Files
optimclaw/src/tools/builder/core.rs
T
33ef0a6ea5 fix: security hardening across all layers (#35)
* fix: comprehensive security hardening across all layers

Critical:
- Replace --dangerously-skip-permissions with explicit tool allowlist
  via settings.json (Claude Code bridge)
- Constant-time token comparison (subtle crate) in web auth and
  orchestrator auth to prevent timing attacks

High:
- Revoke tokens and clean up handles on container creation failure
- Drop SETUID/SETGID capabilities from containers (keep only CHOWN)
- Disable redirect following in HTTP tool and WASM wrapper (SSRF)
- Reject URL userinfo (@) in WASM allowlist parser (host confusion)
- Fix binary body bypassing leak detection (from_utf8 -> from_utf8_lossy)
- Protect identity files from LLM overwrites (prompt injection defense)
- Prevent tool shadowing: built-in tools cannot be replaced dynamically
- User-scoped job APIs: list/detail/cancel/restart/prompt/events/files
- CORS restricted to localhost origins, WebSocket origin validation
- Sandbox shell fail-closed: no silent fallback to unsandboxed execution
- Scrub secrets from log broadcaster before SSE broadcast
- XSS sanitization on rendered markdown in web UI
- WASM epoch ticker thread so timeout deadlines actually fire

Medium:
- Cap state transition history at 200 entries
- SSE/WebSocket connection limit (100 max)
- Request body size limit (1MB)
- Response body size limit enforcement in WASM HTTP
- UTF-8 safe string truncation (routine engine, shell tool)
- Fix PolicyAction::Sanitize to actually run the sanitizer
- TOCTOU fix in scheduler and context manager (hold write lock)
- Project file serving moved behind auth
- Path traversal guard on project_id
- Session file permissions set to 0600 on unix
- AtomicUsize for routine running_count (panic-safe)
- Completion detection hardened against false positives and tool injection
- Tool output no longer drives job completion (only LLM response)

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address security review findings across all layers

- Fix path traversal sandbox bypass via lexical normalization (file.rs)
- Fix SSRF via DNS rebinding with pre-request hostname resolution (http.rs)
- Add token budget enforcement on LLM calls (reasoning.rs, state.rs)
- Fix cross-user chat history leak with ownership verification (store.rs, server.rs)
- Add sliding-window rate limiter on gateway chat endpoint (server.rs)
- Harden extension install: HTTPS-only, 50MB cap, WASM magic validation (manager.rs)
- Add destructive command blocklist that overrides shell auto-approval (shell.rs)
- Add 5MB response body size cap to HTTP tool (http.rs)

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* refactor: deduplicate shared helpers and remove dead code

Extract floor_char_boundary and llm_signals_completion into src/util.rs,
unifying diverging phrase lists from agent/worker.rs and worker/runtime.rs.
Remove dead RespondResult::usage(), duplicate PROTECTED_IDENTITY_FILES
constant, double LeakDetector scanning in WebLogLayer, and invalid
0.0.0.0 origin from WebSocket allow list.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address PR review findings and CI test failures

- Fix record_failed_approve: .truncate(true) wiped the attempts file
  before reading, so failed pairing attempts never accumulated and
  rate limiting never triggered.
- Guard wizard WASM test: skip gracefully when channel build artifacts
  are absent (CI doesn't compile wasm32-wasip2 targets).
- Fix DNS rebinding check: use port 0 instead of hardcoded 443, since
  the port is irrelevant for hostname resolution.
- Remove hardcoded CORS port 3001: the dynamic addr.port() entries
  already cover the actual server port.
- Require WebSocket Origin header: reject connections that omit it
  entirely, since browsers always send Origin for WS upgrades and a
  missing header indicates a non-browser client bypassing the check.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address second round of PR review findings

- store.rs: reintroduce file locking around read-modify-write in
  record_failed_approve (concurrent callers could clobber each other).
- sse.rs: replace load+check+fetch_add with atomic fetch_update in both
  subscribe_raw() and subscribe() to prevent overshooting max_connections.
- ws.rs: decrement WS tracker before early return when subscribe_raw()
  returns None (connection limit reached), fixing a counter leak.
- server.rs: parse WS Origin host exactly instead of prefix matching,
  preventing bypass via crafted origins like http://localhost.evil.com.
- workspace_integration.rs: skip tests gracefully when Postgres is
  unreachable instead of panicking (fixes 10 CI failures).

