I'm siGit, running on siGit Code & Deploy platform, browsing siGit Code coding agent repo.
feature/skill-slash-commands
| 1 | //! Headless programmatic mode: `sigit run "<prompt>"` runs one prompt and exits. |
| 2 | //! |
| 3 | //! This is the entry point for CI, scripts, cron, and the Cloud Agent sandbox |
| 4 | //! runner: no client, no TTY, just plain stdio. Assistant text streams to |
| 5 | //! stdout as it is generated (`--quiet` restricts stdout to the final message); |
| 6 | //! logs and tool progress go to stderr. |
| 7 | //! |
| 8 | //! Cross-platform by design — unlike the ratatui TUI this is NOT gated on |
| 9 | //! `#[cfg(unix)]`, so Windows gets it too. |
| 10 | //! |
| 11 | //! Permission model: nobody is around to answer `ask`, so `ask` collapses to a |
| 12 | //! denial telling the model the tool was not pre-approved (mentioning |
| 13 | //! `--allow-tool`). `--allow-tool <name>` pre-grants a tool for the run's |
| 14 | //! session; `--deny-tool <name>` blocks a tool even if settings would allow it. |
| 15 | //! `SIGIT_PERMISSIONS=allow` remains the blunt instrument. |
| 16 | //! |
| 17 | //! Exit codes: 0 — turn completed; 1 — inference/tool-loop or backend |
| 18 | //! resolution error; 2 — bad invocation (handled by the caller in `main`). |
| 19 | |
| 20 | use std::collections::HashSet; |
| 21 | use std::io::Write; |
| 22 | use std::path::PathBuf; |
| 23 | use std::sync::Arc; |
| 24 | |
| 25 | use crate::backend::{ |
| 26 | self, InferenceBackend, OpenAiBackend, TokenChunk, ToolResult as BackendToolResult, TurnResult, |
| 27 | }; |
| 28 | use crate::{permissions, provider, session_store, settings, tools}; |
| 29 | |
| 30 | pub const USAGE: &str = "Usage: sigit run \"<prompt>\" [--cwd <dir>] [--add-dir <dir>]... \ |
| 31 | [--output text|jsonl] [--quiet] [--resume <session-id>] \ |
| 32 | [--allow-tool <name>]... [--deny-tool <name>]... \ |
| 33 | [--max-tool-rounds <n>] [--model <tier>]\n \ |
| 34 | sigit -p \"<prompt>\" [same options]"; |
| 35 | |
| 36 | #[derive(Debug, Clone, Copy, Default, PartialEq, Eq)] |
| 37 | pub enum OutputFormat { |
| 38 | #[default] |
| 39 | Text, |
| 40 | Jsonl, |
| 41 | } |
| 42 | |
| 43 | /// Parsed `sigit run` or legacy `sigit -p` invocation. |
| 44 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 45 | pub struct HeadlessConfig { |
| 46 | pub prompt: String, |
| 47 | /// Durable session id. A fresh UUID is generated unless `--resume` names |
| 48 | /// an existing session. |
| 49 | pub session_id: String, |
| 50 | /// Restore the saved conversation before sending `prompt`. |
| 51 | pub resume: bool, |
| 52 | /// Working directory to enter before anything loads (instruction files and |
| 53 | /// project-local MCP config then resolve from it). |
| 54 | pub cwd: Option<PathBuf>, |
| 55 | /// Extra project roots beyond `cwd`, the headless counterpart of a |
| 56 | /// multi-root project in the editor (see `workspace.rs`). |
| 57 | pub add_dirs: Vec<PathBuf>, |
| 58 | /// Print only the final assistant message to stdout (no streaming). |
| 59 | pub quiet: bool, |
| 60 | /// Machine-readable event output for CI and factory clients. |
| 61 | pub output: OutputFormat, |
| 62 | /// Tools pre-approved for the run (fed to `permissions::grant_for_session`). |
| 63 | pub allow_tools: Vec<String>, |
| 64 | /// Tools blocked for the run, overriding even settings-level allow. |
| 65 | pub deny_tools: Vec<String>, |
| 66 | /// How many rounds of tool calls the run may make before the model is |
| 67 | /// asked for its final message. `--max-tool-rounds`, else |
| 68 | /// `SIGIT_MAX_TOOL_ROUNDS`, else the built-in cap. |
| 69 | pub max_tool_rounds: usize, |
| 70 | /// siGit Code Cloud tier named by `--model`. When set it wins over the |
| 71 | /// provider override and the local-inference setting. |
| 72 | pub cloud_tier: Option<&'static str>, |
| 73 | } |
| 74 | |
| 75 | /// Upper bound for `--max-tool-rounds` / `SIGIT_MAX_TOOL_ROUNDS`. |
| 76 | pub const MAX_TOOL_ROUNDS_LIMIT: usize = 500; |
| 77 | |
| 78 | /// Environment override for the tool-round cap of a headless run or an ACP |
| 79 | /// prompt turn. |
| 80 | pub const MAX_TOOL_ROUNDS_ENV: &str = "SIGIT_MAX_TOOL_ROUNDS"; |
| 81 | |
| 82 | fn parse_tool_rounds(value: &str, source: &str) -> Result<usize, String> { |
| 83 | match value.trim().parse::<usize>() { |
| 84 | Ok(rounds) if (1..=MAX_TOOL_ROUNDS_LIMIT).contains(&rounds) => Ok(rounds), |
| 85 | _ => Err(format!( |
| 86 | "{source} must be a whole number from 1 to {MAX_TOOL_ROUNDS_LIMIT}" |
| 87 | )), |
| 88 | } |
| 89 | } |
| 90 | |
| 91 | /// The tool-round cap for a run: the flag wins, then the environment, then |
| 92 | /// the built-in cap. An empty environment value counts as unset. |
| 93 | fn resolve_max_tool_rounds(flag: Option<usize>, env: Option<&str>) -> Result<usize, String> { |
| 94 | if let Some(rounds) = flag { |
| 95 | return Ok(rounds); |
| 96 | } |
