fix: make paired evidence loss visible (#202)

2026-05-31 16:20:09 +09:00
parent 6eca648c47
commit 239c7ff1e6
19 changed files with 644 additions and 67 deletions
--- a/src/db-paired-turn-outputs.test.ts
+++ b/src/db-paired-turn-outputs.test.ts
@@ -8,6 +8,33 @@ import {
 } from './db/paired-turn-outputs.js';

 describe('paired turn output attachments', () => {
+  it('marks truncated output text so reviewers know evidence was clipped', () => {
+    const database = new Database(':memory:');
+    try {
+      applyBaseSchema(database);
+
+      insertPairedTurnOutputInDatabase(
+        database,
+        'paired-task-turn-output-truncated',
+        1,
+        'owner',
+        'x'.repeat(60_000),
+      );
+
+      const [output] = getPairedTurnOutputsFromDatabase(
+        database,
+        'paired-task-turn-output-truncated',
+      );
+
+      expect(output.output_text).toHaveLength(50_000);
+      expect(output.output_text).toMatch(
+        /\[Output truncated: 60000 > 50000 chars\]/,
+      );
+    } finally {
+      database.close();
+    }
+  });
+
  it('preserves attachments for reviewer evidence prompts', () => {
    const database = new Database(':memory:');
    try {
--- a/src/db/paired-turn-outputs.ts
+++ b/src/db/paired-turn-outputs.ts
@@ -14,6 +14,15 @@ import {

 const MAX_TURN_OUTPUT_CHARS = 50_000;

+function storedOutputText(outputText: string): string {
+  if (outputText.length <= MAX_TURN_OUTPUT_CHARS) {
+    return outputText;
+  }
+
+  const notice = `\n\n[Output truncated: ${outputText.length} > ${MAX_TURN_OUTPUT_CHARS} chars]`;
+  return `${outputText.slice(0, MAX_TURN_OUTPUT_CHARS - notice.length)}${notice}`;
+}
+
 export function insertPairedTurnOutputInDatabase(
  database: Database,
  taskId: string,
@@ -48,7 +57,7 @@ export function insertPairedTurnOutputInDatabase(
      taskId,
      turnNumber,
      role,
-      outputText.slice(0, MAX_TURN_OUTPUT_CHARS),
+      storedOutputText(outputText),
      serializeAttachmentPayload(options.attachments),
      parseVisibleVerdict(outputText),
      options.createdAt ?? new Date().toISOString(),
--- a/src/message-agent-executor-paired.test.ts
+++ b/src/message-agent-executor-paired.test.ts
@@ -33,11 +33,11 @@ const log = {
  warn: vi.fn(),
 };

-describe('createPairedExecutionLifecycle', () => {
-  beforeEach(() => {
-    vi.clearAllMocks();
-  });
+beforeEach(() => {
+  vi.clearAllMocks();
+});

+describe('createPairedExecutionLifecycle output persistence', () => {
  it('stores final output attachments with the paired turn output', () => {
    const lifecycle = createPairedExecutionLifecycle({
      pairedExecutionContext: {
@@ -101,7 +101,67 @@ describe('createPairedExecutionLifecycle', () => {
      },
    );
  });
+});

