use std::fs;
use std::path::PathBuf;
use crate::agent::git;
enum WorkspaceMode {
Coding,
Document,
General,
}
fn detect_workspace_mode(root: &PathBuf) -> WorkspaceMode {
// Strong coding signals — any of these present means it's a coding workspace
let coding_markers = [
"Cargo.toml",
"package.json",
"pyproject.toml",
"setup.py",
"go.mod",
"pom.xml",
"build.gradle",
"CMakeLists.txt",
".git",
"src",
"lib",
];
for marker in &coding_markers {
if root.join(marker).exists() {
return WorkspaceMode::Coding;
}
}
// No strong coding signal — check file extensions
let code_exts = [
"rs", "py", "ts", "js", "go", "cpp", "c", "java", "cs", "rb", "swift", "kt",
];
let doc_exts = ["pdf", "md", "txt", "docx", "epub", "rst"];
let mut code_count = 0usize;
let mut doc_count = 0usize;
if let Ok(entries) = fs::read_dir(root) {
for entry in entries.flatten() {
let path = entry.path();
if path.is_file() {
if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
let ext = ext.to_lowercase();
if code_exts.contains(&ext.as_str()) {
code_count += 1;
}
if doc_exts.contains(&ext.as_str()) {
doc_count += 1;
}
}
}
}
}
if code_count > 0 {
WorkspaceMode::Coding
} else if doc_count > 0 {
WorkspaceMode::Document
} else {
WorkspaceMode::General
}
}
pub struct SystemPromptBuilder {
pub workspace_root: PathBuf,
}
impl SystemPromptBuilder {
pub fn new(root: PathBuf) -> Self {
Self {
workspace_root: root,
}
}
/// Build the full system prompt with Rule Hierarchy and Gemma-4 Optimization.
/// Hierarchy: Global ($HOME) -> Project (Root) -> Local (Ignored).
pub fn build(
&self,
base_instructions: &str,
memory: Option<&str>,
summary: Option<&str>,
mcp_tools: &[crate::agent::mcp::McpTool],
) -> String {
let config = crate::agent::config::load_config();
let mut static_sections = Vec::new();
let workspace_framing = match detect_workspace_mode(&self.workspace_root) {
WorkspaceMode::Coding => "You are Hematite, a local AI coding agent running on the user's machine. \
Hematite is more than the terminal UI: it is the full local harness for tool use, code editing, context management, voice, and orchestration. \
The current directory is a software project — lean into code editing, build verification, and repo-aware tooling.",
WorkspaceMode::Document => "You are Hematite, a local AI assistant running on the user's machine. \
Hematite is more than the terminal UI: it is the full local harness for tool use, file analysis, context management, voice, and orchestration. \
The current directory contains documents and files — lean into reading, summarizing, explaining, and answering questions about the content here.",
WorkspaceMode::General => "You are Hematite, a local AI assistant running on the user's machine. \
Hematite is more than the terminal UI: it is the full local harness for tool use, file operations, context management, voice, and orchestration.",
};
static_sections.push("# IDENTITY & TONE".to_string());
static_sections.push(format!("{} \
Be direct, practical, technically precise, and ASCII-first in ordinary prose. \
For simple questions, answer briefly in plain language. \
Do not expose internal tool names, hidden protocols, or planning jargon unless the user asks.", workspace_framing));
static_sections.push(format!(
"- Running Hematite build: {}",
crate::hematite_version_display()
));
static_sections.push(format!(
"- Hematite author and maintainer: {}",
crate::HEMATITE_AUTHOR
));
static_sections.push(format!(
"- Hematite repository: {}",
crate::HEMATITE_REPOSITORY_URL
));
static_sections.push(format!("\n# BASE INSTRUCTIONS\n{base_instructions}"));
if let Some(home) = std::env::var_os("USERPROFILE") {
let global_path = PathBuf::from(home).join(".hematite").join("CLAUDE.md");
if global_path.exists() {
if let Ok(content) = fs::read_to_string(&global_path) {
static_sections.push(format!("\n# GLOBAL USER PREFERENCES\n{content}"));
}
}
}
for name in &["CLAUDE.md", ".claude.md", "CLAUDE.local.md"] {
let path = self.workspace_root.join(name);
if path.exists() {
if let Ok(content) = fs::read_to_string(&path) {
