fix(cli): word review gap disclosures so they cannot read as contradictions

packages/cli/src/commands/review/compose-review.test.ts

@@ -193,6 +193,7 @@ function transcript(
     toolCalls?: number;
     text?: string;
     opens?: string[];
+    toolPath?: string;
     /** `[offset, limit]` making the diff reads ranged, as a compliant agent's are. */
     range?: [number, number];
   } = {},
@@ -227,7 +228,7 @@ function transcript(
                       offset: opts.range[0],
                       limit: opts.range[1],
                     }
-                  : { file_path: DIFF },
+                  : { file_path: opts.toolPath ?? DIFF },
               },
             },
           ],
@@ -920,7 +921,12 @@ describe('composeReview — event caps (round-7 Critical #2: caps must reach eve
       base({ suggestionsInline: 1, unreviewedDimensions: ['security'] }),
     );
     expect(r.event).toBe('COMMENT');
-    expect(r.body).toContain('Reviewed. Suggestions are inline.');
+    // The gap disclosure follows, so the opener says the review is partial —
+    // any "Reviewed…" opener above "Not reviewed:" read as the body
+    // contradicting itself (#8811).
+    expect(r.body).toContain(
+      'Partially reviewed — gaps disclosed. Suggestions are inline.',
+    );
     expect(r.body).not.toContain('no blockers');
   });
 });
@@ -1089,7 +1095,7 @@ describe('composeReview — stacked states compose, none erased', () => {
     // downgradeApprove did not fire (base event was COMMENT), so no sentence…
     expect(r.body).not.toContain('Downgraded');
     // …but every disclosure is present exactly once, and nothing certifies.
-    expect(r.body).toContain('Reviewed.');
+    expect(r.body).toContain('Partially reviewed — gaps disclosed.');
     expect(r.body).toContain('Suggestions are inline.');
     expect(r.body).toContain('1 Suggestion-level finding(s)');
     expect(r.body).toContain('Unresolved, please confirm:');
@@ -1300,6 +1306,8 @@ describe('composeReview — budget-gap disclosures (a channel, never a cap)', ()
     });
     expect(r.body).toContain('Not explored to full depth');
     expect(r.body).not.toContain('no blockers');
+    expect(r.body).toContain('Reviewed.');
+    expect(r.body).not.toContain('Partially reviewed');
   });
 });
 
@@ -2089,16 +2097,18 @@ describe('coverage is recomputed, never accepted', () => {
     expect(r.body).not.toContain('Reviewed.');
   });
 
-  it('keeps the "Reviewed." opener while any chunk is certified', () => {
+  it('opens partial, not zero-certified, while any chunk is certified — and names the gaps it carries', () => {
     // chunk 1 built and never launched; chunk 2 reviewed properly. A partial
-    // gap is a disclosure, not a zero-certification.
+    // gap is a disclosure, not a zero-certification — and the opener says
+    // the review is partial, so no "Reviewed…" opener ever sits beside
+    // "Not reviewed:" (#8811).
     const p = plan();
     recordBuilt(p, 1);
     recordBuilt(p, 2);
     recordMatrix(p);
     transcript('a2', goodPrompt(2), { toolCalls: 2 });
     const r = composeReview({ planPath: p, env: ENV, modelId: MODEL });
-    expect(r.body).toContain('Reviewed.');
+    expect(r.body).toContain('Partially reviewed — gaps disclosed.');
     expect(r.body).not.toContain('could not certify');
   });
 
@@ -2157,6 +2167,84 @@ describe('coverage is recomputed, never accepted', () => {
     );
   });
 
