development
claude/github-issue-40-0au5ua
claude/linked-issue-acp-tui-u4vvpa
claude/minor-release-cut-vkugow
claude/sigit-acp-local-chat-cx6380
claude/sigit-cloud-agent-expansion-reox0a
claude/tool-permission-system
claude/zen-feynman-0u78dk
development
feature/acp-cancel-request
feature/acp-session-list-cursor
feature/acp-v1-conformance
feature/agent-tools-multiedit-glob-todos-remember
feature/cargo-workspace
feature/hooks
feature/load-local-model-explicitly
feature/sigit-agent-profile
feature/sigit-code-cloud
feature/xcode-mcp-server
fix/acp-replay-loaded-session-history
fix/acp-replay-reasoning-and-attachments
fix/dsml-doubled-bar-markers
main
release/v1.5.10
release/v1.5.6
release/v1.6.1
| 1 | //! Recovery for tool calls a model emits as literal text instead of the |
| 2 | //! endpoint's structured `tool_calls` field. |
| 3 | //! |
| 4 | //! Several open-weight families siGit Code can drive are fine-tuned on chat |
| 5 | //! templates that render a tool call inside assistant content. A serving stack |
| 6 | //! is supposed to parse that back into the OpenAI-shaped `tool_calls` field |
| 7 | //! before it reaches us. When it doesn't, the raw tag arrives as ordinary |
| 8 | //! content. |
| 9 | //! |
| 10 | //! Nothing then executes it. Worse, `consume_stream` forwards content to the |
| 11 | //! UI as it arrives, so the tag is rendered verbatim in the editor and the |
| 12 | //! turn ends with no tool calls, which reads to the agent loop as "the model |
| 13 | //! chose to answer in prose". Whatever it was actually trying to do is |
| 14 | //! dropped. |
| 15 | //! |
| 16 | //! This lives here rather than only in the gateway because siGit Code talks to |
| 17 | //! more than one kind of endpoint: siGit Code Cloud, an arbitrary |
| 18 | //! OpenAI-compatible base URL from `providers.toml` or `OPENAI_BASE_URL`, and |
| 19 | //! on-device models. A fix in any single upstream leaves the others exposed. |
| 20 | //! |
| 21 | //! Only well-formed shapes are recovered. A block that doesn't match is never |
| 22 | //! guessed at: reissuing a `run_command` is cheap, but guessing wrong at a |
| 23 | //! half-parsed `edit_file` would write the wrong change to a file. It isn't |
| 24 | //! shown either. A tool-call block that fails to parse, or that is still open |
| 25 | //! when the reply ends, is reported as [`ScanEvent::Malformed`] and kept out of |
| 26 | //! both the UI and the conversation history. Left in the history, the model |
| 27 | //! reads back a call it "made" with no result attached, and later turns start |
| 28 | //! inventing `<function_results>` blocks of their own. The backend tells the |
| 29 | //! model the call didn't run and asks for it again instead. |
| 30 | //! |
| 31 | //! A marker the model merely mentions in prose, say while explaining this very |
| 32 | //! scanner, is not a call. So a block that fails to parse is only treated as |
| 33 | //! malformed when what follows the marker looks like the start of a call (see |
| 34 | //! [`looks_like_xml_call`] and [`looks_like_k3_call`]); otherwise it stays |
| 35 | //! text, as it always did. |
| 36 | |
| 37 | use std::collections::HashMap; |
| 38 | |
| 39 | use crate::backend::ToolSpec; |
| 40 | |
| 41 | const XML_OPEN_TAG: &str = "<tool_call>"; |
| 42 | const XML_CLOSE_TAG: &str = "</tool_call>"; |
| 43 | |
| 44 | const K3_TOOLS_OPEN: &str = "<|open|>tools<|sep|>"; |
| 45 | const K3_TOOLS_CLOSE: &str = "<|close|>tools<|sep|>"; |
| 46 | const K3_CALL_OPEN: &str = "<|open|>call"; |
| 47 | const K3_CALL_CLOSE: &str = "<|close|>call<|sep|>"; |
| 48 | const K3_ARGUMENT_OPEN: &str = "<|open|>argument"; |
| 49 | const K3_ARGUMENT_CLOSE: &str = "<|close|>argument<|sep|>"; |
| 50 | const K3_RESPONSE_OPEN: &str = "<|open|>response<|sep|>"; |
| 51 | const K3_RESPONSE_CLOSE: &str = "<|close|>response<|sep|>"; |
| 52 | const K3_THINK_OPEN: &str = "<|open|>think<|sep|>"; |
| 53 | const K3_THINK_CLOSE: &str = "<|close|>think<|sep|>"; |
| 54 | const K3_SEP: &str = "<|sep|>"; |
| 55 | |
| 56 | const DSML_TOOL_CALLS_OPEN: &str = "<｜DSML｜tool_calls>"; |
| 57 | const DSML_TOOL_CALLS_CLOSE: &str = "<｜DSML｜/tool_calls>"; |
| 58 | const DSML_INVOKE_OPEN: &str = "<｜DSML｜invoke"; |
| 59 | const DSML_INVOKE_CLOSE: &str = "<｜DSML｜/invoke>"; |
| 60 | const DSML_PARAMETER_OPEN: &str = "<｜DSML｜parameter"; |
| 61 | const DSML_PARAMETER_CLOSE: &str = "<｜DSML｜/parameter>"; |
| 62 | |
| 63 | const START_MARKERS: &[&str] = &[ |
| 64 | XML_OPEN_TAG, |
| 65 | K3_TOOLS_OPEN, |
| 66 | K3_RESPONSE_OPEN, |
| 67 | K3_THINK_OPEN, |
| 68 | DSML_TOOL_CALLS_OPEN, |
| 69 | ]; |
| 70 | |
| 71 | /// One tool call recovered from inline text. |
| 72 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 73 | pub struct Recovered { |
| 74 | pub name: String, |
| 75 | /// Arguments as a JSON-encoded string, matching `ToolCall::arguments`. |
| 76 | pub arguments: String, |
| 77 | } |
| 78 | |
| 79 | /// What [`extract`] found in a complete text blob. |
| 80 | #[derive(Debug, Default, PartialEq, Eq)] |
| 81 | pub struct Extracted { |
| 82 | /// The text with every tool-call block removed, recovered or not. |
| 83 | pub text: String, |
| 84 | /// Calls recovered from well-formed blocks, in order. |
| 85 | pub calls: Vec<Recovered>, |
| 86 | /// Tool-call blocks that were dropped because they couldn't be parsed. |
| 87 | pub malformed: usize, |
| 88 | } |
| 89 | |
| 90 | /// Scan a complete text blob for inline tool-call blocks. Recovered blocks |
| 91 | /// become calls; blocks that don't parse cleanly, including an unterminated |
| 92 | /// trailing one, are dropped and counted in [`Extracted::malformed`]. |
| 93 | pub fn extract(text: &str, tools: &[ToolSpec]) -> Extracted { |
| 94 | let mut scanner = StreamScanner::new(tools); |
| 95 | let mut extracted = Extracted { |
| 96 | text: String::with_capacity(text.len()), |
| 97 | ..Extracted::default() |
| 98 | }; |
| 99 | |
| 100 | for event in scanner.push(text).into_iter().chain(scanner.finish()) { |
| 101 | match event { |
| 102 | ScanEvent::Text(text) => extracted.text.push_str(&text), |
| 103 | ScanEvent::ToolCall(call) => extracted.calls.push(call), |
| 104 | ScanEvent::Malformed(_) => extracted.malformed += 1, |
| 105 | } |
| 106 | } |
| 107 | |
| 108 | extracted |
| 109 | } |
| 110 | |
| 111 | /// Parse one legacy XML block's inner text: an offered tool name, then zero or |
| 112 | /// more `<arg_key>K</arg_key><arg_value>V</arg_value>` pairs. The malformed |
| 113 | /// opening marker observed from GLM is accepted only for the first argument; |
| 114 | /// later arguments stay strict so text that merely resembles a call cannot be |
| 115 | /// recovered as one. |
| 116 | fn parse_xml_block(inner: &str, tools: &[ToolSpec]) -> Option<Recovered> { |
| 117 | let key_idx = inner.find("<arg_key>"); |
| 118 | let malformed_key_idx = inner.find(XML_OPEN_TAG); |
| 119 | let check_status_name = key_idx.and_then(|idx| parse_check_status_preamble(&inner[..idx])); |
| 120 | let (name, mut rest, malformed_first_key) = |
| 121 | if let (Some(idx), Some(name)) = (key_idx, check_status_name) { |
| 122 | (name, &inner[idx..], false) |
| 123 | } else { |
| 124 | match (key_idx, malformed_key_idx) { |
| 125 | (Some(key_idx), Some(malformed_key_idx)) if malformed_key_idx < key_idx => ( |
| 126 | inner[..malformed_key_idx].trim(), |
| 127 | &inner[malformed_key_idx..], |
| 128 | true, |
| 129 | ), |
| 130 | (Some(idx), _) => (inner[..idx].trim(), &inner[idx..], false), |
| 131 | (None, Some(idx)) => (inner[..idx].trim(), &inner[idx..], true), |
