Files
NeuroSploit/neurosploit-rs/crates/harness/src/models.rs
T
CyberSecurityUP d957429c09 feat(models): add Azure OpenAI provider + GOOGLE_API_KEY alias for Gemini
Resolves the only two open issues that still apply to the Rust build:
- #21 Azure OpenAI: new `azure` provider (OpenAI-compatible). Endpoint comes
  from AZURE_OPENAI_ENDPOINT, api-version from AZURE_OPENAI_API_VERSION
  (default 2024-10-21); the model name is the Azure deployment; auth uses the
  `api-key` header instead of Bearer. Use `--model azure:<deployment>`.
- #25 Gemini key confusion: GEMINI_API_KEY now also accepts GOOGLE_API_KEY
  (Google's standard env var) as an alias; local providers (ollama/litellm)
  require no key. .env.example documents both.

Kept under the v3.5.2 line (additive provider support).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-26 14:17:25 -03:00

471 lines
22 KiB
Rust

use anyhow::{anyhow, Result};
use serde::Serialize;
use std::process::Stdio;
use std::time::Duration;
use tokio::io::{AsyncBufReadExt, AsyncWriteExt, BufReader};
use tokio::process::Command;
/// A model provider exposing an OpenAI-compatible `/chat/completions` endpoint.
#[derive(Clone, Debug, Serialize)]
pub struct Provider {
pub key: &'static str,
pub label: &'static str,
pub base_url: &'static str,
pub env_key: &'static str,
/// "cli" (also drivable by an agentic CLI) | "api"
pub kind: &'static str,
pub models: Vec<&'static str>,
}
/// The full provider registry. Every entry speaks the OpenAI chat schema
/// (Anthropic, xAI, NVIDIA NIM, DeepSeek, Mistral, Qwen, Groq, Together,
/// OpenRouter, Gemini-compat, Ollama).
pub fn providers() -> Vec<Provider> {
vec![
Provider { key: "anthropic", label: "Anthropic Claude", base_url: "https://api.anthropic.com/v1", env_key: "ANTHROPIC_API_KEY", kind: "cli",
models: vec!["claude-opus-4-8", "claude-sonnet-4-6", "claude-haiku-4-5"] },
Provider { key: "openai", label: "OpenAI (ChatGPT)", base_url: "https://api.openai.com/v1", env_key: "OPENAI_API_KEY", kind: "cli",
models: vec!["gpt-5.5", "gpt-5.4", "gpt-5.4-mini", "gpt-5.3-codex", "gpt-5.2", "gpt-5.1", "gpt-5.1-codex", "o4"] },
Provider { key: "xai", label: "xAI Grok", base_url: "https://api.x.ai/v1", env_key: "XAI_API_KEY", kind: "cli",
models: vec!["grok-4", "grok-4-fast"] },
Provider { key: "gemini", label: "Google Gemini", base_url: "https://generativelanguage.googleapis.com/v1beta/openai", env_key: "GEMINI_API_KEY", kind: "cli",
models: vec!["gemini-3-pro", "gemini-2.5-pro", "gemini-2.5-flash"] },
Provider { key: "nvidia_nim", label: "NVIDIA NIM", base_url: "https://integrate.api.nvidia.com/v1", env_key: "NVIDIA_NIM_API_KEY", kind: "api",
models: vec!["nvidia/llama-3.3-nemotron-super-49b-v1", "deepseek-ai/deepseek-r1", "qwen/qwen2.5-coder-32b-instruct"] },
Provider { key: "deepseek", label: "DeepSeek", base_url: "https://api.deepseek.com/v1", env_key: "DEEPSEEK_API_KEY", kind: "api",
models: vec!["deepseek-reasoner", "deepseek-chat"] },
Provider { key: "mistral", label: "Mistral", base_url: "https://api.mistral.ai/v1", env_key: "MISTRAL_API_KEY", kind: "api",
models: vec!["mistral-large-latest", "codestral-latest"] },
Provider { key: "qwen", label: "Qwen (DashScope)", base_url: "https://dashscope-intl.aliyuncs.com/compatible-mode/v1", env_key: "DASHSCOPE_API_KEY", kind: "api",
models: vec!["qwen-max", "qwen2.5-coder-32b-instruct", "qwq-plus"] },
Provider { key: "groq", label: "Groq", base_url: "https://api.groq.com/openai/v1", env_key: "GROQ_API_KEY", kind: "api",
models: vec!["llama-3.3-70b-versatile", "qwen-2.5-coder-32b"] },
Provider { key: "together", label: "Together AI", base_url: "https://api.together.xyz/v1", env_key: "TOGETHER_API_KEY", kind: "api",
models: vec!["Qwen/Qwen2.5-Coder-32B-Instruct", "deepseek-ai/DeepSeek-R1", "meta-llama/Llama-3.3-70B-Instruct-Turbo"] },
// LiteLLM proxy (OpenAI-compatible). Point at your gateway with
// LITELLM_BASE_URL (default http://localhost:4000/v1); key = LITELLM_API_KEY.
