| 1 | //! Compatibility persistent-RLM session tools. |
| 2 | //! |
| 3 | //! v0.8.33 replaces the old one-shot `rlm` tool with a head/hands surface: |
| 4 | //! `rlm_open` creates a named Python kernel over a large context, |
| 5 | //! `rlm_eval` runs bounded probes against it, `rlm_configure` adjusts runtime |
| 6 | //! feedback, and `rlm_close` tears it down. |
| 7 | //! |
| 8 | //! The normal Agent path now owns one session-persistent `repl` kernel. This |
| 9 | //! action-shaped surface stays registered for explicit compatibility and saved |
| 10 | //! transcript replay, but is hidden from new model turns. Its `rlm_*` aliases |
| 11 | //! force the action so old transcripts replay correctly — the pattern |
| 12 | //! `BashTool` established for `exec_shell*` in #4625. |
| 13 | |
| 14 | use std::sync::Arc; |
| 15 | use std::time::{Duration, Instant}; |
| 16 | |
| 17 | use async_trait::async_trait; |
| 18 | use serde_json::{Value, json}; |
| 19 | |
| 20 | use crate::client::CodewhaleClient; |
| 21 | use crate::repl::PythonRuntime; |
| 22 | use crate::rlm::RlmBridge; |
| 23 | use crate::rlm::session::{ |
| 24 | ContextMeta, OutputFeedback, RlmSession, derive_session_name, write_context_file, |
| 25 | }; |
| 26 | use crate::tools::fetch_url::FetchUrlTool; |
| 27 | use crate::tools::handle::VarHandle; |
| 28 | use crate::tools::spec::{ |
| 29 | ApprovalRequirement, ToolCapability, ToolContext, ToolError, ToolResult, ToolSpec, |
| 30 | }; |
| 31 | |
| 32 | const DEFAULT_CHILD_MODEL: &str = "deepseek-v4-flash"; |
| 33 | const MAX_INLINE_CONTENT_CHARS: usize = 200_000; |
| 34 | const FULL_STDOUT_HEAD_CHARS: usize = 4_096; |
| 35 | const FULL_STDOUT_TAIL_CHARS: usize = 1_024; |
| 36 | |
| 37 | /// When `rlm_eval` stdout exceeds this many characters the full body is |
| 38 | /// stored as a `var_handle` instead of inlined into the parent transcript. |
| 39 | /// The model retrieves the body via `handle_read` using the returned handle. |
| 40 | const STDOUT_HANDLE_THRESHOLD_CHARS: usize = 1_000; |
| 41 | const HARD_SUB_RLM_DEPTH_CAP: u32 = 3; |
| 42 | |
| 43 | const ALL_ACTIONS: &[&str] = &["session_objects", "open", "eval", "configure", "close"]; |
| 44 | |
| 45 | fn rlm_kernel_error_result( |
| 46 | error: &str, |
| 47 | elapsed: Duration, |
| 48 | usage_batch: &crate::cost_status::RuntimeUsageBatch, |
| 49 | ) -> ToolResult { |
| 50 | let mut metadata = json!({ |
| 51 | // The registered tool is `rlm`; `eval` is its action. Naming a |
| 52 | // retired `rlm_eval` tool here taught the model a call it cannot |
| 53 | // make (2026-08-04 audit). |
| 54 | "tool": "rlm", |
| 55 | "action": "eval", |
| 56 | "duration_ms": elapsed.as_millis() as u64, |
| 57 | "kernel_error": true, |
| 58 | }); |
| 59 | crate::cost_status::attach_child_usage_batch_metadata(&mut metadata, usage_batch); |
| 60 | ToolResult::error(format!("rlm action='eval': {error}")).with_metadata(metadata) |
| 61 | } |
| 62 | |
| 63 | /// Unified RLM session tool. |
| 64 | /// |
| 65 | /// One struct and one input schema for the canonical `rlm` tool. `client` is |
| 66 | /// only exercised by the `eval` action (child sub-RLM queries); other actions |
| 67 | /// ignore it. |
| 68 | pub struct RlmTool { |
| 69 | name: &'static str, |
| 70 | forced_action: Option<&'static str>, |
| 71 | client: Option<CodewhaleClient>, |
| 72 | /// Kept only for replay-compatible explicit RLM sessions. New normal |
| 73 | /// agent work uses the session kernel and inherits its route there. |
| 74 | root_model: String, |
| 75 | } |
| 76 | |
| 77 | impl RlmTool { |
| 78 | #[must_use] |
| 79 | pub fn new(name: &'static str, client: Option<CodewhaleClient>) -> Self { |
| 80 | Self { |
| 81 | name, |
| 82 | forced_action: None, |
| 83 | client, |
| 84 | root_model: DEFAULT_CHILD_MODEL.to_string(), |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | /// Bind an explicit compatibility session to the active parent route. |
| 89 | /// This prevents a saved/manual RLM invocation from silently falling back |
| 90 | /// to an unrelated legacy child model. |
| 91 | #[must_use] |
| 92 | pub fn with_root_model(mut self, root_model: String) -> Self { |
| 93 | self.root_model = root_model; |
| 94 | self |
| 95 | } |
| 96 | |
| 97 | #[cfg(test)] |
| 98 | #[must_use] |
| 99 | pub fn alias( |
| 100 | name: &'static str, |
| 101 | action: &'static str, |
| 102 | client: Option<CodewhaleClient>, |
| 103 | ) -> Self { |
| 104 | Self { |
| 105 | name, |
| 106 | forced_action: Some(action), |
| 107 | client, |
| 108 | root_model: DEFAULT_CHILD_MODEL.to_string(), |
| 109 | } |
| 110 | } |
| 111 | |
| 112 | fn resolve_action<'a>(&'a self, input: &'a Value) -> Result<&'a str, ToolError> { |
| 113 | let action = match self.forced_action { |
| 114 | Some(action) => action, |
| 115 | None => input.get("action").and_then(Value::as_str).ok_or_else(|| { |
| 116 | ToolError::invalid_input(format!( |
| 117 | "rlm: missing `action` (one of: {})", |
| 118 | ALL_ACTIONS.join(", ") |
| 119 | )) |
| 120 | })?, |
| 121 | }; |
| 122 | if ALL_ACTIONS.contains(&action) { |
| 123 | Ok(action) |
| 124 | } else { |
| 125 | Err(ToolError::invalid_input(format!( |
| 126 | "rlm: invalid action `{action}` (one of: {})", |
| 127 | ALL_ACTIONS.join(", ") |
| 128 | ))) |
| 129 | } |
| 130 | } |
| 131 | |
| 132 | /// Without concrete input, open may fetch a URL. Input-specific approval |
| 133 | /// below keeps local/inline reads automatic and inherits fetch_url's |
| 134 | /// outbound-payload approval for URL sources. |
| 135 | fn action_requires_approval(action: &str) -> bool { |
| 136 | matches!(action, "eval" | "open") |
| 137 | } |
| 138 | |
| 139 | /// Mirror of the legacy per-tool read-only contract (capability-derived): |
| 140 | /// `rlm_open` carries `ExecutesCode`, so only session_objects / configure / |
| 141 | /// close counted as read-only. |
| 142 | fn action_is_read_only(action: &str) -> bool { |
| 143 | matches!(action, "session_objects" | "configure" | "close") |
| 144 | } |
| 145 | |
| 146 | fn action_capabilities(action: &str) -> Vec<ToolCapability> { |
| 147 | match action { |
| 148 | "session_objects" => vec![ToolCapability::ReadOnly], |
| 149 | "open" => vec![ |
| 150 | ToolCapability::ReadOnly, |
| 151 | ToolCapability::Network, |
| 152 | ToolCapability::ExecutesCode, |
| 153 | ToolCapability::RequiresApproval, |
| 154 | ], |
| 155 | "eval" => vec![ |
| 156 | ToolCapability::Network, |
| 157 | ToolCapability::ExecutesCode, |
| 158 | ToolCapability::RequiresApproval, |
| 159 | ], |
| 160 | // configure / close |
| 161 | _ => vec![ToolCapability::ReadOnly], |
| 162 | } |
| 163 | } |
| 164 | } |
| 165 | |
| 166 | #[async_trait] |
| 167 | impl ToolSpec for RlmTool { |
| 168 | fn name(&self) -> &'static str { |
| 169 | self.name |
| 170 | } |
| 171 | |
| 172 | fn model_visible(&self) -> bool { |
| 173 | // The normal Agent path owns a session-scoped `repl` kernel. Keep the |
| 174 | // old action fan-out registered for replay and explicit compatibility, |
| 175 | // but do not teach a second RLM workflow to new model turns. |
| 176 | false |
| 177 | } |