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: add Origin header to WS integration tests

The Origin header requirement added in a3b0190 broke the WS gateway
integration tests. Test clients now send Origin: http://127.0.0.1:{port}
to match the server's localhost validation.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

---------

Co-authored-by: Claude Opus 4.6 <[email protected]>
2026-02-13 05:25:20 +00:00

1050 lines
36 KiB
Rust

//! Software builder for creating programs and tools using LLM-driven code generation.
//!
//! This module provides a general-purpose software building capability that:
//! - Uses an agent loop similar to Codex for iterative development
//! - Can build any software (binaries, libraries, scripts)
//! - Has special context injection when building WASM tools
//! - Integrates with existing tool loading infrastructure
//!
//! # Architecture
//!
//! ```text
//! ┌─────────────────────────────────────────────────────────────────────────────┐
//! │ Software Build Loop │
//! │ │
//! │ 1. Analyze requirement ─▶ Determine project type, language, structure │
//! │ 2. Generate scaffold ─▶ Create initial project files │
//! │ 3. Implement code ─▶ Write the actual implementation │
//! │ 4. Build/compile ─▶ Run build commands (cargo, npm, etc.) │
//! │ 5. Fix errors ─▶ Parse errors, modify code, retry │
//! │ 6. Test ─▶ Run tests, fix failures │
//! │ 7. Package ─▶ Produce final artifact │
//! └─────────────────────────────────────────────────────────────────────────────┘
//! ```
//!
//! For WASM tools specifically:
//! - Injects Tool trait interface documentation
//! - Injects WASM host function documentation
//! - Compiles to wasm32-wasip2 target
//! - Validates against tool interface
//! - Registers with ToolRegistry
use std::path::{Path, PathBuf};
use std::sync::Arc;
use std::time::Duration;
use async_trait::async_trait;
use chrono::{DateTime, Utc};
use serde::{Deserialize, Serialize};
use uuid::Uuid;
use crate::context::JobContext;
use crate::error::ToolError as AgentToolError;
use crate::llm::{
ChatMessage, LlmProvider, Reasoning, ReasoningContext, RespondResult, ToolDefinition,
};
use crate::safety::SafetyLayer;
use crate::tools::ToolRegistry;
use crate::tools::tool::{Tool, ToolError, ToolOutput};
/// Requirement specification for building software.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct BuildRequirement {
/// Name for the software.
pub name: String,
/// Description of what it should do.
pub description: String,
/// Type of software to build.
pub software_type: SoftwareType,
/// Target language/runtime.
pub language: Language,
/// Expected input format (for tools/CLIs).
pub input_spec: Option<String>,
/// Expected output format.
pub output_spec: Option<String>,
/// External dependencies needed.
pub dependencies: Vec<String>,
/// Security/capability requirements (for WASM tools).
pub capabilities: Vec<String>,
}
/// Type of software being built.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
#[serde(rename_all = "snake_case")]
pub enum SoftwareType {
/// A WASM tool for the agent.
WasmTool,
/// A standalone CLI application.
CliBinary,
/// A library/crate.
Library,
/// A script (Python, Bash, etc.).
Script,
/// A web service/API.
WebService,
}
/// Programming language for the build.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
#[serde(rename_all = "snake_case")]
pub enum Language {
Rust,
Python,
TypeScript,
JavaScript,
Go,
Bash,
}
impl Language {
/// Get the file extension for this language.
pub fn extension(&self) -> &'static str {
match self {
Language::Rust => "rs",
Language::Python => "py",
Language::TypeScript => "ts",
Language::JavaScript => "js",
Language::Go => "go",
Language::Bash => "sh",
}
}
/// Get the build command for this language.
pub fn build_command(&self, project_dir: &str) -> Option<String> {
match self {