+  it('quotes a prose agent label — it is the agent’s name, not a claim about the PR', () => {
+    // #8811: a whole-diff agent (no `chunk N of M` in its prompt) was
+    // disclosed by the truncated first line of its launch prompt, rendered
+    // bare — "Not reviewed: This PR narrows the daemon-marker check from a
+    // truthy tes..." read as a sentence about the whole PR, not the name of
+    // the one agent that failed. Quotes say which it is, and the truncation
+    // stops at a word boundary instead of mid-word.
+    const p = plan({ han: true });
+    transcript('a2', goodPrompt(2), { toolCalls: 2 });
+    recordBuilt(p, 2);
+    recordMatrix(p);
+    const brief = briefPath(p, 'chunk-1');
+    writeFileSync(brief, 'The chunk-1 brief.');
+    const launch =
+      'This PR narrows the daemon-marker check from a truthy test to an exact one\n' +
+      `read_file(file_path="${brief}")\n` +
+      `read_file(file_path="${DIFF}", offset=0, limit=100)`;
+    writeFileSync(join(promptRecordDir(p), 'chunk-1.txt'), launch);
+    transcript('p1', launch, {
+      toolCalls: 1,
+      toolPath: join(dir, 'other.ts'),
+      opens: [brief],
+    });
+    const r = composeReview({
+      criticalsInline: 0,
+      suggestionsInline: 0,
+      planPath: p,
+      env: ENV,
+      modelId: MODEL,
+    });
+    expect(r.body).toContain(
+      'Not reviewed: `"This PR narrows the daemon-marker check from a truthy test…"`',
+    );
+    expect(r.body).not.toContain('truthy tes...');
+    expect(r.body).toContain(
+      '启动 prompt 为它指定了 diff 中的行，但它从未打开',
+    );
+  });
+
+  it('keeps long agent labels distinct when their first word matches', () => {
+    transcript(
+      'p1',
+      `Verify the daemon marker rename does not break macos-build behavior\n${DIFF}`,
+    );
+    transcript(
+      'p2',
+      `Verify the daemon marker rename does not break linux-build behavior\n${DIFF}`,
+    );
+    const r = composeReview({ planPath: plan(), env: ENV, modelId: MODEL });
+
+    expect(r.body).toContain('macos-build…');
+    expect(r.body).toContain('linux-build…');
+  });
+
+  it('renders prompt-derived labels as inert Markdown', () => {
+    transcript(
+      'p1',
+      `Fix the "daemon marker" regression for @owner from #123\n${DIFF}`,
+    );
+    const r = composeReview({ planPath: plan(), env: ENV, modelId: MODEL });
+
+    expect(r.body).toContain(
+      'Not reviewed: `"Fix the \\"daemon marker\\" regression for @owner from #123"`',
+    );
+  });
+
+  it('collapses spaces after removing backticks from agent labels', () => {
+    transcript(
+      'p1',
+      `You are review agent \`security\` — inspect auth\n${DIFF}`,
+    );
+    const r = composeReview({ planPath: plan(), env: ENV, modelId: MODEL });
+
+    expect(r.body).toContain(
+      'Not reviewed: `"You are review agent security — inspect auth"`',
+    );
+  });
+
   it('names a blind launch as itself, not as a whiff', () => {
     // An agent whose prompt never named the diff could not have read it, and
     // relaunching it produces another agent that cannot either. The prompt is the
@@ -2934,6 +3022,9 @@ describe('bilingual body — the PR author writes Chinese (prDescriptionHasHan)'
     expect(r.body).toContain('未审查：全 diff 测试覆盖检查——');
     // The zh sentence carries the translated reason, not the English one.
     expect(r.body).toContain('没有记录表明它的 brief 到达过任何 agent');
+    // The partial opener, in both halves (#8811).
+    expect(r.body).toContain('Partially reviewed — gaps disclosed.');
+    expect(r.body).toContain('仅完成部分审查，审查缺口已披露。');
   });
 
   it('quotes untranslatable caller text as-is in both halves', () => {
@@ -3983,6 +4074,9 @@ describe('composeReview — the findings file tag check', () => {
       '1 finding(s) still carried the `— [unverified]` tag when the loop ' +
         'ended',
     );
+    expect(r.body).toContain(
+      'Review incomplete — unverified findings disclosed.',
+    );
     // The opener may not certify over a loop that ended mid-verification.
     expect(r.body).not.toContain('no blockers');
     expect(r.remediation.join(' ')).toContain('--role verify');
@@ -4052,6 +4146,8 @@ describe('composeReview — the findings file tag check', () => {
     expect(r.event).toBe('COMMENT');
     expect(r.cappedBy).toContain('findings-unverified-at-compose');
     expect(r.body).toContain('findings file could not be read at compose time');
+    expect(r.body).toContain('Review incomplete — findings unavailable.');
+    expect(r.body).not.toContain('unverified findings disclosed');
     expect(r.remediation.join(' ')).toContain('findingsPath');
   });
 