| 178 | |
| 179 | fn description(&self) -> &'static str { |
| 180 | match self.forced_action { |
| 181 | Some("session_objects") => { |
| 182 | "List active prompt/history/session symbolic objects as compact cards. \ |
| 183 | Pass one of the returned `id` values to `rlm_open` as \ |
| 184 | `session_object` to inspect it inside an RLM REPL without copying the \ |
| 185 | full prompt or transcript into the parent context." |
| 186 | } |
| 187 | Some("open") => { |
| 188 | "Open a persistent RLM context. Loads `file_path`, `content`, `url`, \ |
| 189 | or `session_object` into a named Python kernel and returns only \ |
| 190 | metadata: name, length, preview, and sha256. Use this for large or \ |
| 191 | unfamiliar inputs so the parent transcript holds a handle, not the \ |
| 192 | body." |
| 193 | } |
| 194 | Some("eval") => { |
| 195 | "Run one Python REPL block against a named RLM context. Returns a \ |
| 196 | bounded projection of stdout/stderr plus metadata. If the code calls \ |
| 197 | FINAL/finalize, the final value is stored as a var_handle retrievable \ |
| 198 | with handle_read instead of copied unbounded into the parent context. \ |
| 199 | Large stdout/stderr payloads (>1k chars) are also stored as \ |
| 200 | var_handles (returned in stdout_handle / stderr_handle) to keep the \ |
| 201 | parent transcript lean. Batch child helpers require \ |
| 202 | dependency_mode='independent'; use sub_query_sequence or a \ |
| 203 | sequential loop for dependent work." |
| 204 | } |
| 205 | Some("configure") => { |
| 206 | "Configure a named RLM context: output feedback, child query timeout, \ |
| 207 | recursive sub-RLM depth, and explicit session sharing." |
| 208 | } |
| 209 | Some("close") => { |
| 210 | "Close a named RLM context, tear down its Python kernel, and return \ |
| 211 | usage/lifecycle metadata." |
| 212 | } |
| 213 | _ => { |
| 214 | "Persistent RLM sessions over large contexts. Actions: \"session_objects\" \ |
| 215 | (list active prompt/history/session symbolic objects as compact cards), \ |
| 216 | \"open\" (load file_path/content/url/session_object into a named Python \ |
| 217 | kernel; returns only metadata so the parent transcript holds a handle, \ |
| 218 | not the body), \"eval\" (run one bounded Python REPL block against a \ |
| 219 | named context; approval required; FINAL/finalize values and large \ |
| 220 | stdout/stderr become var_handles retrievable with handle_read), \ |
| 221 | \"configure\" (output feedback, child timeout, sub-RLM depth, session \ |
| 222 | sharing), \"close\" (tear down the kernel and return usage metadata)." |
| 223 | } |
| 224 | } |
| 225 | } |
| 226 | |
| 227 | fn input_schema(&self) -> Value { |
| 228 | if let Some(action) = self.forced_action { |
| 229 | return legacy_action_schema(action); |
| 230 | } |
| 231 | json!({ |
| 232 | "type": "object", |
| 233 | "properties": { |
| 234 | "action": { |
| 235 | "type": "string", |
| 236 | "enum": ALL_ACTIONS, |
| 237 | "description": "Action to perform." |
| 238 | }, |
| 239 | "name": { |
| 240 | "type": "string", |
| 241 | "description": "RLM context name, unique within this parent session (action=open: optional, defaults to a slug from the source). Required for action=eval/configure/close." |
| 242 | }, |
| 243 | "file_path": { |
| 244 | "type": "string", |
| 245 | "description": "Workspace-relative file to load (action=open; exactly one of file_path/content/url/session_object)." |
| 246 | }, |
| 247 | "content": { |
| 248 | "type": "string", |
| 249 | "description": "Inline content to load. Capped at 200k chars. (action=open)" |
| 250 | }, |
| 251 | "url": { |
| 252 | "type": "string", |
| 253 | "description": "HTTP/HTTPS URL to fetch (through the same path as Web action=\"fetch\") and load. (action=open)" |
| 254 | }, |
| 255 | "session_object": { |
| 256 | "type": "string", |
| 257 | "description": "Stable symbolic active-session ref from action=session_objects, for example session://active/system_prompt or session://active/messages/0. (action=open)" |
| 258 | }, |
| 259 | "code": { |
| 260 | "type": "string", |
| 261 | "description": "Raw Python executed against the context (no markdown fences). The loaded source is in scope as `content`; call FINAL(value)/finalize(...) to return a result handle. Example: print(len(content)). (action=eval)" |
| 262 | }, |
| 263 | "output_feedback": { |
| 264 | "type": "string", |
| 265 | "enum": ["full", "metadata"], |
| 266 | "description": "(action=configure)" |
| 267 | }, |
| 268 | "sub_query_timeout_secs": { |
| 269 | "type": "integer", |
| 270 | "description": "(action=configure)" |
| 271 | }, |
| 272 | "sub_rlm_max_depth": { |
| 273 | "type": "integer", |
| 274 | "minimum": 0, |
| 275 | "maximum": 3, |
| 276 | "description": "(action=configure)" |
| 277 | }, |
| 278 | "share_session": { |
| 279 | "type": "boolean", |
| 280 | "description": "(action=configure)" |
| 281 | } |
| 282 | }, |
| 283 | "additionalProperties": false |
| 284 | }) |
| 285 | } |
| 286 | |
| 287 | fn capabilities(&self) -> Vec<ToolCapability> { |
| 288 | match self.forced_action { |
| 289 | Some(action) => Self::action_capabilities(action), |
| 290 | None => vec![ |
| 291 | ToolCapability::Network, |
| 292 | ToolCapability::ExecutesCode, |
| 293 | ToolCapability::RequiresApproval, |
| 294 | ], |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | fn approval_requirement(&self) -> ApprovalRequirement { |
| 299 | match self.forced_action { |
| 300 | Some(action) if Self::action_requires_approval(action) => ApprovalRequirement::Required, |
| 301 | Some(_) => ApprovalRequirement::Auto, |
| 302 | None => ApprovalRequirement::Required, |
| 303 | } |
| 304 | } |
| 305 | |
| 306 | fn approval_requirement_for(&self, input: &Value) -> ApprovalRequirement { |
| 307 | match self.resolve_action(input) { |
| 308 | Ok("open") if rlm_open_source_field(input, "url").is_some() => { |
| 309 | FetchUrlTool.approval_requirement_for(&json!({"url": input["url"]})) |
| 310 | } |
| 311 | Ok("open") => ApprovalRequirement::Auto, |
| 312 | Ok(action) if Self::action_requires_approval(action) => ApprovalRequirement::Required, |
| 313 | Ok(_) => ApprovalRequirement::Auto, |
| 314 | Err(_) => self.approval_requirement(), |
| 315 | } |
| 316 | } |
| 317 | |
| 318 | fn is_read_only_for(&self, input: &Value) -> bool { |
| 319 | match self.resolve_action(input) { |
| 320 | Ok(action) => Self::action_is_read_only(action), |
| 321 | Err(_) => self.is_read_only(), |
| 322 | } |
| 323 | } |
| 324 | |
| 325 | fn supports_parallel(&self) -> bool { |
| 326 | matches!(self.forced_action, Some("session_objects")) |
| 327 | } |
| 328 | |
| 329 | fn supports_parallel_for(&self, input: &Value) -> bool { |
| 330 | matches!(self.resolve_action(input), Ok("session_objects")) |
| 331 | } |
| 332 | |
| 333 | async fn execute(&self, input: Value, context: &ToolContext) -> Result<ToolResult, ToolError> { |
| 334 | match self.resolve_action(&input)? { |
| 335 | "session_objects" => self.execute_session_objects(context).await, |
| 336 | "open" => self.execute_open(&input, context).await, |