| 97 | match env.map(str::trim).filter(|value| !value.is_empty()) { |
| 98 | Some(value) => parse_tool_rounds(value, MAX_TOOL_ROUNDS_ENV), |
| 99 | None => Ok(crate::MAX_TOOL_ROUNDS), |
| 100 | } |
| 101 | } |
| 102 | |
| 103 | /// The tool-round cap for a surface with no flag of its own (an ACP prompt |
| 104 | /// turn). A value that does not parse is logged and ignored: the editor |
| 105 | /// launched the agent, so there is nobody to show a startup error to. |
| 106 | pub fn max_tool_rounds_from_env() -> usize { |
| 107 | resolve_max_tool_rounds(None, std::env::var(MAX_TOOL_ROUNDS_ENV).ok().as_deref()) |
| 108 | .unwrap_or_else(|error| { |
| 109 | log::warn!( |
| 110 | "{error}; using the built-in cap of {}", |
| 111 | crate::MAX_TOOL_ROUNDS |
| 112 | ); |
| 113 | crate::MAX_TOOL_ROUNDS |
| 114 | }) |
| 115 | } |
| 116 | |
| 117 | /// What the model is told when a run would otherwise end with no final |
| 118 | /// message. `cap_reached` says the tool-round cap is why. |
| 119 | fn wrap_up_prompt(cap_reached: bool, max_tool_rounds: usize) -> String { |
| 120 | let lead = if cap_reached { |
| 121 | format!("You have used all {max_tool_rounds} tool rounds available for this run. ") |
| 122 | } else { |
| 123 | String::new() |
| 124 | }; |
| 125 | format!( |
| 126 | "{lead}Write your final message now, without calling any tools: what you did, \ |
| 127 | what you found, and anything left unfinished." |
| 128 | ) |
| 129 | } |
| 130 | |
| 131 | /// Parse the process arguments (without argv[0]) for headless mode. |
| 132 | /// |
| 133 | /// Returns `Ok(None)` when neither `run` nor `-p`/`--prompt` is present — the |
| 134 | /// invocation is not headless and falls through to the TTY/ACP dispatch. Once |
| 135 | /// headless mode is selected, |
| 136 | /// every remaining argument must be a recognized flag (position-insensitive); |
| 137 | /// anything else is a usage error the caller reports on stderr with exit 2. |
| 138 | pub fn parse_args(args: &[String]) -> Result<Option<HeadlessConfig>, String> { |
| 139 | let run_command = args.first().is_some_and(|arg| arg == "run"); |
| 140 | if !run_command && !args.iter().any(|arg| arg == "-p" || arg == "--prompt") { |
| 141 | return Ok(None); |
| 142 | } |
| 143 | |
| 144 | let mut prompt: Option<String> = None; |
| 145 | let mut cwd: Option<PathBuf> = None; |
| 146 | let mut add_dirs: Vec<PathBuf> = Vec::new(); |
| 147 | let mut quiet = false; |
| 148 | let mut output = OutputFormat::Text; |
| 149 | let mut resume_session: Option<String> = None; |
| 150 | let mut allow_tools: Vec<String> = Vec::new(); |
| 151 | let mut deny_tools: Vec<String> = Vec::new(); |
| 152 | let mut max_tool_rounds: Option<usize> = None; |
| 153 | let mut cloud_tier: Option<&'static str> = None; |
| 154 | |
| 155 | let mut iter = args.iter().skip(usize::from(run_command)); |
| 156 | while let Some(arg) = iter.next() { |
| 157 | match arg.as_str() { |
| 158 | "-p" | "--prompt" => { |
| 159 | let value = iter |
| 160 | .next() |
| 161 | .ok_or_else(|| format!("{arg} requires a prompt argument"))?; |
| 162 | if prompt.is_some() { |
| 163 | return Err("the prompt was given more than once".to_string()); |
| 164 | } |
| 165 | prompt = Some(value.clone()); |
| 166 | } |
| 167 | "--cwd" => { |
| 168 | let value = iter |
| 169 | .next() |
| 170 | .ok_or_else(|| "--cwd requires a directory argument".to_string())?; |
| 171 | cwd = Some(PathBuf::from(value)); |
| 172 | } |
| 173 | "--add-dir" => { |
| 174 | let value = iter |
| 175 | .next() |
| 176 | .ok_or_else(|| "--add-dir requires a directory argument".to_string())?; |
| 177 | add_dirs.push(PathBuf::from(value)); |
| 178 | } |
| 179 | "--quiet" => quiet = true, |
| 180 | "--output" => { |
| 181 | let value = iter |
| 182 | .next() |
| 183 | .ok_or_else(|| "--output requires text or jsonl".to_string())?; |
| 184 | output = match value.as_str() { |
| 185 | "text" => OutputFormat::Text, |
| 186 | "jsonl" => OutputFormat::Jsonl, |
| 187 | _ => return Err("--output must be text or jsonl".to_string()), |
| 188 | }; |
| 189 | } |
| 190 | "--resume" => { |
| 191 | let value = iter |
| 192 | .next() |
| 193 | .ok_or_else(|| "--resume requires a session id".to_string())?; |
| 194 | if resume_session.is_some() { |
| 195 | return Err("--resume was given more than once".to_string()); |
| 196 | } |
| 197 | validate_session_id(value)?; |
| 198 | resume_session = Some(value.clone()); |
| 199 | } |
| 200 | "--allow-tool" => { |
| 201 | let value = iter |
| 202 | .next() |
| 203 | .ok_or_else(|| "--allow-tool requires a tool name".to_string())?; |
| 204 | allow_tools.push(value.clone()); |
| 205 | } |
| 206 | "--deny-tool" => { |
| 207 | let value = iter |
| 208 | .next() |
| 209 | .ok_or_else(|| "--deny-tool requires a tool name".to_string())?; |
| 210 | deny_tools.push(value.clone()); |
| 211 | } |
| 212 | "--max-tool-rounds" => { |
| 213 | let value = iter |
| 214 | .next() |