| 132 | (None, None) => (inner.trim(), "", false), |
| 133 | } |
| 134 | }; |
| 135 | if name.is_empty() || name.contains(['<', '>']) || !offered_tool(tools, name) { |
| 136 | return None; |
| 137 | } |
| 138 | |
| 139 | let mut args = serde_json::Map::new(); |
| 140 | let mut first_key = true; |
| 141 | while !rest.is_empty() { |
| 142 | // The GLM alias is permitted exactly once, at the first key. A later |
| 143 | // `<tool_call>` must reject the complete block rather than letting a |
| 144 | // partial call escape recovery. |
| 145 | if !first_key && rest.starts_with(XML_OPEN_TAG) { |
| 146 | return None; |
| 147 | } |
| 148 | let key_open = if first_key && malformed_first_key { |
| 149 | XML_OPEN_TAG |
| 150 | } else { |
| 151 | "<arg_key>" |
| 152 | }; |
| 153 | let (key, after_key) = split_once(rest.strip_prefix(key_open)?, "</arg_key>")?; |
| 154 | let (value, after_value) = |
| 155 | split_once(after_key.strip_prefix("<arg_value>")?, "</arg_value>")?; |
| 156 | let key = key.trim(); |
| 157 | if key.is_empty() { |
| 158 | return None; |
| 159 | } |
| 160 | args.insert( |
| 161 | key.to_string(), |
| 162 | coerce(value, declared_type(tools, name, key).as_deref()), |
| 163 | ); |
| 164 | rest = after_value; |
| 165 | first_key = false; |
| 166 | } |
| 167 | |
| 168 | Some(Recovered { |
| 169 | name: name.to_string(), |
| 170 | arguments: serde_json::Value::Object(args).to_string(), |
| 171 | }) |
| 172 | } |
| 173 | |
| 174 | /// Some GLM-family responses prefix the real arguments with generation-control |
| 175 | /// metadata shaped like `tool_name CheckStatus=value</arg_value>`. It is not a |
| 176 | /// tool argument: accept that exact bounded preamble and discard it, while |
| 177 | /// leaving every other malformed prefix untouched. |
| 178 | fn parse_check_status_preamble(prefix: &str) -> Option<&str> { |
| 179 | let (name, status) = prefix.trim().split_once(" CheckStatus=")?; |
| 180 | let status = status.strip_suffix("</arg_value>")?; |
| 181 | if name.is_empty() |
| 182 | || status.is_empty() |
| 183 | || !status |
| 184 | .chars() |
| 185 | .all(|character| character.is_ascii_alphanumeric() || matches!(character, '_' | '-')) |
| 186 | { |
| 187 | return None; |
| 188 | } |
| 189 | Some(name) |
| 190 | } |
| 191 | |
| 192 | fn parse_k3_tools_block(inner: &str, tools: &[ToolSpec]) -> Option<Vec<Recovered>> { |
| 193 | let mut calls = Vec::new(); |
| 194 | let mut rest = inner; |
| 195 | |
| 196 | while !rest.trim().is_empty() { |
| 197 | rest = rest.trim_start(); |
| 198 | let header = rest.strip_prefix(K3_CALL_OPEN)?; |
| 199 | let (attr_text, after_header) = split_once(header, K3_SEP)?; |
| 200 | let attrs = parse_k3_attributes(attr_text)?; |
| 201 | let name = attrs.get("tool")?; |
| 202 | if name.is_empty() || !offered_tool(tools, name) { |
| 203 | return None; |
| 204 | } |
| 205 | |
| 206 | let (call_inner, after_call) = split_once(after_header, K3_CALL_CLOSE)?; |
| 207 | let arguments = parse_k3_arguments(call_inner, tools, name)?; |
| 208 | calls.push(Recovered { |
| 209 | name: name.to_string(), |
| 210 | arguments: serde_json::Value::Object(arguments).to_string(), |
| 211 | }); |
| 212 | rest = after_call; |
| 213 | } |
| 214 | |
| 215 | Some(calls) |
| 216 | } |
| 217 | |
| 218 | fn parse_k3_arguments( |
| 219 | mut rest: &str, |
| 220 | tools: &[ToolSpec], |
| 221 | tool_name: &str, |
| 222 | ) -> Option<serde_json::Map<String, serde_json::Value>> { |
| 223 | let mut args = serde_json::Map::new(); |
| 224 | |
| 225 | while !rest.trim().is_empty() { |
| 226 | rest = rest.trim_start(); |
| 227 | let header = rest.strip_prefix(K3_ARGUMENT_OPEN)?; |
| 228 | let (attr_text, after_header) = split_once(header, K3_SEP)?; |
| 229 | let attrs = parse_k3_attributes(attr_text)?; |
| 230 | let key = attrs.get("key")?.trim(); |
| 231 | if key.is_empty() { |
| 232 | return None; |
| 233 | } |
| 234 | // K3 normally emits `type` for every argument. If an endpoint omits |
| 235 | // it, retain the legacy recovery path's schema-based typing instead of |
| 236 | // turning e.g. an integer `task_id` into a JSON string. |
| 237 | let schema_type = declared_type(tools, tool_name, key); |
| 238 | let value_type = attrs |
| 239 | .get("type") |
| 240 | .map(String::as_str) |
| 241 | .or(schema_type.as_deref()) |
| 242 | .unwrap_or("string"); |
| 243 | |
| 244 | let (raw_value, after_argument) = split_once(after_header, K3_ARGUMENT_CLOSE)?; |
| 245 | let value = if value_type == "string" { |
| 246 | serde_json::Value::String(raw_value.to_string()) |
| 247 | } else { |
| 248 | serde_json::from_str(raw_value).ok()? |
| 249 | }; |
| 250 | args.insert(key.to_string(), value); |
| 251 | rest = after_argument; |
| 252 | } |
| 253 | |
| 254 | Some(args) |
| 255 | } |
| 256 | |
| 257 | fn parse_k3_attributes(mut text: &str) -> Option<HashMap<String, String>> { |
| 258 | let mut attrs = HashMap::new(); |
| 259 | text = text.trim(); |
| 260 | |
| 261 | while !text.is_empty() { |
| 262 | let eq_idx = text.find('=')?; |
| 263 | let key = text[..eq_idx].trim(); |
| 264 | if key.is_empty() |
| 265 | || !key |
| 266 | .chars() |
| 267 | .all(|ch| ch.is_ascii_alphanumeric() || ch == '_' || ch == '-') |
| 268 | { |
| 269 | return None; |
| 270 | } |
| 271 | |
| 272 | let mut value_part = text[eq_idx + 1..].trim_start(); |
| 273 | value_part = value_part.strip_prefix('"')?; |
| 274 | let end_idx = value_part.find('"')?; |
| 275 | let raw_value = &value_part[..end_idx]; |
| 276 | attrs.insert(key.to_string(), decode_k3_attribute(raw_value)); |
| 277 | |
| 278 | text = value_part[end_idx + 1..].trim_start(); |
| 279 | } |
| 280 | |
| 281 | Some(attrs) |
| 282 | } |
| 283 | |
| 284 | fn decode_k3_attribute(value: &str) -> String { |
| 285 | // K3's encoder escapes only `&` and `"` in attributes. These replacements |
| 286 | // deliberately mirror that format rather than implementing general HTML |
| 287 | // entity decoding (which could alter literal argument names). |
| 288 | value.replace(""", "\"").replace("&", "&") |
| 289 | } |
| 290 | |
| 291 | /// Parse a DeepSeek DSML `<｜DSML｜tool_calls>` envelope's inner text: zero or |
| 292 | /// more `<｜DSML｜invoke name="$TOOL">...</｜DSML｜invoke>` blocks, each |
| 293 | /// holding `<｜DSML｜parameter name="$KEY" string="true|false">...` entries. |
| 294 | fn parse_dsml_block(inner: &str, tools: &[ToolSpec]) -> Option<Vec<Recovered>> { |
| 295 | let mut calls = Vec::new(); |
| 296 | let mut rest = inner; |
| 297 | |
| 298 | while !rest.trim().is_empty() { |
| 299 | rest = rest.trim_start(); |
| 300 | let header = rest.strip_prefix(DSML_INVOKE_OPEN)?; |
| 301 | let (attr_text, after_header) = split_once(header, ">")?; |
| 302 | let attrs = parse_k3_attributes(attr_text)?; |
| 303 | let name = attrs.get("name")?; |
| 304 | if name.is_empty() || !offered_tool(tools, name) { |
| 305 | return None; |
| 306 | } |
| 307 | |
| 308 | let (call_inner, after_call) = split_once(after_header, DSML_INVOKE_CLOSE)?; |
| 309 | let arguments = parse_dsml_parameters(call_inner, tools, name)?; |
| 310 | calls.push(Recovered { |
| 311 | name: name.to_string(), |
| 312 | arguments: serde_json::Value::Object(arguments).to_string(), |
| 313 | }); |
| 314 | rest = after_call; |
| 315 | } |
| 316 | |
| 317 | Some(calls) |
| 318 | } |
| 319 | |
| 320 | fn parse_dsml_parameters( |
| 321 | mut rest: &str, |
| 322 | tools: &[ToolSpec], |
| 323 | tool_name: &str, |
| 324 | ) -> Option<serde_json::Map<String, serde_json::Value>> { |
| 325 | let mut args = serde_json::Map::new(); |
| 326 | |
| 327 | while !rest.trim().is_empty() { |
