main is red: two commits landed without CI, breaking the capability-doc check and Deep's prompt budget #1053

New issue

Closed

opened 2026-08-19 04:44:57 +00:00 by coilyco-ops · 1 comment

coilyco-ops commented

2026-08-19 04:44:57 +00:00

Member

Both failures on #1018 reproduce on main at 446170e with none of that PR's changes, so #1018 inherits a red main rather than causing it.

--- FAIL: TestTheCapabilityDocsFollowTheHarnessBounds
    coilyco-general has no sentence matching At most \d+ tool rounds
    coilyco-general has no sentence matching A budget of \d+ model calls
--- FAIL: TestRenderedPromptsStayInsideTheirBudget/agents/deep/rendered/prompt.txt
    is 14336 bytes against a 12852 byte budget

Bisected against the merge before them, which is green:

434063f - green
2b4305e fix(skills): tell the service that refusal is an acceptable outcome - breaks the prompt budget
1f6e246 fix(skills): stop capability.md naming a round budget it does not own - breaks the capability-doc check

Both landed directly on main at 04:23Z and 04:24Z, between #1044's merge and #1047's, so neither passed CI before landing. Every red thing downstream of 04:24Z is these two.

Both source changes are correct and should stay. 1f6e246 removed the figures because coilyco-general loads on the Dowel lane and on the domain-neutral HTTP profile, whose deployments set different ceilings, and Dowel was observed live stating 6 rounds and 9 model calls against a deployed 12 and 2 (#978). Restoring the sentences to satisfy the test would restore a confirmed live misstatement. 2b4305e added the bounded-refusal doctrine, which is a deliberate prompt growth and wants a recorded raise.

Fix in #1052: the test moves rather than the docs. A copy that names the bounds still has to name the table's, and a copy that says its ceilings are deployment-set is failed for naming a figure anywhere, because the loophole in naming none is naming one somewhere else. Deep's budget ratchets 12852 to 14336.

Worth separating from the fix: two commits reached main without CI. That is the reusable finding here, and it is a merge-lane question rather than a code one.

Both failures on #1018 reproduce on `main` at 446170e with none of that PR's changes, so #1018 inherits a red main rather than causing it. ``` --- FAIL: TestTheCapabilityDocsFollowTheHarnessBounds coilyco-general has no sentence matching At most \d+ tool rounds coilyco-general has no sentence matching A budget of \d+ model calls --- FAIL: TestRenderedPromptsStayInsideTheirBudget/agents/deep/rendered/prompt.txt is 14336 bytes against a 12852 byte budget ``` Bisected against the merge before them, which is green: * `434063f` - green * `2b4305e` *fix(skills): tell the service that refusal is an acceptable outcome* - breaks the prompt budget * `1f6e246` *fix(skills): stop capability.md naming a round budget it does not own* - breaks the capability-doc check Both landed **directly on main** at 04:23Z and 04:24Z, between #1044's merge and #1047's, so neither passed CI before landing. Every red thing downstream of 04:24Z is these two. **Both source changes are correct and should stay.** `1f6e246` removed the figures because `coilyco-general` loads on the Dowel lane and on the domain-neutral HTTP profile, whose deployments set different ceilings, and Dowel was observed live stating 6 rounds and 9 model calls against a deployed 12 and 2 (#978). Restoring the sentences to satisfy the test would restore a confirmed live misstatement. `2b4305e` added the bounded-refusal doctrine, which is a deliberate prompt growth and wants a recorded raise. Fix in #1052: the test moves rather than the docs. A copy that names the bounds still has to name the table's, and a copy that says its ceilings are deployment-set is failed for naming a figure anywhere, because the loophole in naming none is naming one somewhere else. Deep's budget ratchets 12852 to 14336. Worth separating from the fix: **two commits reached main without CI.** That is the reusable finding here, and it is a merge-lane question rather than a code one.

coilyco-ops referenced this issue from a pull request that will close it,

2026-08-19 04:45:12 +00:00

fix(test): unbreak main - a shared capability doc may name no bound #1052

coilysiren closed this issue

2026-08-19 04:46:40 +00:00

coilyco-ops referenced this issue

2026-08-19 04:49:36 +00:00

feat(dowel): compose the lane without boundaries it has no seat for #1018

coilyco-ops commented

2026-08-19 04:50:29 +00:00

Author

Member

Correction to the body above. I checked the commit statuses rather than inferring from the merge shape, and "neither passed CI before landing" understates one half and overstates the other.

2b4305e has zero statuses. No run at all, because it was pushed together with the commit on top of it and only the tip got one.
1f6e246 ran and went red: ci / test failing after 44s, ci / publish-observed failing after 5s, image-build green.

So CI was not bypassed. It fired on the tip, reported the failure correctly, and the red was left standing on main - through five subsequent merges, including mine. The instrument worked and nothing acted on it.

That moves the reusable finding. It is not "commits reach main without CI", it is a red main does not stop anything: the next PR's author sees their own build fail and has no way to tell an inherited failure from one they caused, which is the twenty minutes this cost. A push straight to main with no PR is still worth a separate look, since it is what left one of the two commits unmeasured, but it is the smaller half.

Correction to the body above. I checked the commit statuses rather than inferring from the merge shape, and "neither passed CI before landing" understates one half and overstates the other. * `2b4305e` has **zero statuses**. No run at all, because it was pushed together with the commit on top of it and only the tip got one. * `1f6e246` **ran and went red**: `ci / test` failing after 44s, `ci / publish-observed` failing after 5s, `image-build` green. So CI was not bypassed. It fired on the tip, reported the failure correctly, and the red was left standing on `main` - through five subsequent merges, including mine. The instrument worked and nothing acted on it. That moves the reusable finding. It is not "commits reach main without CI", it is **a red main does not stop anything**: the next PR's author sees their own build fail and has no way to tell an inherited failure from one they caused, which is the twenty minutes this cost. A push straight to main with no PR is still worth a separate look, since it is what left one of the two commits unmeasured, but it is the smaller half.

coilyco-ops referenced this issue from a pull request that will close it,

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

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

coilyco-ops referenced this issue from a commit

2026-08-19 05:12:48 +00:00

Merge main into feat/dowel-no-boundaries, and consolidate site-work's opening

coilyco-ops referenced this issue

2026-08-19 05:13:03 +00:00

feat(dowel): compose the lane without boundaries it has no seat for #1018

coilyco-ops referenced this issue from a pull request that will close it,

2026-08-19 06:18:40 +00:00

chore(compose): rebake ai and exec, whose personality roster moved upstream #1061