| 215 | .ok_or_else(|| "--max-tool-rounds requires a number".to_string())?; |
| 216 | max_tool_rounds = Some(parse_tool_rounds(value, "--max-tool-rounds")?); |
| 217 | } |
| 218 | "--model" => { |
| 219 | let value = iter |
| 220 | .next() |
| 221 | .ok_or_else(|| "--model requires a cloud tier".to_string())?; |
| 222 | if cloud_tier.is_some() { |
| 223 | return Err("--model was given more than once".to_string()); |
| 224 | } |
| 225 | let tier = provider::cloud_tier_named(value).ok_or_else(|| { |
| 226 | format!( |
| 227 | "unknown cloud tier: {value} (expected one of: {})", |
| 228 | provider::CLOUD_TIERS.join(", ") |
| 229 | ) |
| 230 | })?; |
| 231 | cloud_tier = Some(tier); |
| 232 | } |
| 233 | // A mistyped flag must not become the prompt. A prompt that |
| 234 | // really starts with '-' can still go through -p. |
| 235 | other if other.starts_with('-') => return Err(format!("unknown argument: {other}")), |
| 236 | other if run_command && prompt.is_none() => prompt = Some(other.to_string()), |
| 237 | other if run_command => { |
| 238 | return Err(format!( |
| 239 | "unexpected extra argument: {other} (quote the prompt as one argument)" |
| 240 | )); |
| 241 | } |
| 242 | other => return Err(format!("unknown argument: {other}")), |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | // Reaching this without a prompt means `run` had no positional prompt, or |
| 247 | // a legacy prompt flag was consumed as another flag's value. |
| 248 | let prompt = prompt.ok_or_else(|| { |
| 249 | if run_command { |
| 250 | "missing prompt".to_string() |
| 251 | } else { |
| 252 | "missing -p/--prompt".to_string() |
| 253 | } |
| 254 | })?; |
| 255 | if prompt.trim().is_empty() { |
| 256 | return Err("the prompt must not be empty".to_string()); |
| 257 | } |
| 258 | if quiet && output == OutputFormat::Jsonl { |
| 259 | return Err("--quiet cannot be combined with --output jsonl".to_string()); |
| 260 | } |
| 261 | |
| 262 | let resume = resume_session.is_some(); |
| 263 | let session_id = resume_session.unwrap_or_else(|| uuid::Uuid::new_v4().to_string()); |
| 264 | let max_tool_rounds = resolve_max_tool_rounds( |
| 265 | max_tool_rounds, |
| 266 | std::env::var(MAX_TOOL_ROUNDS_ENV).ok().as_deref(), |
| 267 | )?; |
| 268 | |
| 269 | Ok(Some(HeadlessConfig { |
| 270 | prompt, |
| 271 | session_id, |
| 272 | resume, |
| 273 | cwd, |
| 274 | add_dirs, |
| 275 | quiet, |
| 276 | output, |
| 277 | allow_tools, |
| 278 | deny_tools, |
| 279 | max_tool_rounds, |
| 280 | cloud_tier, |
| 281 | })) |
| 282 | } |
| 283 | |
| 284 | fn validate_session_id(session_id: &str) -> Result<(), String> { |
| 285 | if session_id.is_empty() |
| 286 | || !session_id.chars().all(|character| { |
| 287 | character.is_ascii_alphanumeric() || matches!(character, '.' | '_' | '-') |
| 288 | }) |
| 289 | { |
| 290 | return Err( |
| 291 | "session ids may contain only ASCII letters, numbers, '.', '_', and '-'".to_string(), |
| 292 | ); |
| 293 | } |
| 294 | Ok(()) |
| 295 | } |
| 296 | |
| 297 | /// Denial fed to the model when a tool at `ask` level fires in a headless run. |
| 298 | fn not_preapproved_denial(tool_name: &str) -> String { |
| 299 | format!( |
| 300 | "`{tool_name}` was not executed: this is a non-interactive headless run and \ |
| 301 | nobody can answer a permission prompt, so tools at the `ask` level are denied \ |
| 302 | unless pre-approved. The user can re-run with `--allow-tool {tool_name}` to \ |
| 303 | approve it (or set SIGIT_PERMISSIONS=allow). Do not retry the same call; \ |
| 304 | continue without this tool or report what remains to be done." |
| 305 | ) |
| 306 | } |
| 307 | |
| 308 | /// Denial fed to the model when a tool was blocked with `--deny-tool`. |
| 309 | fn deny_flag_denial(tool_name: &str) -> String { |
| 310 | format!( |
| 311 | "`{tool_name}` is blocked for this headless run (--deny-tool). Do not retry it; \ |
| 312 | continue without this tool or report what remains to be done." |
| 313 | ) |
| 314 | } |
| 315 | |
| 316 | /// Run one headless prompt to completion. Returns the process exit code. |
| 317 | /// |
| 318 | /// The caller (`main`) has already applied `--cwd` and `--add-dir`, initialized |
| 319 | /// logging to stderr, set up the model cache, and run MCP discovery. |
| 320 | pub async fn run(config: HeadlessConfig) -> i32 { |
| 321 | // Fresh permission state for the run, then apply the flag grants. |
| 322 | permissions::reset_session(&config.session_id); |
| 323 | tools::set_active_session(Some(&config.session_id)); |
| 324 | for tool in &config.allow_tools { |
| 325 | // No call arguments at grant time: the empty string records the bare |
| 326 | // tool name, granting the whole tool — the --allow-tool contract. |
| 327 | permissions::grant_for_session(&config.session_id, tool, ""); |
| 328 | } |
| 329 | let denied: HashSet<&str> = config.deny_tools.iter().map(String::as_str).collect(); |
| 330 | |