| 328 | rest = rest.trim_start(); |
| 329 | let header = rest.strip_prefix(DSML_PARAMETER_OPEN)?; |
| 330 | let (attr_text, after_header) = split_once(header, ">")?; |
| 331 | let attrs = parse_k3_attributes(attr_text)?; |
| 332 | let key = attrs.get("name")?.trim(); |
| 333 | if key.is_empty() { |
| 334 | return None; |
| 335 | } |
| 336 | // DSML marks each parameter's encoding with `string="true|false"`. If |
| 337 | // an endpoint omits it, fall back to the schema-based typing the |
| 338 | // other recovery paths use instead of mistyping the value. |
| 339 | let is_string = match attrs.get("string").map(String::as_str) { |
| 340 | Some("true") => true, |
| 341 | Some("false") => false, |
| 342 | _ => { |
| 343 | declared_type(tools, tool_name, key) |
| 344 | .as_deref() |
| 345 | .unwrap_or("string") |
| 346 | == "string" |
| 347 | } |
| 348 | }; |
| 349 | |
| 350 | let (raw_value, after_param) = split_once(after_header, DSML_PARAMETER_CLOSE)?; |
| 351 | let value = if is_string { |
| 352 | serde_json::Value::String(raw_value.to_string()) |
| 353 | } else { |
| 354 | serde_json::from_str(raw_value).ok()? |
| 355 | }; |
| 356 | args.insert(key.to_string(), value); |
| 357 | rest = after_param; |
| 358 | } |
| 359 | |
| 360 | Some(args) |
| 361 | } |
| 362 | |
| 363 | fn split_once<'a>(s: &'a str, delim: &str) -> Option<(&'a str, &'a str)> { |
| 364 | let idx = s.find(delim)?; |
| 365 | Some((&s[..idx], &s[idx + delim.len()..])) |
| 366 | } |
| 367 | |
| 368 | fn offered_tool(tools: &[ToolSpec], tool_name: &str) -> bool { |
| 369 | tools.iter().any(|spec| spec.name == tool_name) |
| 370 | } |
| 371 | |
| 372 | /// The JSON Schema `type` declared for `tool_name`'s `key` parameter, e.g. |
| 373 | /// `"integer"` for `command_output`'s `task_id`. [`ToolSpec`] carries the |
| 374 | /// schema as an unparsed string, so this parses it per lookup; the argument |
| 375 | /// counts involved are tiny and this only runs on the recovery path. |
| 376 | fn declared_type(tools: &[ToolSpec], tool_name: &str, key: &str) -> Option<String> { |
| 377 | let spec = tools.iter().find(|spec| spec.name == tool_name)?; |
| 378 | let schema: serde_json::Value = serde_json::from_str(&spec.parameters_schema).ok()?; |
| 379 | Some( |
| 380 | schema |
| 381 | .get("properties")? |
| 382 | .get(key)? |
| 383 | .get("type")? |
| 384 | .as_str()? |
| 385 | .to_string(), |
| 386 | ) |
| 387 | } |
| 388 | |
| 389 | /// Convert a raw legacy text value to JSON per its declared schema type. |
| 390 | /// Falls back to a plain string for `"string"`, an unknown type, or a value |
| 391 | /// that doesn't parse as what it claims to be. |
| 392 | fn coerce(raw: &str, declared: Option<&str>) -> serde_json::Value { |
| 393 | match declared { |
| 394 | Some("integer") => raw |
| 395 | .trim() |
| 396 | .parse::<i64>() |
| 397 | .map(serde_json::Value::from) |
| 398 | .unwrap_or_else(|_| serde_json::Value::String(raw.to_string())), |
| 399 | Some("number") => raw |
| 400 | .trim() |
| 401 | .parse::<f64>() |
| 402 | .ok() |
| 403 | .and_then(serde_json::Number::from_f64) |
| 404 | .map(serde_json::Value::Number) |
| 405 | .unwrap_or_else(|| serde_json::Value::String(raw.to_string())), |
| 406 | Some("boolean") => match raw.trim() { |
| 407 | "true" => serde_json::Value::Bool(true), |
| 408 | "false" => serde_json::Value::Bool(false), |
| 409 | _ => serde_json::Value::String(raw.to_string()), |
| 410 | }, |
| 411 | _ => serde_json::Value::String(raw.to_string()), |
| 412 | } |
| 413 | } |
| 414 | |
| 415 | /// One event from incrementally scanning a stream of content deltas. |
| 416 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 417 | pub enum ScanEvent { |
| 418 | /// Plain text, safe to forward to the UI. |
| 419 | Text(String), |
| 420 | /// A tool call recovered from a completed block. |
| 421 | ToolCall(Recovered), |
| 422 | /// A tool-call block that couldn't be parsed, or never closed. Carries the |
| 423 | /// raw block for logging; it must not be shown or executed. |
| 424 | Malformed(String), |
| 425 | } |
| 426 | |
| 427 | /// Incremental scanner for a live delta stream. |
| 428 | /// |
| 429 | /// Holds back only the text that could still turn into a known inline protocol |
| 430 | /// marker, so ordinary answers still reach the UI token by token. Buffering |
| 431 | /// starts once an opening marker appears, and lasts only until the matching |
| 432 | /// close arrives. |
| 433 | pub struct StreamScanner<'a> { |
| 434 | tools: &'a [ToolSpec], |
| 435 | pending: String, |
| 436 | } |
| 437 | |
| 438 | impl<'a> StreamScanner<'a> { |
| 439 | pub fn new(tools: &'a [ToolSpec]) -> Self { |
| 440 | Self { |
| 441 | tools, |
| 442 | pending: String::new(), |
| 443 | } |
| 444 | } |
| 445 | |
| 446 | /// Feed the next content delta. Returns events to emit now, in order; may |
| 447 | /// be empty if the chunk was absorbed into a tag that hasn't closed yet. |
| 448 | pub fn push(&mut self, chunk: &str) -> Vec<ScanEvent> { |
| 449 | self.pending.push_str(chunk); |
| 450 | self.drain() |
| 451 | } |
| 452 | |
| 453 | /// Call once the stream has ended to flush whatever is still held back. |
| 454 | /// A tool-call block that opened but never closed is [`ScanEvent::Malformed`]: |
| 455 | /// it can't be executed safely, and it isn't prose either. Anything else, |
| 456 | /// such as a partial marker prefix or an unclosed Kimi response wrapper, is |
| 457 | /// handed back as text. |
| 458 | pub fn finish(&mut self) -> Option<ScanEvent> { |
| 459 | if self.pending.is_empty() { |
| 460 | return None; |
| 461 | } |
| 462 | let rest = std::mem::take(&mut self.pending); |
| 463 | Some( |
| 464 | if rest |
| 465 | .strip_prefix(K3_TOOLS_OPEN) |
| 466 | .is_some_and(looks_like_k3_call) |
| 467 | || rest |
| 468 | .strip_prefix(XML_OPEN_TAG) |
| 469 | .is_some_and(looks_like_xml_call) |
| 470 | || rest |
| 471 | .strip_prefix(DSML_TOOL_CALLS_OPEN) |
| 472 | .is_some_and(looks_like_dsml_call) |
| 473 | { |
| 474 | ScanEvent::Malformed(rest) |
| 475 | } else { |
| 476 | ScanEvent::Text(rest) |
| 477 | }, |
| 478 | ) |
| 479 | } |
| 480 | |
| 481 | fn drain(&mut self) -> Vec<ScanEvent> { |
| 482 | let mut events = Vec::new(); |
| 483 | loop { |
| 484 | let Some((open_idx, marker)) = find_next_marker(&self.pending) else { |
| 485 | let keep = max_partial_prefix_len(&self.pending, START_MARKERS); |
| 486 | let flush_len = self.pending.len() - keep; |
| 487 | if flush_len > 0 { |
| 488 | events.push(ScanEvent::Text( |
| 489 | self.pending.drain(..flush_len).collect::<String>(), |
| 490 | )); |
| 491 | } |
| 492 | break; |
| 493 | }; |
| 494 | |
| 495 | if open_idx > 0 { |
| 496 | events.push(ScanEvent::Text( |
| 497 | self.pending.drain(..open_idx).collect::<String>(), |
| 498 | )); |
| 499 | } |
| 500 | |
| 501 | match marker { |
| 502 | XML_OPEN_TAG => { |
| 503 | let Some(event) = self.drain_xml() else { |
| 504 | break; |
| 505 | }; |
| 506 | events.push(event); |
| 507 | } |
| 508 | K3_TOOLS_OPEN => match self.drain_k3_tools() { |
| 509 | Some(mut recovered) => events.append(&mut recovered), |
| 510 | None => break, |
| 511 | }, |
| 512 | DSML_TOOL_CALLS_OPEN => match self.drain_dsml() { |
| 513 | Some(mut recovered) => events.append(&mut recovered), |
| 514 | None => break, |
| 515 | }, |
| 516 | K3_RESPONSE_OPEN => { |
| 517 | let Some(mut unwrapped) = |
| 518 | self.drain_wrapped_text(K3_RESPONSE_OPEN, K3_RESPONSE_CLOSE, true) |
| 519 | else { |
| 520 | break; |
| 521 | }; |
| 522 | events.append(&mut unwrapped); |
| 523 | } |
