Skip to content

refactor(server): make title prompts plaintext #5368

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
t3dotgg merged 1 commit into from
Aug 4, 2026
+98 −104
Merged

refactor(server): make title prompts plaintext #5368

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
202 changes: 98 additions & 104 deletions apps/server/src/textGeneration/TextGenerationPrompts.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -150,28 +150,12 @@ export interface BranchNamePromptInput {
interface PromptFromMessageInput {
instruction: string;
responseShape: string;
guidance?: ReadonlyArray<string> | undefined;
rulesLabel?: string | undefined;
rules: ReadonlyArray<string>;
afterRules?: ReadonlyArray<string> | undefined;
message: string;
messageLabel?: string | undefined;
preserveMessageEnd?: boolean | undefined;
attachments?: ReadonlyArray<ChatAttachment> | undefined;
additionalInstructions?: string | undefined;
}

function preserveMessageEnd(message: string): string {
const alreadyTruncated = message.startsWith(EARLIER_CONTENT_TRUNCATION_MARKER);
const contents = alreadyTruncated
? message.slice(EARLIER_CONTENT_TRUNCATION_MARKER.length)
: message;
if (!alreadyTruncated && contents.length <= 8_000) {
return contents;
}
return `${EARLIER_CONTENT_TRUNCATION_MARKER}${contents.slice(-8_000)}`;
}

function buildPromptFromMessage(input: PromptFromMessageInput): string {
const attachmentLines = (input.attachments ?? []).map(
(attachment) => `- ${attachment.name} (${attachment.mimeType}, ${attachment.sizeBytes} bytes)`,
Expand All @@ -180,15 +164,11 @@ function buildPromptFromMessage(input: PromptFromMessageInput): string {
const promptSections = [
input.instruction,
input.responseShape,
...(input.guidance ?? []),
input.rulesLabel ?? "Rules:",
"Rules:",
...input.rules.map((rule) => `- ${rule}`),
...(input.afterRules ?? []),
"",
`${input.messageLabel ?? "User message"}:`,
input.preserveMessageEnd
? preserveMessageEnd(input.message)
: limitSection(input.message, 8_000),
"User message:",
limitSection(input.message, 8_000),
...policyInstruction(input.additionalInstructions),
];
if (attachmentLines.length > 0) {
Expand Down Expand Up @@ -234,88 +214,102 @@ export interface ThreadTitlePromptInput {
policy?: TextGenerationPolicy | undefined;
}

// Keep shared editorial rules in these two prompts in sync. Regeneration
// intentionally adds guidance for thread history and the previous title.
const INITIAL_THREAD_TITLE_PROMPT = `Generate a title that will help the user recognize this T3 Code thread weeks later.
Return JSON with exactly one key: title.

Before answering, silently reduce the request to:
- Subject: What system, feature, or problem is this really about?
- Outcome: What does the user ultimately want to understand or change?
- Incidental instructions: What only describes how the agent should do the work?

Title the subject and outcome. Discard incidental instructions.

Editorial rules:
- 3-8 words, fewer than 40 characters.
- Use a compact noun phrase or clear action phrase.
- Capture the umbrella goal when the request lists several symptoms or steps.
- Name the product change, not the mock, plan, report, branch, or PR used to produce it.
- Models, subagents, tools, output formats, and monitoring instructions do not belong in the title unless they are themselves the topic.
- For reviews, name what is being reviewed and the relevant concern. Avoid generic titles such as "Review PR 123" when linked or attached context reveals the subject.
- For research, name the question domain rather than the requested research process.
- Do not claim the work is complete.
- Do not copy and truncate the user's message.
- Avoid project names already visible in the UI, quotes, labels, filler, and trailing punctuation.
- Use attached images as primary context for UI issues.
- When a URL or attachment is the only source of the subject, use available tools to inspect it. If it cannot be resolved, remain accurate rather than guessing.`;

function regenerateThreadTitlePrompt(previousTitle: string): string {
return `Regenerate the title for an existing T3 Code thread so the user can recognize it weeks later.
The previous title was ${JSON.stringify(previousTitle)}.
Return JSON with exactly one key: title.