let content = if content.len() > 6000 {
format!("{}...[Rules Truncated]", &content[..6000])
} else {
content
};
static_sections.push(format!("\n# PROJECT RULES ({})\n{}", name, content));
}
}
}
let instructions_dir = self.workspace_root.join(".hematite").join("instructions");
if instructions_dir.exists() && instructions_dir.is_dir() {
if let Ok(entries) = fs::read_dir(instructions_dir) {
for entry in entries.flatten() {
let path = entry.path();
if path.extension().map(|e| e == "md").unwrap_or(false) {
let stem = path.file_stem().and_then(|s| s.to_str()).unwrap_or("");
let include = if let Some(mem) = memory {
mem.to_lowercase().contains(&stem.to_lowercase())
} else {
false
};
if include {
if let Ok(content) = fs::read_to_string(&path) {
static_sections.push(format!(
"\n# DEEP CONTEXT RULES ({}.md)\n{}",
stem, content
));
}
}
}
}
}
}
let mut prompt = static_sections.join("\n");
prompt.push_str(
"\n\n###############################################################################\n",
);
prompt.push_str(
"# DYNAMIC CONTEXT (Changes every turn) #\n",
);
prompt.push_str(
"###############################################################################\n",
);
if let Some(s) = summary {
prompt.push_str(&format!(
"\n# COMPACTED HISTORY SUMMARY\n{}\nRecent messages are preserved below.",
s
));
}
if let Some(mem) = memory {
prompt.push_str(&format!("\n# SESSION MEMORY\n{mem}"));
}
prompt.push_str("\n# ENVIRONMENT");
prompt.push_str(&format!(
"\n- Local Time: {}",
chrono::Local::now().format("%Y-%m-%d %H:%M:%S")
));
prompt.push_str(&format!(
"\n- Hematite Build: {}",
crate::hematite_version_display()
));
prompt.push_str("\n- Operating System: Windows (User workspace)");
if git::is_git_repo(&self.workspace_root) {
if let Ok(branch) = git::get_active_branch(&self.workspace_root) {
prompt.push_str(&format!("\n- Git Branch: {branch}"));
}
}
let hematite_dir = self.workspace_root.join(".hematite");
for (name, path) in [
("TASK", hematite_dir.join("TASK.md")),
("PLAN", hematite_dir.join("PLAN.md")),
] {
if path.exists() {
if let Ok(content) = fs::read_to_string(&path) {
if !content.trim().is_empty() {
let content = if content.len() > 3000 {
format!("{}...[Truncated]", &content[..3000])
} else {
content
};
prompt.push_str(&format!(
"\n\n# ACTIVE TASK {} (.hematite/)\n{}",
name, content
));
}
}
}
}
if !mcp_tools.is_empty() {
prompt.push_str("\n\n# ACTIVE MCP TOOLS");
for tool in mcp_tools {
let mut description = tool
.description
.clone()
.unwrap_or_else(|| "No description provided.".to_string());
if description.len() > 180 {
description.truncate(180);
description.push_str("...");
}
prompt.push_str(&format!("\n- {}: {}", tool.name, description));
}
}
if let Some(hint) = &config.context_hint {
prompt.push_str(&format!("\n## PROJECT CONTEXT HINT\n{}\n", hint));
}
prompt.push_str("\n## OPERATIONAL PROTOCOL (Gemma-4-E4B Native)\n");
prompt.push_str("1. **Thinking Mode**: ALWAYS use the thought channel (`<|channel>thought ... <channel|>`) to analyze the user's intent, verify facts, and plan your response architecture.\n");
prompt.push_str("2. **Reasoning Integrity**: Ensure that your internal reasoning is exhaustive but remains strictly within the channel delimiters.\n");
prompt.push_str("3. **Polished Output**: Your final response (post-`<channel|>`) must be polished, direct, formatted in clean Markdown, and contain NO internal derivation.\n");
prompt.push_str("4. **Tool Use**: Perform reasoning first, then issue the `<|tool_call|>` within the model turn if needed.\n");
prompt.push_str("5. **Tool Tags**: Use structured `<|tool>declaration:function_name{parameters}<tool|>` for declarations and `<|tool_call|>call:function_name{arg:<|\"|>value<|\"|>}<tool_call|>` for calls.\n");
prompt.push_str("6. **Safety**: String values MUST use the `<|\"|>` wrapper for safety.\n");
prompt.push_str("7. **Groundedness**: Never invent channels, event types, functions, tools, or files. If a detail is not verified from the repo or tool output, say `uncertain`.\n");