// Use `litellm:<any-model-the-proxy-routes>` — model names pass through.
Provider { key: "litellm", label: "LiteLLM (proxy)", base_url: "http://localhost:4000/v1", env_key: "LITELLM_API_KEY", kind: "api",
models: vec!["gpt-4o", "claude-3-7-sonnet", "gemini/gemini-2.5-pro"] },
Provider { key: "openrouter", label: "OpenRouter", base_url: "https://openrouter.ai/api/v1", env_key: "OPENROUTER_API_KEY", kind: "api",
models: vec!["anthropic/claude-opus-4-8", "qwen/qwen-2.5-coder-32b-instruct", "deepseek/deepseek-r1", "meta-llama/llama-3.3-70b-instruct"] },
// Azure OpenAI (OpenAI-compatible). Set AZURE_OPENAI_ENDPOINT (e.g.
// https://<resource>.openai.azure.com), optionally AZURE_OPENAI_API_VERSION
// (default 2024-10-21), and use `azure:<your-deployment-name>` as the model.
// base_url is resolved from the endpoint at call time; auth uses an api-key header.
Provider { key: "azure", label: "Azure OpenAI", base_url: "", env_key: "AZURE_OPENAI_API_KEY", kind: "api",
models: vec!["gpt-4o", "gpt-4o-mini", "gpt-5.1", "o4-mini"] },
Provider { key: "ollama", label: "Ollama (local)", base_url: "http://localhost:11434/v1", env_key: "OLLAMA_API_KEY", kind: "api",
models: vec!["qwen2.5-coder:32b", "qwq:32b", "deepseek-r1:32b", "llama3.3:70b"] },
]
}
pub fn provider_for(key: &str) -> Option<Provider> {
providers().into_iter().find(|p| p.key == key)
}
/// Resolve a provider's API key from the environment, honoring common aliases.
/// For Gemini we also accept `GOOGLE_API_KEY` (Google's standard env var name)
/// when `GEMINI_API_KEY` is unset.
fn resolve_key(p: &Provider) -> String {
let mut k = std::env::var(p.env_key).unwrap_or_default();
if k.is_empty() && p.key == "gemini" {
k = std::env::var("GOOGLE_API_KEY").unwrap_or_default();
}
k
}
/// A `provider:model` selection.
#[derive(Clone, Debug)]
pub struct ModelRef {
pub provider: String,
pub model: String,
}
impl ModelRef {
pub fn parse(s: &str) -> ModelRef {
match s.split_once(':') {
Some((p, m)) => ModelRef { provider: p.to_string(), model: m.to_string() },
None => ModelRef { provider: "anthropic".into(), model: s.to_string() },
}
}
pub fn label(&self) -> String {
format!("{}:{}", self.provider, self.model)
}
}
/// OpenAI-compatible chat client shared across the model pool.