| 337 | "eval" => self.execute_eval(&input, context).await, |
| 338 | "configure" => self.execute_configure(&input, context).await, |
| 339 | "close" => self.execute_close(&input, context).await, |
| 340 | action => Err(ToolError::invalid_input(format!( |
| 341 | "rlm: invalid action `{action}`" |
| 342 | ))), |
| 343 | } |
| 344 | } |
| 345 | } |
| 346 | |
| 347 | impl RlmTool { |
| 348 | async fn execute_session_objects( |
| 349 | &self, |
| 350 | context: &ToolContext, |
| 351 | ) -> Result<ToolResult, ToolError> { |
| 352 | let snapshot = context.session_objects.as_ref().ok_or_else(|| { |
| 353 | ToolError::not_available("rlm_session_objects: active session snapshot unavailable") |
| 354 | })?; |
| 355 | ToolResult::json(&json!({ |
| 356 | "objects": snapshot.object_cards(), |
| 357 | "open_with": { |
| 358 | "tool": "rlm", |
| 359 | "action": "open", |
| 360 | "field": "session_object", |
| 361 | "example": { |
| 362 | "name": "active_prompt", |
| 363 | "session_object": "session://active/system_prompt" |
| 364 | } |
| 365 | }, |
| 366 | "redaction": "Large tool results and thinking blocks are represented by compact metadata in transcript objects; use returned handles and handle_read for bounded payload projections." |
| 367 | })) |
| 368 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 369 | } |
| 370 | |
| 371 | async fn execute_open( |
| 372 | &self, |
| 373 | input: &Value, |
| 374 | context: &ToolContext, |
| 375 | ) -> Result<ToolResult, ToolError> { |
| 376 | let source_count = rlm_open_source_count(input); |
| 377 | if source_count != 1 { |
| 378 | let mut msg = String::from( |
| 379 | "rlm_open: provide exactly one of `file_path` (local file), `content` (inline text), `url`, or `session_object`", |
| 380 | ); |
| 381 | // "did you mean" for common misnamings (#2655). |
| 382 | if let Some(obj) = input.as_object() { |
| 383 | let seen: Vec<&str> = [ |
| 384 | "prompt", |
| 385 | "resident_file", |
| 386 | "text", |
| 387 | "body", |
| 388 | "path", |
| 389 | "file", |
| 390 | "source", |
| 391 | ] |
| 392 | .into_iter() |
| 393 | .filter(|k| obj.contains_key(*k)) |
| 394 | .collect(); |
| 395 | if !seen.is_empty() { |
| 396 | msg.push_str(&format!( |
| 397 | ". Saw {seen:?} — did you mean file_path/content/url/session_object? (to evaluate against an existing context, pass its name to rlm action='eval', or use `session_object`)" |
| 398 | )); |
| 399 | } |
| 400 | } |
| 401 | return Err(ToolError::invalid_input(msg)); |
| 402 | } |
| 403 | |
| 404 | let (body, source_type, source_hint) = load_source(input, context).await?; |
| 405 | if body.trim().is_empty() { |
| 406 | return Err(ToolError::invalid_input( |
| 407 | "rlm_open: input is empty after loading", |
| 408 | )); |
| 409 | } |
| 410 | |
| 411 | let name = input |
| 412 | .get("name") |
| 413 | .and_then(Value::as_str) |
| 414 | .map(str::trim) |
| 415 | .filter(|name| !name.is_empty()) |
| 416 | .map(ToOwned::to_owned) |
| 417 | .unwrap_or_else(|| derive_session_name(source_hint.as_deref())); |
| 418 | |
| 419 | { |
| 420 | let sessions = context.runtime.rlm_sessions.lock().await; |
| 421 | if sessions.contains_key(&name) { |
| 422 | return Err(ToolError::invalid_input(format!( |
| 423 | "rlm_open: context name `{name}` already exists" |
| 424 | ))); |
| 425 | } |
| 426 | } |
| 427 | |
| 428 | let context_path = write_context_file(&body).map_err(|e| { |
| 429 | ToolError::execution_failed(format!("rlm_open: failed to stage context: {e}")) |
| 430 | })?; |
| 431 | let kernel = PythonRuntime::spawn_with_context(&context_path) |
| 432 | .await |
| 433 | .map_err(|e| ToolError::execution_failed(format!("rlm_open: {e}")))?; |
| 434 | let context_meta = ContextMeta::from_body(&body, source_type); |
| 435 | let session = RlmSession::new(name.clone(), kernel, context_meta.clone(), context_path); |
| 436 | let id = session.id.clone(); |
| 437 | |
| 438 | let mut sessions = context.runtime.rlm_sessions.lock().await; |
| 439 | sessions.insert(name.clone(), Arc::new(tokio::sync::Mutex::new(session))); |
| 440 | |
| 441 | ToolResult::json(&json!({ |
| 442 | "name": name, |
| 443 | "id": id, |
| 444 | "length": context_meta.length, |
| 445 | "type": context_meta.type_name, |
| 446 | "preview_500": context_meta.preview_500, |
| 447 | "sha256": context_meta.sha256, |
| 448 | })) |
| 449 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 450 | } |
| 451 | |
| 452 | async fn execute_eval( |
| 453 | &self, |
| 454 | input: &Value, |
| 455 | context: &ToolContext, |
| 456 | ) -> Result<ToolResult, ToolError> { |
| 457 | crate::core::engine::tool_catalog::enforce_tool_denial(context, "rlm_eval", input)?; |
| 458 | let name = required_non_empty_str(input, "name")?; |
| 459 | let code = required_non_empty_str(input, "code").map_err(|_| { |
| 460 | ToolError::invalid_input( |
| 461 | "rlm_eval: `code` is required and runs raw Python against the RLM context (no markdown fences). \ |
| 462 | Example: {\"name\": \"<ctx>\", \"code\": \"print(len(content))\"}; call FINAL(value) to return a result handle.", |
| 463 | ) |
| 464 | })?; |
| 465 | let session = get_session(context, name).await?; |
| 466 | let mut session = session.lock().await; |
| 467 | let config = session.config.clone(); |
| 468 | |
| 469 | let Some(kernel) = session.kernel.as_mut() else { |
| 470 | return Err(ToolError::invalid_input(format!( |
| 471 | "rlm_eval: context `{name}` is closed" |
| 472 | ))); |
| 473 | }; |
| 474 | |
| 475 | let started = Instant::now(); |
| 476 | let (round, child_usage_batch) = if let Some(client) = self.client.clone() { |
| 477 | let bridge = RlmBridge::new( |
| 478 | Arc::new(client), |
| 479 | self.root_model.clone(), |
| 480 | config.sub_rlm_max_depth.min(HARD_SUB_RLM_DEPTH_CAP), |
| 481 | ); |
| 482 | let round_result = kernel.run(code, Some(&bridge)).await; |
| 483 | let usage = bridge.usage_snapshot().await; |
| 484 | let round = match round_result { |
| 485 | Ok(round) => round, |
| 486 | Err(error) => { |
| 487 | // A bridge request may have completed and accrued usage |
| 488 | // before the Python kernel times out or closes stdout. |
| 489 | // Return a failed ToolResult (rather than a bare ToolError) |
| 490 | // so ToolCallComplete still carries the immutable child |
| 491 | // receipt and the runtime can durably account for it. |
| 492 | session.last_used_at = Instant::now(); |
| 493 | return Ok(rlm_kernel_error_result( |
| 494 | &error.to_string(), |
| 495 | started.elapsed(), |
| 496 | &crate::cost_status::RuntimeUsageBatch { |
| 497 | records: usage.records, |
| 498 | drop_records: usage.drop_records, |
| 499 | dropped_records: usage.dropped_records, |
| 500 | }, |
| 501 | )); |
| 502 | } |
| 503 | }; |
| 504 | ( |
| 505 | round, |
| 506 | crate::cost_status::RuntimeUsageBatch { |
| 507 | records: usage.records, |
| 508 | drop_records: usage.drop_records, |
| 509 | dropped_records: usage.dropped_records, |
| 510 | }, |
| 511 | ) |
| 512 | } else { |
| 513 | let round = kernel |
| 514 | .run(code, None::<&RlmBridge>) |
| 515 | .await |
| 516 | .map_err(|e| ToolError::execution_failed(format!("rlm_eval: {e}")))?; |