Language::Rust => Some(format!("cd {} && cargo build --release", project_dir)),
Language::TypeScript => Some(format!("cd {} && npm run build", project_dir)),
Language::Go => Some(format!("cd {} && go build ./...", project_dir)),
Language::Python | Language::JavaScript | Language::Bash => None, // Interpreted
}
}
/// Get the test command for this language.
pub fn test_command(&self, project_dir: &str) -> String {
match self {
Language::Rust => format!("cd {} && cargo test", project_dir),
Language::Python => format!("cd {} && python -m pytest", project_dir),
Language::TypeScript | Language::JavaScript => {
format!("cd {} && npm test", project_dir)
}
Language::Go => format!("cd {} && go test ./...", project_dir),
Language::Bash => format!("cd {} && shellcheck *.sh", project_dir),
}
}
}
/// Result of a build operation.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct BuildResult {
/// Unique ID for this build.
pub build_id: Uuid,
/// The requirement that was built.
pub requirement: BuildRequirement,
/// Path to the output artifact.
pub artifact_path: PathBuf,
/// Build logs.
pub logs: Vec<BuildLog>,
/// Whether the build succeeded.
pub success: bool,
/// Error message if failed.
pub error: Option<String>,
/// When the build started.
pub started_at: DateTime<Utc>,
/// When the build completed.
pub completed_at: DateTime<Utc>,
/// Number of iterations to complete.
pub iterations: u32,
/// Validation warnings (for WASM tools).
#[serde(default)]
pub validation_warnings: Vec<String>,
/// Test results summary.
#[serde(default)]
pub tests_passed: u32,
/// Number of tests that failed.
#[serde(default)]
pub tests_failed: u32,
/// Whether the tool was auto-registered (for WASM tools).
#[serde(default)]
pub registered: bool,
}
/// A log entry from the build process.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct BuildLog {
pub timestamp: DateTime<Utc>,
pub phase: BuildPhase,
pub message: String,
pub details: Option<String>,
}
/// Phases of the build process.
#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq)]
#[serde(rename_all = "snake_case")]
pub enum BuildPhase {
Analyzing,
Scaffolding,
Implementing,
Building,
Testing,
Fixing,
Validating,
Registering,
Packaging,
Complete,
Failed,
}
/// Configuration for the software builder.
#[derive(Debug, Clone)]
pub struct BuilderConfig {
/// Directory where builds happen.
pub build_dir: PathBuf,
/// Maximum iterations before giving up.
pub max_iterations: u32,
/// Timeout for the entire build.
pub timeout: Duration,
/// Whether to clean up failed builds.
pub cleanup_on_failure: bool,
/// Whether to validate WASM tools after building.
pub validate_wasm: bool,
/// Whether to run tests after building.
pub run_tests: bool,
/// Whether to auto-register successful WASM tool builds.
pub auto_register: bool,
/// Directory to copy successful WASM tools for persistence.
pub wasm_output_dir: Option<PathBuf>,
}
impl Default for BuilderConfig {
fn default() -> Self {
Self {
build_dir: std::env::temp_dir().join("ironclaw-builds"),
max_iterations: 10,
timeout: Duration::from_secs(600), // 10 minutes
cleanup_on_failure: false, // Keep for debugging
validate_wasm: true,
run_tests: true,
auto_register: true,
wasm_output_dir: None,
}
}
}
/// Trait for building software.
#[async_trait]
pub trait SoftwareBuilder: Send + Sync {
/// Analyze a natural language description and extract a structured requirement.
async fn analyze(&self, description: &str) -> Result<BuildRequirement, AgentToolError>;
/// Build software from a requirement.
async fn build(&self, requirement: &BuildRequirement) -> Result<BuildResult, AgentToolError>;
/// Attempt to repair a failed build.
async fn repair(
&self,
result: &BuildResult,
error: &str,
) -> Result<BuildResult, AgentToolError>;
}
/// LLM-powered software builder.
pub struct LlmSoftwareBuilder {
config: BuilderConfig,
llm: Arc<dyn LlmProvider>,
safety: Arc<SafetyLayer>,