+describe('createPairedExecutionLifecycle verdict routing', () => {
+  it('uses the full final output for paired verdict routing', () => {
+    const lifecycle = createPairedExecutionLifecycle({
+      pairedExecutionContext: {
+        task: {
+          id: 'paired-task-long-final',
+          chat_jid: 'group@test',
+          group_folder: 'test-group',
+          owner_service_id: 'codex-main',
+          reviewer_service_id: 'claude',
+          title: null,
+          source_ref: 'HEAD',
+          plan_notes: null,
+          round_trip_count: 0,
+          review_requested_at: null,
+          status: 'active',
+          arbiter_verdict: null,
+          arbiter_requested_at: null,
+          completion_reason: null,
+          created_at: '2026-04-09T00:00:00.000Z',
+          updated_at: '2026-04-09T00:00:00.000Z',
+        },
+        workspace: null,
+        envOverrides: {},
+      },
+      pairedTurnIdentity: {
+        turnId: 'paired-task-long-final:2026-04-09T00:00:00.000Z:owner-turn',
+        taskId: 'paired-task-long-final',
+        taskUpdatedAt: '2026-04-09T00:00:00.000Z',
+        intentKind: 'owner-turn',
+        role: 'owner',
+      },
+      completedRole: 'owner',
+      chatJid: 'group@test',
+      runId: 'run-long-final',
+      enqueueMessageCheck: vi.fn(),
+      log,
+    });
+    const longPreface = '검증 증거 '.repeat(100);
+    const finalOutput = `${longPreface}\nTASK_DONE\n뒤쪽 상태줄도 라우팅에 반영되어야 합니다.`;
+
+    lifecycle.recordFinalOutputBeforeDelivery(finalOutput);
+
+    expect(
+      pairedExecutionContextModule.completePairedExecutionContext,
+    ).toHaveBeenCalledWith(
+      expect.objectContaining({
+        taskId: 'paired-task-long-final',
+        role: 'owner',
+        status: 'succeeded',
+        runId: 'run-long-final',
+        summary: finalOutput,
+      }),
+    );
+    expect(finalOutput.indexOf('TASK_DONE')).toBeGreaterThan(500);
+  });
+});
+
+describe('createPairedExecutionLifecycle completion handling', () => {
  it('does not emit a second public notification after arbiter ESCALATE', async () => {
    const outputs: AgentOutput[] = [];

--- a/src/message-agent-executor-paired.ts
+++ b/src/message-agent-executor-paired.ts
@@ -422,7 +422,7 @@ class PairedExecutionLifecycleController implements PairedExecutionLifecycle {
      this.pairedFinalAttachments = attachments;
    }
    if (!this.pairedSummaryLocked) {
-      this.pairedExecutionSummary = outputText.slice(0, 500);
+      this.pairedExecutionSummary = outputText;
      this.pairedSummaryLocked = true;
    }
    this.pairedSawOutput = true;
@@ -451,7 +451,7 @@ class PairedExecutionLifecycleController implements PairedExecutionLifecycle {
        'Adopted direct terminal delivery as paired final output',
      );
    } else if (!this.pairedSummaryLocked) {
-      this.pairedExecutionSummary = this.pairedFinalOutput.slice(0, 500);
+      this.pairedExecutionSummary = this.pairedFinalOutput;
      this.pairedSummaryLocked = true;
    }
    return outputText;
--- a/src/message-runtime-prompts.test.ts
+++ b/src/message-runtime-prompts.test.ts
@@ -323,6 +323,9 @@ describe('message-runtime-prompts arbiter output context', () => {

    expect(prompt).not.toContain('arbiter-context-output-01');
    expect(prompt).not.toContain('arbiter-context-output-02');
+    expect(prompt).toContain(
+      '[Earlier paired turn outputs omitted: 2 older outputs not shown]',
+    );
    expect(prompt).toContain('arbiter-context-output-03');
    expect(prompt).toContain('arbiter-context-output-08');
    expect(prompt).not.toContain('예전 유저 지시');
@@ -362,6 +365,9 @@ describe('message-runtime-prompts arbiter output context', () => {
    expect(prompt).not.toContain('이전 작업의 유저 지시');
    expect(prompt).not.toContain('arbiter-context-output-01');
    expect(prompt).not.toContain('arbiter-context-output-02');
+    expect(prompt).toContain(
+      '[Earlier paired turn outputs omitted: 2 older outputs not shown]',
+    );
    expect(prompt).toContain('arbiter-context-output-03');
    expect(prompt).toContain('arbiter-context-output-08');
  });
--- a/src/message-runtime-prompts.ts
+++ b/src/message-runtime-prompts.ts
@@ -60,6 +60,45 @@ function turnOutputsToMessages(
  }));
 }

