Constants Reference

Source: Amp Code v0.0.1769212917 Master reference for all magic numbers and configuration values.

---

Table of Contents

1. Token & Context Constants
2. Handoff Constants
3. Tool Limits
4. Retry & Timeout Constants
5. Streaming Constants
6. Course Correction Constants
7. Cache Control Constants
8. File System Constants

---

1. Token & Context Constants

Primary Token Limits

Constant	Value	Purpose
COMPACTION_THRESHOLD	100,000	Token count triggering summary compaction
MAX_OUTPUT_TOKENS	32,000	Default max output tokens
SUBAGENT_MAX_TOKENS	8,000	Max output for subagents
CHARS_PER_TOKEN	4	Fallback token estimation ratio
BYTES_PER_TOKEN	3.5	Bytes per token for file estimation

Token Estimation

// Characters per token (fallback)
const CHARS_PER_TOKEN = 4;

function estimateTokens(text: string): number {
  return Math.ceil(text.length / CHARS_PER_TOKEN);
}

// File size estimation (for token budgeting)
const BYTES_PER_TOKEN = 3.5;
const MAX_FILE_SIZE_FOR_ESTIMATION = 32768;  // 32KB (conservative for budgeting)

function estimateFileTokens(sizeBytes: number): number {
  const capped = Math.min(sizeBytes, MAX_FILE_SIZE_FOR_ESTIMATION);
  return Math.ceil(capped / BYTES_PER_TOKEN);
}

// Note: Read tool uses a separate limit of 65,536 bytes (64KB)
// The 32KB cap here is intentionally smaller for conservative budgeting

Context Window by Model

Model Family	Context Window	Max Output	Effective Input
Claude Sonnet 4/4.5	1,000,000	32,000	968,000
Claude Opus 4/4.1/4.5	200,000	32,000	168,000
Claude Haiku 4.5	200,000	64,000	136,000
GPT-5/5.1/5.2	400,000	128,000	272,000
Gemini 2.5/3	1,048,576	65,535	983,041

---

2. Handoff Constants

Constant	Value	Purpose
HANDOFF_TOKEN_BUDGET	25,000	Max tokens for handoff context
MAX_HANDOFF_FILES	10	Maximum files in handoff
MAX_FILE_SIZE	32,768	Max file size in bytes
HANDOFF_TIMEOUT	30,000	Handoff operation timeout (ms)
FALLBACK_MODEL	"vertexai/gemini-2.5-flash"	Model for context overflow

Budget Calculation

const HANDOFF_TOKEN_BUDGET = 25000;
const MAX_HANDOFF_FILES = 10;

// Effective per-file budget
const perFileBudget = HANDOFF_TOKEN_BUDGET / MAX_HANDOFF_FILES; // ~2,500 tokens

---

3. Tool Limits

Output Truncation

Tool	Limit	Purpose
Bash	50,000 chars	Max bash output
Grep	100 matches	Max matches returned
Glob	1,000 files	Max files returned
MCP	65,536 bytes	Max MCP results
Read	2,000 lines	Hard cap on returned lines (Read default is 500)
Read	2,000 chars/line	Max chars per line

Tool Timeout

Constant	Value (ms)	Purpose
DEFAULT_TOOL_TIMEOUT	120,000	Default tool timeout
BASH_TIMEOUT	600,000	Bash command timeout
MCP_CALL_TIMEOUT	60,000	MCP tool call timeout

Truncation Implementation

const BASH_OUTPUT_LIMIT = 50000;

function truncateBashOutput(output: string): string {
  if (output.length > BASH_OUTPUT_LIMIT) {
    return output.slice(0, BASH_OUTPUT_LIMIT) +
      `\n\n[Output truncated at ${BASH_OUTPUT_LIMIT} characters]`;
  }
  return output;
}

---

4. Retry & Timeout Constants

SDK-Level Retry

Constant	Value	Purpose
MAX_RETRIES	2	Maximum retry attempts
ANTHROPIC_TIMEOUT	600,000ms	Anthropic request timeout
OPENAI_TIMEOUT	60,000ms	OpenAI request timeout
GOOGLE_TIMEOUT	60,000ms	Google request timeout
BASE_RETRY_DELAY	500ms	Initial retry delay
MAX_RETRY_DELAY	8,000ms	Maximum retry delay

UI-Level Auto-Retry

Constant	Value	Purpose
BASE_RETRY_SECONDS	5	Initial auto-retry delay
MAX_RETRY_SECONDS	60	Maximum auto-retry delay
MAX_AUTO_RETRIES	5	Maximum auto-retry attempts

Retry Progression

// UI Auto-Retry progression
class RetryHandler {
  static BASE_RETRY_SECONDS = 5;
  static MAX_RETRY_SECONDS = 60;
  static MAX_AUTO_RETRIES = 5;

  getRetryDelaySeconds(attempt: number): number | undefined {
    if (attempt >= RetryHandler.MAX_AUTO_RETRIES) return undefined;
    const delay = RetryHandler.BASE_RETRY_SECONDS * Math.pow(2, attempt);
    return Math.min(delay, RetryHandler.MAX_RETRY_SECONDS);
  }
}