| 331 | // Backend resolution mirrors the ACP server: the explicit provider override |
| 332 | // first, else the signed-in cloud tier when local inference is off (what |
| 333 | // `apply_startup_inference_mode` does at every ACP session entry). There is |
| 334 | // never an implicit on-device load — a fresh process has no model in memory |
| 335 | // and headless mode must not silently download gigabytes. `--model` is the |
| 336 | // exception to that order: a tier named on the command line wins over the |
| 337 | // override and the setting, the way a flag wins over the environment. |
| 338 | let provider_cfg = match config.cloud_tier { |
| 339 | Some(tier) => provider::cloud_tier_provider(tier), |
| 340 | None => provider::active_provider().or_else(|| { |
| 341 | if settings::local_inference_enabled() { |
| 342 | None |
| 343 | } else { |
| 344 | provider::cloud_tier_provider(provider::DEFAULT_CLOUD_TIER) |
| 345 | } |
| 346 | }), |
| 347 | }; |
| 348 | let Some(cfg) = provider_cfg else { |
| 349 | let message = match config.cloud_tier { |
| 350 | Some(tier) => format!( |
| 351 | "--model {tier} needs a siGit Code Cloud session. Sign in with `sigit login`." |
| 352 | ), |
| 353 | None => "headless mode needs a remote inference provider — running on-device would \ |
| 354 | require loading (and possibly downloading) a local model, which headless \ |
| 355 | mode never does implicitly. Set OPENAI_BASE_URL and OPENAI_API_KEY (or \ |
| 356 | configure providers.toml), or sign in with `sigit login` and turn local \ |
| 357 | inference off." |
| 358 | .to_string(), |
| 359 | }; |
| 360 | emit_error(&config, &message); |
| 361 | return 1; |
| 362 | }; |
| 363 | |
| 364 | log::info!( |
| 365 | "headless: using {} (model {}) at {}", |
| 366 | cfg.display_name, |
| 367 | cfg.model, |
| 368 | cfg.base_url |
| 369 | ); |
| 370 | crate::register_subagent_factory_for(&cfg); |
| 371 | |
| 372 | // Same always-on project context the other surfaces inject: cwd guidance |
| 373 | // plus AGENTS.md / CLAUDE.md instruction files. |
| 374 | let cwd = std::env::current_dir().unwrap_or_else(|_| PathBuf::from(".")); |
| 375 | let mut system_prompt = crate::system_prompt_for_model(true); |
| 376 | system_prompt.push_str("\n\n"); |
| 377 | system_prompt.push_str(&crate::session_context_message( |
| 378 | &cwd, |
| 379 | &crate::workspace::additional_roots(), |
| 380 | )); |
| 381 | |
| 382 | let backend: Arc<dyn InferenceBackend> = Arc::new( |
| 383 | OpenAiBackend::new( |
| 384 | cfg.base_url, |
| 385 | cfg.api_key, |
| 386 | cfg.model.clone(), |
| 387 | Some(system_prompt), |
| 388 | ) |
| 389 | .with_session_token(cfg.session_token), |
| 390 | ); |
| 391 | |
| 392 | if config.resume { |
| 393 | let Some(history) = session_store::load(&config.session_id) else { |
| 394 | emit_error(&config, "saved session not found"); |
| 395 | return 1; |
| 396 | }; |
| 397 | backend::adopt_carryover(backend.as_ref(), backend::carryover_history(history)).await; |
| 398 | } |
| 399 | |
| 400 | emit_event( |
| 401 | &config, |
| 402 | serde_json::json!({ |
| 403 | "type": "session", |
| 404 | "session_id": config.session_id, |
| 405 | "resumed": config.resume, |
| 406 | }), |
| 407 | ); |
| 408 | if config.output == OutputFormat::Text { |
| 409 | eprintln!("Session: {}", config.session_id); |
| 410 | } |
| 411 | |
| 412 | let outcome = run_prompt(backend.as_ref(), &config, &denied).await; |
| 413 | |
| 414 | // Persist the conversation like the other surfaces, so a follow-up feature |
| 415 | // can resume it. Saved even on error: a partial transcript beats none. |
| 416 | let snapshot = backend.history_snapshot().await; |
| 417 | if let Err(error) = session_store::save(&config.session_id, &snapshot) { |
| 418 | log::warn!("headless: session save failed: {error}"); |
| 419 | } |
| 420 | session_store::save_meta( |
| 421 | &config.session_id, |
| 422 | &cwd, |
| 423 | &crate::workspace::additional_roots(), |
| 424 | Some(&cfg.model), |
| 425 | ); |
| 426 | |
| 427 | match outcome { |
| 428 | Ok(()) => 0, |
| 429 | Err(error) => { |
| 430 | emit_error(&config, &error.to_string()); |
| 431 | 1 |
| 432 | } |
| 433 | } |
| 434 | } |
| 435 | |
| 436 | fn emit_event(config: &HeadlessConfig, event: serde_json::Value) { |
| 437 | if config.output == OutputFormat::Jsonl { |
| 438 | println!("{event}"); |
| 439 | let _ = std::io::stdout().flush(); |
| 440 | } |
| 441 | } |
| 442 | |
| 443 | fn emit_error(config: &HeadlessConfig, message: &str) { |
| 444 | if config.output == OutputFormat::Jsonl { |
| 445 | emit_event( |
| 446 | config, |
| 447 | serde_json::json!({ |
| 448 | "type": "error", |
| 449 | "session_id": config.session_id, |
| 450 | "message": message, |
| 451 | }), |
| 452 | ); |
| 453 | } else { |
| 454 | eprintln!("sigit: {message}"); |
| 455 | } |
| 456 | } |
| 457 | |
| 458 | /// The turn loop: send the prompt, execute tool calls under the permission |