tools: Arc<ToolRegistry>,
}
impl LlmSoftwareBuilder {
/// Create a new LLM-based software builder.
pub fn new(
config: BuilderConfig,
llm: Arc<dyn LlmProvider>,
safety: Arc<SafetyLayer>,
tools: Arc<ToolRegistry>,
) -> Self {
// Ensure build directory exists
if let Err(e) = std::fs::create_dir_all(&config.build_dir) {
tracing::warn!("Failed to create build directory: {}", e);
}
Self {
config,
llm,
safety,
tools,
}
}
/// Get the build tools available for the build loop.
async fn get_build_tools(&self) -> Vec<ToolDefinition> {
// Only include tools useful for building software
self.tools
.tool_definitions_for(&[
"shell",
"read_file",
"write_file",
"list_dir",
"apply_patch",
"http", // For fetching docs/deps
])
.await
}
/// Create the system prompt for the build agent.
fn build_system_prompt(&self, requirement: &BuildRequirement) -> String {
let mut prompt = format!(
r#"You are a software developer building a program.
## Task
Build: {name}
Description: {description}
Type: {software_type:?}
Language: {language:?}
## Process
1. Create the project structure with necessary files
2. Implement the code based on the requirements
3. Build/compile if needed
4. Run tests to verify correctness
5. Fix any errors and iterate
## Guidelines
- Write clean, well-structured code
- Handle errors appropriately
- Add minimal but useful comments
- Follow idiomatic patterns for the language
- Test edge cases
## Tools Available
- shell: Run build commands, tests, install dependencies
- read_file: Read existing files
- write_file: Create new files
- apply_patch: Edit existing files surgically
- list_dir: Explore project structure
"#,
name = requirement.name,
description = requirement.description,
software_type = requirement.software_type,
language = requirement.language,
);
// Add tool-specific context when building WASM tools
if requirement.software_type == SoftwareType::WasmTool {
prompt.push_str(&self.wasm_tool_context());
}
prompt
}
/// Get additional context for building WASM tools.
fn wasm_tool_context(&self) -> String {
r#"
## WASM Tool Requirements
You are building a WASM Component tool for an autonomous agent using the WASM Component Model.
The tool MUST use `wit_bindgen` and `cargo-component` to build.
## Available Host Functions (from WIT interface)
The host provides these functions via `near::agent::host`:
```rust
// Logging (always available)
host::log(level: LogLevel, message: &str); // LogLevel: Trace, Debug, Info, Warn, Error
// Time (always available)
host::now_millis() -> u64; // Unix timestamp in milliseconds
// Workspace (if capability granted)
host::workspace_read(path: &str) -> Option<String>;
// HTTP (if capability granted)
host::http_request(method: &str, url: &str, headers_json: &str, body: Option<Vec<u8>>)
-> Result<HttpResponse, String>;
// HttpResponse has: status: u16, headers_json: String, body: Vec<u8>
// Tool invocation (if capability granted)
host::tool_invoke(alias: &str, params_json: &str) -> Result<String, String>;
// Secrets (if capability granted) - can only CHECK existence, not read values
host::secret_exists(name: &str) -> bool;
```
## Project Structure
```
my_tool/
├── Cargo.toml
├── wit/
│ └── tool.wit # Copy from agent's wit/tool.wit
└── src/
└── lib.rs
```
## Cargo.toml Template
```toml
[package]
name = "my_tool"
version = "0.1.0"
edition = "2021"
[lib]
crate-type = ["cdylib"]
[dependencies]
wit-bindgen = "0.41"
serde = { version = "1", features = ["derive"] }
serde_json = "1"
```
## src/lib.rs Template
```rust
// Generate bindings from the WIT interface
wit_bindgen::generate!({
world: "sandboxed-tool",
path: "wit/tool.wit",
});
use serde::{Deserialize, Serialize};
use exports::near::agent::tool::{Guest, Request, Response};
use near::agent::host::{self, LogLevel};
// Your input/output types
#[derive(Deserialize)]
struct MyInput {
// Define parameters here
}
#[derive(Serialize)]