+function omittedTurnOutputsMessage(args: {
+  chatJid: string;
+  omittedCount: number;
+  firstVisibleOutput?: PairedTurnOutput;
+}): NewMessage {
+  return {
+    id: `turn-output-omitted-${args.firstVisibleOutput?.task_id ?? 'unknown'}-${args.omittedCount}`,
+    chat_jid: args.chatJid,
+    sender: 'system',
+    sender_name: 'system',
+    content: `[Earlier paired turn outputs omitted: ${args.omittedCount} older outputs not shown]`,
+    timestamp:
+      args.firstVisibleOutput?.created_at ?? '1970-01-01T00:00:00.000Z',
+    is_bot_message: true as const,
+    is_from_me: false as const,
+  };
+}
+
+function latestTurnOutputMessagesWithOmissionMarker(
+  outputs: PairedTurnOutput[],
+  chatJid: string,
+  limit: number,
+): NewMessage[] {
+  const latestOutputs = latestTurnOutputs(outputs, limit);
+  const messages = turnOutputsToMessages(latestOutputs, chatJid);
+  const omittedCount = outputs.length - latestOutputs.length;
+  if (omittedCount <= 0) {
+    return messages;
+  }
+  return [
+    omittedTurnOutputsMessage({
+      chatJid,
+      omittedCount,
+      firstVisibleOutput: latestOutputs[0],
+    }),
+    ...messages,
+  ];
+}
+
 function mergeHumanAndTurnOutputMessages(
  chatJid: string,
  humanMessages: NewMessage[],
@@ -221,12 +260,10 @@ export function buildArbiterPromptForTask(args: {
    args.turnOutputs.length > 0
      ? [
          ...taskHumanMessages,
-          ...turnOutputsToMessages(
-            latestTurnOutputs(
-              args.turnOutputs,
-              ARBITER_TURN_OUTPUT_CONTEXT_LIMIT,
-            ),
+          ...latestTurnOutputMessagesWithOmissionMarker(
+            args.turnOutputs,
            args.chatJid,
+            ARBITER_TURN_OUTPUT_CONTEXT_LIMIT,
          ),
        ]
      : args.labeledRecentMessages;
--- a/src/paired-execution-context-arbiter.ts
+++ b/src/paired-execution-context-arbiter.ts
@@ -113,22 +113,17 @@ export function handleArbiterCompletion(args: {
      transitionPairedTaskStatus({
        taskId,
        currentStatus: 'in_arbitration',
-        nextStatus: 'review_ready',
+        nextStatus: 'completed',
        expectedUpdatedAt: task.updated_at,
        updatedAt: now,
        patch: {
-          round_trip_count: ARBITER_RESOLUTION_ROUND_TRIP_COUNT,
-          owner_failure_count: 0,
-          owner_step_done_streak: 0,
-          finalize_step_done_count: 0,
-          empty_step_done_streak: 0,
          arbiter_verdict: 'unknown',
-          arbiter_requested_at: null,
+          completion_reason: 'arbiter_escalated',
        },
      });
      logger.warn(
        { taskId, summary: summary?.slice(0, 200) },
-        'Arbiter verdict unrecognized — returning task to reviewer as proceed fallback',
+        'Arbiter verdict unrecognized — escalating instead of treating it as approval',
      );
      return;
  }
--- a/src/paired-execution-context-carry-forward.test.ts
+++ b/src/paired-execution-context-carry-forward.test.ts
@@ -0,0 +1,164 @@
+import { describe, expect, it, vi } from 'vitest';
+
+vi.mock('./config.js', async () => {
+  const actual =
+    await vi.importActual<typeof import('./config.js')>('./config.js');
+  return {
+    ...actual,
+    PAIRED_CARRY_FORWARD_LATEST_OWNER_FINAL: true,
+  };
+});
+
+vi.mock('./db.js', () => {
+  const updatePairedTask = vi.fn();
+  return {
+    cancelPairedTurn: vi.fn(),
+    createPairedTask: vi.fn(),
+    getLatestPairedTaskForChat: vi.fn(),
+    getLatestOpenPairedTaskForChat: vi.fn(),
+    getPairedTaskById: vi.fn(),
+    getPairedTurnById: vi.fn(),
+    getPairedTurnOutputs: vi.fn(() => []),
+    getPairedWorkspace: vi.fn(),
+    insertPairedTurnOutput: vi.fn(),
+    updatePairedTask,
+    updatePairedTaskIfUnchanged: vi.fn((id, _expectedUpdatedAt, updates) => {
+      updatePairedTask(id, updates);
+      return true;
+    }),
+    upsertPairedProject: vi.fn(),
+    hasActiveCiWatcherForChat: vi.fn(() => false),