| 517 | (round, crate::cost_status::RuntimeUsageBatch::default()) |
| 518 | }; |
| 519 | |
| 520 | session.rpc_count = session.rpc_count.saturating_add(round.rpc_count); |
| 521 | session.total_duration += round.elapsed; |
| 522 | session.last_used_at = Instant::now(); |
| 523 | |
| 524 | let final_handle = if let Some(value_json) = round.final_json.clone() { |
| 525 | session.final_count = session.final_count.saturating_add(1); |
| 526 | let handle_name = format!("final_{}", session.final_count); |
| 527 | let handle = { |
| 528 | let mut store = context.runtime.handle_store.lock().await; |
| 529 | match value_json { |
| 530 | Value::String(value) => { |
| 531 | store.insert_text(session.id.clone(), handle_name, value) |
| 532 | } |
| 533 | other => store.insert_json(session.id.clone(), handle_name, other), |
| 534 | } |
| 535 | }; |
| 536 | Some(handle) |
| 537 | } else { |
| 538 | None |
| 539 | }; |
| 540 | |
| 541 | let had_error = round.has_error; |
| 542 | let rpc_count = round.rpc_count; |
| 543 | let duration_ms = round.elapsed.as_millis() as u64; |
| 544 | // Route large stdout/stderr into a var_handle to avoid bloat in |
| 545 | // the parent transcript. The model calls handle_read for bounded |
| 546 | // projections; a short inline note describes availability. |
| 547 | fn route_output( |
| 548 | text: &str, |
| 549 | feedback: &OutputFeedback, |
| 550 | store: &mut crate::tools::handle::HandleStore, |
| 551 | session_id: &str, |
| 552 | tag: &str, |
| 553 | ) -> (Option<String>, Option<crate::tools::handle::VarHandle>) { |
| 554 | let threshold = STDOUT_HANDLE_THRESHOLD_CHARS; |
| 555 | match (feedback, text.len()) { |
| 556 | (OutputFeedback::Full, len) if len <= threshold => { |
| 557 | (Some(preview_output(text)), None) |
| 558 | } |
| 559 | (OutputFeedback::Full, _) if !text.trim().is_empty() => { |
| 560 | // Store full body as a handle for out-of-band retrieval |
| 561 | let name = format!("{tag}_{}", 0); // single counter is fine |
| 562 | let handle = store.insert_text(session_id, name, text); |
| 563 | ( |
| 564 | Some(format!("{} chars; retrieve via handle_read", text.len())), |
| 565 | Some(handle), |
| 566 | ) |
| 567 | } |
| 568 | _ => (None, None), |
| 569 | } |
| 570 | } |
| 571 | |
| 572 | let (stdout_preview, stdout_handle) = route_output( |
| 573 | &round.full_stdout, |
| 574 | &config.output_feedback, |
| 575 | &mut *context.runtime.handle_store.lock().await, |
| 576 | &session.id, |
| 577 | "stdout", |
| 578 | ); |
| 579 | let (stderr_preview, stderr_handle) = route_output( |
| 580 | &round.stderr, |
| 581 | &config.output_feedback, |
| 582 | &mut *context.runtime.handle_store.lock().await, |
| 583 | &session.id, |
| 584 | "stderr", |
| 585 | ); |
| 586 | |
| 587 | let mut output = json!({ |
| 588 | "name": session.name, |
| 589 | "id": session.id, |
| 590 | "duration_ms": duration_ms, |
| 591 | "rpc_count": rpc_count, |
| 592 | "had_error": had_error, |
| 593 | "new_vars": [], |
| 594 | "final": final_handle, |
| 595 | }); |
| 596 | if let Some(ref stdout_preview) = stdout_preview { |
| 597 | output["stdout_preview"] = json!(stdout_preview); |
| 598 | } |
| 599 | if let Some(ref stderr_preview) = stderr_preview { |
| 600 | output["stderr_preview"] = json!(stderr_preview); |
| 601 | } |
| 602 | if let (Some(h), Some(_)) = (stdout_handle, &stdout_preview) { |
| 603 | output["stdout_handle"] = json!(h); |
| 604 | } |
| 605 | if let (Some(h), Some(_)) = (stderr_handle, &stderr_preview) { |
| 606 | output["stderr_handle"] = json!(h); |
| 607 | } |
| 608 | if let Some(confidence) = round.final_confidence.clone() { |
| 609 | output["confidence"] = confidence; |
| 610 | } |
| 611 | |
| 612 | let mut metadata = json!({ |
| 613 | "tool": "rlm_eval", |
| 614 | "duration_ms": started.elapsed().as_millis() as u64, |
| 615 | }); |
| 616 | // Every RLM provider call keeps its own dispatch timestamp and frozen |
| 617 | // quote. The preferred batch format prevents a fan-out from being |
| 618 | // retroactively priced as one aggregate call on the first route. |
| 619 | crate::cost_status::attach_child_usage_batch_metadata(&mut metadata, &child_usage_batch); |
| 620 | |
| 621 | Ok(ToolResult::json(&output) |
| 622 | .map_err(|e| ToolError::execution_failed(e.to_string()))? |
| 623 | .with_metadata(metadata)) |
| 624 | } |
| 625 | |
| 626 | async fn execute_configure( |
| 627 | &self, |
| 628 | input: &Value, |
| 629 | context: &ToolContext, |
| 630 | ) -> Result<ToolResult, ToolError> { |
| 631 | let name = required_non_empty_str(input, "name")?; |
| 632 | let session = get_session(context, name).await?; |
| 633 | let mut session = session.lock().await; |
| 634 | |
| 635 | if let Some(value) = input.get("output_feedback").and_then(Value::as_str) { |
| 636 | session.config.output_feedback = match value { |
| 637 | "full" => OutputFeedback::Full, |
| 638 | "metadata" => OutputFeedback::Metadata, |
| 639 | other => { |
| 640 | return Err(ToolError::invalid_input(format!( |
| 641 | "rlm_configure: invalid output_feedback `{other}`" |
| 642 | ))); |
| 643 | } |
| 644 | }; |
| 645 | } |
| 646 | if let Some(timeout) = input.get("sub_query_timeout_secs").and_then(Value::as_u64) { |
| 647 | session.config.sub_query_timeout_secs = timeout.clamp(1, 600); |
| 648 | } |
| 649 | if let Some(depth) = input.get("sub_rlm_max_depth").and_then(Value::as_u64) { |
| 650 | session.config.sub_rlm_max_depth = (depth as u32).min(HARD_SUB_RLM_DEPTH_CAP); |
| 651 | } |
| 652 | if let Some(share) = input.get("share_session").and_then(Value::as_bool) { |
| 653 | session.config.share_session = share; |
| 654 | } |
| 655 | |
| 656 | ToolResult::json(&json!({ |
| 657 | "name": session.name, |
| 658 | "current_config": session.config, |
| 659 | })) |
| 660 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 661 | } |
| 662 | |
| 663 | async fn execute_close( |
| 664 | &self, |
| 665 | input: &Value, |
| 666 | context: &ToolContext, |
| 667 | ) -> Result<ToolResult, ToolError> { |
| 668 | let name = required_non_empty_str(input, "name")?; |
| 669 | let removed = { |
| 670 | let mut sessions = context.runtime.rlm_sessions.lock().await; |
| 671 | sessions.remove(name) |
| 672 | }; |
| 673 | let Some(session) = removed else { |
| 674 | return Err(ToolError::invalid_input(format!( |
| 675 | "rlm_close: unknown context `{name}`" |
| 676 | ))); |
| 677 | }; |
| 678 | |
| 679 | let mut session = session.lock().await; |
| 680 | let kernel = session.kernel.take(); |
| 681 | let output = json!({ |
| 682 | "name": session.name, |
| 683 | "id": session.id, |
| 684 | "rpc_count": session.rpc_count, |
| 685 | "total_duration_ms": session.total_duration.as_millis() as u64, |
| 686 | "peak_var_count": session.peak_var_count, |
| 687 | "created_ms_ago": session.created_at.elapsed().as_millis() as u64, |
| 688 | "context_path": session.context_path, |
| 689 | }); |
| 690 | drop(session); |
| 691 | |
| 692 | if let Some(kernel) = kernel { |
| 693 | kernel.shutdown().await; |
| 694 | } |
| 695 | |
| 696 | ToolResult::json(&output).map_err(|e| ToolError::execution_failed(e.to_string())) |