| 524 | K3_THINK_OPEN => { |
| 525 | let Some(mut unwrapped) = |
| 526 | self.drain_wrapped_text(K3_THINK_OPEN, K3_THINK_CLOSE, false) |
| 527 | else { |
| 528 | break; |
| 529 | }; |
| 530 | events.append(&mut unwrapped); |
| 531 | } |
| 532 | _ => unreachable!("unknown inline marker"), |
| 533 | } |
| 534 | } |
| 535 | events |
| 536 | } |
| 537 | |
| 538 | fn drain_xml(&mut self) -> Option<ScanEvent> { |
| 539 | let close_rel = self.pending[XML_OPEN_TAG.len()..].find(XML_CLOSE_TAG)?; |
| 540 | let close_idx = XML_OPEN_TAG.len() + close_rel; |
| 541 | let block: String = self |
| 542 | .pending |
| 543 | .drain(..close_idx + XML_CLOSE_TAG.len()) |
| 544 | .collect(); |
| 545 | let inner = &block[XML_OPEN_TAG.len()..block.len() - XML_CLOSE_TAG.len()]; |
| 546 | Some(match parse_xml_block(inner, self.tools) { |
| 547 | Some(call) => ScanEvent::ToolCall(call), |
| 548 | None if looks_like_xml_call(inner) => ScanEvent::Malformed(block), |
| 549 | None => ScanEvent::Text(block), |
| 550 | }) |
| 551 | } |
| 552 | |
| 553 | fn drain_k3_tools(&mut self) -> Option<Vec<ScanEvent>> { |
| 554 | let close_rel = self.pending[K3_TOOLS_OPEN.len()..].find(K3_TOOLS_CLOSE)?; |
| 555 | let close_idx = K3_TOOLS_OPEN.len() + close_rel; |
| 556 | let block: String = self |
| 557 | .pending |
| 558 | .drain(..close_idx + K3_TOOLS_CLOSE.len()) |
| 559 | .collect(); |
| 560 | let inner = &block[K3_TOOLS_OPEN.len()..block.len() - K3_TOOLS_CLOSE.len()]; |
| 561 | |
| 562 | Some(match parse_k3_tools_block(inner, self.tools) { |
| 563 | Some(calls) => calls.into_iter().map(ScanEvent::ToolCall).collect(), |
| 564 | None if looks_like_k3_call(inner) => vec![ScanEvent::Malformed(block)], |
| 565 | None => vec![ScanEvent::Text(block)], |
| 566 | }) |
| 567 | } |
| 568 | |
| 569 | fn drain_dsml(&mut self) -> Option<Vec<ScanEvent>> { |
| 570 | let close_rel = self.pending[DSML_TOOL_CALLS_OPEN.len()..].find(DSML_TOOL_CALLS_CLOSE)?; |
| 571 | let close_idx = DSML_TOOL_CALLS_OPEN.len() + close_rel; |
| 572 | let block: String = self |
| 573 | .pending |
| 574 | .drain(..close_idx + DSML_TOOL_CALLS_CLOSE.len()) |
| 575 | .collect(); |
| 576 | let inner = &block[DSML_TOOL_CALLS_OPEN.len()..block.len() - DSML_TOOL_CALLS_CLOSE.len()]; |
| 577 | |
| 578 | Some(match parse_dsml_block(inner, self.tools) { |
| 579 | Some(calls) => calls.into_iter().map(ScanEvent::ToolCall).collect(), |
| 580 | None if looks_like_dsml_call(inner) => vec![ScanEvent::Malformed(block)], |
| 581 | None => vec![ScanEvent::Text(block)], |
| 582 | }) |
| 583 | } |
| 584 | |
| 585 | fn drain_wrapped_text( |
| 586 | &mut self, |
| 587 | open: &str, |
| 588 | close: &str, |
| 589 | emit_inner: bool, |
| 590 | ) -> Option<Vec<ScanEvent>> { |
| 591 | let after_open = &self.pending[open.len()..]; |
| 592 | let close_rel = match after_open.find(close) { |
| 593 | Some(close_rel) => close_rel, |
| 594 | None => { |
| 595 | // A malformed response/think wrapper must not swallow a later |
| 596 | // valid tool block forever. Preserve the malformed prefix as |
| 597 | // text, leave the next marker pending, and resume scanning. |
| 598 | let (next_rel, _) = find_next_marker(after_open)?; |
| 599 | return Some(vec![ScanEvent::Text( |
| 600 | self.pending |
| 601 | .drain(..open.len() + next_rel) |
| 602 | .collect::<String>(), |
| 603 | )]); |
| 604 | } |
| 605 | }; |
| 606 | let close_idx = open.len() + close_rel; |
| 607 | let block: String = self.pending.drain(..close_idx + close.len()).collect(); |
| 608 | |
| 609 | if emit_inner { |
| 610 | let inner = &block[open.len()..block.len() - close.len()]; |
| 611 | Some(scan_complete_text(inner, self.tools)) |
| 612 | } else { |
| 613 | Some(Vec::new()) |
| 614 | } |
| 615 | } |
| 616 | } |
| 617 | |
| 618 | /// Scan text from a wrapper whose closing marker has already arrived. This |
| 619 | /// preserves event order when a response body itself contains an inline block. |
| 620 | /// Whether the text after a `<tool_call>` tag reads as the start of a call: |
| 621 | /// a tool name straight after the tag or on the next line, a JSON object or |
| 622 | /// another tag after any whitespace, or nothing at all (a reply cut off right |
| 623 | /// after the tag). Prose that mentions the tag, like "`<tool_call>` wraps..." |
| 624 | /// or "<tool_call> and </tool_call> tags", fails every one of those. |
| 625 | fn looks_like_xml_call(after_tag: &str) -> bool { |
| 626 | let after_newlines = after_tag.trim_start_matches(['\n', '\r']); |
| 627 | if after_newlines |
| 628 | .chars() |
| 629 | .next() |
| 630 | .is_some_and(|c| c.is_ascii_alphanumeric() || c == '_') |
| 631 | { |
| 632 | return true; |
| 633 | } |
| 634 | let rest = after_newlines.trim_start(); |
| 635 | rest.is_empty() || rest.starts_with(['{', '<']) |
| 636 | } |
| 637 | |
| 638 | /// Whether the text after a Kimi `<|open|>tools<|sep|>` marker reads as the |
| 639 | /// start of a call: another `<|open|>call` marker after any whitespace, or |
| 640 | /// nothing at all. A quoted marker, like the constant in this file, fails. |
| 641 | fn looks_like_k3_call(after_marker: &str) -> bool { |
| 642 | let rest = after_marker.trim_start(); |
| 643 | rest.is_empty() || rest.starts_with(K3_CALL_OPEN) |
| 644 | } |
| 645 | |
| 646 | /// Whether the text after a DeepSeek `<｜DSML｜tool_calls>` marker reads as |
| 647 | /// the start of a call: a `<｜DSML｜invoke` marker after any whitespace, or |
| 648 | /// nothing at all. A quoted marker, like the constant in this file, fails. |
| 649 | fn looks_like_dsml_call(after_marker: &str) -> bool { |
| 650 | let rest = after_marker.trim_start(); |
| 651 | rest.is_empty() || rest.starts_with(DSML_INVOKE_OPEN) |
| 652 | } |
| 653 | |
| 654 | fn scan_complete_text(text: &str, tools: &[ToolSpec]) -> Vec<ScanEvent> { |
| 655 | let mut scanner = StreamScanner::new(tools); |
| 656 | let mut events = scanner.push(text); |
| 657 | events.extend(scanner.finish()); |
| 658 | events |
| 659 | } |
| 660 | |
| 661 | fn find_next_marker<'a>(text: &str) -> Option<(usize, &'a str)> { |
| 662 | START_MARKERS |
| 663 | .iter() |
| 664 | .filter_map(|marker| text.find(marker).map(|idx| (idx, *marker))) |
| 665 | .min_by_key(|(idx, _)| *idx) |
| 666 | } |
| 667 | |
| 668 | /// Length of the longest suffix of `buf` that is a prefix of any known opening |
| 669 | /// marker. Lets a split opening tag be caught without delaying text that can't |
| 670 | /// possibly be part of one. |
| 671 | fn max_partial_prefix_len(buf: &str, needles: &[&str]) -> usize { |
| 672 | needles |
| 673 | .iter() |
| 674 | .map(|needle| partial_prefix_len(buf, needle)) |
| 675 | .max() |
| 676 | .unwrap_or(0) |
| 677 | } |
| 678 | |
| 679 | fn partial_prefix_len(buf: &str, needle: &str) -> usize { |
| 680 | let max = buf.len().min(needle.len() - 1); |
| 681 | for len in (1..=max).rev() { |
| 682 | let start = buf.len() - len; |
| 683 | if buf.is_char_boundary(start) && needle.starts_with(&buf[start..]) { |
| 684 | return len; |
| 685 | } |
| 686 | } |
| 687 | 0 |
| 688 | } |
| 689 | |
| 690 | #[cfg(test)] |
| 691 | mod tests { |
| 692 | use super::*; |
| 693 | |
| 694 | fn spec(name: &str, schema: serde_json::Value) -> ToolSpec { |
| 695 | ToolSpec { |
| 696 | name: name.to_string(), |
| 697 | description: String::new(), |
| 698 | parameters_schema: schema.to_string(), |
| 699 | } |
| 700 | } |
| 701 | |
| 702 | fn command_output_spec() -> ToolSpec { |
| 703 | spec( |
| 704 | "command_output", |
| 705 | serde_json::json!({ |
| 706 | "type": "object", |
| 707 | "properties": { "task_id": { "type": "integer" } } |