+    releasePairedTaskExecutionLease: vi.fn(),
+  };
+});
+
+vi.mock('./paired-workspace-manager.js', () => ({
+  isOwnerWorkspaceRepairNeededError: vi.fn(() => false),
+  markPairedTaskReviewReady: vi.fn(),
+  prepareReviewerWorkspaceForExecution: vi.fn(() => ({
+    workspace: null,
+    autoRefreshed: false,
+  })),
+  provisionOwnerWorkspaceForPairedTask: vi.fn(() => ({
+    id: 'task-new:owner',
+    task_id: 'task-new',
+    role: 'owner',
+    workspace_dir: '/tmp/paired/task-new/owner',
+    snapshot_source_dir: null,
+    snapshot_ref: null,
+    status: 'ready',
+    snapshot_refreshed_at: null,
+    created_at: '2026-03-28T00:00:00.000Z',
+    updated_at: '2026-03-28T00:00:00.000Z',
+  })),
+}));
+
+vi.mock('./logger.js', () => ({
+  logger: {
+    debug: vi.fn(),
+    info: vi.fn(),
+    warn: vi.fn(),
+    error: vi.fn(),
+  },
+}));
+
+import * as db from './db.js';
+import { resolveOwnerTaskForHumanMessage } from './paired-execution-context.js';
+import type { PairedTask, RegisteredGroup, RoomRoleContext } from './types.js';
+
+const group: RegisteredGroup = {
+  name: 'Paired Room',
+  folder: 'paired-room',
+  trigger: '@codex',
+  added_at: '2026-03-28T00:00:00.000Z',
+  agentType: 'codex',
+  workDir: '/repo/canonical',
+};
+
+const ownerContext: RoomRoleContext = {
+  serviceId: 'codex-main',
+  role: 'owner',
+  ownerServiceId: 'codex-main',
+  reviewerServiceId: 'claude-review',
+  failoverOwner: false,
+};
+
+function buildTask(overrides: Partial<PairedTask> = {}): PairedTask {
+  return {
+    id: 'task-superseded',
+    chat_jid: 'dc:test',
+    group_folder: group.folder,
+    owner_service_id: 'codex-main',
+    reviewer_service_id: 'claude-review',
+    owner_agent_type: 'codex',
+    reviewer_agent_type: 'claude-code',
+    arbiter_agent_type: 'codex',
+    title: null,
+    source_ref: 'HEAD',
+    plan_notes: null,
+    review_requested_at: null,
+    round_trip_count: 0,
+    owner_failure_count: 0,
+    owner_step_done_streak: 0,
+    finalize_step_done_count: 0,
+    task_done_then_user_reopen_count: 0,
+    empty_step_done_streak: 0,
+    status: 'merge_ready',
+    arbiter_verdict: null,
+    arbiter_requested_at: null,
+    completion_reason: null,
+    created_at: '2026-03-28T00:00:00.000Z',
+    updated_at: '2026-03-28T00:05:00.000Z',
+    ...overrides,
+  };
+}
+
+describe('paired execution carry-forward attachments', () => {
+  it('preserves latest owner final attachments when carrying context into a superseding task', () => {
+    const supersededTask = buildTask();
+    const attachments = [
+      {
+        path: '/data/attachments/paired-turn-outputs/task/1/owner/result.png',
+        name: 'result.png',
+        mime: 'image/png',
+      },
+    ];
+    vi.mocked(db.getPairedTurnOutputs).mockReturnValue([
+      {
+        id: 1,
+        task_id: supersededTask.id,
+        turn_number: 1,
+        role: 'owner',
+        output_text: 'TASK_DONE\n새 렌더 이미지입니다.',
+        attachments,
+        created_at: '2026-03-28T00:01:00.000Z',
+      },
+      {
+        id: 2,
+        task_id: supersededTask.id,
+        turn_number: 2,
+        role: 'reviewer',
+        output_text: 'TASK_DONE\nreview approved',
+        created_at: '2026-03-28T00:02:00.000Z',
+      },
+    ]);
+
+    resolveOwnerTaskForHumanMessage({
+      group,
+      chatJid: 'dc:test',
+      roomRoleContext: ownerContext,
+      existingTask: supersededTask,
+    });
+
+    expect(db.insertPairedTurnOutput).toHaveBeenCalledWith(
+      expect.any(String),
+      0,
+      'owner',
+      expect.stringContaining('TASK_DONE\n새 렌더 이미지입니다.'),
+      {
+        createdAt: '2026-03-28T00:01:00.000Z',
+        attachments,
+      },
+    );
+  });
+});
--- a/src/paired-execution-context-routing.test.ts
+++ b/src/paired-execution-context-routing.test.ts
@@ -169,7 +169,7 @@ describe('paired execution routing loop guards', () => {
    );
  });