| 459 | /// policy, feed results back, repeat up to `config.max_tool_rounds` — the same |
| 460 | /// shape as the ACP `handle_prompt`, minus the ACP notifications. |
| 461 | async fn run_prompt( |
| 462 | backend: &dyn InferenceBackend, |
| 463 | config: &HeadlessConfig, |
| 464 | denied: &HashSet<&str>, |
| 465 | ) -> Result<(), backend::BackendError> { |
| 466 | let tools = crate::agent_tools_as_specs(); |
| 467 | |
| 468 | // Token sink: assistant text streams through this while a turn runs; the |
| 469 | // drain loop forwards the visible portion to stdout live. In quiet mode no |
| 470 | // sink is passed and only the final message is printed. |
| 471 | let (sink, mut sink_rx) = tokio::sync::mpsc::unbounded_channel::<TokenChunk>(); |
| 472 | let sink_opt = if config.quiet { None } else { Some(&sink) }; |
| 473 | let mut reply = crate::StreamedReply::default(); |
| 474 | |
| 475 | let mut result = drain_to_stdout( |
| 476 | backend.send_message_with_tools(&config.prompt, &tools, sink_opt), |
| 477 | &mut sink_rx, |
| 478 | &mut reply, |
| 479 | config, |
| 480 | ) |
| 481 | .await?; |
| 482 | |
| 483 | let mut round = 0; |
| 484 | |
| 485 | while !result.tool_calls.is_empty() && round < config.max_tool_rounds { |
| 486 | round += 1; |
| 487 | |
| 488 | // Auto-compaction: long tool runs grow history fast; fold it into a |
| 489 | // summary before the next round rather than blowing the window. |
| 490 | let estimate = backend::estimate_tokens(&backend.history_snapshot().await); |
| 491 | if estimate > backend::DEFAULT_CONTEXT_TOKEN_BUDGET { |
| 492 | log::info!( |
| 493 | "headless: history ≈{estimate} tokens exceeds budget {} — compacting", |
| 494 | backend::DEFAULT_CONTEXT_TOKEN_BUDGET |
| 495 | ); |
| 496 | if let Err(error) = backend.compact_history(backend::COMPACT_KEEP_LAST).await { |
| 497 | log::warn!("headless: compaction failed: {error}"); |
| 498 | } |
| 499 | } |
| 500 | |
| 501 | let mut tool_results = Vec::new(); |
| 502 | |
| 503 | for tc in &result.tool_calls { |
| 504 | let args_preview: String = tc.arguments.chars().take(120).collect(); |
| 505 | if config.output == OutputFormat::Jsonl { |
| 506 | emit_event( |
| 507 | config, |
| 508 | serde_json::json!({ |
| 509 | "type": "tool_call", |
| 510 | "session_id": config.session_id, |
| 511 | "tool_call_id": tc.id, |
| 512 | "name": tc.name, |
| 513 | "arguments": tc.arguments, |
| 514 | }), |
| 515 | ); |
| 516 | } else { |
| 517 | eprintln!("→ {}({args_preview})", tc.name); |
| 518 | } |
| 519 | |
| 520 | // The headless deny set outranks everything, including a |
| 521 | // settings-level allow and a --allow-tool grant for the same name. |
| 522 | let output = if denied.contains(tc.name.as_str()) { |
| 523 | log::info!("headless: {} blocked by --deny-tool", tc.name); |
| 524 | deny_flag_denial(&tc.name) |
| 525 | } else { |
| 526 | match permissions::decision_for(&config.session_id, &tc.name, &tc.arguments) { |
| 527 | permissions::Decision::Allow => { |
| 528 | tools::execute_tool(&tc.name, &tc.arguments).await |
| 529 | } |
| 530 | permissions::Decision::Deny(reason) => { |
| 531 | log::info!("headless: {} denied by policy", tc.name); |
| 532 | reason |
| 533 | } |
| 534 | // Nobody can answer an interactive prompt here: ask |
| 535 | // collapses to a denial pointing at --allow-tool. |
| 536 | permissions::Decision::Ask => { |
| 537 | log::info!("headless: {} not pre-approved — denied", tc.name); |
| 538 | not_preapproved_denial(&tc.name) |
| 539 | } |
| 540 | } |
| 541 | }; |
| 542 | |
| 543 | emit_event( |
| 544 | config, |
| 545 | serde_json::json!({ |
| 546 | "type": "tool_result", |
| 547 | "session_id": config.session_id, |
| 548 | "tool_call_id": tc.id, |
| 549 | "name": tc.name, |
| 550 | "content": output, |
| 551 | }), |
| 552 | ); |
| 553 | |
| 554 | tool_results.push(BackendToolResult { |
| 555 | tool_call_id: tc.id.clone(), |
| 556 | content: output, |
| 557 | }); |
| 558 | } |
| 559 | |
| 560 | let allow_tool_calls = round < config.max_tool_rounds; |
| 561 | |
| 562 | // The round that reaches the cap is forced to answer in text. Say so, |
| 563 | // or the model spends its last reply narrating the step it can no |
| 564 | // longer take. |
| 565 | if !allow_tool_calls && let Some(last) = tool_results.last_mut() { |
| 566 | last.content |
| 567 | .push_str(&crate::headless_round_cap_note(config.max_tool_rounds)); |
| 568 | } |
| 569 | |
| 570 | // Whatever this round says starts a new paragraph rather than |
| 571 | // continuing the sentence the tool calls interrupted. |
| 572 | reply.interrupt(); |
| 573 | |
| 574 | result = drain_to_stdout( |
| 575 | backend.send_tool_results(tool_results, &tools, allow_tool_calls, sink_opt), |
| 576 | &mut sink_rx, |
| 577 | &mut reply, |
| 578 | config, |
| 579 | ) |
| 580 | .await?; |
| 581 | } |
| 582 | |
| 583 | // A run can reach this point with nothing to show: the last round was |