| 697 | } |
| 698 | } |
| 699 | |
| 700 | /// The exact schema the legacy per-action tool exposed, kept so hidden alias |
| 701 | /// registrations report an identical contract to the pre-unification tools. |
| 702 | fn legacy_action_schema(action: &str) -> Value { |
| 703 | match action { |
| 704 | "session_objects" => json!({ |
| 705 | "type": "object", |
| 706 | "properties": {} |
| 707 | }), |
| 708 | "open" => json!({ |
| 709 | "type": "object", |
| 710 | "properties": { |
| 711 | "name": { |
| 712 | "type": "string", |
| 713 | "description": "Caller-chosen context name, unique within this parent session. Defaults to a slug from the source." |
| 714 | }, |
| 715 | "file_path": { |
| 716 | "type": "string", |
| 717 | "description": "Workspace-relative file to load." |
| 718 | }, |
| 719 | "content": { |
| 720 | "type": "string", |
| 721 | "description": "Inline content to load. Capped at 200k chars." |
| 722 | }, |
| 723 | "url": { |
| 724 | "type": "string", |
| 725 | "description": "HTTP/HTTPS URL to fetch (through the same path as Web action=\"fetch\") and load." |
| 726 | }, |
| 727 | "session_object": { |
| 728 | "type": "string", |
| 729 | "description": "Stable symbolic active-session ref from rlm_session_objects, for example session://active/system_prompt or session://active/messages/0." |
| 730 | } |
| 731 | } |
| 732 | }), |
| 733 | "eval" => json!({ |
| 734 | "type": "object", |
| 735 | "required": ["name", "code"], |
| 736 | "properties": { |
| 737 | "name": { "type": "string", "description": "RLM context name returned by rlm_open." }, |
| 738 | "code": { "type": "string", "description": "Raw Python executed against the context (no markdown fences). The loaded source is in scope as `content`; call FINAL(value)/finalize(...) to return a result handle. Example: print(len(content))." } |
| 739 | } |
| 740 | }), |
| 741 | "configure" => json!({ |
| 742 | "type": "object", |
| 743 | "required": ["name"], |
| 744 | "properties": { |
| 745 | "name": { "type": "string" }, |
| 746 | "output_feedback": { "type": "string", "enum": ["full", "metadata"] }, |
| 747 | "sub_query_timeout_secs": { "type": "integer" }, |
| 748 | "sub_rlm_max_depth": { "type": "integer", "minimum": 0, "maximum": 3 }, |
| 749 | "share_session": { "type": "boolean" } |
| 750 | } |
| 751 | }), |
| 752 | // close |
| 753 | _ => json!({ |
| 754 | "type": "object", |
| 755 | "required": ["name"], |
| 756 | "properties": { |
| 757 | "name": { "type": "string", "description": "RLM context name from rlm_open." } |
| 758 | } |
| 759 | }), |
| 760 | } |
| 761 | } |
| 762 | |
| 763 | async fn load_source( |
| 764 | input: &Value, |
| 765 | context: &ToolContext, |
| 766 | ) -> Result<(String, String, Option<String>), ToolError> { |
| 767 | if let Some(path) = rlm_open_source_field(input, "file_path").map(str::trim) { |
| 768 | let resolved = context.resolve_path(path)?; |
| 769 | let body = tokio::fs::read_to_string(&resolved).await.map_err(|e| { |
| 770 | ToolError::execution_failed(format!("rlm_open: read {}: {e}", resolved.display())) |
| 771 | })?; |
| 772 | return Ok((body, "file".to_string(), Some(path.to_string()))); |
| 773 | } |
| 774 | |
| 775 | if let Some(content) = rlm_open_source_field(input, "content") { |
| 776 | if content.chars().count() > MAX_INLINE_CONTENT_CHARS { |
| 777 | return Err(ToolError::invalid_input(format!( |
| 778 | "rlm_open: inline content is {} chars (cap {MAX_INLINE_CONTENT_CHARS})", |
| 779 | content.chars().count() |
| 780 | ))); |
| 781 | } |
| 782 | return Ok((content.to_string(), "content".to_string(), None)); |
| 783 | } |
| 784 | |
| 785 | if let Some(object_ref) = rlm_open_source_field(input, "session_object") { |
| 786 | let snapshot = context.session_objects.as_ref().ok_or_else(|| { |
| 787 | ToolError::not_available("rlm_open: active session snapshot unavailable") |
| 788 | })?; |
| 789 | let object = snapshot.resolve(object_ref).ok_or_else(|| { |
| 790 | ToolError::invalid_input(format!("rlm_open: unknown session object `{object_ref}`")) |
| 791 | })?; |
| 792 | return Ok(( |
| 793 | object.body, |
| 794 | format!("session_object:{}", object.kind), |
| 795 | Some(object.id), |
| 796 | )); |
| 797 | } |
| 798 | |
| 799 | let url = rlm_open_source_field(input, "url") |
| 800 | .map(str::trim) |
| 801 | .ok_or_else(|| ToolError::invalid_input("rlm_open: missing source"))?; |
| 802 | crate::core::engine::tool_catalog::enforce_tool_denial(context, "fetch_url", input)?; |
| 803 | let result = FetchUrlTool |
| 804 | .execute(json!({"url": url, "format": "raw"}), context) |
| 805 | .await?; |
| 806 | let parsed: Value = serde_json::from_str(&result.content).map_err(|e| { |
| 807 | ToolError::execution_failed(format!("rlm_open: fetch_url returned invalid JSON: {e}")) |
| 808 | })?; |
| 809 | let body = parsed |
| 810 | .get("content") |
| 811 | .and_then(Value::as_str) |
| 812 | .ok_or_else(|| ToolError::execution_failed("rlm_open: fetched body missing content"))? |
| 813 | .to_string(); |
| 814 | let source_type = parsed |
| 815 | .get("content_type") |
| 816 | .and_then(Value::as_str) |
| 817 | .unwrap_or("url") |
| 818 | .to_string(); |
| 819 | Ok((body, source_type, Some(url.to_string()))) |
| 820 | } |
| 821 | |
| 822 | fn rlm_open_source_count(input: &Value) -> usize { |
| 823 | ["file_path", "content", "url", "session_object"] |
| 824 | .iter() |
| 825 | .filter(|field| rlm_open_source_field(input, field).is_some()) |
| 826 | .count() |
| 827 | } |
| 828 | |
| 829 | fn rlm_open_source_field<'a>(input: &'a Value, field: &str) -> Option<&'a str> { |
| 830 | input |
| 831 | .get(field) |
| 832 | .and_then(Value::as_str) |
| 833 | .filter(|value| !value.trim().is_empty()) |
| 834 | } |
| 835 | |
| 836 | async fn get_session( |
| 837 | context: &ToolContext, |
| 838 | name: &str, |
| 839 | ) -> Result<Arc<tokio::sync::Mutex<RlmSession>>, ToolError> { |
| 840 | let sessions = context.runtime.rlm_sessions.lock().await; |
| 841 | sessions.get(name).cloned().ok_or_else(|| { |
| 842 | ToolError::invalid_input(format!( |
| 843 | "unknown RLM context `{name}`; open it first with rlm action='open'" |
| 844 | )) |
| 845 | }) |
| 846 | } |
| 847 | |
| 848 | fn required_non_empty_str<'a>(input: &'a Value, field: &str) -> Result<&'a str, ToolError> { |
| 849 | let value = input |
| 850 | .get(field) |
| 851 | .and_then(Value::as_str) |
| 852 | .ok_or_else(|| ToolError::missing_field(field))? |
| 853 | .trim(); |
| 854 | if value.is_empty() { |
| 855 | return Err(ToolError::invalid_input(format!( |
| 856 | "rlm: `{field}` must not be empty" |
| 857 | ))); |
| 858 | } |
| 859 | Ok(value) |
| 860 | } |
| 861 | |
| 862 | fn preview_output(text: &str) -> String { |
| 863 | let total = text.chars().count(); |
| 864 | if total <= FULL_STDOUT_HEAD_CHARS + FULL_STDOUT_TAIL_CHARS { |
| 865 | return text.to_string(); |
| 866 | } |
| 867 | let head: String = text.chars().take(FULL_STDOUT_HEAD_CHARS).collect(); |
| 868 | let tail: String = text |
| 869 | .chars() |
| 870 | .skip(total.saturating_sub(FULL_STDOUT_TAIL_CHARS)) |
| 871 | .collect(); |
| 872 | format!( |