| 708 | }), |
| 709 | ) |
| 710 | } |
| 711 | |
| 712 | fn run_command_spec() -> ToolSpec { |
| 713 | spec( |
| 714 | "run_command", |
| 715 | serde_json::json!({ |
| 716 | "type": "object", |
| 717 | "properties": { |
| 718 | "command": { "type": "string" }, |
| 719 | "cwd": { "type": "string" }, |
| 720 | "run_in_background": { "type": "boolean" } |
| 721 | } |
| 722 | }), |
| 723 | ) |
| 724 | } |
| 725 | |
| 726 | #[test] |
| 727 | fn an_integer_argument_is_recovered_as_a_json_number() { |
| 728 | let tools = vec![command_output_spec()]; |
| 729 | let Extracted { text, calls, .. } = extract( |
| 730 | "Checking.<tool_call>command_output<arg_key>task_id</arg_key><arg_value>2</arg_value></tool_call>", |
| 731 | &tools, |
| 732 | ); |
| 733 | assert_eq!(text, "Checking."); |
| 734 | assert_eq!(calls.len(), 1); |
| 735 | assert_eq!(calls[0].name, "command_output"); |
| 736 | let args: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 737 | assert_eq!(args["task_id"], 2); |
| 738 | assert!(args["task_id"].is_number()); |
| 739 | } |
| 740 | |
| 741 | #[test] |
| 742 | fn recovers_multiple_string_arguments_in_order() { |
| 743 | let tools = vec![run_command_spec()]; |
| 744 | let Extracted { text, calls, .. } = extract( |
| 745 | "<tool_call>run_command<arg_key>command</arg_key><arg_value>cargo test --locked 2>&1 | tail -20</arg_value><arg_key>cwd</arg_key><arg_value>/Users/x/sigit</arg_value><arg_key>run_in_background</arg_key><arg_value>true</arg_value></tool_call>", |
| 746 | &tools, |
| 747 | ); |
| 748 | assert_eq!(text, ""); |
| 749 | let args: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 750 | assert_eq!(args["command"], "cargo test --locked 2>&1 | tail -20"); |
| 751 | assert_eq!(args["cwd"], "/Users/x/sigit"); |
| 752 | assert_eq!(args["run_in_background"], true); |
| 753 | } |
| 754 | |
| 755 | #[test] |
| 756 | fn recovers_glm_call_with_a_malformed_first_argument_marker() { |
| 757 | let tools = vec![run_command_spec()]; |
| 758 | let Extracted { text, calls, .. } = extract( |
| 759 | "<tool_call>run_command<tool_call>command</arg_key><arg_value>pwd</arg_value><arg_key>cwd</arg_key><arg_value>/tmp</arg_value></tool_call>", |
| 760 | &tools, |
| 761 | ); |
| 762 | |
| 763 | assert_eq!(text, ""); |
| 764 | assert_eq!(calls.len(), 1); |
| 765 | assert_eq!(calls[0].name, "run_command"); |
| 766 | let args: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 767 | assert_eq!(args["command"], "pwd"); |
| 768 | assert_eq!(args["cwd"], "/tmp"); |
| 769 | } |
| 770 | |
| 771 | #[test] |
| 772 | fn recovers_glm_call_with_check_status_before_real_arguments() { |
| 773 | let tools = vec![command_output_spec()]; |
| 774 | let Extracted { text, calls, .. } = extract( |
| 775 | "polling <tool_call>command_output CheckStatus=true_or_poll_again_with_different_params</arg_value><arg_key>task_id</arg_key><arg_value>1</arg_value></tool_call>", |
| 776 | &tools, |
| 777 | ); |
| 778 | |
| 779 | assert_eq!(text, "polling "); |
| 780 | assert_eq!(calls.len(), 1); |
| 781 | assert_eq!(calls[0].name, "command_output"); |
| 782 | let args: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 783 | assert_eq!(args["task_id"], 1); |
| 784 | assert!(args["task_id"].is_number()); |
| 785 | assert!(args.get("CheckStatus").is_none()); |
| 786 | } |
| 787 | |
| 788 | #[test] |
| 789 | fn malformed_check_status_preamble_is_dropped() { |
| 790 | let text = "<tool_call>command_output CheckStatus=keep going</arg_value><arg_key>task_id</arg_key><arg_value>1</arg_value></tool_call>"; |
| 791 | let Extracted { |
| 792 | text: out, |
| 793 | calls, |
| 794 | malformed, |
| 795 | } = extract(text, &[command_output_spec()]); |
| 796 | assert_eq!(out, ""); |
| 797 | assert_eq!(malformed, 1); |
| 798 | assert!(calls.is_empty()); |
| 799 | } |
| 800 | |
| 801 | #[test] |
| 802 | fn check_status_after_a_real_argument_is_dropped() { |
| 803 | let text = "<tool_call>command_output<arg_key>task_id</arg_key><arg_value>1</arg_value> CheckStatus=true_or_poll_again_with_different_params</arg_value></tool_call>"; |
| 804 | let Extracted { |
| 805 | text: out, |
| 806 | calls, |
| 807 | malformed, |
| 808 | } = extract(text, &[command_output_spec()]); |
| 809 | assert_eq!(out, ""); |
| 810 | assert_eq!(malformed, 1); |
| 811 | assert!(calls.is_empty()); |
| 812 | } |
| 813 | |
| 814 | #[test] |
| 815 | fn check_status_for_an_unknown_tool_is_dropped() { |
| 816 | let text = "<tool_call>not_offered CheckStatus=true_or_poll_again_with_different_params</arg_value><arg_key>task_id</arg_key><arg_value>1</arg_value></tool_call>"; |
| 817 | let Extracted { |
| 818 | text: out, |
| 819 | calls, |
| 820 | malformed, |
| 821 | } = extract(text, &[command_output_spec()]); |
| 822 | assert_eq!(out, ""); |
| 823 | assert_eq!(malformed, 1); |
| 824 | assert!(calls.is_empty()); |
| 825 | } |
| 826 | |
| 827 | #[test] |
| 828 | fn unknown_glm_tool_is_dropped() { |
| 829 | let text = "<tool_call>not_offered<tool_call>command</arg_key><arg_value>pwd</arg_value></tool_call>"; |
| 830 | let Extracted { |
| 831 | text: out, |
| 832 | calls, |
| 833 | malformed, |
| 834 | } = extract(text, &[run_command_spec()]); |
| 835 | assert_eq!(out, ""); |
| 836 | assert_eq!(malformed, 1); |
| 837 | assert!(calls.is_empty()); |
| 838 | } |
| 839 | |
| 840 | #[test] |
| 841 | fn malformed_glm_marker_is_accepted_only_for_the_first_argument() { |
| 842 | let text = "<tool_call>run_command<arg_key>command</arg_key><arg_value>pwd</arg_value><tool_call>cwd</arg_key><arg_value>/tmp</arg_value></tool_call>"; |
| 843 | let Extracted { |
| 844 | text: out, |
| 845 | calls, |
| 846 | malformed, |
| 847 | } = extract(text, &[run_command_spec()]); |
| 848 | assert_eq!(out, ""); |
| 849 | assert_eq!(malformed, 1); |
| 850 | assert!(calls.is_empty()); |
| 851 | } |
| 852 | |
| 853 | #[test] |
| 854 | fn a_malformed_block_is_dropped_rather_than_guessed_at() { |
| 855 | let text = "<tool_call>edit_file<arg_value>new_text</arg_key><arg_value>def x; end</arg_value></tool_call>"; |
| 856 | let Extracted { |
| 857 | text: out, |
| 858 | calls, |
| 859 | malformed, |
| 860 | } = extract(text, &[run_command_spec()]); |
| 861 | assert_eq!(out, ""); |
| 862 | assert_eq!(malformed, 1); |
| 863 | assert!(calls.is_empty()); |
| 864 | } |
| 865 | |
| 866 | #[test] |
| 867 | fn an_unterminated_tag_is_dropped() { |
| 868 | let text = "working on it <tool_call>run_command<arg_key>command</arg_key>"; |
| 869 | let Extracted { |
| 870 | text: out, |
| 871 | calls, |
| 872 | malformed, |
| 873 | } = extract(text, &[run_command_spec()]); |
| 874 | assert_eq!(out, "working on it "); |
| 875 | assert_eq!(malformed, 1); |
| 876 | assert!(calls.is_empty()); |
| 877 | } |
| 878 | |
| 879 | #[test] |
| 880 | fn ordinary_text_passes_through_untouched() { |
| 881 | let Extracted { |
| 882 | text: out, calls, .. |
| 883 | } = extract("if x < y then a<b, nothing to recover", &[]); |
| 884 | assert_eq!(out, "if x < y then a<b, nothing to recover"); |
| 885 | assert!(calls.is_empty()); |
| 886 | } |
| 887 | |
| 888 | #[test] |
| 889 | fn kimi_k3_run_command_block_is_recovered_and_removed() { |
| 890 | let tools = vec![run_command_spec()]; |
| 891 | let Extracted { text, calls, .. } = extract( |
| 892 | "Working <|open|>tools<|sep|><|open|>call tool=\"run_command\" index=\"1\"<|sep|><|open|>argument key=\"command\" type=\"string\"<|sep|>sleep 180; echo waited<|close|>argument<|sep|><|open|>argument key=\"cwd\" type=\"string\"<|sep|>/Users/setoelkahfi/Repositories/onde-ed<|close|>argument<|sep|><|close|>call<|sep|><|close|>tools<|sep|>", |
