Files
nibiru-framework.com/application/module/ai/plugins/agent.php
stephan 48c839d927 Initial public push: docs cosmos v4 + AI module + framework groundwork
This is the snapshot the production landing site (nibiru-framework.com) is
deployed from. Brings together the recent splash + docs migration to the v4
"Cosmos" design system, the new in-framework AI module, and the framework
groundwork that backs the framework-reference extraction.

What lands:
- docs/: Astro + Starlight site with the v4 dark cosmic palette, GalaxyHero
  canvas constellation, Mission Control chat (wired to /api/oracle →
  api.neuronetz.ai via providers.mjs Ollama), 5-panel MMVC stage
  (Model · AI · Module · Controller · View), translated EN/DE/JA/ES/FR
  content, PWA + sitemap + llms.txt + Umami analytics.
- docs/design-system/: canonical mockup bundle (source/index-v2.html for
  splash, source/docs-system.html + preview/ for docs, SPEC.md, tokens).
- docs/scripts/extraction/framework-reference-v2.md: deep framework
  reference (~1.6k lines, file:line citations, every public factory and
  idiom — basis for the LoRA training corpus.
- application/module/ai/: AI module with chat / embed / RAG / agent
  plugins, plus pdoQuery / httpGet / fileRead tools and Modelfile +
  smoke-test in training/.
- application/module/users/: user / ACL / form-factory traits used as the
  reference plugin pattern for the framework docs.
- application/settings/config/database/: schema + seed migrations
  including the AI module tables (200–203).
- Form factory + autogenerator changes the framework-reference-v2 covers.

Production secrets stay out: docs/.env, settings.production.ini and
ai.production.ini are all gitignored (.example files are in tree).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-08 15:22:18 +02:00

163 lines
4.9 KiB
PHP

<?php
namespace Nibiru\Module\Ai\Plugins;
/**
* Tool-using agent. Runs an LLM in a ReAct-style loop: it reads the
* user's task, picks a tool, runs it, observes the result, repeats —
* until it emits a final answer.
*
* $agent = $ai->agent()->withTools([
* new Plugin\Tools\PdoQuery(),
* new Plugin\Tools\HttpGet(),
* new Plugin\Tools\FileRead(),
* ]);
* echo $agent->run('How many active users do we have?');
*
* Tool-call protocol uses a simple JSON sentinel — no model-specific
* tool-calling APIs needed, so it works on every Ollama model. Models that
* support native tool-calling can be plugged in via a subclass override
* of {@link parseToolCall()}.
*/
class Agent
{
protected \stdClass $cfg;
protected Chat $chat;
/** @var Tool[] */
protected array $tools = [];
/** @var array<int, array{step:int, action:string, observation:string}> */
protected array $trace = [];
public function __construct(\stdClass $cfg, Chat $chat)
{
$this->cfg = $cfg;
$this->chat = $chat;
}
/** @param Tool[] $tools */
public function withTools(array $tools): self
{
$this->tools = $tools;
return $this;
}
/**
* Run the agent against a task. Returns the final answer text.
*/
public function run(string $task): string
{
$maxIter = (int) ($this->cfg->agent_max_iterations ?? 6);
$this->trace = [];
$system = $this->systemPrompt();
$this->chat->reset()->system($system)->user($task);
for ($step = 1; $step <= $maxIter; $step++) {
$reply = $this->chat->complete();
$call = $this->parseToolCall($reply);
if ($call === null) {
// Final answer — no tool call detected.
return $this->stripFinalMarker($reply);
}
$tool = $this->findTool($call['tool']);
if ($tool === null) {
$obs = "ERROR: no tool named \"{$call['tool']}\". Available: "
. implode(', ', array_map(fn($t) => $t->name(), $this->tools));
} else {
try {
$obs = (string) $tool->execute($call['args']);
} catch (\Throwable $e) {
$obs = 'ERROR: ' . $e->getMessage();
}
}
$this->trace[] = ['step' => $step, 'action' => json_encode($call), 'observation' => $obs];
// Feed the observation back into the conversation.
$this->chat->user("Observation:\n" . $obs . "\n\nContinue.");
}
return $this->chat->complete(); // final attempt after max iterations
}
/**
* @return array<int, array{step:int, action:string, observation:string}>
*/
public function trace(): array
{
return $this->trace;
}
// -----------------------------------------------------------------------
// Internals
// -----------------------------------------------------------------------
protected function systemPrompt(): string
{
$toolBlock = '';
foreach ($this->tools as $t) {
$toolBlock .= $t->asPrompt() . "\n\n";
}
return <<<PROMPT
You are a Nibiru AI agent. You answer the user's task by either calling a tool
or producing a final answer.
When you need a tool, output a single fenced JSON block on its own line:
```tool
{"tool": "<name>", "args": { ... }}
```
After you receive the observation, decide whether to call another tool or to
produce a final answer. When done, write the final answer prefixed with
"FINAL:" on its own paragraph. Do not call a tool and produce a final answer
in the same turn.
Available tools:
$toolBlock
Be concise. Prefer real tool data over guesses. If a tool errors twice, give a
final answer that says you couldn't find out, and why.
PROMPT;
}
/**
* Look for a ```tool { ... }``` JSON block. Returns null if absent.
*
* @return array{tool:string,args:array}|null
*/
protected function parseToolCall(string $reply): ?array
{
if (preg_match('/```(?:tool|json)\s*\n(.*?)\n```/s', $reply, $m)) {
$decoded = json_decode($m[1], true);
if (is_array($decoded) && isset($decoded['tool'])) {
return [
'tool' => (string) $decoded['tool'],
'args' => is_array($decoded['args'] ?? null) ? $decoded['args'] : [],
];
}
}
// Fallback: explicit final marker
if (preg_match('/^FINAL:\s*/m', $reply)) return null;
// No tool call detected — treat as final.
return null;
}
protected function stripFinalMarker(string $reply): string
{
return preg_replace('/^FINAL:\s*/m', '', $reply) ?? $reply;
}
protected function findTool(string $name): ?Tool
{
foreach ($this->tools as $t) {
if ($t->name() === $name) return $t;
}
return null;
}
}