| 873 | "{head}\n... [{} chars truncated, retrieve via handle_read when returned as a handle] ...\n{tail}", |
| 874 | total.saturating_sub(FULL_STDOUT_HEAD_CHARS + FULL_STDOUT_TAIL_CHARS) |
| 875 | ) |
| 876 | } |
| 877 | |
| 878 | fn _assert_var_handle_shape(_: Option<VarHandle>) {} |
| 879 | |
| 880 | #[cfg(test)] |
| 881 | mod tests { |
| 882 | use super::*; |
| 883 | use crate::rlm::session::SessionObjectSnapshot; |
| 884 | use crate::tools::handle::HandleReadTool; |
| 885 | use crate::tools::spec::ToolContext; |
| 886 | use codewhale_models::Role; |
| 887 | use codewhale_models::{ContentBlock, Message, SystemPrompt}; |
| 888 | use std::path::PathBuf; |
| 889 | |
| 890 | fn ctx() -> ToolContext { |
| 891 | ToolContext::new(".") |
| 892 | } |
| 893 | |
| 894 | fn ctx_with_session_objects() -> ToolContext { |
| 895 | ToolContext::new(".").with_session_objects(SessionObjectSnapshot::new( |
| 896 | "session-1".to_string(), |
| 897 | "deepseek-v4-pro".to_string(), |
| 898 | PathBuf::from("."), |
| 899 | Some(SystemPrompt::Text("You are CodeWhale.".to_string())), |
| 900 | vec![ |
| 901 | Message { |
| 902 | role: Role::User, |
| 903 | content: vec![ContentBlock::Text { |
| 904 | text: "Please inspect the RLM surface.".to_string(), |
| 905 | cache_control: None, |
| 906 | }], |
| 907 | }, |
| 908 | Message { |
| 909 | role: Role::Assistant, |
| 910 | content: vec![ContentBlock::Text { |
| 911 | text: "I will use symbolic session objects.".to_string(), |
| 912 | cache_control: None, |
| 913 | }], |
| 914 | }, |
| 915 | ], |
| 916 | )) |
| 917 | } |
| 918 | |
| 919 | #[test] |
| 920 | fn schema_uses_new_tool_names() { |
| 921 | assert_eq!( |
| 922 | RlmTool::alias("rlm_session_objects", "session_objects", None).name(), |
| 923 | "rlm_session_objects" |
| 924 | ); |
| 925 | assert_eq!(RlmTool::alias("rlm_open", "open", None).name(), "rlm_open"); |
| 926 | assert_eq!(RlmTool::alias("rlm_eval", "eval", None).name(), "rlm_eval"); |
| 927 | assert_eq!( |
| 928 | RlmTool::alias("rlm_configure", "configure", None).name(), |
| 929 | "rlm_configure" |
| 930 | ); |
| 931 | assert_eq!( |
| 932 | RlmTool::alias("rlm_close", "close", None).name(), |
| 933 | "rlm_close" |
| 934 | ); |
| 935 | } |
| 936 | |
| 937 | #[test] |
| 938 | fn rlm_tool_is_compatibility_only_not_model_visible() { |
| 939 | let canonical = RlmTool::new("rlm", None); |
| 940 | assert!(!canonical.model_visible()); |
| 941 | assert_eq!(canonical.name(), "rlm"); |
| 942 | let actions = canonical.input_schema()["properties"]["action"]["enum"] |
| 943 | .as_array() |
| 944 | .expect("action enum") |
| 945 | .clone(); |
| 946 | for action in ["session_objects", "open", "eval", "configure", "close"] { |
| 947 | assert!( |
| 948 | actions.iter().any(|value| value.as_str() == Some(action)), |
| 949 | "canonical schema must offer action {action}" |
| 950 | ); |
| 951 | } |
| 952 | |
| 953 | for alias in [ |
| 954 | RlmTool::alias("rlm_session_objects", "session_objects", None), |
| 955 | RlmTool::alias("rlm_open", "open", None), |
| 956 | RlmTool::alias("rlm_eval", "eval", None), |
| 957 | RlmTool::alias("rlm_configure", "configure", None), |
| 958 | RlmTool::alias("rlm_close", "close", None), |
| 959 | ] { |
| 960 | assert!( |
| 961 | !alias.model_visible(), |
| 962 | "compatibility alias {} must stay hidden", |
| 963 | alias.name() |
| 964 | ); |
| 965 | } |
| 966 | } |
| 967 | |
| 968 | #[test] |
| 969 | fn kernel_failure_result_retains_child_usage_receipt() { |
| 970 | let route = crate::cost_status::EffectiveRouteEnvelope::capture( |
| 971 | None, |
| 972 | crate::config::ApiProvider::Deepseek, |
| 973 | "deepseek-rlm", |
| 974 | DEFAULT_CHILD_MODEL, |
| 975 | Some(crate::config::ApiProvider::Deepseek.default_base_url()), |
| 976 | chrono::Utc::now(), |
| 977 | ); |
| 978 | let usage = codewhale_models::Usage { |
| 979 | input_tokens: 23, |
| 980 | output_tokens: 5, |
| 981 | reasoning_replay_tokens: Some(7), |
| 982 | ..Default::default() |
| 983 | }; |
| 984 | let record = crate::cost_status::RuntimeUsageRecord { |
| 985 | source_id: "rlm:test:request:0".to_string(), |
| 986 | usage: crate::cost_status::EffectiveRouteUsage { |
| 987 | route: route.clone(), |
| 988 | usage: usage.clone(), |
| 989 | }, |
| 990 | }; |
| 991 | let drop_record = crate::cost_status::RuntimeUsageDropRecord { |
| 992 | source_id: "rlm:test:request:1".to_string(), |
| 993 | route: route.clone(), |
| 994 | }; |
| 995 | let result = rlm_kernel_error_result( |
| 996 | "kernel stdout closed", |
| 997 | Duration::from_millis(11), |
| 998 | &crate::cost_status::RuntimeUsageBatch { |
| 999 | records: vec![record], |
| 1000 | drop_records: vec![drop_record], |
| 1001 | dropped_records: 1, |
| 1002 | }, |
| 1003 | ); |
| 1004 | |
| 1005 | assert!(!result.success); |
| 1006 | let metadata = result |
| 1007 | .metadata |
| 1008 | .expect("usage metadata on failed tool result"); |
| 1009 | let batch = crate::cost_status::child_usage_records_from_metadata(&metadata) |
| 1010 | .expect("preferred routed batch"); |
| 1011 | assert_eq!(batch.dropped_records, 1); |
| 1012 | assert_eq!(batch.records.len(), 1); |
| 1013 | assert_eq!(batch.drop_records.len(), 1); |
| 1014 | assert_eq!(batch.records[0].usage.route, route); |
| 1015 | assert_eq!(batch.records[0].usage.usage, usage); |
| 1016 | assert_eq!(batch.drop_records[0].route, route); |
| 1017 | } |
| 1018 | |
| 1019 | #[test] |
| 1020 | fn rlm_eval_requires_approval() { |
| 1021 | let tool = RlmTool::alias("rlm_eval", "eval", None); |
| 1022 | assert_eq!(tool.approval_requirement(), ApprovalRequirement::Required); |
| 1023 | assert!( |
| 1024 | tool.capabilities() |
| 1025 | .contains(&ToolCapability::RequiresApproval) |
| 1026 | ); |
| 1027 | |
| 1028 | // Evaluation requires approval; concrete open inputs are classified below. |
| 1029 | let canonical = RlmTool::new("rlm", None); |
| 1030 | assert_eq!( |
| 1031 | canonical.approval_requirement_for(&json!({"action": "eval"})), |
| 1032 | ApprovalRequirement::Required |
| 1033 | ); |
| 1034 | assert_eq!( |
| 1035 | canonical.approval_requirement_for(&json!({"action": "open"})), |
| 1036 | ApprovalRequirement::Auto |
| 1037 | ); |
| 1038 | assert_eq!( |
| 1039 | canonical.approval_requirement_for(&json!({"action": "session_objects"})), |
| 1040 | ApprovalRequirement::Auto |
| 1041 | ); |
| 1042 | } |
| 1043 | |
| 1044 | #[test] |
| 1045 | fn rlm_open_requires_outbound_approval_but_keeps_local_reads_automatic() { |
| 1046 | for tool in [ |
| 1047 | RlmTool::new("rlm", None), |
| 1048 | RlmTool::alias("rlm_open", "open", None), |
| 1049 | ] { |
| 1050 | assert_eq!(tool.approval_requirement(), ApprovalRequirement::Required); |
| 1051 | for source in [ |
| 1052 | json!({"url": "https://example.com/document"}), |
| 1053 | json!({"url": " https://example.com/document ", "content": ""}), |
| 1054 | ] { |
| 1055 | let mut input = source; |
| 1056 | input["action"] = json!("open"); |
| 1057 | assert_eq!( |
| 1058 | tool.approval_requirement_for(&input), |
| 1059 | ApprovalRequirement::Required |
| 1060 | ); |
| 1061 | } |
| 1062 | for source in [ |
| 1063 | json!({"content": "local fixture"}), |