#[derive(Clone)]
pub struct ChatClient {
http: reqwest::Client,
}
impl ChatClient {
pub fn new() -> Self {
let http = reqwest::Client::builder()
.timeout(Duration::from_secs(120))
.build()
.unwrap_or_else(|_| reqwest::Client::new());
ChatClient { http }
}
/// One chat completion. Errors (missing key, network, non-2xx) propagate so
/// the pool can fail over to the next candidate model.
pub async fn chat(&self, m: &ModelRef, system: &str, user: &str) -> Result<String> {
let p = provider_for(&m.provider)
.ok_or_else(|| anyhow!("unknown provider '{}'", m.provider))?;
let key = resolve_key(&p);
if key.is_empty() && p.key != "ollama" && p.key != "litellm" {
let hint = if p.key == "gemini" { format!("{} (or GOOGLE_API_KEY)", p.env_key) } else { p.env_key.to_string() };
return Err(anyhow!("no API key ({}) for provider '{}'", hint, p.key));
}
// Azure OpenAI uses a per-resource endpoint + deployment + api-version,
// and authenticates with an `api-key` header instead of Bearer.
let azure = p.key == "azure";
let url = if azure {
let endpoint = std::env::var("AZURE_OPENAI_ENDPOINT").unwrap_or_default();
if endpoint.is_empty() {
return Err(anyhow!("set AZURE_OPENAI_ENDPOINT (e.g. https://<resource>.openai.azure.com) for the azure provider"));
}
let ver = std::env::var("AZURE_OPENAI_API_VERSION").unwrap_or_else(|_| "2024-10-21".to_string());
// `model` is the Azure DEPLOYMENT name (use `azure:<deployment>`).
format!("{}/openai/deployments/{}/chat/completions?api-version={}",
endpoint.trim_end_matches('/'), m.model, ver)
} else {
// Allow an env base-URL override (LiteLLM gateway, self-hosted proxies, …).
let base = match p.key {
"litellm" => std::env::var("LITELLM_BASE_URL").unwrap_or_else(|_| p.base_url.to_string()),
"ollama" => std::env::var("OLLAMA_BASE_URL").unwrap_or_else(|_| p.base_url.to_string()),
_ => p.base_url.to_string(),
};
format!("{}/chat/completions", base.trim_end_matches('/'))
};
let body = serde_json::json!({
"model": m.model,
"max_tokens": 4096,
"temperature": 0.2,
"messages": [
{"role": "system", "content": system},
{"role": "user", "content": user}
]
});
let mut req = self.http.post(&url).json(&body);
if !key.is_empty() {
if azure { req = req.header("api-key", &key); } else { req = req.bearer_auth(&key); }
}
let resp = req.send().await?;
let status = resp.status();
let text = resp.text().await.unwrap_or_default();
if !status.is_success() {
return Err(anyhow!("{} returned {}: {}", p.key, status, truncate(&text, 200)));
}
let v: serde_json::Value = serde_json::from_str(&text)?;
let content = v["choices"][0]["message"]["content"]
.as_str()
.ok_or_else(|| anyhow!("no content in response"))?;
Ok(content.to_string())
}
}
impl ChatClient {
/// Complete via a locally-installed **agentic CLI subscription** (Claude
/// Code / Codex / Grok / Gemini) instead of an API key. This uses the user's
/// logged-in subscription, so no provider key is required.