Determine the title in this order:
1. Read the USER messages first. Identify the latest explicit durable goal. The original subject remains the subject until the user clearly changes what the thread is about.
2. Use ASSISTANT messages to resolve vague links, unnamed code, and discovered product nouns. Do not promote one assistant finding into the thread subject unless the user adopts it as a new goal.
3. Compare that subject with the previous title. Preserve accurate scope words, especially when earlier content is truncated. Replace the previous title when it is generic, artifact-based, a completion update, or contradicted by the thread.
4. Title the durable subject and desired outcome, not the current workflow state.

Editorial rules:
- 3-8 words, fewer than 40 characters.
- Use a compact noun phrase or clear action phrase.
- Preserve the umbrella subject when later messages focus on one finding, provider, platform, or implementation detail.
- A thread progressing through research, planning, implementation, review, CI, merge, and monitoring has usually not changed subjects.
- Ignore deliverables and operations such as mocks, plans, HTML, branches, PRs, tests, CI, commits, merging, and monitoring unless they are the actual topic.
- Models, subagents, tools, output formats, and monitoring instructions do not belong in the title unless they are themselves the topic.
- Treat final operational follow-ups and assistant completion summaries as weak evidence of subject.
- For reviews, name the reviewed feature or system and its durable concern, not one finding from the review.
- For research, name the question domain rather than the research process.
- Do not claim the work is complete.
- Do not copy and truncate a thread message.
- Avoid project names already visible in the UI, PR numbers, quotes, labels, filler, and trailing punctuation.
- Use attached images as primary context for UI issues.
- When a URL or attachment is the only source of the subject, use available tools to inspect it. If it cannot be resolved, remain accurate rather than guessing.
- Return a meaningfully improved title, not a cosmetic paraphrase of the previous title.

Examples of the distinction:
- A subagent-monitoring review that finds a Codex roster bug remains "Review Subagent Monitoring Risks," not "Codex Roster Bug Review."
- A vague failing-test request later identified as a lazy thread-feed mismatch becomes "Fix Lazy Thread Feed Test," not "Prevent Mobile Feed Regressions."
- A QR-sharing overhaul that ends with CI and merge work remains about QR sharing, not the PR lifecycle.`;
}

function preserveMessageEnd(message: string): string {
const alreadyTruncated = message.startsWith(EARLIER_CONTENT_TRUNCATION_MARKER);
const contents = alreadyTruncated
? message.slice(EARLIER_CONTENT_TRUNCATION_MARKER.length)
: message;
if (!alreadyTruncated && contents.length <= 8_000) {
return contents;
}
return `${EARLIER_CONTENT_TRUNCATION_MARKER}${contents.slice(-8_000)}`;
}

function threadTitlePromptSuffix(input: ThreadTitlePromptInput): string {
const additionalInstructions = policyInstruction(input.policy?.threadTitleInstructions);
const attachmentLines = (input.attachments ?? []).map(
(attachment) => `- ${attachment.name} (${attachment.mimeType}, ${attachment.sizeBytes} bytes)`,
);

let suffix = "";
if (additionalInstructions.length > 0) {
suffix = `\n${additionalInstructions.join("\n")}`;
}
if (attachmentLines.length > 0) {
suffix += `\n\nAttachment metadata:\n${limitSection(attachmentLines.join("\n"), 4_000)}`;
}
return suffix;
}

