Skip to content

Soul-driven AI agent with permission-hardened tools, token budgets, and multi-channel access. Runs 24/7 from CLI, Telegram or More.

refactor(core): extract system-prompt assembly and repetition detection from Agent

Salman Qureshi6720fc499c76
diff --git a/src/core/agent.ts b/src/core/agent.ts
index c4430898..357a6127 100644
--- a/src/core/agent.ts
+++ b/src/core/agent.ts
@@ -85,7 +85,8 @@ import { compactConversation, memoryGovernorThresholds, memoryGovernorVerdict }
import { classifyStreamCompletion, isLengthTruncation, truncationContinuationPrompt, toolTruncationContinuationPrompt } from './stream-completion.js';
import { MAX_EXECUTE_CONTINUATIONS, MAX_VERIFICATION_CONTINUATIONS, executeContinuationPrompt, shouldForceExecuteContinuation, isFailedToolResult, shouldRequireVerification, isVerificationOutputOk, verificationPrompt, responseAsksUser, EXECUTE_MUTATING_TOOLS, VERIFICATION_COMMAND_PATTERN, wakeUpPrompt } from './execute-guard.js';
import { classifyTurnEnd, stepsExhaustedPrompt, STEPS_PAUSED_BANNER, WORK_NOT_STARTED_BANNER, VERIFICATION_FAILED_BANNER, type LoopEndCause } from './completion-verdict.js';
-import { HISTORY_TOKEN_BUDGET, HISTORY_TOKEN_BUDGET_SAVER, TOOL_TRACE_KIND, selectHistoryWindow, toModelMessage, withReasoningParts, formatToolTrace, formatToolTraceLine, summarizeToolArgs, summarizeToolResult } from './context-window.js';
+import { HISTORY_TOKEN_BUDGET, HISTORY_TOKEN_BUDGET_SAVER, TOOL_TRACE_KIND, selectHistoryWindow, toModelMessage, withReasoningParts, formatToolTrace, formatToolTraceLine, summarizeToolArgs, summarizeToolResult, detectRepetitionLoop } from './context-window.js';
+import { buildSystemPrompt as assembleSystemPrompt } from './system-prompt.js';
import { buildStatusVerbPrompt, parseStatusVerbs, shouldRefreshStatusVerbs } from './status-verbs.js';
import { verbPoolFor } from '../ui/status-word.js';
import { StallWatchdog } from './stall-watchdog.js';
@@ -3094,45 +3095,8 @@ export class Agent {
// mind." turns primed terse acknowledgements and polluted history.
const contextBlocks: string[] = [];
- const recentSteps = recentMemory.slice(-6);
- let loopWarning: string | null = null;
- if (recentSteps.length >= 3) {
- const toolCallPattern = /\[Using: (.+?)\]/g;
- const toolCalls: string[] = [];
- for (const m of recentSteps) {
- if (m.role === 'assistant') {
- let match;
- while ((match = toolCallPattern.exec(m.content)) !== null) {
- toolCalls.push(match[1]);
- }
- }
- }
- if (toolCalls.length >= 3) {
- const last3 = toolCalls.slice(-3);
- if (last3[0] === last3[1] && last3[1] === last3[2]) {
- loopWarning = `[SYSTEM WARNING] You have called ${last3[0]} 3+ times in a row with the same result. Stop repeating this call. Try a different approach — if you're failing on permissions, try a different path. If you're failing on git push auth, use github_api with PUT /repos/{owner}/{repo}/contents/{path} to push files directly through the API.`;
- }
- }
-
- if (!loopWarning) {
- const assistantMessages = recentSteps.filter(m => m.role === 'assistant' && m.content.length > 20);
- if (assistantMessages.length >= 3) {
- const last3 = assistantMessages.slice(-3);
- const normalizeText = (t: string) => t.toLowerCase().replace(/[^\w\s]/g, '').replace(/\s+/g, ' ').trim().slice(0, 150);
- const normalized = last3.map(m => normalizeText(m.content));
- const words0 = new Set(normalized[0].split(' '));