| 893 | &tools, |
| 894 | ); |
| 895 | assert_eq!(text, "Working "); |
| 896 | assert_eq!(calls.len(), 1); |
| 897 | assert_eq!(calls[0].name, "run_command"); |
| 898 | let args: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 899 | assert_eq!(args["command"], "sleep 180; echo waited"); |
| 900 | assert_eq!(args["cwd"], "/Users/setoelkahfi/Repositories/onde-ed"); |
| 901 | } |
| 902 | |
| 903 | #[test] |
| 904 | fn kimi_k3_non_string_arguments_are_json_decoded() { |
| 905 | let tools = vec![command_output_spec()]; |
| 906 | let Extracted { calls, .. } = extract( |
| 907 | "<|open|>tools<|sep|><|open|>call tool=\"command_output\" index=\"1\"<|sep|><|open|>argument key=\"task_id\" type=\"integer\"<|sep|>2<|close|>argument<|sep|><|close|>call<|sep|><|close|>tools<|sep|>", |
| 908 | &tools, |
| 909 | ); |
| 910 | let args: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 911 | assert_eq!(args["task_id"], 2); |
| 912 | assert!(args["task_id"].is_number()); |
| 913 | } |
| 914 | |
| 915 | #[test] |
| 916 | fn kimi_k3_missing_argument_type_uses_the_tool_schema() { |
| 917 | let tools = vec![command_output_spec()]; |
| 918 | let Extracted { calls, .. } = extract( |
| 919 | "<|open|>tools<|sep|><|open|>call tool=\"command_output\" index=\"1\"<|sep|><|open|>argument key=\"task_id\"<|sep|>2<|close|>argument<|sep|><|close|>call<|sep|><|close|>tools<|sep|>", |
| 920 | &tools, |
| 921 | ); |
| 922 | let args: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 923 | assert_eq!(args["task_id"], 2); |
| 924 | assert!(args["task_id"].is_number()); |
| 925 | } |
| 926 | |
| 927 | #[test] |
| 928 | fn kimi_k3_multiple_calls_are_recovered_in_order() { |
| 929 | let tools = vec![command_output_spec()]; |
| 930 | let Extracted { calls, .. } = extract( |
| 931 | "<|open|>tools<|sep|><|open|>call tool=\"command_output\" index=\"1\"<|sep|><|open|>argument key=\"task_id\" type=\"integer\"<|sep|>1<|close|>argument<|sep|><|close|>call<|sep|><|open|>call tool=\"command_output\" index=\"2\"<|sep|><|open|>argument key=\"task_id\" type=\"integer\"<|sep|>2<|close|>argument<|sep|><|close|>call<|sep|><|close|>tools<|sep|>", |
| 932 | &tools, |
| 933 | ); |
| 934 | let first: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 935 | let second: serde_json::Value = serde_json::from_str(&calls[1].arguments).unwrap(); |
| 936 | assert_eq!(calls.len(), 2); |
| 937 | assert_eq!(first["task_id"], 1); |
| 938 | assert_eq!(second["task_id"], 2); |
| 939 | } |
| 940 | |
| 941 | #[test] |
| 942 | fn kimi_k3_response_is_unwrapped_and_think_is_hidden() { |
| 943 | let Extracted { text, calls, .. } = extract( |
| 944 | "a<|open|>think<|sep|>private<|close|>think<|sep|>b<|open|>response<|sep|>visible<|close|>response<|sep|>c", |
| 945 | &[], |
| 946 | ); |
| 947 | assert_eq!(text, "abvisiblec"); |
| 948 | assert!(calls.is_empty()); |
| 949 | } |
| 950 | |
| 951 | #[test] |
| 952 | fn kimi_k3_response_content_is_scanned_for_nested_tool_calls() { |
| 953 | let tools = vec![command_output_spec()]; |
| 954 | let Extracted { text, calls, .. } = extract( |
| 955 | "<|open|>response<|sep|>before <|open|>think<|sep|>private<|close|>think<|sep|><|open|>tools<|sep|><|open|>call tool=\"command_output\" index=\"1\"<|sep|><|open|>argument key=\"task_id\" type=\"integer\"<|sep|>2<|close|>argument<|sep|><|close|>call<|sep|><|close|>tools<|sep|> after<|close|>response<|sep|>", |
| 956 | &tools, |
| 957 | ); |
| 958 | assert_eq!(text, "before after"); |
| 959 | assert_eq!(calls.len(), 1); |
| 960 | assert_eq!(calls[0].name, "command_output"); |
| 961 | } |
| 962 | |
| 963 | #[test] |
| 964 | fn kimi_k3_attribute_entities_are_decoded() { |
| 965 | let attrs = parse_k3_attributes(r#"tool="run"command" key="a&b""#).unwrap(); |
| 966 | assert_eq!(attrs["tool"], "run\"command"); |
| 967 | assert_eq!(attrs["key"], "a&b"); |
| 968 | assert_eq!(decode_k3_attribute("line break"), "line break"); |
| 969 | } |
| 970 | |
| 971 | #[test] |
| 972 | fn unterminated_kimi_k3_tools_block_is_dropped() { |
| 973 | let text = "<|open|>tools<|sep|><|open|>call tool=\"run_command\" index=\"1\"<|sep|><|open|>argument key=\"command\" type=\"string\"<|sep|>echo hi<|close|>argument<|sep|>"; |
| 974 | let Extracted { |
| 975 | text: out, |
| 976 | calls, |
| 977 | malformed, |
| 978 | } = extract(text, &[run_command_spec()]); |
| 979 | assert_eq!(out, ""); |
| 980 | assert_eq!(malformed, 1); |
| 981 | assert!(calls.is_empty()); |
| 982 | } |
| 983 | |
| 984 | #[test] |
| 985 | fn unknown_kimi_k3_tool_block_is_dropped() { |
| 986 | let text = "<|open|>tools<|sep|><|open|>call tool=\"delete_everything\" index=\"1\"<|sep|><|close|>call<|sep|><|close|>tools<|sep|>"; |
| 987 | let Extracted { |
| 988 | text: out, |
| 989 | calls, |
| 990 | malformed, |
| 991 | } = extract(text, &[run_command_spec()]); |
| 992 | assert_eq!(out, ""); |
| 993 | assert_eq!(malformed, 1); |
| 994 | assert!(calls.is_empty()); |
| 995 | } |
| 996 | |
| 997 | #[test] |
| 998 | fn invalid_kimi_k3_non_string_argument_is_dropped() { |
| 999 | let text = "<|open|>tools<|sep|><|open|>call tool=\"command_output\" index=\"1\"<|sep|><|open|>argument key=\"task_id\" type=\"integer\"<|sep|>not-json<|close|>argument<|sep|><|close|>call<|sep|><|close|>tools<|sep|>"; |
| 1000 | let Extracted { |
| 1001 | text: out, |
| 1002 | calls, |
| 1003 | malformed, |
| 1004 | } = extract(text, &[command_output_spec()]); |
| 1005 | assert_eq!(out, ""); |
| 1006 | assert_eq!(malformed, 1); |
| 1007 | assert!(calls.is_empty()); |
| 1008 | } |
| 1009 | |
| 1010 | #[test] |
| 1011 | fn scanner_recovers_a_legacy_tag_split_across_chunk_boundaries() { |
| 1012 | let tools = vec![command_output_spec()]; |
| 1013 | let mut scanner = StreamScanner::new(&tools); |
| 1014 | let mut text = String::new(); |
| 1015 | let mut calls = Vec::new(); |
| 1016 | |
| 1017 | for chunk in [ |
| 1018 | "polling ", |
| 1019 | "<tool_c", |
| 1020 | "all>command_output<arg_", |
| 1021 | "key>task_id</arg_key><arg_value>2</arg_v", |
| 1022 | "alue></tool_call>", |
| 1023 | " done", |
| 1024 | ] { |
| 1025 | for event in scanner.push(chunk) { |
| 1026 | match event { |
| 1027 | ScanEvent::Text(t) => text.push_str(&t), |
| 1028 | ScanEvent::ToolCall(c) => calls.push(c), |
| 1029 | ScanEvent::Malformed(block) => panic!("unexpected malformed block: {block}"), |
| 1030 | } |
| 1031 | } |
| 1032 | } |
| 1033 | if let Some(ScanEvent::Text(rest)) = scanner.finish() { |
| 1034 | text.push_str(&rest); |
| 1035 | } |
| 1036 | |
| 1037 | assert_eq!(text, "polling done", "no tag fragment may reach the UI"); |
| 1038 | assert_eq!(calls.len(), 1); |
| 1039 | assert_eq!(calls[0].name, "command_output"); |
| 1040 | } |
| 1041 | |
| 1042 | #[test] |
| 1043 | fn scanner_recovers_glm_call_with_malformed_marker_split_across_chunks() { |
| 1044 | let tools = vec![run_command_spec()]; |
| 1045 | let mut scanner = StreamScanner::new(&tools); |
| 1046 | let mut text = String::new(); |
| 1047 | let mut calls = Vec::new(); |
| 1048 | |
| 1049 | for chunk in [ |
| 1050 | "before ", |
| 1051 | "<tool_call>run_command<tool_", |
| 1052 | "call>command</arg_key><arg_value>pwd</arg_value></tool_call>", |
| 1053 | " after", |
| 1054 | ] { |
| 1055 | for event in scanner.push(chunk) { |
| 1056 | match event { |
| 1057 | ScanEvent::Text(value) => text.push_str(&value), |
| 1058 | ScanEvent::ToolCall(call) => calls.push(call), |
| 1059 | ScanEvent::Malformed(block) => panic!("unexpected malformed block: {block}"), |
| 1060 | } |
| 1061 | } |
| 1062 | } |
