 and 18356 - and was filed at 04:52Z, before this. Closing in favour of it. Moved the scope check that this one carried onto #1055, since it is the part that is not duplicated.

coilyco-ops closed this pull request

2026-08-19 04:55:29 +00:00

coilyco-ops referenced this pull request

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

fix(test): ratchet both prompt budgets to what main now renders #1055