| 584 | // forced to answer in text and the model answered with nothing, most |
| 585 | // often because it was cut off at the tool-round cap mid-task. A caller |
| 586 | // reading stdout could not tell that from a run with nothing to report, |
| 587 | // so ask once, plainly, for the final message. |
| 588 | if crate::chat::strip_think_blocks(result.text.trim()) |
| 589 | .1 |
| 590 | .is_empty() |
| 591 | { |
| 592 | let cap_reached = round >= config.max_tool_rounds; |
| 593 | if cap_reached { |
| 594 | log::warn!( |
| 595 | "headless: reached the cap of {} tool round(s) without a final message", |
| 596 | config.max_tool_rounds |
| 597 | ); |
| 598 | } |
| 599 | reply.interrupt(); |
| 600 | match drain_to_stdout( |
| 601 | backend.send_message_with_tools( |
| 602 | &wrap_up_prompt(cap_reached, config.max_tool_rounds), |
| 603 | &[], |
| 604 | sink_opt, |
| 605 | ), |
| 606 | &mut sink_rx, |
| 607 | &mut reply, |
| 608 | config, |
| 609 | ) |
| 610 | .await |
| 611 | { |
| 612 | Ok(wrap_up) => result = wrap_up, |
| 613 | Err(error) => log::warn!("headless: asking for a final message failed: {error}"), |
| 614 | } |
| 615 | } |
| 616 | |
| 617 | // Final text: in quiet mode nothing streamed, so print the final assistant |
| 618 | // message now; otherwise print only what streaming did not already cover. |
| 619 | let (_think, final_visible) = crate::chat::strip_think_blocks(result.text.trim()); |
| 620 | let mut stdout = std::io::stdout(); |
| 621 | if config.output == OutputFormat::Jsonl { |
| 622 | emit_event( |
| 623 | config, |
| 624 | serde_json::json!({ |
| 625 | "type": "result", |
| 626 | "session_id": config.session_id, |
| 627 | "text": final_visible, |
| 628 | "tool_rounds": round, |
| 629 | }), |
| 630 | ); |
| 631 | } else if config.quiet { |
| 632 | if !final_visible.is_empty() { |
| 633 | let _ = writeln!(stdout, "{final_visible}"); |
| 634 | } |
| 635 | } else if reply.streamed_any { |
| 636 | // The reply is already on stdout; end the line for the shell. |
| 637 | let _ = writeln!(stdout); |
| 638 | } else if !final_visible.is_empty() { |
| 639 | let _ = writeln!(stdout, "{final_visible}"); |
| 640 | } |
| 641 | let _ = stdout.flush(); |
| 642 | |
| 643 | log::info!("headless: prompt complete — {round} tool round(s)"); |
| 644 | Ok(()) |
| 645 | } |
| 646 | |
| 647 | /// Run one inference turn while forwarding streamed tokens to stdout as they |
| 648 | /// arrive — the stdio counterpart of `SiGitAgent::drain_turn`. |
| 649 | async fn drain_to_stdout<F>( |
| 650 | fut: F, |
| 651 | sink_rx: &mut tokio::sync::mpsc::UnboundedReceiver<TokenChunk>, |
| 652 | reply: &mut crate::StreamedReply, |
| 653 | config: &HeadlessConfig, |
| 654 | ) -> Result<TurnResult, backend::BackendError> |
| 655 | where |
| 656 | F: std::future::Future<Output = Result<TurnResult, backend::BackendError>>, |
| 657 | { |
| 658 | tokio::pin!(fut); |
| 659 | let result = loop { |
| 660 | tokio::select! { |
| 661 | done = &mut fut => break done, |
| 662 | Some(chunk) = sink_rx.recv() => { |
| 663 | emit_token_chunk(chunk, reply, config); |
| 664 | } |
| 665 | } |
| 666 | }; |
| 667 | // Flush tokens that landed between the last poll and the future resolving. |
| 668 | while let Ok(chunk) = sink_rx.try_recv() { |
| 669 | emit_token_chunk(chunk, reply, config); |
| 670 | } |
| 671 | result |
| 672 | } |
| 673 | |
| 674 | /// Fold a streamed fragment into `reply` and print whatever it newly reveals — |
| 675 | /// the stdio counterpart of `SiGitAgent::emit_token_chunk`. Reasoning (from a |
| 676 | /// dedicated field or an inline `<think>` block) never joins the stdout answer |
| 677 | /// stream; under `--output jsonl` it surfaces as a separate `reasoning_delta` |
| 678 | /// event, and is dropped in plain mode. |
| 679 | fn emit_token_chunk(chunk: TokenChunk, reply: &mut crate::StreamedReply, config: &HeadlessConfig) { |
| 680 | match chunk { |
| 681 | TokenChunk::Reasoning(text) => emit_reasoning_delta(&text, config), |
| 682 | TokenChunk::Visible(piece) => { |
| 683 | let revealed = reply.push(&piece); |
| 684 | if let Some(reasoning) = revealed.reasoning { |
| 685 | emit_reasoning_delta(&reasoning, config); |
| 686 | } |
| 687 | if let Some(extra) = revealed.visible { |
| 688 | if config.output == OutputFormat::Jsonl { |
| 689 | emit_event( |
| 690 | config, |
| 691 | serde_json::json!({ |
| 692 | "type": "assistant_delta", |
| 693 | "session_id": config.session_id, |
| 694 | "text": extra, |
| 695 | }), |
| 696 | ); |
| 697 | } else { |
| 698 | print!("{extra}"); |
| 699 | } |
| 700 | let _ = std::io::stdout().flush(); |
| 701 | } |
| 702 | } |
| 703 | } |
| 704 | } |
| 705 | |
| 706 | /// Emit model reasoning as a structured event under `--output jsonl`. In plain |
| 707 | /// mode reasoning is not part of the answer, so it is dropped. |
| 708 | fn emit_reasoning_delta(text: &str, config: &HeadlessConfig) { |
| 709 | if config.output == OutputFormat::Jsonl { |