- const overlap01 = normalized[0] && normalized[1] ? [...words0].filter(w => new Set(normalized[1].split(' ')).has(w)).length / Math.max(words0.size, 1) : 0;
- const overlap12 = normalized[1] && normalized[2] ? [...new Set(normalized[1].split(' '))].filter(w => new Set(normalized[2].split(' ')).has(w)).length / Math.max(new Set(normalized[1].split(' ')).size, 1) : 0;
- if (overlap01 > 0.75 && overlap12 > 0.75) {
- loopWarning = `[SYSTEM WARNING] Your last 3 responses are nearly identical. You are stuck in a text repetition loop. Stop immediately and give a completely different response. If you cannot complete the task, tell the user clearly why.`;
- }
- }
- }
- }
-
- if (loopWarning) {
- contextBlocks.push(loopWarning);
- }
+ const loopWarning = detectRepetitionLoop(recentMemory);
+ if (loopWarning) contextBlocks.push(loopWarning);
if (this.userMemory) {
const memoryContext = this.userMemory.retrieveRelevant(msg.content, { maxRecords: 5, maxChars: 900 });
@@ -4859,112 +4823,26 @@ export class Agent {
];
}
+ /** Stable system prompt; assembly lives in core/system-prompt.ts. */
private buildSystemPrompt(): string {
- let prompt = this.identity.getSystemPrompt(this.config.identity);
- const skillContext = this.capabilities.getSkillContext();
- if (skillContext) {
- prompt += '\n\n' + skillContext;
- }
- // Mercury Bots: the main agent must know the fleet exists — what each
- // bot does, its live state, and how to dispatch to it. Without this the
- // conversational agent answers bot questions blindly.
- if (this.botManager) {
- prompt += this.botManager.getSystemPromptSection();
- }
- const programmingSuffix = this.programmingMode.getSystemPromptSuffix();
- if (programmingSuffix) {
- prompt += programmingSuffix;
- }
- const researchSuffix = this.researchMode.getSystemPromptSuffix();
- if (researchSuffix) {
- prompt += researchSuffix;
- }
- const budgetStatus = this.tokenBudget.getStatusText();
- prompt += '\n\n' + budgetStatus;
- if (this.tokenBudget.getUsagePercentage() > 70) {
- prompt += '\nBe concise to conserve tokens.';
- }
- const saverSuffix = this.saverMode.getSystemPromptSuffix();
- if (saverSuffix) {
- prompt += saverSuffix;
- }
-
- const now = new Date();
- const dateStr = now.toLocaleDateString('en-US', { weekday: 'long', year: 'numeric', month: 'long', day: 'numeric' });
- // Hour resolution on purpose: a minute-precise clock changed the system
- // prompt every turn and defeated prompt caching for the ~10K-token
- // prefix that follows. Tools report exact timestamps when it matters.
- const timeStr = `about ${now.toLocaleTimeString('en-US', { hour: 'numeric', hour12: true })}`;
- const timezone = Intl.DateTimeFormat().resolvedOptions().timeZone;
- prompt += `\n\nEnvironment:\n- Date: ${dateStr}, ${timeStr} (${timezone})\n- Platform: ${process.platform}\n- Working directory: ${this.capabilities.getCwd()}`;
-
- prompt += `\n\n**Tool Usage Guidelines:**
-- Use write_file, create_file, and edit_file tools DIRECTLY to create and modify files. Do NOT create intermediary scripts (Python, bash, Node.js) whose sole purpose is to generate other files — you have native file tools for this.
-- Use run_command for: building, testing, installing dependencies, running the project, git operations, and other system tasks that require a shell.
-- Do NOT use run_command with echo/cat/tee/heredoc to write files. Use write_file or create_file instead.
-- Do NOT create one-time-use helper scripts. If the user asks you to create a file, create it directly with create_file or write_file.