struct MyOutput {
// Define output here
}
struct MyTool;
impl Guest for MyTool {
fn execute(req: Request) -> Response {
// Parse input
let input: MyInput = match serde_json::from_str(&req.params) {
Ok(i) => i,
Err(e) => return Response {
output: None,
error: Some(format!("Invalid input: {}", e)),
},
};
host::log(LogLevel::Info, &format!("Processing request..."));
// Your implementation here
let output = MyOutput { /* ... */ };
// Return success
Response {
output: Some(serde_json::to_string(&output).unwrap()),
error: None,
}
}
fn schema() -> String {
serde_json::json!({
"type": "object",
"properties": {
// Define your JSON Schema here
},
"required": []
}).to_string()
}
fn description() -> String {
"Description of what this tool does".to_string()
}
}
export!(MyTool);
```
## Build Commands
```bash
# Install cargo-component (one time)
cargo install cargo-component
# Build the WASM component
cargo component build --release
# Output: target/wasm32-wasip2/release/my_tool.wasm
```
## Capabilities File (my_tool.capabilities.json)
Create alongside the .wasm file to grant capabilities:
```json
{
"http": {
"allowed_endpoints": [
{"host": "api.example.com", "path_prefix": "/v1/"}
]
},
"workspace": true,
"secrets": {
"allowed": ["API_KEY"]
}
}
```
## Important Notes
1. NEVER panic - always return Response with error field set
2. Secrets are NEVER exposed to WASM - use placeholders like `{API_KEY}` in URLs
and the host will inject the real value
3. HTTP requests are rate-limited and only allowed to endpoints in capabilities
4. Keep the tool focused on one thing - small, composable tools are better
"#
.to_string()
}
/// Execute the build loop.
async fn execute_build_loop(
&self,
requirement: &BuildRequirement,
project_dir: &Path,
) -> Result<BuildResult, AgentToolError> {
let build_id = Uuid::new_v4();
let started_at = Utc::now();
let mut logs = Vec::new();
let mut iteration = 0;
// Create reasoning engine
let reasoning = Reasoning::new(self.llm.clone(), self.safety.clone());
// Build initial context
let tool_defs = self.get_build_tools().await;
let mut reason_ctx = ReasoningContext::new().with_tools(tool_defs);
// Add system prompt
reason_ctx
.messages
.push(ChatMessage::system(self.build_system_prompt(requirement)));
// Add initial user message - directive to force immediate tool use
reason_ctx.messages.push(ChatMessage::user(format!(
"Build the {} in directory: {}\n\n\
Requirements:\n- {}\n\n\
IMPORTANT: Use the write_file tool NOW to create Cargo.toml. \
Do not explain, plan, or output JSON—immediately call write_file.",
requirement.name,
project_dir.display(),
requirement.description
)));
logs.push(BuildLog {
timestamp: Utc::now(),
phase: BuildPhase::Analyzing,
message: "Starting build process".into(),
details: None,
});
// Main build loop
let mut current_phase = BuildPhase::Scaffolding;
let mut last_error: Option<String> = None;
let mut tools_executed = false;
let mut consecutive_text_responses = 0;
loop {
iteration += 1;
if iteration > self.config.max_iterations {
logs.push(BuildLog {
timestamp: Utc::now(),
phase: BuildPhase::Failed,
message: "Maximum iterations exceeded".into(),
details: last_error.clone(),
});
return Ok(BuildResult {
build_id,
requirement: requirement.clone(),
artifact_path: project_dir.to_path_buf(),
logs,
success: false,
error: Some("Maximum iterations exceeded".into()),
started_at,
completed_at: Utc::now(),
iterations: iteration,
validation_warnings: Vec::new(),
tests_passed: 0,
tests_failed: 0,
registered: false,
});
}
// Refresh tool definitions each iteration
reason_ctx.available_tools = self.get_build_tools().await;
// Get response from LLM (may be text or tool calls)
let result = reasoning
.respond_with_tools(&reason_ctx)
.await
.map_err(|e| {
AgentToolError::BuilderFailed(format!("LLM response failed: {}", e))
})?;
match result.result {
RespondResult::Text(response) => {
