fix: preserve paired input evidence context (#204)

This commit is contained in:
Eyejoker
2026-05-31 17:17:38 +09:00
committed by GitHub
parent c0703836e1
commit 778ed9b94a
12 changed files with 287 additions and 23 deletions

View File

@@ -2,6 +2,7 @@ import fs from 'fs';
import path from 'path';
import {
expandImagePromptReferences,
extractImageTagPaths,
imageTagCaption,
missingImageTagCaption,
@@ -79,7 +80,8 @@ export function buildMultimodalContent(
text: string,
log: LogFn,
): StreamContent {
const { imagePaths } = extractImageTagPaths(text);
const expandedText = expandImagePromptReferences(text);
const { imagePaths } = extractImageTagPaths(expandedText);
if (imagePaths.length === 0) return text;
const blocks: ContentBlock[] = [];
@@ -88,7 +90,7 @@ export function buildMultimodalContent(
if (trimmed) blocks.push({ type: 'text', text: trimmed });
};
for (const part of splitImageTagPromptParts(text)) {
for (const part of splitImageTagPromptParts(expandedText)) {
if (part.type === 'text') {
pushText(part.text);
continue;

View File

@@ -94,4 +94,57 @@ describe('agent-runner multimodal prompts', () => {
]);
expect(logs).toContain(`Unsupported image type, skipping: ${imagePath}`);
});
it('loads MEDIA image directives as Claude image blocks', () => {
const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'ejclaw-media-image-'));
cleanupDirs.push(dir);
const imagePath = path.join(dir, 'media-render.png');
fs.writeFileSync(imagePath, ONE_PIXEL_PNG);
const content = buildMultimodalContent(`증거\nMEDIA:${imagePath}`, () => {
// no-op
});
expect(content).toEqual([
{ type: 'text', text: '증거' },
{ type: 'text', text: 'Image evidence: media-render.png' },
{
type: 'image',
source: {
type: 'base64',
media_type: 'image/png',
data: ONE_PIXEL_PNG.toString('base64'),
},
},
]);
});
it('loads markdown image links as Claude image blocks', () => {
const dir = fs.mkdtempSync(
path.join(os.tmpdir(), 'ejclaw-markdown-image-'),
);
cleanupDirs.push(dir);
const imagePath = path.join(dir, 'markdown-render.png');
fs.writeFileSync(imagePath, ONE_PIXEL_PNG);
const content = buildMultimodalContent(
`증거 ![render](${imagePath})`,
() => {
// no-op
},
);
expect(content).toEqual([
{ type: 'text', text: '증거' },
{ type: 'text', text: 'Image evidence: markdown-render.png' },
{
type: 'image',
source: {
type: 'base64',
media_type: 'image/png',
data: ONE_PIXEL_PNG.toString('base64'),
},
},
]);
});
});

View File

@@ -2,6 +2,7 @@ import fs from 'fs';
import path from 'path';
import {
expandImagePromptReferences,
extractImageTagPaths,
imageTagCaption,
missingImageTagCaption,
@@ -22,7 +23,8 @@ export function parseAppServerInput(
text: string,
log: (message: string) => void = () => undefined,
): AppServerInputItem[] {
const { imagePaths } = extractImageTagPaths(text);
const expandedText = expandImagePromptReferences(text);
const { imagePaths } = extractImageTagPaths(expandedText);
const input: AppServerInputItem[] = [];
const pushText = (value: string) => {
const trimmed = value.trim();
@@ -30,7 +32,7 @@ export function parseAppServerInput(
};
if (imagePaths.length > 0) {
for (const part of splitImageTagPromptParts(text)) {
for (const part of splitImageTagPromptParts(expandedText)) {
if (part.type === 'text') {
pushText(part.text);
continue;