-  it('routes unknown arbiter verdicts to reviewer as a proceed fallback', () => {
+  it('escalates unknown arbiter verdicts instead of treating them as approval', () => {
    vi.mocked(db.getPairedTaskById).mockReturnValue(
      buildPairedTask({
        status: 'in_arbitration',
@@ -190,12 +190,9 @@ describe('paired execution routing loop guards', () => {
    expect(db.updatePairedTask).toHaveBeenCalledWith(
      'task-1',
      expect.objectContaining({
-        status: 'review_ready',
-        round_trip_count: 0,
-        owner_step_done_streak: 0,
-        empty_step_done_streak: 0,
+        status: 'completed',
        arbiter_verdict: 'unknown',
-        arbiter_requested_at: null,
+        completion_reason: 'arbiter_escalated',
      }),
    );
  });
--- a/src/paired-execution-context.ts
+++ b/src/paired-execution-context.ts
@@ -259,13 +259,17 @@ function carryForwardLatestOwnerFinal(args: {
    0,
    'owner',
    `[Carried forward context from the previous task: latest owner final]\n${latestOwnerFinal.output_text}`,
-    latestOwnerFinal.created_at,
+    {
+      createdAt: latestOwnerFinal.created_at,
+      attachments: latestOwnerFinal.attachments,
+    },
  );
  logger.info(
    {
      sourceTaskId: args.sourceTask.id,
      targetTaskId: args.targetTask.id,
      carriedChars: latestOwnerFinal.output_text.length,
+      attachmentCount: latestOwnerFinal.attachments?.length ?? 0,
    },
    'Carried forward latest owner final into superseding paired task',
  );
--- a/src/paired-verdict.test.ts
+++ b/src/paired-verdict.test.ts
@@ -43,7 +43,7 @@ describe('paired verdict parser', () => {
    ).toBe('continue');
  });