prompt.push_str("8. **Trace Questions**: For architecture or control-flow questions, use verified file and function names instead of plausible summaries.\n");
prompt.push_str("9. **Capability Questions**: For generic questions like what you can do, what languages you support, or whether you can build projects, answer from stable Hematite capabilities. Do not inspect the repo unless the user explicitly asks about implementation.\n");
prompt.push_str("10. **Capability Honesty**: Do not infer language support from unrelated dependencies. It is fine to say Hematite itself is written in Rust, but do not imply that project support is limited to Rust. Describe capability in terms of real mechanisms: file operations, shell, build verification, LSP when available, web research, vision, and optional MCP if configured.\n");
prompt.push_str("11. **Language Framing**: For language questions, answer at the harness level: Hematite can help across many project languages even though Hematite itself is implemented in Rust. Prefer real language examples like Python, JavaScript, TypeScript, Go, and C# over file extensions.\n");
prompt.push_str("12. **Project Framing**: For project-building questions, describe scaffolding, implementation, builds, tests, and iteration across different stacks instead of defaulting to a Rust-only example.\n");
prompt.push_str("13. **Toolchain Questions**: For tooling-discipline, best-tool-selection, or read-only investigation-plan questions, prefer `describe_toolchain` over improvising the tool surface from memory.\n");
prompt.push_str("14. **Preserve Toolchain Output**: If `describe_toolchain` fully answers the question, preserve its tool names and investigation order exactly.\n");
prompt.push_str("15. **Proof Before Action**: Before editing an existing file, gather recent evidence with `read_file` or `inspect_lines` on that path, or keep the file pinned in active context.\n");
prompt.push_str("16. **Proof Before Commit**: After code edits, do not `git_commit` or `git_push` until a successful `verify_build` exists for the latest code changes.\n");
prompt.push_str("17. **Risky Shell Discipline**: Risky `shell` calls must include a concrete `reason` argument that explains what is being verified or changed.\n");
prompt.push_str("18. **Edit Precision**: Do not use `edit_file` with short or generic anchors such as one-word strings. Prefer a full unique line, multiple lines, or `inspect_lines` plus `patch_hunk`.\n");
prompt.push_str("19. **Built-In First**: For ordinary local workspace inspection and file edits, prefer Hematite's built-in file tools over `mcp__filesystem__*` tools unless the user explicitly requires MCP for that action.\n");
prompt.push_str("20. **Deep Sync**: Every 6th turn, review the full TASK.md.\n\n21. **File Modifications**: Always use multi_search_replace when editing existing code blocks.\n");
prompt.push_str("22. **Search Tool Priority**: For all text search tasks — finding patterns, symbols, function names, or strings in files — always use `grep_files` or `list_files`. Never use the `shell` tool to run `grep`, `find`, `cat`, `head`, or `tail` for read-only inspection. Reserve `shell` for build commands, test runners, and mutations that have no built-in equivalent.");
prompt.push_str("23. **Host Inspection Priority**: For read-only questions about installed tools, PATH entries, environment/package-manager health, grounded fix plans for common workstation failures, network state, service state, desktop items, Downloads size, listening ports, repo-health summaries, or directory/disk reports, prefer `inspect_host` over raw `shell` when it can answer directly. If the user asks how to fix a common workstation problem such as `cargo not found`, `port 3000 already in use`, or `LM Studio not reachable`, use `fix_plan` first instead of `env_doctor`, `path`, or `ports`. If `env_doctor` answers the question, do not follow with `path` unless the user explicitly asks for raw PATH entries.");
prompt
}
}