View File

@@ -86,4 +86,36 @@ describe('codex app-server input', () => {
]);
expect(logs).toContain(`Unsupported image type, skipping: ${imagePath}`);
});
it('loads MEDIA image directives as local image input items', () => {
const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'ejclaw-codex-media-'));
cleanupDirs.push(dir);
const imagePath = path.join(dir, 'media-render.png');
fs.writeFileSync(imagePath, ONE_PIXEL_PNG);
const input = parseAppServerInput(`증거\nMEDIA:${imagePath}`);
expect(input).toEqual([
{ type: 'text', text: '증거' },
{ type: 'text', text: 'Image evidence: media-render.png' },
{ type: 'localImage', path: imagePath },
]);
});
it('loads markdown image links as local image input items', () => {
const dir = fs.mkdtempSync(
path.join(os.tmpdir(), 'ejclaw-codex-markdown-'),
);
cleanupDirs.push(dir);
const imagePath = path.join(dir, 'markdown-render.png');
fs.writeFileSync(imagePath, ONE_PIXEL_PNG);
const input = parseAppServerInput(`증거 ![render](${imagePath})`);
expect(input).toEqual([
{ type: 'text', text: '증거' },
{ type: 'text', text: 'Image evidence: markdown-render.png' },
{ type: 'localImage', path: imagePath },
]);
});
});

View File

@@ -92,6 +92,36 @@ export function extractImageTagPaths(text: string): {
};
}
export function expandImagePromptReferences(text: string): string {
const codeSpans = fencedCodeSpans(text);
const withMediaImages = text.replace(
MEDIA_TAG_RE,
(full: string, doubleQuoted, singleQuoted, backticked, bare, offset) => {
if (isInsideSpans(offset, codeSpans)) return full;
const filePath = String(
doubleQuoted ?? singleQuoted ?? backticked ?? bare ?? '',
).trim();
if (!filePath.startsWith('/') || !IMAGE_EXT_RE.test(filePath)) {
return full;
}
const name = attachmentName(filePath) ?? filePath;
return `[Image: ${name}${filePath}]`;
},
);
const markdownCodeSpans = fencedCodeSpans(withMediaImages);
return withMediaImages.replace(
MARKDOWN_IMAGE_ABSOLUTE_LINK_RE,
(full: string, rawPath: string, offset: number) => {
if (isInsideSpans(offset, markdownCodeSpans)) return full;
const filePath = rawPath.trim();
if (!IMAGE_EXT_RE.test(filePath)) return full;
const name = attachmentName(filePath) ?? filePath;
return `[Image: ${name}${filePath}]`;
},
);
}
export type ImageTagPromptPart =
| { type: 'text'; text: string }
| { type: 'image'; label: string | null; path: string; raw: string };

View File

@@ -38,6 +38,7 @@ export {
type TaskContextMode,
} from './task-runtime.js';
export {
expandImagePromptReferences,
extractMarkdownImageAttachments,
extractMediaAttachments,
extractImageTagPaths,

View File

@@ -1,6 +1,7 @@
import { describe, expect, it } from 'vitest';
import {
expandImagePromptReferences,
extractImageTagPaths,
imageTagCaption,
missingImageTagCaption,
@@ -59,6 +60,23 @@ describe('shared agent protocol helpers', () => {
);
});
it('expands MEDIA image directives into image prompt tags', () => {
expect(expandImagePromptReferences('증거\nMEDIA:/tmp/render.png\n끝')).toBe(
'증거\n[Image: render.png → /tmp/render.png]\n끝',
);
});
it('expands markdown image links into image prompt tags', () => {
expect(
expandImagePromptReferences('증거 ![render](/tmp/render.png) 끝'),
).toBe('증거 [Image: render.png → /tmp/render.png] 끝');
});
it('keeps non-image media and fenced media text unchanged', () => {
const text = 'MEDIA:/tmp/demo.mp4\n```text\nMEDIA:/tmp/render.png\n```';
expect(expandImagePromptReferences(text)).toBe(text);
});
it('normalizes plain text runner output as public text', () => {
expect(normalizePublicTextOutput('DONE')).toEqual({
result: 'DONE',