| 1063 | if let Some(ScanEvent::Text(rest)) = scanner.finish() { |
| 1064 | text.push_str(&rest); |
| 1065 | } |
| 1066 | |
| 1067 | assert_eq!(text, "before after"); |
| 1068 | assert_eq!(calls.len(), 1); |
| 1069 | assert_eq!(calls[0].name, "run_command"); |
| 1070 | let args: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 1071 | assert_eq!(args["command"], "pwd"); |
| 1072 | } |
| 1073 | |
| 1074 | #[test] |
| 1075 | fn scanner_recovers_check_status_call_split_across_chunks() { |
| 1076 | let tools = vec![command_output_spec()]; |
| 1077 | let mut scanner = StreamScanner::new(&tools); |
| 1078 | let mut text = String::new(); |
| 1079 | let mut calls = Vec::new(); |
| 1080 | |
| 1081 | for chunk in [ |
| 1082 | "before <tool_call>command_output CheckStatus=true_or_poll_", |
| 1083 | "again_with_different_params</arg_", |
| 1084 | "value><arg_key>task_id</arg_key><arg_value>1</arg_value></tool_", |
| 1085 | "call> after", |
| 1086 | ] { |
| 1087 | for event in scanner.push(chunk) { |
| 1088 | match event { |
| 1089 | ScanEvent::Text(value) => text.push_str(&value), |
| 1090 | ScanEvent::ToolCall(call) => calls.push(call), |
| 1091 | ScanEvent::Malformed(block) => panic!("unexpected malformed block: {block}"), |
| 1092 | } |
| 1093 | } |
| 1094 | } |
| 1095 | if let Some(ScanEvent::Text(rest)) = scanner.finish() { |
| 1096 | text.push_str(&rest); |
| 1097 | } |
| 1098 | |
| 1099 | assert_eq!(text, "before after"); |
| 1100 | assert_eq!(calls.len(), 1); |
| 1101 | assert_eq!(calls[0].name, "command_output"); |
| 1102 | } |
| 1103 | |
| 1104 | #[test] |
| 1105 | fn scanner_recovers_kimi_k3_tools_split_across_chunk_boundaries() { |
| 1106 | let tools = vec![run_command_spec()]; |
| 1107 | let mut scanner = StreamScanner::new(&tools); |
| 1108 | let mut text = String::new(); |
| 1109 | let mut calls = Vec::new(); |
| 1110 | |
| 1111 | for chunk in [ |
| 1112 | "pre ", |
| 1113 | "<|open|>too", |
| 1114 | "ls<|sep|><|open|>call tool=\"run_command\" index=\"1\"<|sep|>", |
| 1115 | "<|open|>argument key=\"command\" type=\"string\"<|sep|>echo hi", |
| 1116 | "<|close|>argument<|sep|><|close|>call<|sep|><|close|>tools<|sep|>", |
| 1117 | " post", |
| 1118 | ] { |
| 1119 | for event in scanner.push(chunk) { |
| 1120 | match event { |
| 1121 | ScanEvent::Text(t) => text.push_str(&t), |
| 1122 | ScanEvent::ToolCall(c) => calls.push(c), |
| 1123 | ScanEvent::Malformed(block) => panic!("unexpected malformed block: {block}"), |
| 1124 | } |
| 1125 | } |
| 1126 | } |
| 1127 | if let Some(ScanEvent::Text(rest)) = scanner.finish() { |
| 1128 | text.push_str(&rest); |
| 1129 | } |
| 1130 | |
| 1131 | assert_eq!(text, "pre post"); |
| 1132 | assert_eq!(calls.len(), 1); |
| 1133 | assert_eq!(calls[0].name, "run_command"); |
| 1134 | } |
| 1135 | |
| 1136 | #[test] |
| 1137 | fn scanner_recovers_tools_after_an_unclosed_k3_wrapper() { |
| 1138 | let tools = vec![command_output_spec()]; |
| 1139 | let mut scanner = StreamScanner::new(&tools); |
| 1140 | let mut text = String::new(); |
| 1141 | let mut calls = Vec::new(); |
| 1142 | |
| 1143 | for chunk in [ |
| 1144 | "before <|open|>response<|sep|>broken ", |
| 1145 | "<|open|>tools<|sep|><|open|>call tool=\"command_output\" index=\"1\"<|sep|><|open|>argument key=\"task_id\" type=\"integer\"<|sep|>2<|close|>argument<|sep|><|close|>call<|sep|><|close|>tools<|sep|> after", |
| 1146 | ] { |
| 1147 | for event in scanner.push(chunk) { |
| 1148 | match event { |
| 1149 | ScanEvent::Text(chunk) => text.push_str(&chunk), |
| 1150 | ScanEvent::ToolCall(call) => calls.push(call), |
| 1151 | ScanEvent::Malformed(block) => panic!("unexpected malformed block: {block}"), |
| 1152 | } |
| 1153 | } |
| 1154 | } |
| 1155 | if let Some(ScanEvent::Text(rest)) = scanner.finish() { |
| 1156 | text.push_str(&rest); |
| 1157 | } |
| 1158 | |
| 1159 | assert_eq!(text, "before <|open|>response<|sep|>broken after"); |
| 1160 | assert_eq!(calls.len(), 1); |
| 1161 | assert_eq!(calls[0].name, "command_output"); |
| 1162 | } |
| 1163 | |
| 1164 | #[test] |
| 1165 | fn scanner_recovers_tools_after_an_unclosed_k3_think_wrapper() { |
| 1166 | let tools = vec![command_output_spec()]; |
| 1167 | let Extracted { text, calls, .. } = extract( |
| 1168 | "<|open|>think<|sep|>broken <|open|>tools<|sep|><|open|>call tool=\"command_output\" index=\"1\"<|sep|><|open|>argument key=\"task_id\" type=\"integer\"<|sep|>2<|close|>argument<|sep|><|close|>call<|sep|><|close|>tools<|sep|>", |
| 1169 | &tools, |
| 1170 | ); |
| 1171 | assert_eq!(text, "<|open|>think<|sep|>broken "); |
| 1172 | assert_eq!(calls.len(), 1); |
| 1173 | assert_eq!(calls[0].name, "command_output"); |
| 1174 | } |
| 1175 | |
| 1176 | #[test] |
| 1177 | fn scanner_forwards_ordinary_text_without_holding_it_back() { |
| 1178 | let mut scanner = StreamScanner::new(&[]); |
| 1179 | let events = scanner.push("a normal streamed answer"); |
| 1180 | assert_eq!( |
| 1181 | events, |
| 1182 | vec![ScanEvent::Text("a normal streamed answer".to_string())] |
| 1183 | ); |
| 1184 | assert!(scanner.finish().is_none()); |
| 1185 | } |
| 1186 | |
| 1187 | #[test] |
| 1188 | fn scanner_releases_a_near_miss_prefix_once_it_cannot_match() { |
| 1189 | let mut scanner = StreamScanner::new(&[]); |
| 1190 | assert_eq!( |
| 1191 | scanner.push("see <tool"), |
| 1192 | vec![ScanEvent::Text("see ".to_string())] |
| 1193 | ); |
| 1194 | assert_eq!( |
| 1195 | scanner.push("box for details"), |
| 1196 | vec![ScanEvent::Text("<toolbox for details".to_string())] |
| 1197 | ); |
| 1198 | assert!(scanner.finish().is_none()); |
| 1199 | } |
| 1200 | |
| 1201 | #[test] |
| 1202 | fn scanner_reports_an_unclosed_tag_at_stream_end_as_malformed() { |
| 1203 | let tools = vec![run_command_spec()]; |
| 1204 | let mut scanner = StreamScanner::new(&tools); |
| 1205 | let events = scanner.push("partial <tool_call>run_command<arg_key>cmd"); |
| 1206 | assert_eq!(events, vec![ScanEvent::Text("partial ".to_string())]); |
| 1207 | assert_eq!( |
| 1208 | scanner.finish(), |
| 1209 | Some(ScanEvent::Malformed( |
| 1210 | "<tool_call>run_command<arg_key>cmd".to_string() |
| 1211 | )) |
| 1212 | ); |
| 1213 | } |
| 1214 | |
| 1215 | #[test] |
| 1216 | fn scanner_hands_back_a_partial_marker_prefix_at_stream_end_as_text() { |
| 1217 | let mut scanner = StreamScanner::new(&[]); |
| 1218 | assert_eq!( |
| 1219 | scanner.push("see <tool_"), |
| 1220 | vec![ScanEvent::Text("see ".to_string())] |
| 1221 | ); |
| 1222 | assert_eq!( |
| 1223 | scanner.finish(), |
| 1224 | Some(ScanEvent::Text("<tool_".to_string())) |
| 1225 | ); |
| 1226 | |
| 1227 | let mut scanner = StreamScanner::new(&[]); |
| 1228 | assert_eq!( |
| 1229 | scanner.push("see <|open|>to"), |
| 1230 | vec![ScanEvent::Text("see ".to_string())] |
| 1231 | ); |
| 1232 | assert_eq!( |
| 1233 | scanner.finish(), |
| 1234 | Some(ScanEvent::Text("<|open|>to".to_string())) |
| 1235 | ); |
| 1236 | } |
| 1237 | |
| 1238 | #[test] |
| 1239 | fn a_tag_mentioned_in_prose_stays_text() { |
| 1240 | let tools = vec![run_command_spec()]; |
| 1241 | for text in [ |
| 1242 | "Wrap it in `<tool_call>` and `</tool_call>` tags.", |
| 1243 | "The scanner holds <tool_call> and </tool_call> blocks back.", |
| 1244 | "The model opens with `<tool_call>` and never closes it.", |
| 1245 | ] { |
| 1246 | assert_eq!( |
| 1247 | extract(text, &tools), |
| 1248 | Extracted { |
| 1249 | text: text.to_string(), |
| 1250 | ..Extracted::default() |
| 1251 | }, |
| 1252 | "{text}" |
| 1253 | ); |
| 1254 | } |
| 1255 | } |
| 1256 | |
| 1257 | #[test] |
| 1258 | fn a_kimi_marker_mentioned_in_prose_stays_text() { |