| 1064 | json!({"file_path": "fixture.txt"}), |
| 1065 | json!({"session_object": "fixture-object"}), |
| 1066 | json!({"content": "local fixture", "url": " "}), |
| 1067 | ] { |
| 1068 | let mut input = source; |
| 1069 | input["action"] = json!("open"); |
| 1070 | assert_eq!( |
| 1071 | tool.approval_requirement_for(&input), |
| 1072 | ApprovalRequirement::Auto |
| 1073 | ); |
| 1074 | } |
| 1075 | } |
| 1076 | } |
| 1077 | |
| 1078 | #[test] |
| 1079 | fn read_only_and_parallel_flags_match_legacy_contract() { |
| 1080 | // Legacy: session_objects was parallel-friendly read-only; open carried |
| 1081 | // ExecutesCode (not read-only). Open now classifies the concrete source. |
| 1082 | let session_objects = RlmTool::alias("rlm_session_objects", "session_objects", None); |
| 1083 | assert!(session_objects.supports_parallel()); |
| 1084 | assert!(session_objects.is_read_only_for(&json!({}))); |
| 1085 | |
| 1086 | let open = RlmTool::alias("rlm_open", "open", None); |
| 1087 | assert!(!open.is_read_only_for(&json!({}))); |
| 1088 | assert_eq!(open.approval_requirement(), ApprovalRequirement::Required); |
| 1089 | |
| 1090 | let canonical = RlmTool::new("rlm", None); |
| 1091 | assert!(canonical.supports_parallel_for(&json!({"action": "session_objects"}))); |
| 1092 | assert!(!canonical.supports_parallel_for(&json!({"action": "eval"}))); |
| 1093 | assert!(canonical.is_read_only_for(&json!({"action": "configure"}))); |
| 1094 | assert!(!canonical.is_read_only_for(&json!({"action": "open"}))); |
| 1095 | assert!(!canonical.is_read_only_for(&json!({"action": "eval"}))); |
| 1096 | } |
| 1097 | |
| 1098 | #[test] |
| 1099 | fn canonical_rejects_unknown_or_missing_action() { |
| 1100 | let tool = RlmTool::new("rlm", None); |
| 1101 | let err = tool |
| 1102 | .resolve_action(&json!({})) |
| 1103 | .expect_err("missing action must fail"); |
| 1104 | assert!(err.to_string().contains("missing `action`")); |
| 1105 | let err = tool |
| 1106 | .resolve_action(&json!({"action": "explode"})) |
| 1107 | .expect_err("unknown action must fail"); |
| 1108 | assert!(err.to_string().contains("invalid action")); |
| 1109 | } |
| 1110 | |
| 1111 | #[test] |
| 1112 | fn rlm_open_source_count_ignores_empty_string_defaults() { |
| 1113 | assert_eq!( |
| 1114 | rlm_open_source_count( |
| 1115 | &json!({"name": "url-doc", "file_path": "", "content": "", "url": "https://example.com/doc"}) |
| 1116 | ), |
| 1117 | 1 |
| 1118 | ); |
| 1119 | assert_eq!( |
| 1120 | rlm_open_source_count( |
| 1121 | &json!({"name": "inline-doc", "file_path": "", "content": "body", "url": ""}) |
| 1122 | ), |
| 1123 | 1 |
| 1124 | ); |
| 1125 | assert_eq!( |
| 1126 | rlm_open_source_count(&json!({"content": "body", "url": "https://example.com/doc"})), |
| 1127 | 2 |
| 1128 | ); |
| 1129 | assert_eq!( |
| 1130 | rlm_open_source_count( |
| 1131 | &json!({"content": "body", "session_object": "session://active/system_prompt"}) |
| 1132 | ), |
| 1133 | 2 |
| 1134 | ); |
| 1135 | } |
| 1136 | |
| 1137 | #[tokio::test] |
| 1138 | async fn rlm_session_objects_lists_active_prompt_object() { |
| 1139 | let ctx = ctx_with_session_objects(); |
| 1140 | let result = RlmTool::alias("rlm_session_objects", "session_objects", None) |
| 1141 | .execute(json!({}), &ctx) |
| 1142 | .await |
| 1143 | .expect("list session objects"); |
| 1144 | let body: Value = serde_json::from_str(&result.content).expect("json"); |
| 1145 | let objects = body["objects"].as_array().expect("objects array"); |
| 1146 | |
| 1147 | assert!(objects.iter().any(|object| { |
| 1148 | object["id"] == "session://active/system_prompt" && object["kind"] == "system_prompt" |
| 1149 | })); |
| 1150 | assert!(objects.iter().any(|object| { |
| 1151 | object["id"] == "session://active/messages/0" && object["kind"] == "message" |
| 1152 | })); |
| 1153 | } |
| 1154 | |
| 1155 | #[tokio::test] |
| 1156 | async fn rlm_open_loads_active_session_prompt_object() { |
| 1157 | let ctx = ctx_with_session_objects(); |
| 1158 | let open = RlmTool::alias("rlm_open", "open", None) |
| 1159 | .execute( |
| 1160 | json!({"name": "active_prompt", "session_object": "session://active/system_prompt"}), |
| 1161 | &ctx, |
| 1162 | ) |
| 1163 | .await |
| 1164 | .expect("open prompt object"); |
| 1165 | let open_json: Value = serde_json::from_str(&open.content).expect("open json"); |
| 1166 | assert_eq!(open_json["type"], "session_object:system_prompt"); |
| 1167 | assert!( |
| 1168 | open_json["preview_500"] |
| 1169 | .as_str() |
| 1170 | .unwrap() |
| 1171 | .contains("CodeWhale") |
| 1172 | ); |
| 1173 | |
| 1174 | RlmTool::alias("rlm_close", "close", None) |
| 1175 | .execute(json!({"name": "active_prompt"}), &ctx) |
| 1176 | .await |
| 1177 | .expect("close"); |
| 1178 | } |
| 1179 | |
| 1180 | #[tokio::test] |
| 1181 | async fn rlm_open_loads_transcript_message_object() { |
| 1182 | let ctx = ctx_with_session_objects(); |
| 1183 | let open = RlmTool::alias("rlm_open", "open", None) |
| 1184 | .execute( |
| 1185 | json!({"name": "first_message", "session_object": "session://active/messages/0"}), |
| 1186 | &ctx, |
| 1187 | ) |
| 1188 | .await |
| 1189 | .expect("open transcript slice"); |
| 1190 | let open_json: Value = serde_json::from_str(&open.content).expect("open json"); |
| 1191 | assert_eq!(open_json["type"], "session_object:message"); |
| 1192 | assert!( |
| 1193 | open_json["preview_500"] |
| 1194 | .as_str() |
| 1195 | .unwrap() |
| 1196 | .contains("RLM surface") |
| 1197 | ); |
| 1198 | |
| 1199 | RlmTool::alias("rlm_close", "close", None) |
| 1200 | .execute(json!({"name": "first_message"}), &ctx) |
| 1201 | .await |
| 1202 | .expect("close"); |
| 1203 | } |
| 1204 | |
| 1205 | #[tokio::test] |
| 1206 | async fn rlm_open_ignores_blank_source_defaults_from_schema_fillers() { |
| 1207 | let ctx = ctx(); |
| 1208 | RlmTool::alias("rlm_open", "open", None) |
| 1209 | .execute( |
| 1210 | json!({"name": "blank-defaults", "file_path": "", "content": "body", "url": ""}), |
| 1211 | &ctx, |
| 1212 | ) |
| 1213 | .await |
| 1214 | .expect("open with blank sibling source fields"); |
| 1215 | |
| 1216 | RlmTool::alias("rlm_close", "close", None) |
| 1217 | .execute(json!({"name": "blank-defaults"}), &ctx) |
| 1218 | .await |
| 1219 | .expect("close"); |
| 1220 | } |
| 1221 | |
| 1222 | #[tokio::test] |
| 1223 | async fn rlm_open_misnamed_source_field_gets_did_you_mean_hint() { |
| 1224 | // #2655: a wrong source field name yields actionable guidance, not just |
| 1225 | // the canonical "provide exactly one" message. |
| 1226 | let ctx = ctx(); |
| 1227 | let err = RlmTool::alias("rlm_open", "open", None) |
| 1228 | .execute(json!({"name": "doc", "prompt": "summarize this"}), &ctx) |
| 1229 | .await |
| 1230 | .expect_err("misnamed source field should fail"); |
| 1231 | let msg = err.to_string(); |
| 1232 | assert!(msg.contains("file_path"), "names the real fields: {msg}"); |
| 1233 | assert!( |
| 1234 | msg.contains("`url`, or `session_object`"), |
| 1235 | "names session_object in the valid source field list: {msg}" |
| 1236 | ); |
