From cecd76ec88f4faad00461cc8feb0f8ee37d7880e Mon Sep 17 00:00:00 2001 From: garrytan Date: Tue, 3 Nov 2026 15:37:18 +0000 Subject: [PATCH] docs(changelog): name 41247097050 as a PR screen Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> --- changelog.d/gsta262-pty-fixtures.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/changelog.d/gsta262-pty-fixtures.md b/changelog.d/gsta262-pty-fixtures.md index 94fbe7fdb..a749aceaa 100644 --- a/changelog.d/gsta262-pty-fixtures.md +++ b/changelog.d/gsta262-pty-fixtures.md @@ -1,4 +1,4 @@ ### Fixed -- **The autoplan-guard-pty eval no longer fails when its test actor mistypes a hash.** The live guard test has a Haiku model act out an /autoplan phase boundary, and it had to copy each reviewer dispatch, including two 64-character hashes, into its Agent call. Every so often it got a hash wrong (in train gate 41247097050 it made up a tail for the Design dispatch's hash), and the guard rightly refused the dispatch. Now the actor sends a short declared token, `DISPATCH ceo` or `DISPATCH design`, and a loopback relay in the test driver puts that phase's exact dispatch in place of the token before Claude Code sees the tool call. The relay also puts the token back in the history the actor sees, because in the first panel's trial 1 the actor copied the substituted CEO dispatch for its Design call. The guard, the hash in the dispatch, the Haiku actor and every assertion are unchanged. The first preregistered panel stopped at that trial 1 failure. After the history fix, the second panel passed 3 of 3 (4 of 4 guarded allows in each trial, no permission cards). +- **The autoplan-guard-pty eval no longer fails when its test actor mistypes a hash.** The live guard test has a Haiku model act out an /autoplan phase boundary, and it had to copy each reviewer dispatch, including two 64-character hashes, into its Agent call. Every so often it got a hash wrong (in PR screen 41247097050 it made up a tail for the Design dispatch's hash), and the guard rightly refused the dispatch. Now the actor sends a short declared token, `DISPATCH ceo` or `DISPATCH design`, and a loopback relay in the test driver puts that phase's exact dispatch in place of the token before Claude Code sees the tool call. The relay also puts the token back in the history the actor sees, because in the first panel's trial 1 the actor copied the substituted CEO dispatch for its Design call. The guard, the hash in the dispatch, the Haiku actor and every assertion are unchanged. The first preregistered panel stopped at that trial 1 failure. After the history fix, the second panel passed 3 of 3 (4 of 4 guarded allows in each trial, no permission cards). - **The plan-design-with-ui-scope eval accepts both ways the skill writes the outside-voices menu.** The question format says option descriptions use literal ✅/❌ bullets, and models read that two ways: earlier captures kept the bullets in the descriptions, while train gate 41470212877 moved them into the question text and dropped them from the descriptions, so the test's fixture user refused the menu. The fixture now declares both spellings, and the picker accepts exactly the declared description, or that description with every bullet removed. The red question is saved as a fixture. Restoring only some bullets, flipping one, swapping or rewording a description still fail. A focused paid run passed.