-- When creating multiple files, call create_file or write_file for each one individually. Do not batch them into a script.`;
-
- if (this.userMemory) {
- const summary = this.userMemory.getSummary();
- prompt += `\n\nSecond Brain (SQLite-backed long-term memory) is ENABLED. You have ${summary.total} persistent memories about this user.`;
- prompt += `\nMemory types: identity, preference, goal, project, habit, decision, constraint, relationship, episode, reflection.`;
- prompt += `\n\nCRITICAL — Memory storage rules:`;
- prompt += `\n- ALL persistent user knowledge lives in the Second Brain SQLite database — this is the single source of truth.`;
- prompt += `\n- NEVER use create_file, write_file, edit_file, or any file tool to store memories, notes, facts, preferences, or brain data. Files are for code and documents, not for knowledge storage.`;
- prompt += `\n- New memories are extracted AUTOMATICALLY after each conversation turn. You do not need to ask the user if they want to save something.`;
- prompt += `\n- When the user explicitly asks you to "save/remember/note/keep this," use the save_memory tool to store it directly — no follow-up questions needed.`;
- prompt += `\n- When you need to actively recall something beyond auto-injected context (e.g. "do you remember...", "what do I know about..."), use the search_memory tool.`;
- prompt += `\n- Relevant memories are auto-injected before each message. You can reference them naturally (e.g. "I remember you prefer TypeScript").`;
- prompt += `\n- Users can manage memory with: /memory (overview, search, pause learning, clear).`;
- if (summary.learningPaused) {
- prompt += `\n\nLearning is currently PAUSED — no new memories will be extracted or saved until resumed.`;
- }
- } else {
- prompt += '\n\nSecond Brain is DISABLED. Basic long-term memory (text search over facts) is still active.';
- }
-
- // Notification routing guidance for tweet-notifier skill
- const skillNames = this.capabilities.getSkillContext();
- if (skillNames.includes('tweet-notifier')) {
- prompt += `\n\n**Tweet Notification System Available** — The tweet-notifier skill is installed.
-When you need to schedule tweets, manage approvals, or notify founders/supporters:
-1. Use the \`use_skill\` tool to invoke the \`tweet-notifier\` skill for detailed instructions
-2. The skill provides templates for scheduling tweets, notifying founders (via send_message), and alerting supporters (approved Telegram users)
-3. Key tools used by this system: schedule_task (for timing), send_message (for notifications to Telegram), save_memory (for tweet state tracking), search_memory (for checking existing tweets)
-4. Supporters are all approved Telegram users — send_message will reach them
-5. The founder (Optimus Prime) receives notifications via send_message (Telegram)`;
- }
-
- const toolNames = this.capabilities.getToolNames();
- const githubTools = ['create_pr', 'review_pr', 'list_issues', 'create_issue', 'github_api'];
- const hasGitHub = githubTools.some(t => toolNames.includes(t));
- if (hasGitHub) {
- let githubHint = '\n\nGitHub companion is active.';
- const { defaultOwner, defaultRepo } = this.config.github;
- if (defaultOwner && defaultRepo) {
- githubHint += ` Default repo: ${defaultOwner}/${defaultRepo}. Use this when the user doesn't specify a repo.`;
- }
-
- githubHint += `
-
-Available GitHub tools and when to use them:
-- git_add, git_commit, git_push: LOCAL git operations (stage, commit, push to a remote you have SSH/auth access to). All commits include "Co-authored-by: Mercury <[email protected]>".
-- create_pr: Create a pull request on GitHub. The head branch must already exist on the remote.
-- review_pr: Get PR details and optionally post a review comment.
-- list_issues, create_issue: Browse and file issues.
-- github_api: Raw GitHub API access. IMPORTANT USE CASES:
- - Push files directly to GitHub via PUT /repos/{owner}/{repo}/contents/{path} when git push fails due to auth. The body must include "message" and "content" (base64-encoded file content). This creates a commit on GitHub with Mercury as co-author.
- - Delete files via DELETE /repos/{owner}/{repo}/contents/{path} with a "message" and "sha" in the body.
- - Any other GitHub API operation not covered by the other tools.
-
-When the user asks to "push to GitHub" or "upload files" and git push fails, use github_api with PUT /repos/{owner}/{repo}/contents/{path} to push content directly through the API. This bypasses local git entirely.
