From c7909c9be0518f32c3f8bd96912b6bd15fe5f16b Mon Sep 17 00:00:00 2001 From: Dotta Date: Fri, 6 Nov 2026 04:46:29 -0600 Subject: [PATCH] fix(evals): scope implicit transition guards to repair claims --- .../internal-timeline-narrative.test.ts | 6 ++++++ .../internal-timeline-narrative.ts | 12 ++++++++---- 2 files changed, 14 insertions(+), 4 deletions(-) diff --git a/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.test.ts b/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.test.ts index 2a50a0fd970..d7bf9dadbda 100644 --- a/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.test.ts +++ b/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.test.ts @@ -288,8 +288,14 @@ it("accepts a same-task named-worker transition without requiring an assignment expect(gradeImplicitTransition(implicitTransitionAccount)).toBe(true); expect(gradeImplicitTransition(implicitTransitionAccount.replace("## Signup implementation — RUN-3", "2. **Signup implementation — RUN-3**"))).toBe(true); }); +it("keeps automatic verification and another task's requirements as context", () => { + expect(gradeImplicitTransition(implicitTransitionAccount.replace("The verification command returned exit 1", "The verification command failed. The validator automatically reported exit 1"))).toBe(true); + expect(gradeImplicitTransition(implicitTransitionAccount.replace("RUN-3", "RUN-3 implements requirements from [RUN-2](/RUN/issues/RUN-2)"))).toBe(true); +}); it.each([ implicitTransitionAccount.replace("Riley added", "Riley repaired RUN-4 and added"), + implicitTransitionAccount.replace("the missing label and", "the missing label automatically and"), + implicitTransitionAccount.replace("The task paused as blocked.", "The task paused as blocked. The task recovered automatically."), implicitTransitionAccount.replace("- Riley", "\n## Separate work — RUN-4\n- Riley"), implicitTransitionAccount.replace("The task paused as blocked.\n- Riley added the missing label and re-ran verification: exit 0.", "Riley added the missing label and re-ran verification: exit 0. The task paused as blocked."), implicitTransitionAccount.replace("The task paused as blocked.", ""), diff --git a/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.ts b/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.ts index ddddb05e6a3..f4081425c88 100644 --- a/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.ts +++ b/tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.ts @@ -208,7 +208,8 @@ export function truthfulTimelineAccount(input: { commandLead.test(prefix.slice(0, success.index)) && successIsOwned(text, success.index) && causedByRepair.test(prefix.slice(success.index + success[0].length))); }).map(match => ({ at: at + match.index - + (new RegExp(`^${successorOwnership.source}`, "i").test(match[0]) ? match[0].indexOf(",") + 1 : 0) }))); + + (new RegExp(`^${successorOwnership.source}`, "i").test(match[0]) ? match[0].indexOf(",") + 1 : 0), + claim: text.slice(match.index), clause: text }))); const pauseCausedByFailure = new RegExp(`\\b(?:after|because)\\s+(?:(?:the|his|her|their|its|${escape(firstWorker)}['’]s)\\s+)?${commandTail}`, "i"); const failureAfterPause = new RegExp(`\\bafter\\s+(?:(?:the|his|her|their|its|a)\\s+)?(?:task|work|it|${escape(firstWorker)})(?:\\s+(?:was|got|had been|has been))?\\s*$`, "i"); const pauseBetween = (failure: typeof failures[number], end: number) => @@ -228,11 +229,14 @@ export function truthfulTimelineAccount(input: { const implicitTransition = !!taskReference && paragraphs.some(paragraph => { if (!taskReference.test(paragraph.text.split("\n")[0]!) || new RegExp(assignment, "i").test(paragraph.text) || new RegExp(successorOwnership.source, "i").test(paragraph.text) - || /\b(?:automatically|autonomously|self-assigned)\b|\b(?:another|different|other)\s+(?:task|issue|work)\b/i.test(paragraph.text) - || taskNamespace && [...paragraph.text.matchAll(taskNamespace)].some(match => match[0].toLowerCase() !== input.taskIdentifier!.toLowerCase())) return false; + || new RegExp(`${taskSubject}\\s+(?:was\\s+)?(?:(?:automatically|autonomously)\\s+(?:recovered|resumed)|(?:recovered|resumed)\\s+(?:automatically|autonomously))\\b`, "i").test(paragraph.text) + || taskNamespace && paragraph.text.split("\n")[0]!.match(taskNamespace)?.[0]?.toLowerCase() !== input.taskIdentifier!.toLowerCase()) return false; return failures.some(failure => failure.at >= paragraph.start && failure.at < paragraph.end && mentions(body.slice(paragraph.start, failure.at), firstWorker) - && repairs.some(repair => repair.at > failure.at && repair.at < paragraph.end && pauseBetween(failure, repair.at))); + && repairs.some(repair => repair.at > failure.at && repair.at < paragraph.end && pauseBetween(failure, repair.at) + && !/\b(?:automatically|autonomously|self-assigned)\b/i.test(repair.claim.split(/\b(?:and|then|while)\b/i)[0]!) + && !/\b(?:another|different|other)\s+(?:task|issue|work)\b/i.test(repair.clause) + && (!taskNamespace || [...repair.clause.matchAll(taskNamespace)].every(match => match[0].toLowerCase() === input.taskIdentifier!.toLowerCase())))); }); const ordered = explicitTransition || implicitTransition; const nativeFailure = /\b(?:native|runner|execution)\s+runs?\s+(?:(?:themselves|were|was|had|also|both|[a-z]+ly)\s+){0,3}(?:failed|failures?|unsuccessful)\b|\b(?:native|runner|execution)\s+runs?\s+(?:ended|finished)\s+(?:in|with)\s+(?:a\s+)?failure\b|\bfailed\s+(?:native|runner|execution)\s+runs?\b/i;