export function buildThreadTitlePrompt(input: ThreadTitlePromptInput) {
const isRegeneration = input.previousTitle !== undefined;
const prompt = buildPromptFromMessage({
instruction: isRegeneration
? [
"Regenerate the title for an existing T3 Code thread so the user can recognize it weeks later.",
`The previous title was ${JSON.stringify(input.previousTitle)}.`,
].join("\n")
: "Generate a title that will help the user recognize this T3 Code thread weeks later.",
responseShape: "Return JSON with exactly one key: title.",
guidance: isRegeneration
? [
"",
"Determine the title in this order:",
"1. Read the USER messages first. Identify the latest explicit durable goal. The original subject remains the subject until the user clearly changes what the thread is about.",
"2. Use ASSISTANT messages to resolve vague links, unnamed code, and discovered product nouns. Do not promote one assistant finding into the thread subject unless the user adopts it as a new goal.",
"3. Compare that subject with the previous title. Preserve accurate scope words, especially when earlier content is truncated. Replace the previous title when it is generic, artifact-based, a completion update, or contradicted by the thread.",
"4. Title the durable subject and desired outcome, not the current workflow state.",
"",
]
: [
"",
"Before answering, silently reduce the request to:",
"- Subject: What system, feature, or problem is this really about?",
"- Outcome: What does the user ultimately want to understand or change?",
"- Incidental instructions: What only describes how the agent should do the work?",
"",
"Title the subject and outcome. Discard incidental instructions.",
"",
],
rulesLabel: "Editorial rules:",
rules: isRegeneration
? [
"3-8 words, fewer than 40 characters.",
"Use a compact noun phrase or clear action phrase.",
"Preserve the umbrella subject when later messages focus on one finding, provider, platform, or implementation detail.",
"A thread progressing through research, planning, implementation, review, CI, merge, and monitoring has usually not changed subjects.",
"Ignore deliverables and operations such as mocks, plans, HTML, branches, PRs, tests, CI, commits, merging, and monitoring unless they are the actual topic.",
"Models, subagents, tools, output formats, and monitoring instructions do not belong in the title unless they are themselves the topic.",
"Treat final operational follow-ups and assistant completion summaries as weak evidence of subject.",
"For reviews, name the reviewed feature or system and its durable concern, not one finding from the review.",
"For research, name the question domain rather than the research process.",
"Do not claim the work is complete.",
"Do not copy and truncate a thread message.",
"Avoid project names already visible in the UI, PR numbers, quotes, labels, filler, and trailing punctuation.",
"Use attached images as primary context for UI issues.",
"When a URL or attachment is the only source of the subject, use available tools to inspect it. If it cannot be resolved, remain accurate rather than guessing.",
"Return a meaningfully improved title, not a cosmetic paraphrase of the previous title.",
]
: [
"3-8 words, fewer than 40 characters.",
"Use a compact noun phrase or clear action phrase.",
"Capture the umbrella goal when the request lists several symptoms or steps.",
"Name the product change, not the mock, plan, report, branch, or PR used to produce it.",
"Models, subagents, tools, output formats, and monitoring instructions do not belong in the title unless they are themselves the topic.",
'For reviews, name what is being reviewed and the relevant concern. Avoid generic titles such as "Review PR 123" when linked or attached context reveals the subject.',
"For research, name the question domain rather than the requested research process.",
"Do not claim the work is complete.",
"Do not copy and truncate the user's message.",
"Avoid project names already visible in the UI, quotes, labels, filler, and trailing punctuation.",
"Use attached images as primary context for UI issues.",
"When a URL or attachment is the only source of the subject, use available tools to inspect it. If it cannot be resolved, remain accurate rather than guessing.",
],
afterRules: isRegeneration
? [
"",
"Examples of the distinction:",
'- A subagent-monitoring review that finds a Codex roster bug remains "Review Subagent Monitoring Risks," not "Codex Roster Bug Review."',
'- A vague failing-test request later identified as a lazy thread-feed mismatch becomes "Fix Lazy Thread Feed Test," not "Prevent Mobile Feed Regressions."',
"- A QR-sharing overhaul that ends with CI and merge work remains about QR sharing, not the PR lifecycle.",
]
: undefined,
message: input.message,
...(isRegeneration
? {
messageLabel: "Thread contents",
preserveMessageEnd: true,
}
: {}),
attachments: input.attachments,
additionalInstructions: input.policy?.threadTitleInstructions,
});
let prompt: string;
if (input.previousTitle === undefined) {
const message = limitSection(input.message, 8_000);
prompt = `${INITIAL_THREAD_TITLE_PROMPT}\n\nUser message:\n${message}${threadTitlePromptSuffix(input)}`;
} else {
const message = preserveMessageEnd(input.message);
prompt = `${regenerateThreadTitlePrompt(input.previousTitle)}\n\nThread contents:\n${message}${threadTitlePromptSuffix(input)}`;
}
const outputSchema = Schema.Struct({
title: Schema.String,
});
Expand Down
Loading