-
-Always specify owner and repo parameters on GitHub tools. The user's GitHub username is ${this.config.github.username || 'not set'}.'`;
-
- prompt += githubHint;
- }
- return prompt;
+ const summary = this.userMemory ? this.userMemory.getSummary() : null;
+ return assembleSystemPrompt({
+ identityPrompt: this.identity.getSystemPrompt(this.config.identity),
+ skillContext: this.capabilities.getSkillContext(),
+ botSection: this.botManager ? this.botManager.getSystemPromptSection() : '',
+ programmingSuffix: this.programmingMode.getSystemPromptSuffix(),
+ researchSuffix: this.researchMode.getSystemPromptSuffix(),
+ budgetStatus: this.tokenBudget.getStatusText(),
+ budgetUsagePercentage: this.tokenBudget.getUsagePercentage(),
+ saverSuffix: this.saverMode.getSystemPromptSuffix(),
+ cwd: this.capabilities.getCwd(),
+ now: new Date(),
+ platform: process.platform,
+ timezone: Intl.DateTimeFormat().resolvedOptions().timeZone,
+ memorySummary: summary ? { total: summary.total, learningPaused: summary.learningPaused } : null,
+ toolNames: this.capabilities.getToolNames(),
+ github: this.config.github,
+ });
}
async processInternalPrompt(prompt: string, channelId?: string, channelType?: string): Promise<void> {
diff --git a/src/core/context-window.test.ts b/src/core/context-window.test.ts
index fe71d543..a100a00d 100644
--- a/src/core/context-window.test.ts
+++ b/src/core/context-window.test.ts
@@ -2,7 +2,9 @@ import { describe, expect, it } from 'vitest';
import type { SessionMessage } from '../sessions/types.js';
import {
HISTORY_MIN_ENTRIES,
+ TEXT_REPETITION_WARNING,
TOOL_TRACE_KIND,
+ detectRepetitionLoop,
formatToolTrace,
formatToolTraceLine,
selectHistoryWindow,
@@ -84,6 +86,19 @@ describe('reasoning parts', () => {
});
});
+describe('detectRepetitionLoop', () => {
+ const reply = (text: string) => entry('assistant', text);
+ it('fires when the last three replies are near-identical', () => {
+ const same = 'I will look into the build failure and report back shortly with details.';
+ expect(detectRepetitionLoop([entry('user', 'go'), reply(same), reply(same + '!'), reply(same)])).toBe(TEXT_REPETITION_WARNING);
+ });
+ it('stays quiet for varied replies, short replies, or fewer than three', () => {
+ expect(detectRepetitionLoop([reply('Fixed the test in src/app.ts and re-ran the suite.'), reply('Now updating the docs for the new flag.'), reply('Done — all green, summary below.')])).toBeNull();
+ expect(detectRepetitionLoop([reply('ok'), reply('ok'), reply('ok')])).toBeNull();
+ expect(detectRepetitionLoop([reply('I will look into the build failure and report back shortly.'), reply('I will look into the build failure and report back shortly.')])).toBeNull();
+ });
+});
+
describe('tool trace formatting', () => {
it('summarises args by the most useful field and results by outcome', () => {
expect(summarizeToolArgs({ path: 'src/app.ts', encoding: 'utf8' })).toBe('path=src/app.ts');
diff --git a/src/core/context-window.ts b/src/core/context-window.ts
index a52862c8..435e465f 100644
--- a/src/core/context-window.ts
+++ b/src/core/context-window.ts
@@ -137,3 +137,29 @@ export function formatToolTrace(lines: readonly string[]): string {
function truncate(text: string, max: number): string {
return text.length > max ? `${text.slice(0, max - 1)}…` : text;
}
+
+/** Warning injected when the last three assistant replies are near-identical. */
+export const TEXT_REPETITION_WARNING =
+ '[SYSTEM WARNING] Your last 3 responses are nearly identical. You are stuck in a text repetition loop. Stop immediately and give a completely different response. If you cannot complete the task, tell the user clearly why.';
+
+const normalizeText = (t: string) => t.toLowerCase().replace(/[^\w\s]/g, '').replace(/\s+/g, ' ').trim().slice(0, 150);
+const overlap = (a: string, b: string): number => {
+ const wa = new Set(a.split(' '));
+ const wb = new Set(b.split(' '));
+ return [...wa].filter((w) => wb.has(w)).length / Math.max(wa.size, 1);
+};
+
+/**
+ * Detect a reply-repetition loop in recent history: three consecutive
+ * assistant replies whose word sets overlap by more than 75%. Returns the
+ * warning text to put in the volatile context block, or null.
+ * (An older check for repeated "[Using: tool]" markers was removed: those
+ * markers were never written to history, so it could never fire.)