reason_ctx.messages.push(ChatMessage::assistant(&response));
// If tools haven't been executed yet, we're stuck in planning mode
if !tools_executed {
consecutive_text_responses += 1;
// Fail fast after 2 consecutive text-only responses
if consecutive_text_responses >= 2 {
logs.push(BuildLog {
timestamp: Utc::now(),
phase: BuildPhase::Failed,
message: "Builder stuck in planning mode".into(),
details: Some(format!(
"LLM returned {} consecutive text responses without calling tools. \
Try a more specific requirement.",
consecutive_text_responses
)),
});
return Ok(BuildResult {
build_id,
requirement: requirement.clone(),
artifact_path: project_dir.to_path_buf(),
logs,
success: false,
error: Some(
"LLM not executing tools - stuck in planning mode".into(),
),
started_at,
completed_at: Utc::now(),
iterations: iteration,
validation_warnings: Vec::new(),
tests_passed: 0,
tests_failed: 0,
registered: false,
});
}
tracing::debug!(
"Builder: no tools executed (text response #{}/2), forcing tool use",
consecutive_text_responses
);
reason_ctx.messages.push(ChatMessage::user(
"STOP. Do NOT output text, JSON specs, or explanations. \
Call the write_file tool RIGHT NOW to create Cargo.toml. \
Just call the tool—no commentary.",
));
continue;
}
// Reset counter when tools have been executed (we're in completion phase)
consecutive_text_responses = 0;
// Check for completion signals
let response_lower = response.to_lowercase();
if response_lower.contains("build complete")
|| response_lower.contains("successfully built")
|| response_lower.contains("all tests pass")
|| response_lower.contains("complete")
{
logs.push(BuildLog {
timestamp: Utc::now(),
phase: BuildPhase::Complete,
message: "Build completed successfully".into(),
details: Some(response),
});
// Determine artifact path
let artifact_path = self.find_artifact(requirement, project_dir).await;
return Ok(BuildResult {
build_id,
requirement: requirement.clone(),
artifact_path,
logs,
success: true,
error: None,
started_at,
completed_at: Utc::now(),
iterations: iteration,
validation_warnings: Vec::new(),
tests_passed: 0,
tests_failed: 0,
registered: false,
});
}
// Ask for next steps
reason_ctx
.messages
.push(ChatMessage::user("Continue with the next step."));
}
RespondResult::ToolCalls {
tool_calls,
content,
} => {
tools_executed = true;
// Add assistant message with tool_calls (OpenAI protocol)
reason_ctx
.messages
.push(ChatMessage::assistant_with_tool_calls(
content,
tool_calls.clone(),
));
// Execute each tool call
for tc in tool_calls {
logs.push(BuildLog {
timestamp: Utc::now(),
phase: current_phase,
message: format!("Executing: {}", tc.name),
details: Some(format!("{:?}", tc.arguments)),
});
// Execute tool
let tool_result = self
.execute_build_tool(&tc.name, &tc.arguments, project_dir)
.await;
match tool_result {
Ok(output) => {
let output_str = serde_json::to_string_pretty(&output.result)
.unwrap_or_default();
// Add to context
reason_ctx.messages.push(ChatMessage::tool_result(
&tc.id,
&tc.name,
output_str.clone(),
));
// Update phase based on tool
current_phase = match tc.name.as_str() {
"write_file" => BuildPhase::Implementing,
"shell" if tc.arguments.to_string().contains("build") => {
BuildPhase::Building
}
"shell" if tc.arguments.to_string().contains("test") => {
BuildPhase::Testing
}
_ => current_phase,
};
// Check for build/test errors in output
if output_str.to_lowercase().contains("error:")
|| output_str.to_lowercase().contains("error[")
|| output_str.to_lowercase().contains("failed")
{
last_error = Some(output_str);
current_phase = BuildPhase::Fixing;
}
}
Err(e) => {
let error_msg = format!("Tool error: {}", e);
last_error = Some(error_msg.clone());
reason_ctx.messages.push(ChatMessage::tool_result(
&tc.id,
&tc.name,
format!("Error: {}", e),
));
logs.push(BuildLog {
timestamp: Utc::now(),
phase: BuildPhase::Fixing,
message: "Tool execution failed".into(),