| 710 | emit_event( |
| 711 | config, |
| 712 | serde_json::json!({ |
| 713 | "type": "reasoning_delta", |
| 714 | "session_id": config.session_id, |
| 715 | "text": text, |
| 716 | }), |
| 717 | ); |
| 718 | } |
| 719 | } |
| 720 | |
| 721 | #[cfg(test)] |
| 722 | mod tests { |
| 723 | use super::*; |
| 724 | |
| 725 | fn args(list: &[&str]) -> Vec<String> { |
| 726 | list.iter().map(|s| s.to_string()).collect() |
| 727 | } |
| 728 | |
| 729 | #[test] |
| 730 | fn absent_prompt_flag_is_not_headless() { |
| 731 | assert_eq!(parse_args(&args(&[])), Ok(None)); |
| 732 | assert_eq!(parse_args(&args(&["login"])), Ok(None)); |
| 733 | // Flags alone don't trigger headless mode; only -p/--prompt does. |
| 734 | assert_eq!(parse_args(&args(&["--quiet"])), Ok(None)); |
| 735 | } |
| 736 | |
| 737 | #[test] |
| 738 | fn parses_minimal_invocation() { |
| 739 | let config = parse_args(&args(&["-p", "do the thing"])).unwrap().unwrap(); |
| 740 | assert_eq!(config.prompt, "do the thing"); |
| 741 | assert!(!config.session_id.is_empty()); |
| 742 | assert!(!config.resume); |
| 743 | assert_eq!(config.cwd, None); |
| 744 | assert!(!config.quiet); |
| 745 | assert_eq!(config.output, OutputFormat::Text); |
| 746 | assert!(config.allow_tools.is_empty()); |
| 747 | assert!(config.deny_tools.is_empty()); |
| 748 | } |
| 749 | |
| 750 | #[test] |
| 751 | fn run_subcommand_accepts_a_positional_prompt() { |
| 752 | let config = parse_args(&args(&["run", "do the thing"])) |
| 753 | .unwrap() |
| 754 | .unwrap(); |
| 755 | assert_eq!(config.prompt, "do the thing"); |
| 756 | assert!(!config.resume); |
| 757 | } |
| 758 | |
| 759 | #[test] |
| 760 | fn long_form_prompt_flag_works() { |
| 761 | let config = parse_args(&args(&["--prompt", "hello"])).unwrap().unwrap(); |
| 762 | assert_eq!(config.prompt, "hello"); |
| 763 | } |
| 764 | |
| 765 | #[test] |
| 766 | fn parses_all_flags_position_insensitively() { |
| 767 | let config = parse_args(&args(&[ |
| 768 | "--quiet", |
| 769 | "--allow-tool", |
| 770 | "run_command", |
| 771 | "--cwd", |
| 772 | "/tmp/project", |
| 773 | "-p", |
| 774 | "build it", |
| 775 | "--deny-tool", |
| 776 | "delete_file", |
| 777 | "--output", |
| 778 | "text", |
| 779 | "--allow-tool", |
| 780 | "edit_file", |
| 781 | ])) |
| 782 | .unwrap() |
| 783 | .unwrap(); |
| 784 | assert_eq!(config.prompt, "build it"); |
| 785 | assert_eq!(config.cwd, Some(PathBuf::from("/tmp/project"))); |
| 786 | assert!(config.quiet); |
| 787 | assert_eq!(config.output, OutputFormat::Text); |
| 788 | assert_eq!(config.allow_tools, vec!["run_command", "edit_file"]); |
| 789 | assert_eq!(config.deny_tools, vec!["delete_file"]); |
| 790 | } |
| 791 | |
| 792 | #[test] |
| 793 | fn resume_reuses_the_named_session() { |
| 794 | let config = parse_args(&args(&[ |
| 795 | "run", |
| 796 | "continue", |
| 797 | "--resume", |
| 798 | "session-123", |
| 799 | "--output", |
| 800 | "jsonl", |
| 801 | ])) |
| 802 | .unwrap() |
| 803 | .unwrap(); |
| 804 | assert_eq!(config.session_id, "session-123"); |
| 805 | assert!(config.resume); |
| 806 | assert_eq!(config.output, OutputFormat::Jsonl); |
| 807 | } |
| 808 | |
| 809 | #[test] |
| 810 | fn resume_rejects_unsafe_session_ids() { |
| 811 | let error = parse_args(&args(&["run", "continue", "--resume", "../escape"])).unwrap_err(); |
| 812 | assert!(error.contains("session ids"), "{error}"); |
| 813 | } |
| 814 | |
| 815 | #[test] |
| 816 | fn max_tool_rounds_flag_is_parsed_and_bounded() { |
| 817 | let config = parse_args(&args(&["-p", "work", "--max-tool-rounds", "60"])) |
| 818 | .unwrap() |
| 819 | .unwrap(); |
| 820 | assert_eq!(config.max_tool_rounds, 60); |
| 821 | |
| 822 | for bad in ["0", "501", "-1", "many", ""] { |
| 823 | let error = parse_args(&args(&["-p", "work", "--max-tool-rounds", bad])).unwrap_err(); |
| 824 | assert!(error.contains("--max-tool-rounds"), "{bad}: {error}"); |
| 825 | } |
| 826 | assert!(parse_args(&args(&["-p", "work", "--max-tool-rounds"])).is_err()); |
| 827 | } |
| 828 | |
| 829 | #[test] |
| 830 | fn max_tool_rounds_resolves_flag_then_environment_then_default() { |
| 831 | assert_eq!(resolve_max_tool_rounds(Some(7), Some("60")), Ok(7)); |
| 832 | assert_eq!(resolve_max_tool_rounds(None, Some(" 60 ")), Ok(60)); |
| 833 | assert_eq!( |
| 834 | resolve_max_tool_rounds(None, None), |
| 835 | Ok(crate::MAX_TOOL_ROUNDS) |
| 836 | ); |
| 837 | // An empty variable is unset, not an error. |
| 838 | assert_eq!( |
| 839 | resolve_max_tool_rounds(None, Some("")), |
| 840 | Ok(crate::MAX_TOOL_ROUNDS) |
| 841 | ); |
| 842 | let error = resolve_max_tool_rounds(None, Some("lots")).unwrap_err(); |
| 843 | assert!(error.contains(MAX_TOOL_ROUNDS_ENV), "{error}"); |
| 844 | assert!(resolve_max_tool_rounds(None, Some("0")).is_err()); |
| 845 | } |
| 846 | |
| 847 | #[test] |
| 848 | fn model_flag_names_a_cloud_tier() { |
| 849 | let config = parse_args(&args(&["run", "go"])).unwrap().unwrap(); |