// Progression: 5s → 10s → 20s → 40s → 60s (capped)

---

5. Streaming Constants

SSE Reconnection

Constant	Value	Purpose
initialReconnectionDelay	1,000ms	Initial reconnect delay
maxReconnectionDelay	30,000ms	Maximum reconnect delay
reconnectionDelayGrowFactor	1.5	Exponential growth factor
maxRetries	2	Maximum reconnection attempts

SSE Configuration

const SSE_RECONNECTION_OPTIONS = {
  initialReconnectionDelay: 1000,    // 1 second
  maxReconnectionDelay: 30000,       // 30 seconds
  reconnectionDelayGrowFactor: 1.5,  // 1.5x each retry
  maxRetries: 2                       // 2 retries max
};

// Progression: 1s → 1.5s → 2.25s → ... → 30s (capped)

---

6. Course Correction Constants

Constant	Value	Purpose
MIN_TOOL_CALLS	5	Minimum tool calls before check
REQUIRES_FILE_EDIT	true	Must have file edits
ASSESSMENT_TIMEOUT	1,000ms	Wait time for assessment
TEMPERATURE	0.1	Very low for deterministic decisions
MAX_OUTPUT_TOKENS	1,024	Short responses only

Trigger Logic

function shouldRunCourseCorrection(messages: Message[]): boolean {
  let toolCount = 0;
  let hasFileEdits = false;

  for (const msg of getMessagesSinceLastUser(messages)) {
    for (const content of msg.content) {
      if (content.type === "tool_use") {
        toolCount++;
        if (content.name === "edit_file" || content.name === "create_file") {
          hasFileEdits = true;
        }
      }
    }
  }

  // Trigger: 5+ tools AND at least one file edit
  return toolCount >= 5 && hasFileEdits;
}

---

7. Cache Control Constants

Constant	Value	Purpose
LAST_MESSAGE_TTL	"5m"	Cache TTL for last message
BOUNDARY_TTL	"1h"	Cache TTL for message boundaries
BOUNDARY_INTERVAL	10	Cache marker every N messages

Cache Implementation

function applyCacheControl(messages: Message[]): Message[] {
  if (messages.length === 0) return messages;

  const lastIndex = messages.length - 1;
  const boundaryIndex = Math.floor((messages.length - 1) / 10) * 10;

  return messages.map((msg, i) => {
    if (i === lastIndex) {
      // Last message: 5 minute cache
      return { ...msg, content: addCacheTTL(msg.content, "5m") };
    } else if (i === boundaryIndex - 1 && boundaryIndex > 0) {
      // 10th message boundary: 1 hour cache
      return { ...msg, content: addCacheTTL(msg.content, "1h") };
    }
    return msg;
  });
}

---

8. File System Constants

Constant	Value	Purpose
MAX_FILE_SIZE (Read tool)	65,536	Hard limit for Read tool (64KB)
MAX_FILE_SIZE (estimation)	32,768	Cap for token budgeting (32KB)
MAX_READ_LINES	2,000	Default lines to read
MAX_LINE_LENGTH	2,000	Max chars per line
MAX_GLOB_RESULTS	1,000	Max files from glob
MAX_GREP_RESULTS	100	Max grep matches

Note: Two different MAX_FILE_SIZE values exist intentionally. The Read tool accepts files up to 64KB, but token estimation uses a conservative 32KB cap to prevent single files from dominating context budgets.

Guarded File Patterns

const GUARDED_FILE_PATTERNS = [
  "**/.env",
  "**/.env.*",
  "**/.envrc",
  "**/secrets.*",
  "**/credentials.*",
  "**/*_rsa",
  "**/*_dsa",
  "**/*_ed25519",
  "**/*.pem",
  "**/*.key",
  "**/*.p12",
  "**/*.pfx"
];

---

Summary Tables

All Token Constants

Constant	Value	Category
Compaction threshold	100,000	Context
Max output tokens	32,000	Output
Subagent max tokens	8,000	Subagent
Handoff token budget	25,000	Handoff
Chars per token	4	Estimation
Bytes per token	3.5	Estimation

All Timeout Constants (ms)

Constant	Value	Category
Anthropic timeout	600,000	SDK
OpenAI timeout	60,000	SDK
Google timeout	60,000	SDK
Tool timeout	120,000	Tool
Bash timeout	600,000	Tool
MCP timeout	60,000	MCP
Handoff timeout	30,000	Handoff

All Retry Constants

Constant	Value	Category
SDK max retries	2	SDK
UI max auto-retries	5	UI
SSE max retries	2	Streaming

---

Constants extracted from v0.0.1769212917