details: Some(error_msg),
});
current_phase = BuildPhase::Fixing;
}
}
}
}
}
}
}
/// Execute a build tool.
async fn execute_build_tool(
&self,
tool_name: &str,
params: &serde_json::Value,
_project_dir: &Path,
) -> Result<ToolOutput, ToolError> {
let tool =
self.tools.get(tool_name).await.ok_or_else(|| {
ToolError::ExecutionFailed(format!("Tool not found: {}", tool_name))
})?;
// Execute with a dummy context (build tools don't need job context)
let ctx = JobContext::default();
tool.execute(params.clone(), &ctx).await
}
/// Find the build artifact based on project type.
async fn find_artifact(&self, requirement: &BuildRequirement, project_dir: &Path) -> PathBuf {
match (&requirement.software_type, &requirement.language) {
(SoftwareType::WasmTool, Language::Rust) => {
// WASM output location
project_dir.join(format!(
"target/wasm32-wasip2/release/{}.wasm",
requirement.name.replace('-', "_")
))
}
(SoftwareType::CliBinary, Language::Rust) => project_dir.join(format!(
"target/release/{}",
requirement.name.replace('-', "_")
)),
(SoftwareType::Script, Language::Python) => {
project_dir.join(format!("{}.py", requirement.name))
}
(SoftwareType::Script, Language::Bash) => {
project_dir.join(format!("{}.sh", requirement.name))
}
_ => project_dir.to_path_buf(),
}
}
}
#[async_trait]
impl SoftwareBuilder for LlmSoftwareBuilder {
async fn analyze(&self, description: &str) -> Result<BuildRequirement, AgentToolError> {
// Use LLM to parse the description
let reasoning = Reasoning::new(self.llm.clone(), self.safety.clone());
let prompt = format!(
r#"Analyze this software requirement and extract structured information.
Description: {}
IMPORTANT: If this is a "tool" that the agent will use (e.g., "calendar tool", "email tool",
"API client tool"), you MUST use:
- software_type: "wasm_tool"
- language: "rust"
Only use cli_binary/script/library for software meant for human end-users, not agent tools.
Respond with a JSON object containing:
- name: A short identifier (snake_case)
- description: What the software should do
- software_type: One of "wasm_tool", "cli_binary", "library", "script", "web_service"
(PREFER "wasm_tool" for agent-usable tools)
- language: One of "rust", "python", "typescript", "javascript", "go", "bash"
(PREFER "rust" for wasm_tool)
- input_spec: Expected input format (optional)
- output_spec: Expected output format (optional)
- dependencies: List of external dependencies needed
- capabilities: For WASM tools, list needed capabilities (http, workspace, secrets)
JSON:"#,
description
);
let ctx = ReasoningContext::new().with_message(ChatMessage::user(&prompt));
let response = reasoning
.respond(&ctx)
.await
.map_err(|e| AgentToolError::BuilderFailed(format!("Analysis failed: {}", e)))?;
// Extract JSON from response
let json_start = response.find('{').unwrap_or(0);
let json_end = response.rfind('}').map(|i| i + 1).unwrap_or(response.len());
let json_str = &response[json_start..json_end];
serde_json::from_str(json_str).map_err(|e| {
AgentToolError::BuilderFailed(format!("Failed to parse requirement: {}", e))
})
}
async fn build(&self, requirement: &BuildRequirement) -> Result<BuildResult, AgentToolError> {
// Create project directory
let project_dir = self.config.build_dir.join(&requirement.name);
if project_dir.exists() {
std::fs::remove_dir_all(&project_dir).map_err(|e| {
AgentToolError::BuilderFailed(format!("Failed to clean project dir: {}", e))
})?;
}
std::fs::create_dir_all(&project_dir).map_err(|e| {
AgentToolError::BuilderFailed(format!("Failed to create project dir: {}", e))
})?;
// Run the build loop with timeout
let result = tokio::time::timeout(
self.config.timeout,
self.execute_build_loop(requirement, &project_dir),
)
.await;
match result {
Ok(Ok(build_result)) => Ok(build_result),
Ok(Err(e)) => Err(e),
Err(_) => Err(AgentToolError::BuilderFailed("Build timed out".into())),
}
}
async fn repair(
&self,