///
/// When `mcp_config` is set (a path to an `.mcp.json`), Claude/Codex run with
/// the MCP servers enabled and tool autonomy, so agents can actually drive
/// **Playwright** (browse, execute JS, screenshot) during execution.
pub async fn chat_cli(
&self,
label: &str,
provider: &str,
model: &str,
system: &str,
user: &str,
mcp_config: Option<&str>,
progress: Option<tokio::sync::mpsc::Sender<String>>,
) -> Result<String> {
let bin = cli_binary_for(provider)
.ok_or_else(|| anyhow!("no CLI/subscription backend for provider '{}'", provider))?;
let prompt = format!("{system}\n\n{user}");
// Claude Code can stream structured events (tools, commands, files) which
// we surface live as a categorized activity feed, attributed to `label`.
if bin == "claude" {
return self.chat_claude_stream(label, model, &prompt, mcp_config, progress).await;
}
let mut cmd = Command::new(bin);
match bin {
// Codex non-interactive exec (uses the ChatGPT/Codex login), prompt on stdin.
"codex" => {
cmd.arg("exec").arg("--model").arg(model)
.arg("--dangerously-bypass-approvals-and-sandbox");
if let Some(mcp) = mcp_config {
cmd.arg("--config").arg(format!("mcp_config_file={mcp}"));
}
cmd.arg("-");
}
// Google Gemini CLI (uses the Gemini subscription login).
"gemini" => {
cmd.arg("-m").arg(model);
}
// Grok CLI, prompt on stdin (best-effort flags).
"grok" => {
cmd.arg("--model").arg(model);
}
_ => {}
}
cmd.stdin(Stdio::piped()).stdout(Stdio::piped()).stderr(Stdio::piped()).kill_on_drop(true);
let mut child = cmd.spawn().map_err(|e| anyhow!("spawn {} failed: {}", bin, e))?;
if let Some(mut stdin) = child.stdin.take() {
stdin.write_all(prompt.as_bytes()).await?;
// Drop closes stdin so the CLI processes the prompt and exits.
}
// Cap a single agentic CLI turn so a stuck tool-loop can't hang the run.
let out = match tokio::time::timeout(Duration::from_secs(600), child.wait_with_output()).await {
Ok(r) => r?,
Err(_) => return Err(anyhow!("{} subscription CLI timed out after 600s", bin)),
};
let stdout = String::from_utf8_lossy(&out.stdout).trim().to_string();
let stderr = String::from_utf8_lossy(&out.stderr);
if !out.status.success() {
// The CLI often writes the real reason (rate limit, auth) to stdout,
// not stderr — surface both plus the exit code so the error isn't blank.
let detail = if !stderr.trim().is_empty() {
stderr.trim().to_string()
} else if !stdout.is_empty() {
stdout.clone()
} else {
"no output".to_string()
};
return Err(anyhow!(
"{} subscription CLI exit {}: {}",
bin,
out.status.code().map(|c| c.to_string()).unwrap_or_else(|| "signal".into()),
truncate(&detail, 240)
));
}
if stdout.is_empty() {
return Err(anyhow!("{} subscription CLI returned empty output", bin));
}
Ok(stdout)
}
/// Drive Claude Code with `--output-format stream-json` and surface its
/// activity as a live, categorized feed (states, tools, commands, files).
/// Tagged events are sent to `progress`; the final assistant text is returned.
async fn chat_claude_stream(
&self,
label: &str,
model: &str,
prompt: &str,
mcp_config: Option<&str>,
progress: Option<tokio::sync::mpsc::Sender<String>>,
) -> Result<String> {
let mut cmd = Command::new("claude");
cmd.arg("-p").arg("--model").arg(model)
.arg("--output-format").arg("stream-json").arg("--verbose")
.arg("--dangerously-skip-permissions")
.env("IS_SANDBOX", "1");
if let Some(mcp) = mcp_config {
cmd.arg("--mcp-config").arg(mcp);
}
cmd.stdin(Stdio::piped()).stdout(Stdio::piped()).stderr(Stdio::piped()).kill_on_drop(true);
let mut child = cmd.spawn().map_err(|e| anyhow!("spawn claude failed: {e}"))?;
if let Some(mut stdin) = child.stdin.take() {
stdin.write_all(prompt.as_bytes()).await?;
}
let stdout = child.stdout.take().ok_or_else(|| anyhow!("no stdout"))?;
let mut lines = BufReader::new(stdout).lines();
// Tag every streamed event with the agent label so the feed is attributable.
