context-paging/conversation-runner.php

<?php
/**
 * Conversation Runner — Thin wrapper around index.php
 * 
 * Manages conversation state and prompts, pipes everything through
 * the context-paging pipeline in index.php.
 * 
 * Usage: php conversation-runner.php
 * 
 * Environment variables (passed to index.php):
 *   API_BASE_URL        — Override endpoint (default: from conversation file)
 *   MAX_CONTEXT_TOKENS  — Override context limit (default: from conversation file)
 * 
 * Note: REDIS_URL, VLLM_URL, etc. are read from .env by index.php
 */

require __DIR__ . '/vendor/autoload.php';

use ContextPaging\TokenCounter;

// -----------------------------------------------------
// Configuration
// -----------------------------------------------------

$conversationFile = __DIR__ . '/conversations/coding-session.json';

// -----------------------------------------------------
// Load conversation state
// -----------------------------------------------------

if (!file_exists($conversationFile)) {
    echo "Error: Conversation file not found: {$conversationFile}\n";
    exit(1);
}

$conversation = json_decode(file_get_contents($conversationFile), true);

if (json_last_error() !== JSON_ERROR_NONE) {
    echo "Error: Invalid JSON in conversation file\n";
    exit(1);
}

// -----------------------------------------------------
// Model-specific settings
// -----------------------------------------------------

$modelConfig = [
    'HuggingFaceTB/SmolLM3-3B' => [
        'system_prompt' => '/no_think',
        'reasoning_parser_workaround' => true,
    ],
];

$currentModel = $conversation['model'];
$modelSettings = $modelConfig[$currentModel] ?? [];

// -----------------------------------------------------
// Prompts for each turn (excerpted for brevity - full list below)
// -----------------------------------------------------

$prompts = require __DIR__ . '/prompts.php';

// -----------------------------------------------------
// Determine next turn
// -----------------------------------------------------

$nextTurn = count($conversation['turns']) + 1;

if (!isset($prompts[$nextTurn])) {
    echo "No more prompts defined. Current turn: {$nextTurn}\n";
    echo "Conversation complete.\n";
    exit(0);
}

$nextPrompt = $prompts[$nextTurn];

// -----------------------------------------------------
// Build messages array
// -----------------------------------------------------

$messages = [];

// Inject model-specific system prompt if configured
if (!empty($modelSettings['system_prompt'])) {
    $messages[] = ['role' => 'system', 'content' => $modelSettings['system_prompt']];
}

// Add conversation history
foreach ($conversation['turns'] as $turn) {
    $messages[] = ['role' => 'user', 'content' => $turn['user']];
    
    $assistantMsg = ['role' => 'assistant'];
    if (isset($turn['assistant'])) {
        $assistantMsg['content'] = $turn['assistant'];
    }
    if (isset($turn['tool_calls'])) {
        $assistantMsg['tool_calls'] = $turn['tool_calls'];
    }
    $messages[] = $assistantMsg;
}

// Add the next prompt
$messages[] = ['role' => 'user', 'content' => $nextPrompt];

// -----------------------------------------------------
// Build payload for index.php
// -----------------------------------------------------

$payload = [
    'messages' => $messages,
    'model' => $conversation['model'],
    'max_tokens' => 8000,
    'temperature' => 0.7,
];

// -----------------------------------------------------
// Show turn info
// -----------------------------------------------------

echo "=== TURN {$nextTurn} ===\n";
echo "User: " . substr($nextPrompt, 0, 100) . "...\n\n";

// Count tokens before
$tokenCounter = new TokenCounter();
$originalTokens = $tokenCounter->contextSize($messages);
echo "Input tokens: {$originalTokens}\n";
echo "Context limit: {$conversation['max_context']}\n\n";

// -----------------------------------------------------
// Call index.php with the payload
// -----------------------------------------------------

$payloadJson = json_encode($payload);
$indexPhp = __DIR__ . '/index.php';