| 1237 | assert!(msg.contains("prompt"), "echoes the wrong field: {msg}"); |
| 1238 | } |
| 1239 | |
| 1240 | #[tokio::test] |
| 1241 | async fn rlm_eval_missing_code_explains_raw_python() { |
| 1242 | // #2655: the missing-code error should teach the tool, with an example. |
| 1243 | let ctx = ctx(); |
| 1244 | let err = RlmTool::alias("rlm_eval", "eval", None) |
| 1245 | .execute(json!({"name": "doc"}), &ctx) |
| 1246 | .await |
| 1247 | .expect_err("missing code should fail"); |
| 1248 | let msg = err.to_string(); |
| 1249 | assert!(msg.contains("raw Python"), "explains it runs Python: {msg}"); |
| 1250 | assert!( |
| 1251 | msg.contains("print(len(content))") || msg.contains("FINAL"), |
| 1252 | "includes an example: {msg}" |
| 1253 | ); |
| 1254 | } |
| 1255 | |
| 1256 | #[test] |
| 1257 | fn rlm_eval_schema_names_the_runtime_content_variable() { |
| 1258 | let schema = RlmTool::alias("rlm_eval", "eval", None).input_schema(); |
| 1259 | let description = schema["properties"]["code"]["description"] |
| 1260 | .as_str() |
| 1261 | .expect("rlm_eval code description"); |
| 1262 | |
| 1263 | assert!(description.contains("`content`")); |
| 1264 | assert!(description.contains("print(len(content))")); |
| 1265 | assert!(!description.contains("SOURCE")); |
| 1266 | } |
| 1267 | |
| 1268 | #[tokio::test] |
| 1269 | async fn rlm_session_open_eval_close_lifecycle() { |
| 1270 | let ctx = ctx(); |
| 1271 | RlmTool::alias("rlm_open", "open", None) |
| 1272 | .execute( |
| 1273 | json!({"name": "sample", "content": "alpha\nbeta\ngamma"}), |
| 1274 | &ctx, |
| 1275 | ) |
| 1276 | .await |
| 1277 | .expect("open"); |
| 1278 | |
| 1279 | let eval = RlmTool::alias("rlm_eval", "eval", None) |
| 1280 | .execute(json!({"name": "sample", "code": "print('ok')"}), &ctx) |
| 1281 | .await |
| 1282 | .expect("eval"); |
| 1283 | let eval_json: Value = serde_json::from_str(&eval.content).expect("eval json"); |
| 1284 | let stdout_preview = eval_json["stdout_preview"] |
| 1285 | .as_str() |
| 1286 | .expect("stdout_preview") |
| 1287 | .replace("\r\n", "\n"); |
| 1288 | assert_eq!(stdout_preview, "ok\n"); |
| 1289 | |
| 1290 | let close = RlmTool::alias("rlm_close", "close", None) |
| 1291 | .execute(json!({"name": "sample"}), &ctx) |
| 1292 | .await |
| 1293 | .expect("close"); |
| 1294 | assert!(close.content.contains("sample")); |
| 1295 | } |
| 1296 | |
| 1297 | #[tokio::test] |
| 1298 | async fn rlm_canonical_action_routing_runs_full_lifecycle() { |
| 1299 | // The visible surface: one `rlm` tool, action-parameterized. |
| 1300 | let ctx = ctx(); |
| 1301 | let tool = RlmTool::new("rlm", None); |
| 1302 | tool.execute( |
| 1303 | json!({"action": "open", "name": "canonical", "content": "body"}), |
| 1304 | &ctx, |
| 1305 | ) |
| 1306 | .await |
| 1307 | .expect("open via canonical action"); |
| 1308 | |
| 1309 | let eval = tool |
| 1310 | .execute( |
| 1311 | json!({"action": "eval", "name": "canonical", "code": "print('ok')"}), |
| 1312 | &ctx, |
| 1313 | ) |
| 1314 | .await |
| 1315 | .expect("eval via canonical action"); |
| 1316 | let eval_json: Value = serde_json::from_str(&eval.content).expect("eval json"); |
| 1317 | let stdout_preview = eval_json["stdout_preview"] |
| 1318 | .as_str() |
| 1319 | .expect("stdout_preview") |
| 1320 | .replace("\r\n", "\n"); |
| 1321 | assert_eq!(stdout_preview, "ok\n"); |
| 1322 | |
| 1323 | let close = tool |
| 1324 | .execute(json!({"action": "close", "name": "canonical"}), &ctx) |
| 1325 | .await |
| 1326 | .expect("close via canonical action"); |
| 1327 | assert!(close.content.contains("canonical")); |
| 1328 | } |
| 1329 | |
| 1330 | #[tokio::test] |
| 1331 | async fn rlm_eval_final_returns_handle() { |
| 1332 | let ctx = ctx(); |
| 1333 | RlmTool::alias("rlm_open", "open", None) |
| 1334 | .execute(json!({"name": "finals", "content": "body"}), &ctx) |
| 1335 | .await |
| 1336 | .expect("open"); |
| 1337 | |
| 1338 | let eval = RlmTool::alias("rlm_eval", "eval", None) |
| 1339 | .execute( |
| 1340 | json!({"name": "finals", "code": "finalize('done', confidence=0.8)"}), |
| 1341 | &ctx, |
| 1342 | ) |
| 1343 | .await |
| 1344 | .expect("eval"); |
| 1345 | let eval_json: Value = serde_json::from_str(&eval.content).expect("eval json"); |
| 1346 | assert_eq!(eval_json["final"]["kind"], "var_handle"); |
| 1347 | assert_eq!(eval_json["final"]["name"], "final_1"); |
| 1348 | assert_eq!(eval_json["confidence"], 0.8); |
| 1349 | |
| 1350 | RlmTool::alias("rlm_close", "close", None) |
| 1351 | .execute(json!({"name": "finals"}), &ctx) |
| 1352 | .await |
| 1353 | .expect("close"); |
| 1354 | } |
| 1355 | |
| 1356 | #[tokio::test] |
| 1357 | async fn rlm_eval_final_preserves_json_handle() { |
| 1358 | let ctx = ctx(); |
| 1359 | RlmTool::alias("rlm_open", "open", None) |
| 1360 | .execute(json!({"name": "json-final", "content": "body"}), &ctx) |
| 1361 | .await |
| 1362 | .expect("open"); |
| 1363 | |
| 1364 | let eval = RlmTool::alias("rlm_eval", "eval", None) |
| 1365 | .execute( |
| 1366 | json!({"name": "json-final", "code": "finalize({'answer': 42, 'items': ['a', 'b']})"}), |
| 1367 | &ctx, |
| 1368 | ) |
| 1369 | .await |
| 1370 | .expect("eval"); |
| 1371 | let eval_json: Value = serde_json::from_str(&eval.content).expect("eval json"); |
| 1372 | assert_eq!(eval_json["final"]["kind"], "var_handle"); |
| 1373 | assert_eq!(eval_json["final"]["type"], "dict"); |
| 1374 | assert_eq!(eval_json["final"]["length"], 2); |
| 1375 | |
| 1376 | let read = HandleReadTool |
| 1377 | .execute( |
| 1378 | json!({"handle": eval_json["final"].clone(), "jsonpath": "$.items[*]"}), |
| 1379 | &ctx, |
| 1380 | ) |
| 1381 | .await |
| 1382 | .expect("read final handle"); |
| 1383 | let read_json: Value = serde_json::from_str(&read.content).expect("read json"); |
| 1384 | assert_eq!(read_json["matches"], json!(["a", "b"])); |
| 1385 | |
| 1386 | RlmTool::alias("rlm_close", "close", None) |
| 1387 | .execute(json!({"name": "json-final"}), &ctx) |
| 1388 | .await |
| 1389 | .expect("close"); |
| 1390 | } |
| 1391 | |
| 1392 | #[tokio::test] |
| 1393 | async fn rlm_configure_metadata_omits_stdout() { |
| 1394 | let ctx = ctx(); |
| 1395 | RlmTool::alias("rlm_open", "open", None) |
| 1396 | .execute(json!({"name": "quiet", "content": "body"}), &ctx) |
| 1397 | .await |
| 1398 | .expect("open"); |
| 1399 | RlmTool::alias("rlm_configure", "configure", None) |
| 1400 | .execute( |
| 1401 | json!({"name": "quiet", "output_feedback": "metadata", "sub_rlm_max_depth": 99}), |
| 1402 | &ctx, |
| 1403 | ) |
| 1404 | .await |
| 1405 | .expect("configure"); |
| 1406 | |
| 1407 | let eval = RlmTool::alias("rlm_eval", "eval", None) |
| 1408 | .execute(json!({"name": "quiet", "code": "print('hidden')"}), &ctx) |
| 1409 | .await |
| 1410 | .expect("eval"); |
| 1411 | let eval_json: Value = serde_json::from_str(&eval.content).expect("eval json"); |
| 1412 | assert!(eval_json.get("stdout_preview").is_none()); |
| 1413 | |
| 1414 | RlmTool::alias("rlm_close", "close", None) |
| 1415 | .execute(json!({"name": "quiet"}), &ctx) |
| 1416 | .await |
| 1417 | .expect("close"); |
| 1418 | } |
| 1419 | } |
| 1420 |