View File

@@ -1,8 +1,4 @@
import {
getLastHumanMessageContent,
getPairedTurnOutputs,
getRecentChatMessages,
} from './db.js';
import { getLastHumanMessageContent, getPairedTurnOutputs } from './db.js';
import { logger } from './logger.js';
import {
buildArbiterPromptForTask,
@@ -22,6 +18,7 @@ import {
resolveNextTurnAction,
} from './message-runtime-rules.js';
import type { ExecuteTurnFn } from './message-runtime-types.js';
import { getTaskContextMessages } from './message-runtime-task-context.js';
import { buildPairedTurnIdentity } from './paired-turn-identity.js';
import { resolveStoredVisibleVerdict } from './paired-verdict.js';
import {
@@ -106,7 +103,6 @@ export function buildPendingPairedTurn(args: {
timezone,
task,
rawMissedMessages,
recentHumanMessages,
labeledRecentMessages,
resolveChannel,
} = args;
@@ -124,7 +120,7 @@ export function buildPendingPairedTurn(args: {
}),
ownerFailureCount: task.owner_failure_count,
});
const recentMessages = getRecentChatMessages(chatJid, 20);
const taskContextMessages = getTaskContextMessages(chatJid, task);
const lastHumanMessage = getLastHumanMessageContent(chatJid);
if (nextTurnAction.kind === 'reviewer-turn') {
@@ -133,7 +129,9 @@ export function buildPendingPairedTurn(args: {
chatJid,
timezone,
turnOutputs,
recentHumanMessages,
recentHumanMessages: taskContextMessages.filter(
(message) => !message.is_bot_message,
),
lastHumanMessage,
taskCreatedAt: task.created_at,
}),
@@ -154,7 +152,7 @@ export function buildPendingPairedTurn(args: {
chatJid,
timezone,
turnOutputs,
recentMessages,
recentMessages: taskContextMessages,
labeledRecentMessages,
}),
channel: resolveChannel(taskStatus),
@@ -185,7 +183,9 @@ export function buildPendingPairedTurn(args: {
chatJid,
timezone,
turnOutputs,
recentHumanMessages,
recentHumanMessages: taskContextMessages.filter(
(message) => !message.is_bot_message,
),
lastHumanMessage,
taskCreatedAt: task.created_at,
}),

View File

@@ -1,4 +1,5 @@
import { buildArbiterContextPrompt } from './arbiter-context.js';
import { TASK_USER_CONTEXT_START_SKEW_MS } from './message-runtime-task-context.js';
import { formatMessages } from './router.js';
import type {
NewMessage,
@@ -14,7 +15,6 @@ const CARRIED_FORWARD_OWNER_FINAL_GUIDANCE = `System note:
If you see a message beginning with "${CARRIED_FORWARD_OWNER_FINAL_MARKER}", treat it as background only. Do not repeat, continue, or answer that carried-forward final directly. Respond only to the latest human request and the current task.`;
const ARBITER_TURN_OUTPUT_CONTEXT_LIMIT = 6;
const TASK_USER_CONTEXT_START_SKEW_MS = 5_000;
function isImageAttachment(attachment: OutboundAttachment): boolean {
if (attachment.mime?.toLowerCase().startsWith('image/')) return true;

View File

@@ -1,4 +1,4 @@
import { getPairedTurnOutputs, getRecentChatMessages } from './db.js';
import { getPairedTurnOutputs } from './db.js';
import { logger } from './logger.js';
import {
buildArbiterPromptForTask,
@@ -21,6 +21,7 @@ import {
resolveQueuedPairedTurnRole,
resolveQueuedTurnRole,
} from './message-runtime-rules.js';
import { getTaskContextMessages } from './message-runtime-task-context.js';
import { claimPairedTurnExecution } from './paired-follow-up-scheduler.js';
import type {
ExecuteTurnFn,
@@ -67,7 +68,10 @@ function buildQueuedGroupTurnPrompt(args: {
formatMessages: (messages: NewMessage[], timezone: string) => string;
}): string {
if (args.turnRole === 'arbiter' && args.currentTask) {
const recentMessages = getRecentChatMessages(args.chatJid, 20);
const recentMessages = getTaskContextMessages(
args.chatJid,
args.currentTask,
);
return buildArbiterPromptForTask({
task: args.currentTask,
chatJid: args.chatJid,
@@ -165,13 +169,13 @@ export async function runPendingPairedTurnIfNeeded(args: {
return null;
}
const recentMessages = getRecentChatMessages(chatJid, 20);
const recentHumanMessages = recentMessages.filter(
const taskContextMessages = getTaskContextMessages(chatJid, task);
const recentHumanMessages = taskContextMessages.filter(
(message) => !message.is_bot_message,
);
const labeledRecentMessages = args.labelPairedSenders(
chatJid,
recentMessages,
taskContextMessages,
);
const pendingTurn = buildPendingPairedTurn({
chatJid,
@@ -255,9 +259,11 @@ export async function runQueuedGroupTurn(args: {
});
currentTask = resolvedTask.task;
if (resolvedTask.supersededTask) {
fallbackMessages = getRecentChatMessages(chatJid, 20).filter(
(message) => !message.is_bot_message,
);
fallbackMessages = currentTask
? getTaskContextMessages(chatJid, currentTask).filter(
(message) => !message.is_bot_message,
)
: missedMessages;
}
}
const taskStatus = currentTask?.status;