+ */
+export function detectRepetitionLoop(recent: readonly { role: string; content: string }[]): string | null {
+ const replies = recent.slice(-6).filter((m) => m.role === 'assistant' && m.content.length > 20);
+ if (replies.length < 3) return null;
+ const [a, b, c] = replies.slice(-3).map((m) => normalizeText(m.content));
+ if (!a || !b || !c) return null;
+ return overlap(a, b) > 0.75 && overlap(b, c) > 0.75 ? TEXT_REPETITION_WARNING : null;
+}
diff --git a/src/core/system-prompt.test.ts b/src/core/system-prompt.test.ts
new file mode 100644
index 00000000..f0db3cd8
--- /dev/null
+++ b/src/core/system-prompt.test.ts
@@ -0,0 +1,68 @@
+import { describe, expect, it } from 'vitest';
+import { buildSystemPrompt, environmentSection, type SystemPromptInputs } from './system-prompt.js';
+
+const base: SystemPromptInputs = {
+ identityPrompt: 'You are Mercury.',
+ skillContext: '',
+ botSection: '',
+ programmingSuffix: '',
+ researchSuffix: '',
+ budgetStatus: 'Token budget: 1,000 / 100,000 used today.',
+ budgetUsagePercentage: 1,
+ saverSuffix: '',
+ cwd: '/repo',
+ now: new Date('2026-10-09T18:42:00Z'),
+ platform: 'darwin',
+ timezone: 'UTC',
+ memorySummary: null,
+ toolNames: ['read_file', 'run_command'],
+ github: {},
+};
+
+describe('system prompt assembly', () => {
+ it('starts with the identity and keeps the stable section order', () => {
+ const p = buildSystemPrompt(base);
+ expect(p.startsWith('You are Mercury.')).toBe(true);
+ const order = ['Token budget', 'Environment:', 'Tool Usage Guidelines', 'Second Brain is DISABLED'].map((s) => p.indexOf(s));
+ expect(order.every((v, i) => v > -1 && (i === 0 || v > order[i - 1]))).toBe(true);
+ });
+
+ it('renders the clock at hour resolution so the prefix stays cacheable', () => {
+ const a = environmentSection(new Date('2026-10-09T18:42:00Z'), 'UTC', 'linux', '/w');
+ const b = environmentSection(new Date('2026-10-09T18:57:00Z'), 'UTC', 'linux', '/w');
+ expect(a).toBe(b);
+ expect(a).toContain('about 6 PM');
+ expect(a).toContain('Working directory: /w');
+ });
+
+ it('adds the concision hint above 70% budget and the saver suffix when given', () => {
+ const p = buildSystemPrompt({ ...base, budgetUsagePercentage: 80, saverSuffix: '\n\nSAVER' });
+ expect(p).toContain('Be concise to conserve tokens.');
+ expect(p).toContain('SAVER');
+ expect(buildSystemPrompt(base)).not.toContain('Be concise');
+ });
+
+ it('describes Second Brain when enabled, including the paused state', () => {
+ const on = buildSystemPrompt({ ...base, memorySummary: { total: 276, learningPaused: false } });
+ expect(on).toContain('You have 276 persistent memories');
+ expect(on).not.toContain('PAUSED');
+ const paused = buildSystemPrompt({ ...base, memorySummary: { total: 1, learningPaused: true } });
+ expect(paused).toContain('Learning is currently PAUSED');
+ });
+
+ it('includes the GitHub section only when GitHub tools are registered', () => {
+ expect(buildSystemPrompt(base)).not.toContain('GitHub companion');
+ const p = buildSystemPrompt({
+ ...base,
+ toolNames: ['read_file', 'github_api'],
+ github: { defaultOwner: 'cosmicstack-labs', defaultRepo: 'mercury-agent', username: 'hotheadhacker' },
+ });
+ expect(p).toContain('GitHub companion is active. Default repo: cosmicstack-labs/mercury-agent.');
+ expect(p).toContain("The user's GitHub username is hotheadhacker.");
+ });
+
+ it('includes skills, bots and mode suffixes verbatim', () => {
+ const p = buildSystemPrompt({ ...base, skillContext: 'SKILLS: tweet-notifier', botSection: '\n\nBOTS', programmingSuffix: '\n\nCODE', researchSuffix: '\n\nRESEARCH' });
+ for (const s of ['SKILLS: tweet-notifier', 'BOTS', 'CODE', 'RESEARCH', 'Tweet Notification System Available']) expect(p).toContain(s);
+ });
+});
diff --git a/src/core/system-prompt.ts b/src/core/system-prompt.ts
new file mode 100644
index 00000000..0ead273c
--- /dev/null
+++ b/src/core/system-prompt.ts
@@ -0,0 +1,123 @@
+/**
+ * System prompt assembly, extracted from Agent so it can be read and tested
+ * on its own. The agent gathers the inputs (identity, skills, bots, modes,
+ * budget, memory summary, tool names, GitHub config) and this module turns
+ * them into the prompt text. Everything here is the STABLE prefix of a turn;
+ * per-turn context (retrieved memory, loop warnings) goes into the volatile
+ * system block built in Agent.handleMessage.