packages/cli/src/commands/review/compose-review.ts

@@ -30,6 +30,7 @@ import {
   verificationGaps,
   TranscriptsUnavailableError,
 } from './lib/coverage.js';
+import { compressSummary } from './findings.js';
 import {
   BUDGET_STOP_PHRASE,
   ROUND_CAP_PHRASE,
@@ -584,6 +585,7 @@ function composeReviewBody(
       for (const label of cov.idleAgents) {
         coverageEntries.push({
           subject: label,
+          publicSubject: publicAgentSubject(label),
           reason: 'the agent made no tool call: it read nothing',
           reasonZh: '该 agent 未发起任何工具调用：它什么都没读',
         });
@@ -605,6 +607,7 @@ function composeReviewBody(
       for (const label of cov.blindAgents) {
         coverageEntries.push({
           subject: label,
+          publicSubject: publicAgentSubject(label),
           reason:
             'launched with a prompt that never named the diff file, so it ' +
             'could not have read it',
@@ -626,11 +629,13 @@ function composeReviewBody(
       for (const label of cov.unopenedAgents) {
         coverageEntries.push({
           subject: label,
+          publicSubject: publicAgentSubject(label),
           reason:
             'pointed at diff lines it never opened: it made tool calls, but ' +
             'none of them read the diff',
           reasonZh:
-            '它被指向 diff 的行却从未打开：有工具调用，但没有一次读取 diff',
+            '启动 prompt 为它指定了 diff 中的行，但它从未打开：有工具调用，' +
+            '却没有一次读取 diff',
         });
       }
       if (cov.unopenedAgents.length > 0) {
@@ -1117,11 +1122,19 @@ function composeReviewBody(
     const shown = keptBudgetGaps.slice(0, MAX_BUDGET_GAP_LINES);
     const more = keptBudgetGaps.length - shown.length;
     const enList =
-      shown.map((it) => `${it.agent}: ${mdField(it.gap)}`).join('; ') +
-      (more > 0 ? `, and ${more} more` : '');
+      shown
+        .map(
+          (it) =>
+            `${publicAgentSubject(it.agent) ?? it.agent}: ${mdField(it.gap)}`,
+        )
+        .join('; ') + (more > 0 ? `, and ${more} more` : '');
     const zhList =
-      shown.map((it) => `${it.agent}：${mdField(it.gap)}`).join('；') +
-      (more > 0 ? `，另有 ${more} 条` : '');
+      shown
+        .map(
+          (it) =>
+            `${publicAgentSubject(it.agent) ?? it.agent}：${mdField(it.gap)}`,
+        )
+        .join('；') + (more > 0 ? `，另有 ${more} 条` : '');
     notReviewedParts.push({
       en: `Not explored to full depth (tool budget reached): ${enList}.`,
       zh: `未探索到全部深度（达到工具调用预算）：${zhList}。`,
@@ -1405,6 +1418,16 @@ function composeReviewBody(
       coverageEntries.some((e) => e.subject === 'coverage') ||
       (plannedChunks.length > 0 &&
         coveredChunks.every((id) => disclosedChunkIds.has(id)));
+    const hasCoverageGaps =
+      unreviewed.length + coverageEntries.length > 0 ||
+      missingReceipts.length > 0 ||
+      uncoverable.length > 0;
+    // Any opener starting with "Reviewed" reads as contradicting the
+    // "Not reviewed:" clauses below it — announcing the gaps does not fix
+    // it, as the first cut of this wording showed (#8811). When disclosures
+    // follow, the opener says the review is PARTIAL instead, so the pair
+    // reads in one direction; the certifying and the zero-certified openers
+    // above keep their exact wording.
     clauses.push(
       nothingCertified
         ? {
@@ -1413,7 +1436,22 @@ function composeReviewBody(
           }
         : canCertify
           ? { en: 'Reviewed — no blockers.', zh: '已审查——无阻断问题。' }
-          : { en: 'Reviewed.', zh: '已审查。' },
+          : hasCoverageGaps
+            ? {
+                en: 'Partially reviewed — gaps disclosed.',
+                zh: '仅完成部分审查，审查缺口已披露。',
+              }
+            : findingsFileUnreadable
+              ? {
+                  en: 'Review incomplete — findings unavailable.',
+                  zh: '审查未完成——发现不可用。',
+                }
+              : findingsUnverifiedAtCompose
+                ? {
+                    en: 'Review incomplete — unverified findings disclosed.',
+                    zh: '审查未完成——未验证的发现已披露。',
+                  }
+                : { en: 'Reviewed.', zh: '已审查。' },
     );
   }
 
@@ -1483,6 +1521,22 @@ function composeReviewBody(
   };
 }
 
+/**
+ * The public subject for an agent-derived disclosure label. A `chunk N`
+ * label stays bare — the chunk collapse translates it into the author's
+ * units. Any other label is the truncated first line of a launch prompt:
+ * prose. Rendered bare it reads as a claim about the PR itself —
+ * #8811's posted body carried "Not reviewed: This PR narrows the
+ * daemon-marker check from a truthy tes..." — not as the name of the agent
+ * that failed. Quoted, it reads as a name. The INTERNAL subject stays the
+ * unquoted label: the dedup and certification checks key on it.
+ */
+function publicAgentSubject(label: string): string | undefined {
+  return /^chunk \d+$/.test(label)
+    ? undefined
+    : mdField(JSON.stringify(compressSummary(label.replace(/[`\r\n]+/g, ' '))));
+}
+
 /**
  * A set of unreviewed chunk ids, said in the PR author's units.
  *

packages/cli/src/commands/review/lib/coverage.ts

@@ -359,7 +359,7 @@ function publicRoleLabelZh(req: RequiredAgent): string | undefined {
 function label(rec: AgentRecord, chunk: number | null): string {
   if (chunk !== null) return `chunk ${chunk}`;
   const first = rec.launchPrompt.split('\n')[0]?.trim() ?? '';
-  if (first) return first.length > 60 ? `${first.slice(0, 57)}...` : first;
+  if (first) return first.replace(/\s+/g, ' ');
   return rec.agentName || rec.agentId;
 }