result: &BuildResult,
error: &str,
) -> Result<BuildResult, AgentToolError> {
// Create a new requirement with repair context
let mut requirement = result.requirement.clone();
requirement.description = format!(
"{}\n\nPrevious build failed with error:\n{}\n\nFix the issues and rebuild.",
requirement.description, error
);
// Rebuild (preserving project directory if it exists)
self.build(&requirement).await
}
}
/// Tool that allows the agent to build software on demand.
pub struct BuildSoftwareTool {
builder: Arc<dyn SoftwareBuilder>,
}
impl BuildSoftwareTool {
pub fn new(builder: Arc<dyn SoftwareBuilder>) -> Self {
Self { builder }
}
}
#[async_trait]
impl Tool for BuildSoftwareTool {
fn name(&self) -> &str {
"build_software"
}
fn description(&self) -> &str {
"Build software from a description. IMPORTANT: For tools the agent will use, \
ALWAYS build Rust WASM tools (type: wasm_tool, language: rust). Only use cli_binary, \
script, or other types for software meant for human users. The builder scaffolds, \
implements, compiles, and tests iteratively."
}
fn parameters_schema(&self) -> serde_json::Value {
serde_json::json!({
"type": "object",
"properties": {
"description": {
"type": "string",
"description": "Natural language description of what to build"
},
"type": {
"type": "string",
"enum": ["wasm_tool", "cli_binary", "library", "script"],
"description": "Type of software to build (optional, will be inferred)"
},
"language": {
"type": "string",
"enum": ["rust", "python", "typescript", "bash"],
"description": "Programming language to use (optional, will be inferred)"
}
},
"required": ["description"]
})
}
async fn execute(
&self,
params: serde_json::Value,
_ctx: &JobContext,
) -> Result<ToolOutput, ToolError> {
let description = params
.get("description")
.and_then(|v| v.as_str())
.ok_or_else(|| ToolError::InvalidParameters("missing 'description'".into()))?;
let start = std::time::Instant::now();
// Analyze the requirement
let mut requirement = self
.builder
.analyze(description)
.await
.map_err(|e| ToolError::ExecutionFailed(format!("Analysis failed: {}", e)))?;
// Override type/language if specified
if let Some(type_str) = params.get("type").and_then(|v| v.as_str()) {
requirement.software_type = match type_str {
"wasm_tool" => SoftwareType::WasmTool,
"cli_binary" => SoftwareType::CliBinary,
"library" => SoftwareType::Library,
"script" => SoftwareType::Script,
_ => requirement.software_type,
};
}
if let Some(lang_str) = params.get("language").and_then(|v| v.as_str()) {
requirement.language = match lang_str {
"rust" => Language::Rust,
"python" => Language::Python,
"typescript" => Language::TypeScript,
"bash" => Language::Bash,
_ => requirement.language,
};
}
// Build
let result = self
.builder
.build(&requirement)
.await
.map_err(|e| ToolError::ExecutionFailed(format!("Build failed: {}", e)))?;
let output = serde_json::json!({
"build_id": result.build_id.to_string(),
"name": result.requirement.name,
"success": result.success,
"artifact_path": result.artifact_path.display().to_string(),
"iterations": result.iterations,
"error": result.error,
"phases": result.logs.iter().map(|l| format!("{:?}: {}", l.phase, l.message)).collect::<Vec<_>>()
});
Ok(ToolOutput::success(output, start.elapsed()))
}
fn requires_approval(&self) -> bool {
true // Building software should require approval
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_language_extensions() {
assert_eq!(Language::Rust.extension(), "rs");
assert_eq!(Language::Python.extension(), "py");
assert_eq!(Language::TypeScript.extension(), "ts");
}
#[test]
fn test_build_commands() {
assert!(Language::Rust.build_command("/tmp/project").is_some());
assert!(Language::Python.build_command("/tmp/project").is_none());
}
#[test]
fn test_software_type_serialization() {
let json = serde_json::to_string(&SoftwareType::WasmTool).unwrap();
assert_eq!(json, "\"wasm_tool\"");
}
}