+ */
+
+export interface SystemPromptInputs {
+ /** Soul + guardrails + persona, from Identity.getSystemPrompt(). */
+ identityPrompt: string;
+ /** Installed skill summaries, from CapabilityRegistry.getSkillContext(). */
+ skillContext: string;
+ /** Mercury Bots roster section, or '' when bots are not enabled. */
+ botSection: string;
+ programmingSuffix: string;
+ researchSuffix: string;
+ budgetStatus: string;
+ budgetUsagePercentage: number;
+ saverSuffix: string;
+ cwd: string;
+ now: Date;
+ platform: string;
+ timezone: string;
+ /** Second Brain summary, or null when Second Brain is disabled. */
+ memorySummary: { total: number; learningPaused: boolean } | null;
+ toolNames: readonly string[];
+ github: { defaultOwner?: string; defaultRepo?: string; username?: string };
+}
+
+export const GITHUB_TOOL_NAMES = ['create_pr', 'review_pr', 'list_issues', 'create_issue', 'github_api'] as const;
+
+export const TOOL_USAGE_GUIDELINES = `**Tool Usage Guidelines:**
+- Use write_file, create_file, and edit_file tools DIRECTLY to create and modify files. Do NOT create intermediary scripts (Python, bash, Node.js) whose sole purpose is to generate other files — you have native file tools for this.
+- Use run_command for: building, testing, installing dependencies, running the project, git operations, and other system tasks that require a shell.
+- Do NOT use run_command with echo/cat/tee/heredoc to write files. Use write_file or create_file instead.
+- Do NOT create one-time-use helper scripts. If the user asks you to create a file, create it directly with create_file or write_file.
+- When creating multiple files, call create_file or write_file for each one individually. Do not batch them into a script.`;
+
+/** Date at day resolution and time at hour resolution: the prompt is cached as a prefix. */
+export function environmentSection(now: Date, timezone: string, platform: string, cwd: string): string {
+ const dateStr = now.toLocaleDateString('en-US', { weekday: 'long', year: 'numeric', month: 'long', day: 'numeric', timeZone: timezone });
+ // Hour resolution on purpose: a minute-precise clock changed the system
+ // prompt every turn and defeated prompt caching for the ~10K-token
+ // prefix that follows. Tools report exact timestamps when it matters.
+ const timeStr = `about ${now.toLocaleTimeString('en-US', { hour: 'numeric', hour12: true, timeZone: timezone })}`;
+ return `Environment:\n- Date: ${dateStr}, ${timeStr} (${timezone})\n- Platform: ${platform}\n- Working directory: ${cwd}`;
+}
+
+export function memorySection(summary: SystemPromptInputs['memorySummary']): string {
+ if (!summary) return 'Second Brain is DISABLED. Basic long-term memory (text search over facts) is still active.';
+ let text = `Second Brain (SQLite-backed long-term memory) is ENABLED. You have ${summary.total} persistent memories about this user.`;
+ text += `\nMemory types: identity, preference, goal, project, habit, decision, constraint, relationship, episode, reflection.`;
+ text += `\n\nCRITICAL — Memory storage rules:`;
+ text += `\n- ALL persistent user knowledge lives in the Second Brain SQLite database — this is the single source of truth.`;
+ text += `\n- NEVER use create_file, write_file, edit_file, or any file tool to store memories, notes, facts, preferences, or brain data. Files are for code and documents, not for knowledge storage.`;
+ text += `\n- New memories are extracted AUTOMATICALLY after each conversation turn. You do not need to ask the user if they want to save something.`;
+ text += `\n- When the user explicitly asks you to "save/remember/note/keep this," use the save_memory tool to store it directly — no follow-up questions needed.`;
+ text += `\n- When you need to actively recall something beyond auto-injected context (e.g. "do you remember...", "what do I know about..."), use the search_memory tool.`;
+ text += `\n- Relevant memories are auto-injected before each message. You can reference them naturally (e.g. "I remember you prefer TypeScript").`;
+ text += `\n- Users can manage memory with: /memory (overview, search, pause learning, clear).`;
+ if (summary.learningPaused) {
+ text += `\n\nLearning is currently PAUSED — no new memories will be extracted or saved until resumed.`;
+ }
+ return text;
+}
+
+export function tweetNotifierSection(skillContext: string): string {
+ if (!skillContext.includes('tweet-notifier')) return '';
+ return `**Tweet Notification System Available** — The tweet-notifier skill is installed.