View File

@@ -0,0 +1,85 @@
import { beforeEach, describe, expect, it } from 'vitest';
import { _initTestDatabase, storeChatMetadata, storeMessage } from './db.js';
import { getTaskContextMessages } from './message-runtime-task-context.js';
import type { NewMessage, PairedTask } from './types.js';
function makeMessage(
content: string,
timestamp: string,
overrides: Partial<NewMessage> = {},
): NewMessage {
return {
id: `msg-${content}`,
chat_jid: 'group@test',
sender: 'user@test',
sender_name: 'User',
content,
timestamp,
is_from_me: false,
is_bot_message: false,
...overrides,
};
}
function makeTask(overrides: Partial<PairedTask> = {}): PairedTask {
return {
id: 'task-context',
chat_jid: 'group@test',
group_folder: 'group',
owner_service_id: 'codex-main',
reviewer_service_id: 'claude',
owner_agent_type: 'codex',
reviewer_agent_type: 'claude-code',
arbiter_agent_type: 'codex',
title: null,
source_ref: null,
plan_notes: null,
review_requested_at: null,
round_trip_count: 1,
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: 'review_ready',
arbiter_verdict: null,
arbiter_requested_at: null,
completion_reason: null,
created_at: '2026-04-20T00:00:00.000Z',
updated_at: '2026-04-20T00:30:00.000Z',
...overrides,
};
}
describe('message runtime task context', () => {
beforeEach(() => {
_initTestDatabase();
});
it('keeps the original task request even after it falls outside the recent 20 messages', () => {
storeChatMetadata(
'group@test',
'Test Group',
'2026-04-20T00:00:00.000Z',
'discord',
true,
);
storeMessage(makeMessage('원래 사용자 요청', '2026-04-20T00:00:01.000Z'));
for (let index = 0; index < 25; index += 1) {
storeMessage(
makeMessage(
`후속 노이즈 ${index}`,
`2026-04-20T00:${String(index + 2).padStart(2, '0')}:00.000Z`,
{ id: `noise-${index}`, is_bot_message: index % 2 === 0 },
),
);
}
const messages = getTaskContextMessages('group@test', makeTask());
expect(messages.map((message) => message.content)).toContain(
'원래 사용자 요청',
);
});
});

View File

@@ -0,0 +1,35 @@
import { getMessagesSince, getRecentChatMessages } from './db.js';
import type { NewMessage, PairedTask } from './types.js';
export const TASK_USER_CONTEXT_START_SKEW_MS = 5_000;
const TASK_CONTEXT_MESSAGE_LIMIT = 200;
const NO_BOT_PREFIX_FILTER = '';
function taskContextStartTimestamp(task: PairedTask): string | null {
const taskCreatedMs = Date.parse(task.created_at);
if (!Number.isFinite(taskCreatedMs)) {
return null;
}
return new Date(
taskCreatedMs - TASK_USER_CONTEXT_START_SKEW_MS,
).toISOString();
}
export function getTaskContextMessages(
chatJid: string,
task: PairedTask,
): NewMessage[] {
const sinceTimestamp = taskContextStartTimestamp(task);
if (!sinceTimestamp) {
return getRecentChatMessages(chatJid, 20);
}
const taskMessages = getMessagesSince(
chatJid,
sinceTimestamp,
NO_BOT_PREFIX_FILTER,
TASK_CONTEXT_MESSAGE_LIMIT,
);
return taskMessages.length > 0
? taskMessages
: getRecentChatMessages(chatJid, 20);
}