-  it('does not scan beyond the leading visible line window', () => {
+  it('accepts status tokens after a longer evidence preface', () => {
    expect(
      parseVisibleVerdict(
        [
@@ -55,14 +55,41 @@ describe('paired verdict parser', () => {
          'TASK_DONE',
        ].join('\n'),
      ),
+    ).toBe('task_done');
+  });
+
+  it('does not scan beyond the leading visible line window', () => {
+    expect(
+      parseVisibleVerdict(
+        [
+          '01',
+          '02',
+          '03',
+          '04',
+          '05',
+          '06',
+          '07',
+          '08',
+          '09',
+          '10',
+          '11',
+          '12',
+          'TASK_DONE',
+        ].join('\n'),
+      ),
    ).toBe('continue');
  });

-  it('classifies arbiter verdicts from the first visible line', () => {
+  it('classifies arbiter verdicts from leading visible lines', () => {
    expect(classifyArbiterVerdict('PROCEED\ncontinue')).toBe('proceed');
    expect(classifyArbiterVerdict('**VERDICT: REVISE**\nfix it')).toBe(
      'revise',
    );
+    expect(
+      classifyArbiterVerdict(
+        ['중재 판단을 정리합니다.', 'VERDICT: REVISE', 'fix it'].join('\n'),
+      ),
+    ).toBe('revise');
    expect(
      classifyArbiterVerdict(
        '<internal>private notes</internal>\nVERDICT — RESET\nrestart',
--- a/src/paired-verdict.ts
+++ b/src/paired-verdict.ts
@@ -11,7 +11,8 @@ export type VisibleVerdict =

 export type ArbiterVerdictResult = ArbiterVerdict | 'unknown';

-const VISIBLE_VERDICT_SCAN_LINE_LIMIT = 5;
+const VISIBLE_VERDICT_SCAN_LINE_LIMIT = 12;
+const ARBITER_VERDICT_SCAN_LINE_LIMIT = 12;

 function leadingVisibleLines(text: string, limit: number): string[] {
  const lines: string[] = [];
@@ -35,6 +36,10 @@ function leadingVisibleLines(text: string, limit: number): string[] {
  return lines;
 }

+function stripInternalBlocks(text: string): string {
+  return text.replace(/<internal>[\s\S]*?(?:<\/internal>|$)/g, '');
+}
+
 function parseVisibleVerdictLine(line: string): VisibleVerdict | null {
  if (/^\*{0,2}BLOCKED(?:\*{0,2})?\b/i.test(line)) return 'blocked';
  if (/^\*{0,2}NEEDS_CONTEXT(?:\*{0,2})?\b/i.test(line)) return 'needs_context';
@@ -52,7 +57,7 @@ export function parseVisibleVerdict(
  summary: string | null | undefined,
 ): VisibleVerdict {
  if (!summary) return 'continue';
-  const cleaned = summary.replace(/<internal>[\s\S]*?<\/internal>/g, '').trim();
+  const cleaned = stripInternalBlocks(summary).trim();
  if (!cleaned) return 'continue';
  for (const line of leadingVisibleLines(
    cleaned,
@@ -68,17 +73,21 @@ export function classifyArbiterVerdict(
  summary: string | null | undefined,
 ): ArbiterVerdictResult {
  if (!summary) return 'unknown';
-  const cleaned = summary.replace(/<internal>[\s\S]*?<\/internal>/g, '').trim();
+  const cleaned = stripInternalBlocks(summary).trim();
  if (!cleaned) return 'unknown';
-  const firstLine = cleaned.split('\n')[0].trim();
-  const verdictMatch = firstLine.match(
-    /\*{0,2}(?:VERDICT\s*[:—-]\s*)?(PROCEED|REVISE|RESET|ESCALATE|CONTINUE)\*{0,2}/i,
-  );
-  if (verdictMatch) {
-    const normalized = verdictMatch[1].toLowerCase();
-    return normalized === 'continue'
-      ? 'proceed'
-      : (normalized as ArbiterVerdict);
+  for (const line of leadingVisibleLines(
+    cleaned,
+    ARBITER_VERDICT_SCAN_LINE_LIMIT,
+  )) {
+    const verdictMatch = line.match(
+      /\*{0,2}(?:VERDICT\s*[:—-]\s*)?(PROCEED|REVISE|RESET|ESCALATE|CONTINUE)\*{0,2}/i,
+    );
+    if (verdictMatch) {
+      const normalized = verdictMatch[1].toLowerCase();
+      return normalized === 'continue'
+        ? 'proceed'
+        : (normalized as ArbiterVerdict);
+    }
  }
  return 'unknown';
 }