+When you need to schedule tweets, manage approvals, or notify founders/supporters:
+1. Use the \`use_skill\` tool to invoke the \`tweet-notifier\` skill for detailed instructions
+2. The skill provides templates for scheduling tweets, notifying founders (via send_message), and alerting supporters (approved Telegram users)
+3. Key tools used by this system: schedule_task (for timing), send_message (for notifications to Telegram), save_memory (for tweet state tracking), search_memory (for checking existing tweets)
+4. Supporters are all approved Telegram users — send_message will reach them
+5. The founder (Optimus Prime) receives notifications via send_message (Telegram)`;
+}
+
+export function githubSection(toolNames: readonly string[], github: SystemPromptInputs['github']): string {
+ if (!GITHUB_TOOL_NAMES.some((t) => toolNames.includes(t))) return '';
+ let hint = 'GitHub companion is active.';
+ if (github.defaultOwner && github.defaultRepo) {
+ hint += ` Default repo: ${github.defaultOwner}/${github.defaultRepo}. Use this when the user doesn't specify a repo.`;
+ }
+ hint += `
+
+Available GitHub tools and when to use them:
+- git_add, git_commit, git_push: LOCAL git operations (stage, commit, push to a remote you have SSH/auth access to). All commits include "Co-authored-by: Mercury <[email protected]>".
+- create_pr: Create a pull request on GitHub. The head branch must already exist on the remote.
+- review_pr: Get PR details and optionally post a review comment.
+- list_issues, create_issue: Browse and file issues.
+- github_api: Raw GitHub API access. IMPORTANT USE CASES:
+ - Push files directly to GitHub via PUT /repos/{owner}/{repo}/contents/{path} when git push fails due to auth. The body must include "message" and "content" (base64-encoded file content). This creates a commit on GitHub with Mercury as co-author.
+ - Delete files via DELETE /repos/{owner}/{repo}/contents/{path} with a "message" and "sha" in the body.
+ - Any other GitHub API operation not covered by the other tools.
+
+When the user asks to "push to GitHub" or "upload files" and git push fails, use github_api with PUT /repos/{owner}/{repo}/contents/{path} to push content directly through the API. This bypasses local git entirely.
+
+Always specify owner and repo parameters on GitHub tools. The user's GitHub username is ${github.username || 'not set'}.'`;
+ return hint;
+}
+
+/** Assemble the stable system prompt. Section order is part of the cache key; keep it stable. */
+export function buildSystemPrompt(i: SystemPromptInputs): string {
+ const sections: string[] = [i.identityPrompt];
+ if (i.skillContext) sections.push(i.skillContext);
+ let prompt = sections.join('\n\n');
+ if (i.botSection) prompt += i.botSection;
+ if (i.programmingSuffix) prompt += i.programmingSuffix;
+ if (i.researchSuffix) prompt += i.researchSuffix;
+ prompt += '\n\n' + i.budgetStatus;
+ if (i.budgetUsagePercentage > 70) prompt += '\nBe concise to conserve tokens.';
+ if (i.saverSuffix) prompt += i.saverSuffix;
+ prompt += '\n\n' + environmentSection(i.now, i.timezone, i.platform, i.cwd);
+ prompt += '\n\n' + TOOL_USAGE_GUIDELINES;
+ prompt += '\n\n' + memorySection(i.memorySummary);
+ const tweet = tweetNotifierSection(i.skillContext);
+ if (tweet) prompt += '\n\n' + tweet;
+ const github = githubSection(i.toolNames, i.github);
+ if (github) prompt += '\n\n' + github;
+ return prompt;
+}