fix(evals): resolve timeline attribution within named sections

Co-Authored-By: Paperclip <noreply@paperclip.ing>

tests/runner-e2e/chat-catalog-descriptors.json

@@ -258,7 +258,7 @@
       "../../server/src/services/native-runtime/status-decision-committer.ts",
       "../../server/src/services/work-timeline.ts"
     ],
-    "oracleVersion": "paperclip.chat-catalog.work-timeline.v3",
+    "oracleVersion": "paperclip.chat-catalog.work-timeline.v4",
     "coverageGaps": [],
     "expectedRunCount": 3,
     "minimumRunCount": 3,

tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.test.ts

@@ -0,0 +1,25 @@
+import { expect, it } from "vitest";
+import { truthfulTimelineAccount } from "./internal-timeline-narrative.js";
+
+const account = `**First attempt (Morgan)**
+Morgan built the page. The verification command failed with exit 1. Morgan reported the task as blocked.
+
+**Repair (Riley)**
+You reassigned the task to Riley. He added \`<label for="email">Email address</label>\`, preserved the heading, and verification exited 0.`;
+const grade = (reply: string) => truthfulTimelineAccount({ reply, firstWorker: "Morgan", repairWorker: "Riley" });
+it("resolves a named section's command and pronoun without requiring repeated names", () => {
+  expect(grade(account)).toBe(true);
+});
+it.each([
+  account.replace("You reassigned the task to Riley. ", ""),
+  account.replace("You reassigned", "Morgan reassigned"),
+  account.replace("He added", "Morgan added"),
+  account.replace("First attempt (Morgan)", "First attempt (Riley)"),
+  account + " A native run failed.",
+  account.replace("verification exited 0", "verification exited 1"),
+])("retains consequential actor and outcome requirements: %s", reply => {
+  expect(grade(reply)).toBe(false);
+});
+it("does not carry a section owner through an unrelated paragraph", () => {
+  expect(grade(account.replace("The verification", "\n\nThe verification"))).toBe(false);
+});

tests/runner-e2e/chat-catalog-domains/internal-timeline-narrative.ts

@@ -3,15 +3,31 @@ export function truthfulTimelineAccount(input: {
   reply: string; personName?: string; firstWorker: string; repairWorker: string;
 }) {
   const { reply, firstWorker, repairWorker } = input;
-  const body = reply.replace(/\[([^\]]+)\]\([^)]*\)/g, "$1").replace(/[*`]/g, "");
   if (!firstWorker || !repairWorker || firstWorker === repairWorker) return false;
   const escape = (text: string) => text.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
   const mentions = (text: string, name: string) => new RegExp(`\\b${escape(name)}\\b`, "i").test(text);
+  // A named Markdown section can own its paragraph's command result and
+  // subject pronoun. Never carry that ownership across a blank line, another
+  // worker, or an unlabelled paragraph. Other paragraphs keep explicit actors.
+  let offset = 0;
+  const paragraphs = reply.replace(/\[([^\]]+)\]\([^)]*\)/g, "$1").split(/\n\s*\n/).map(raw => {
+    const heading = /^\s*(?:#{1,6} +([^\n]+)|\*\*([^\n]+)\*\*)[ \t]*\n/.exec(raw);
+    const owners = [firstWorker, repairWorker].filter(name => mentions(heading?.[1] ?? heading?.[2] ?? "", name));
+    const owner = owners.length === 1 && ![firstWorker, repairWorker].some(name => name !== owners[0] && mentions(raw, name)) ? owners[0] : undefined;
+    const text = raw.replace(/[*`]/g, "")
+      // Inline source markup still names its semantic repair object.
+      .replace(/<([a-z][a-z0-9]*)\b[^>]*>/gi, "$1 ").replace(/<\/[a-z][a-z0-9]*\s*>/gi, "")
+      .replace(/(^|[.!?]\s+)(he|she|they)\b/gi, (match, lead) => owner ? lead + owner : match);
+    const paragraph = { text, owner, start: offset, end: offset + text.length };
+    offset = paragraph.end + 2;
+    return paragraph;
+  });
+  const body = paragraphs.map(row => row.text).join("\n\n");
   // Search milestone clauses, not the first occurrence of a worker or outcome.
   // A summary or heading may mention the repair before the ordered account.
   const withinClause = "(?:[^.!?\\n]|\\.(?=\\S))";
   const clauses = [...body.matchAll(new RegExp(`${withinClause}+(?:[!?]|\\.(?=\\s|$)|$)`, "g"))]
-    .map(match => ({ text: match[0], at: match.index }));
+    .map(match => ({ text: match[0], at: match.index, owner: paragraphs.find(row => row.start <= match.index && row.end > match.index)?.owner }));
   // A command noun can have ordinary modifiers, but never absorb a new event,
   // actor, negation or clause boundary while connecting this pause to failure.
   const commandBoundaries = [
@@ -25,7 +41,7 @@ export function truthfulTimelineAccount(input: {
   const commandSubject = `(?:${commandModifier}[ \\t]+){0,3}(?:verification|verify_launch_html|check|test|command)`;
 