let lbl = if label.is_empty() { String::new() } else { format!("@{label} ") };
let emit = |s: String| {
if let Some(tx) = &progress {
let _ = tx.try_send(format!("{lbl}{s}"));
}
};
let mut result = String::new();
let mut had_err = String::new();
let read = async {
while let Ok(Some(line)) = lines.next_line().await {
let Ok(v) = serde_json::from_str::<serde_json::Value>(&line) else { continue };
match v.get("type").and_then(|t| t.as_str()) {
Some("assistant") => {
if let Some(content) = v.pointer("/message/content").and_then(|c| c.as_array()) {
for blk in content {
match blk.get("type").and_then(|t| t.as_str()) {
Some("text") => {
if let Some(t) = blk.get("text").and_then(|x| x.as_str()) {
let t = t.trim();
if !t.is_empty() {
emit(format!("ai: {}", truncate(t, 240)));
}
}
}
Some("tool_use") => {
let name = blk.get("name").and_then(|x| x.as_str()).unwrap_or("tool");
let input = blk.get("input");
emit(tool_event(name, input));
}
_ => {}
}
}
}
}
Some("result") => {
if let Some(r) = v.get("result").and_then(|x| x.as_str()) {
result = r.to_string();
}
// Token/cost telemetry from the final result event.
let ti = v.pointer("/usage/input_tokens").and_then(|x| x.as_u64());
let to = v.pointer("/usage/output_tokens").and_then(|x| x.as_u64());
let cost = v.get("total_cost_usd").and_then(|x| x.as_f64());
if ti.is_some() || to.is_some() || cost.is_some() {
emit(format!("tokens: in={} out={} cost=${:.4}",
ti.unwrap_or(0), to.unwrap_or(0), cost.unwrap_or(0.0)));
}
if v.get("is_error").and_then(|x| x.as_bool()).unwrap_or(false) {
had_err = v.get("result").and_then(|x| x.as_str()).unwrap_or("error").to_string();
}
}
_ => {}
}
}
};
// Bound the whole streamed turn.
if tokio::time::timeout(Duration::from_secs(900), read).await.is_err() {
return Err(anyhow!("claude stream timed out after 900s"));
}
let _ = child.wait().await;
if !had_err.is_empty() && result.is_empty() {
return Err(anyhow!("claude: {}", truncate(&had_err, 240)));
}
if result.is_empty() {
return Err(anyhow!("claude stream produced no result"));
}
Ok(result)
}
}
/// Categorise a Claude tool_use block into a tagged activity-feed event.
fn tool_event(name: &str, input: Option<&serde_json::Value>) -> String {
let s = |k: &str| input.and_then(|i| i.get(k)).and_then(|x| x.as_str()).unwrap_or("");
match name {
"Bash" => {
let c = s("command");
let danger = c.contains("rm -rf") || c.contains("mkfs") || c.contains(":(){")
|| c.contains("dd if=") || c.contains("> /dev/");
format!("{}: {}", if danger { "danger" } else { "exec" }, truncate(c, 200))
}
"Read" => format!("read: {}", s("file_path")),
"Write" | "Edit" => format!("edit: {}", s("file_path")),
"Grep" => format!("tool: grep {}", truncate(s("pattern"), 80)),
"Glob" => format!("tool: glob {}", truncate(s("pattern"), 80)),
"WebFetch" => format!("net: fetch {}", s("url")),
n if n.contains("playwright") || n.contains("browser") => {
let url = s("url");
format!("net: browser {}{}", n.rsplit('_').next().unwrap_or(n), if url.is_empty() { String::new() } else { format!(" {url}") })
}
other => format!("tool: {other}"),
}
}
/// Map a provider to its local agentic CLI binary (subscription backend).