// Set environment variables for index.php
$env = [
    'API_BASE_URL' => $conversation['endpoint'],
    'MAX_CONTEXT_TOKENS' => (string) $conversation['max_context'],
];

// Build the command
$envString = '';
foreach ($env as $key => $value) {
    $envString .= "{$key}=" . escapeshellarg($value) . ' ';
}

$command = "{$envString} php {$indexPhp} " . escapeshellarg($payloadJson) . " 2>&1";

// Execute
$output = shell_exec($command);

if ($output === null) {
    echo "Error: Failed to execute index.php\n";
    exit(1);
}

// -----------------------------------------------------
// Parse response
// -----------------------------------------------------

$response = json_decode($output, true);

if (json_last_error() !== JSON_ERROR_NONE) {
    echo "Error: Invalid JSON response from index.php\n";
    echo "Raw output:\n{$output}\n";
    exit(1);
}

if (isset($response['error'])) {
    echo "ERROR: " . json_encode($response['error'], JSON_PRETTY_PRINT) . "\n";
    exit(1);
}

// Extract response data
$message = $response['choices'][0]['message'] ?? [];
$assistantContent = $message['content'] ?? null;
$toolCalls = $message['tool_calls'] ?? null;
$reasoning = $message['reasoning'] ?? null;
$usage = $response['usage'] ?? [];

// Handle reasoning workaround
if ($assistantContent === null && $reasoning !== null && !empty($modelSettings['reasoning_parser_workaround'])) {
    echo "Note: Using reasoning as content (parser workaround)\n";
    $assistantContent = $reasoning;
    unset($reasoning);
}

// -----------------------------------------------------
// Display results
// -----------------------------------------------------

if ($toolCalls) {
    echo "Tool calls: " . json_encode($toolCalls, JSON_PRETTY_PRINT) . "\n";
}
if ($reasoning) {
    echo "Reasoning: " . substr($reasoning, 0, 200) . "...\n";
}
echo "Assistant: " . substr($assistantContent ?? '(null)', 0, 300) . "...\n\n";

// -----------------------------------------------------
// Usage stats
// -----------------------------------------------------

if ($usage) {
    echo "=== USAGE ===\n";
    echo "Prompt tokens: {$usage['prompt_tokens']}\n";
    echo "Completion tokens: {$usage['completion_tokens']}\n";
    echo "Total tokens: {$usage['total_tokens']}\n";
    echo "Context limit: {$conversation['max_context']}\n";
    echo "Remaining: " . ($conversation['max_context'] - $usage['prompt_tokens']) . "\n\n";
}

// -----------------------------------------------------
// Save turn
// -----------------------------------------------------

$turnData = [
    'turn' => $nextTurn,
    'user' => $nextPrompt,
    'usage' => $usage,
];

if ($assistantContent !== null) {
    $turnData['assistant'] = $assistantContent;
}
if ($toolCalls) {
    $turnData['tool_calls'] = $toolCalls;
}
if ($reasoning) {
    $turnData['reasoning'] = $reasoning;
}

$conversation['turns'][] = $turnData;

file_put_contents($conversationFile, json_encode($conversation, JSON_PRETTY_PRINT | JSON_UNESCAPED_UNICODE));

echo "=== SAVED ===\n";
echo "Turn {$nextTurn} saved.\n";

// -----------------------------------------------------
// Context usage warning
// -----------------------------------------------------

if ($usage) {
    $contextPercent = ($usage['prompt_tokens'] / $conversation['max_context']) * 100;
    echo "Context usage: " . round($contextPercent, 1) . "%\n";
    
    if ($contextPercent > 80) {
        echo "\n*** APPROACHING CONTEXT LIMIT (" . round($contextPercent, 1) . "%) ***\n";
    }
    if ($contextPercent > 95) {
        echo "*** STOP - Context nearly exhausted ***\n";
    }
}