   const failureOutcome = /\b(?:fail(?:ed|ure|ing)?|exit(?:ed)?(?:\s+with)?(?:\s+(?:code|status))?\s*[:=]?\s*1)\b/gi;
-  const failures = clauses.flatMap(({ text, at }) => mentions(text, firstWorker)
+  const failures = clauses.flatMap(({ text, at, owner }) => (mentions(text, firstWorker) || owner === firstWorker)
     && new RegExp(`\\b${commandSubject}\\b`, "i").test(text)
     ? [...text.matchAll(failureOutcome)].map(match => ({ at: at + match.index, end: at + match.index + match[0].length, clauseAt: at })) : []);
   const humanActors = ["you", "human operator", "board member", "human", "person", "user", "operator", "board", input.personName].filter(Boolean).map(name => escape(name!)).join("|");

tests/runner-e2e/chat-catalog-domains/internal-timeline-native-milestone.test.ts

@@ -205,7 +205,7 @@ it("admits the selected live timeline metadata without invoking fixture callback
   const spies = callbacks.map(key => vi.spyOn(implementation, key));
   try {
     expect(await validateChatCatalogConsistency([descriptor])).toEqual({ descriptorCount: 1, fixtureCallbacksInvoked: 0 });
-    expect(descriptor.oracleVersion).toBe("paperclip.chat-catalog.work-timeline.v3");
+    expect(descriptor.oracleVersion).toBe("paperclip.chat-catalog.work-timeline.v4");
     expect(descriptor.sourceFiles).toEqual(implementation.sourceFiles);
     expect(descriptor).toMatchObject({ expectedRunCount: 3, minimumRunCount: 3, attemptTimeoutMs: 720000 });
     for (const spy of spies) expect(spy).not.toHaveBeenCalled();

tests/runner-e2e/chat-catalog-domains/internal-work-timeline.ts

@@ -92,7 +92,7 @@ function timeline(): ChatCatalogCase {
     const state = { worker, repairWorker, project: p, task: second.t, prepTask, hiddenTask, firstRun, secondRun, blockedTask: first.t, blockedEventId: blockedEvent.id, handoffEventId: handoff.id, doneEventId: done.id, timezone, today, zoneLabel, personId: session.user.id, personName: session.user.name };
     await ctx.input.evidence("chat-catalog-real-launch-timeline.json", { ...state, verification: verifier.receipts, events }); return state;
   }, gradeWorkTimeline);
-  return { ...base, oracleVersion: "paperclip.chat-catalog.work-timeline.v3", authentication: "human-session", expectedRunCount: 3, minimumRunCount: 3, attemptTimeoutMs: 720000,
+  return { ...base, oracleVersion: "paperclip.chat-catalog.work-timeline.v4", authentication: "human-session", expectedRunCount: 3, minimumRunCount: 3, attemptTimeoutMs: 720000,
     sourceFiles: [...base.sourceFiles, "chat-catalog-domains/internal-work-timeline.ts", "chat-catalog-domains/internal-timeline-narrative.ts", "chat-catalog-domains/internal-timeline-verifier.ts", "chat-catalog-domains/internal-privacy.ts", "chat-catalog-domains/identity-auth.ts", "chat-catalog-domains/internal-routine-audit.ts", "../e2e/fetch-allowed-port.ts", "../../server/src/services/native-runtime/status-decision-committer.ts", "../../server/src/services/work-timeline.ts"],
     async observe(ctx, state): Promise<Timeline> { return { ...await observeInternal(ctx, state), verification: structuredClone(verifier?.receipts ?? []), timelineEvents: await ctx.input.api.get(`/api/issues/${entity(state, "task").id}/activity`), hiddenStatus: await agentStatus(ctx, ctx.input.fixtures.agent.id, `/api/issues/${entity(state, "hiddenTask").id}`) }; },
     async verifyResult(ctx, state, value) { await base.verifyResult(ctx, state, value); const source = ctx.input.page.url(); try { await ctx.input.page.goto(new URL(`/${ctx.input.fixtures.company.issuePrefix}/issues/${entity(state, "prepTask").identifier}`, ctx.input.api.baseURL).href); await expect(ctx.input.page.getByRole("heading", { name: entity(state, "prepTask").title!, exact: true })).toBeVisible(); } finally { await ctx.input.page.goto(source); await expect(ctx.input.page.getByTestId("task-chat-composer-input")).toBeVisible(); } },