| 850 | assert_eq!(config.cloud_tier, None); |
| 851 | let config = parse_args(&args(&["run", "go", "--model", "prism"])) |
| 852 | .unwrap() |
| 853 | .unwrap(); |
| 854 | assert_eq!(config.cloud_tier, Some("prism")); |
| 855 | // The wire id works too, so a model id copied from logs is accepted. |
| 856 | let config = parse_args(&args(&["-p", "go", "--model", "onde-zenith"])) |
| 857 | .unwrap() |
| 858 | .unwrap(); |
| 859 | assert_eq!(config.cloud_tier, Some("zenith")); |
| 860 | } |
| 861 | |
| 862 | #[test] |
| 863 | fn model_flag_rejects_unknown_and_repeated_tiers() { |
| 864 | let error = parse_args(&args(&["run", "go", "--model", "gpt-4o"])).unwrap_err(); |
| 865 | assert!(error.starts_with("unknown cloud tier: gpt-4o"), "{error}"); |
| 866 | assert!(error.contains("prism"), "{error}"); |
| 867 | assert_eq!( |
| 868 | parse_args(&args(&["run", "go", "--model", "apex", "--model", "flare"])), |
| 869 | Err("--model was given more than once".to_string()) |
| 870 | ); |
| 871 | } |
| 872 | |
| 873 | #[test] |
| 874 | fn wrap_up_prompt_names_the_cap_only_when_it_was_reached() { |
| 875 | let capped = wrap_up_prompt(true, 24); |
| 876 | assert!(capped.contains("all 24 tool rounds"), "{capped}"); |
| 877 | assert!(capped.contains("without calling any tools")); |
| 878 | |
| 879 | let plain = wrap_up_prompt(false, 24); |
| 880 | assert!(!plain.contains("tool rounds"), "{plain}"); |
| 881 | assert!(plain.starts_with("Write your final message now")); |
| 882 | } |
| 883 | |
| 884 | #[test] |
| 885 | fn quiet_and_jsonl_are_mutually_exclusive() { |
| 886 | let error = |
| 887 | parse_args(&args(&["run", "work", "--quiet", "--output", "jsonl"])).unwrap_err(); |
| 888 | assert!(error.contains("cannot be combined"), "{error}"); |
| 889 | } |
| 890 | |
| 891 | #[test] |
| 892 | fn add_dir_is_repeatable_and_keeps_its_order() { |
| 893 | let config = parse_args(&args(&[ |
| 894 | "-p", |
| 895 | "look around", |
| 896 | "--add-dir", |
| 897 | "/tmp/lib", |
| 898 | "--add-dir", |
| 899 | "/tmp/docs", |
| 900 | ])) |
| 901 | .unwrap() |
| 902 | .unwrap(); |
| 903 | assert_eq!( |
| 904 | config.add_dirs, |
| 905 | vec![PathBuf::from("/tmp/lib"), PathBuf::from("/tmp/docs")] |
| 906 | ); |
| 907 | } |
| 908 | |
| 909 | #[test] |
| 910 | fn add_dir_without_a_value_is_a_usage_error() { |
| 911 | assert!(parse_args(&args(&["-p", "x", "--add-dir"])).is_err()); |
| 912 | } |
| 913 | |
| 914 | #[test] |
| 915 | fn prompt_value_may_look_like_a_flag() { |
| 916 | // -p consumes the next argument verbatim. |
| 917 | let config = parse_args(&args(&["-p", "--quiet"])).unwrap().unwrap(); |
| 918 | assert_eq!(config.prompt, "--quiet"); |
| 919 | assert!(!config.quiet); |
| 920 | } |
| 921 | |
| 922 | #[test] |
| 923 | fn unknown_flag_is_a_usage_error() { |
| 924 | let error = parse_args(&args(&["-p", "x", "--frobnicate"])).unwrap_err(); |
| 925 | assert!(error.contains("--frobnicate"), "{error}"); |
| 926 | } |
| 927 | |
| 928 | #[test] |
| 929 | fn missing_values_are_usage_errors() { |
| 930 | assert!(parse_args(&args(&["-p"])).is_err()); |
| 931 | assert!(parse_args(&args(&["-p", "x", "--cwd"])).is_err()); |
| 932 | assert!(parse_args(&args(&["-p", "x", "--allow-tool"])).is_err()); |
| 933 | assert!(parse_args(&args(&["-p", "x", "--deny-tool"])).is_err()); |
| 934 | assert!(parse_args(&args(&["run", "x", "--output"])).is_err()); |
| 935 | assert!(parse_args(&args(&["run", "x", "--resume"])).is_err()); |
| 936 | assert!(parse_args(&args(&["run", "x", "--model"])).is_err()); |
| 937 | } |
| 938 | |
| 939 | #[test] |
| 940 | fn duplicate_prompt_is_a_usage_error() { |
| 941 | assert!(parse_args(&args(&["-p", "a", "--prompt", "b"])).is_err()); |
| 942 | } |
| 943 | |
| 944 | #[test] |
| 945 | fn run_rejects_a_second_prompt_from_either_source() { |
| 946 | let error = parse_args(&args(&["run", "a", "b"])).unwrap_err(); |
| 947 | assert!(error.contains("unexpected extra argument: b"), "{error}"); |
| 948 | let error = parse_args(&args(&["run", "a", "-p", "b"])).unwrap_err(); |
| 949 | assert!(error.contains("more than once"), "{error}"); |
| 950 | let error = parse_args(&args(&["run", "-p", "a", "b"])).unwrap_err(); |
| 951 | assert!(error.contains("unexpected extra argument: b"), "{error}"); |
| 952 | } |
| 953 | |
| 954 | #[test] |
| 955 | fn run_does_not_take_a_mistyped_flag_as_the_prompt() { |
| 956 | let error = parse_args(&args(&["run", "--quite", "fix it"])).unwrap_err(); |
| 957 | assert!(error.contains("unknown argument: --quite"), "{error}"); |
| 958 | } |
| 959 | |
| 960 | #[test] |
| 961 | fn bare_run_reports_a_missing_prompt() { |
| 962 | assert_eq!( |
| 963 | parse_args(&args(&["run"])), |
| 964 | Err("missing prompt".to_string()) |
| 965 | ); |
| 966 | } |
| 967 | |
| 968 | #[test] |
| 969 | fn empty_prompt_is_a_usage_error() { |
| 970 | assert!(parse_args(&args(&["-p", " "])).is_err()); |
| 971 | } |
| 972 | } |