pub fn cli_binary_for(provider: &str) -> Option<&'static str> {
match provider {
"anthropic" => Some("claude"),
"openai" => Some("codex"),
"xai" => Some("grok"),
"gemini" => Some("gemini"),
_ => None,
}
}
/// Is `name` an executable found on PATH?
pub fn binary_in_path(name: &str) -> bool {
std::env::var_os("PATH")
.map(|path| std::env::split_paths(&path).any(|dir| dir.join(name).is_file()))
.unwrap_or(false)
}
/// Which subscription CLI backends are installed locally.
pub fn installed_cli_backends() -> Vec<&'static str> {
["claude", "codex", "grok", "gemini"].into_iter().filter(|b| binary_in_path(b)).collect()
}
/// Does this provider's agentic CLI accept a Playwright MCP config?
/// Claude Code and Codex do; Gemini/Grok CLIs don't take an MCP-config flag, so
/// they fall back to their own built-in tools.
pub fn mcp_supported(provider: &str) -> bool {
matches!(provider, "anthropic" | "openai")
}
/// Best-effort ensure the Playwright MCP server is available locally. Requires
/// `npx`; pre-warms `@playwright/mcp` so the first agent call isn't a cold start.
/// Returns Err with a clear reason when it can't be provisioned (caller then
/// degrades to built-in tools).
pub fn ensure_playwright_mcp() -> Result<()> {
if !binary_in_path("npx") {
return Err(anyhow!("npx (Node.js) not found — install Node to use Playwright MCP"));
}
// `npx -y @playwright/mcp@latest --help` installs the package into the npx
// cache on first run; ignore non-zero exit (some versions lack --help) as long
// as the package resolves.
let out = std::process::Command::new("npx")
.args(["-y", "@playwright/mcp@latest", "--help"])
.stdout(Stdio::null())
.stderr(Stdio::null())
.status();
match out {
Ok(_) => Ok(()),
Err(e) => Err(anyhow!("could not provision @playwright/mcp via npx: {e}")),
}
}
/// Write an `.mcp.json` into `dir` (Playwright by default) and return its path,
/// so the agentic CLI can drive a real browser during execution. If
/// `extra_servers` points at a JSON file shaped like `{ "mcpServers": {...} }`
/// (or just `{...}` of servers), those servers are MERGED in — letting users
/// plug additional MCP tools into the pipeline to potentiate testing.
pub fn write_mcp_config(dir: &std::path::Path, extra_servers: Option<&std::path::Path>) -> std::io::Result<std::path::PathBuf> {
std::fs::create_dir_all(dir)?;
let mut servers = serde_json::json!({
"playwright": { "command": "npx", "args": ["-y", "@playwright/mcp@latest", "--headless", "--isolated"] }
});
if let Some(extra) = extra_servers {
if let Ok(txt) = std::fs::read_to_string(extra) {
if let Ok(v) = serde_json::from_str::<serde_json::Value>(&txt) {
let add = v.get("mcpServers").cloned().unwrap_or(v);
if let (Some(dst), Some(src)) = (servers.as_object_mut(), add.as_object()) {
for (k, val) in src {
dst.insert(k.clone(), val.clone());
}
}
}
}
}
let cfg = serde_json::json!({ "mcpServers": servers });
let path = dir.join(".mcp.json");
std::fs::write(&path, serde_json::to_string_pretty(&cfg).unwrap_or_default())?;
Ok(path)
}
impl Default for ChatClient {
fn default() -> Self {
Self::new()
}
}
fn truncate(s: &str, n: usize) -> String {
// Truncate by CHARACTERS, never bytes — slicing `&s[..n]` panics when `n`
// lands inside a multi-byte char (e.g. '—'). That panic was crashing agent
// tasks and silently dropping their findings.
if s.chars().count() <= n {
s.to_string()
} else {
format!("{}…", s.chars().take(n).collect::<String>())
}
}