| 1259 | let tools = vec![run_command_spec()]; |
| 1260 | for text in [ |
| 1261 | "The marker is `<|open|>tools<|sep|>` and it never closes here.", |
| 1262 | "Quote `<|open|>tools<|sep|>` then `<|close|>tools<|sep|>` in docs.", |
| 1263 | ] { |
| 1264 | assert_eq!( |
| 1265 | extract(text, &tools), |
| 1266 | Extracted { |
| 1267 | text: text.to_string(), |
| 1268 | ..Extracted::default() |
| 1269 | }, |
| 1270 | "{text}" |
| 1271 | ); |
| 1272 | } |
| 1273 | } |
| 1274 | |
| 1275 | #[test] |
| 1276 | fn a_json_shaped_block_is_dropped() { |
| 1277 | let text = "<tool_call>\n{\"name\": \"run_command\", \"arguments\": {}}\n</tool_call>"; |
| 1278 | let Extracted { |
| 1279 | text: out, |
| 1280 | calls, |
| 1281 | malformed, |
| 1282 | } = extract(text, &[run_command_spec()]); |
| 1283 | assert_eq!(out, ""); |
| 1284 | assert_eq!(malformed, 1); |
| 1285 | assert!(calls.is_empty()); |
| 1286 | } |
| 1287 | |
| 1288 | /// Issue #105: a GLM-family reply that opened a call, then wrote an |
| 1289 | /// invented invoke/result transcript instead of closing it. None of that |
| 1290 | /// may reach the editor, and it must not count as a recovered call. |
| 1291 | #[test] |
| 1292 | fn a_hallucinated_invoke_transcript_is_dropped() { |
| 1293 | let tools = vec![command_output_spec()]; |
| 1294 | let reply = "Checking.<tool_call>command_output\n\n<invokeID>reply_A</invokeID>\n\ |
| 1295 | <parameter_name>task_id</parameter_name>\n<parameter>6</parameter>\n\ |
| 1296 | </invoke>\n\n<function_results>PyPI and NuGet are queued.\n\ |
| 1297 | <tool_call>command_output<arg_key>task_id</arg_key><arg_value>5</task_id>\n\ |
| 1298 | </invoke>\n</function_results>"; |
| 1299 | let Extracted { |
| 1300 | text, |
| 1301 | calls, |
| 1302 | malformed, |
| 1303 | } = extract(reply, &tools); |
| 1304 | assert_eq!(text, "Checking."); |
| 1305 | assert!(calls.is_empty()); |
| 1306 | assert_eq!(malformed, 1); |
| 1307 | } |
| 1308 | |
| 1309 | /// Issue #97: the well-formed blocks from that thread are recovered as |
| 1310 | /// real calls, multi-line argument values included. |
| 1311 | #[test] |
| 1312 | fn well_formed_blocks_from_issue_97_are_recovered() { |
| 1313 | let tools = vec![ |
| 1314 | run_command_spec(), |
| 1315 | spec( |
| 1316 | "edit_file", |
| 1317 | serde_json::json!({ |
| 1318 | "type": "object", |
| 1319 | "properties": { |
| 1320 | "path": { "type": "string" }, |
| 1321 | "old_text": { "type": "string" }, |
| 1322 | "new_text": { "type": "string" } |
| 1323 | } |
| 1324 | }), |
| 1325 | ), |
| 1326 | ]; |
| 1327 | let reply = "Now wire `body_path` into the step:<tool_call>edit_file<arg_key>new_text</arg_key><arg_value> - name: Release\n with:\n body_path: release_notes.md</arg_value><arg_key>old_text</arg_key><arg_value> - name: Release\n with:</arg_value><arg_key>path</arg_key><arg_value>/repo/.github/workflows/release-github.yml</arg_value></tool_call>"; |
| 1328 | let Extracted { |
| 1329 | text, |
| 1330 | calls, |
| 1331 | malformed, |
| 1332 | } = extract(reply, &tools); |
| 1333 | assert_eq!(text, "Now wire `body_path` into the step:"); |
| 1334 | assert_eq!(malformed, 0); |
| 1335 | assert_eq!(calls.len(), 1); |
| 1336 | let args: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 1337 | assert_eq!( |
| 1338 | args["new_text"], |
| 1339 | " - name: Release\n with:\n body_path: release_notes.md" |
| 1340 | ); |
| 1341 | assert_eq!(args["path"], "/repo/.github/workflows/release-github.yml"); |
| 1342 | |
| 1343 | let Extracted { calls, .. } = extract( |
| 1344 | "<tool_call>run_command<arg_key>command</arg_key><arg_value>sleep 90</arg_value><arg_key>cwd</arg_key><arg_value>/repo</arg_value></tool_call>", |
| 1345 | &tools, |
| 1346 | ); |
| 1347 | assert_eq!(calls.len(), 1); |
| 1348 | assert_eq!(calls[0].name, "run_command"); |
| 1349 | } |
| 1350 | |
| 1351 | #[test] |
| 1352 | fn dsml_multiple_invokes_are_recovered_in_order() { |
| 1353 | let tools = vec![command_output_spec()]; |
| 1354 | let Extracted { text, calls, .. } = extract( |
| 1355 | "Working <｜DSML｜tool_calls><｜DSML｜invoke name=\"command_output\"><｜DSML｜parameter name=\"task_id\" string=\"false\">1<｜DSML｜/parameter><｜DSML｜/invoke><｜DSML｜invoke name=\"command_output\"><｜DSML｜parameter name=\"task_id\" string=\"false\">2<｜DSML｜/parameter><｜DSML｜/invoke><｜DSML｜/tool_calls>", |
| 1356 | &tools, |
| 1357 | ); |
| 1358 | assert_eq!(text, "Working "); |
| 1359 | assert_eq!(calls.len(), 2); |
| 1360 | let first: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 1361 | let second: serde_json::Value = serde_json::from_str(&calls[1].arguments).unwrap(); |
| 1362 | assert_eq!(first["task_id"], 1); |
| 1363 | assert_eq!(second["task_id"], 2); |
| 1364 | } |
| 1365 | |
| 1366 | #[test] |
| 1367 | fn dsml_string_true_is_kept_verbatim_and_string_false_is_json_typed() { |
| 1368 | let tools = vec![run_command_spec()]; |
| 1369 | let Extracted { calls, .. } = extract( |
| 1370 | "<｜DSML｜tool_calls><｜DSML｜invoke name=\"run_command\"><｜DSML｜parameter name=\"command\" string=\"true\">echo {\"a\":1}<｜DSML｜/parameter><｜DSML｜parameter name=\"run_in_background\" string=\"false\">true<｜DSML｜/parameter><｜DSML｜/invoke><｜DSML｜/tool_calls>", |
| 1371 | &tools, |
| 1372 | ); |
| 1373 | let args: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 1374 | assert_eq!(args["command"], "echo {\"a\":1}"); |
| 1375 | assert_eq!(args["run_in_background"], true); |
| 1376 | } |
| 1377 | |
| 1378 | #[test] |
| 1379 | fn dsml_missing_string_attribute_uses_the_tool_schema() { |
| 1380 | let tools = vec![command_output_spec()]; |
| 1381 | let Extracted { calls, .. } = extract( |
| 1382 | "<｜DSML｜tool_calls><｜DSML｜invoke name=\"command_output\"><｜DSML｜parameter name=\"task_id\">2<｜DSML｜/parameter><｜DSML｜/invoke><｜DSML｜/tool_calls>", |
| 1383 | &tools, |
| 1384 | ); |
| 1385 | let args: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 1386 | assert_eq!(args["task_id"], 2); |
| 1387 | assert!(args["task_id"].is_number()); |
| 1388 | } |
| 1389 | |
| 1390 | #[test] |
| 1391 | fn unterminated_dsml_tool_calls_block_is_dropped() { |
| 1392 | let text = "<｜DSML｜tool_calls><｜DSML｜invoke name=\"run_command\"><｜DSML｜parameter name=\"command\" string=\"true\">echo hi<｜DSML｜/parameter>"; |
| 1393 | let Extracted { |
| 1394 | text: out, |
| 1395 | calls, |
| 1396 | malformed, |
| 1397 | } = extract(text, &[run_command_spec()]); |
| 1398 | assert_eq!(out, ""); |
| 1399 | assert_eq!(malformed, 1); |
| 1400 | assert!(calls.is_empty()); |
| 1401 | } |
| 1402 | |
| 1403 | #[test] |
| 1404 | fn unknown_dsml_tool_is_dropped() { |
| 1405 | let text = "<｜DSML｜tool_calls><｜DSML｜invoke name=\"delete_everything\"><｜DSML｜/invoke><｜DSML｜/tool_calls>"; |
| 1406 | let Extracted { |
| 1407 | text: out, |
| 1408 | calls, |
| 1409 | malformed, |
| 1410 | } = extract(text, &[run_command_spec()]); |
| 1411 | assert_eq!(out, ""); |
| 1412 | assert_eq!(malformed, 1); |
| 1413 | assert!(calls.is_empty()); |
| 1414 | } |
| 1415 | |
| 1416 | #[test] |
| 1417 | fn invalid_dsml_json_typed_parameter_is_dropped() { |
| 1418 | let text = "<｜DSML｜tool_calls><｜DSML｜invoke name=\"command_output\"><｜DSML｜parameter name=\"task_id\" string=\"false\">not-json<｜DSML｜/parameter><｜DSML｜/invoke><｜DSML｜/tool_calls>"; |
| 1419 | let Extracted { |
| 1420 | text: out, |
| 1421 | calls, |
| 1422 | malformed, |
| 1423 | } = extract(text, &[command_output_spec()]); |
| 1424 | assert_eq!(out, ""); |
| 1425 | assert_eq!(malformed, 1); |
| 1426 | assert!(calls.is_empty()); |
| 1427 | } |
| 1428 | |
| 1429 | #[test] |
| 1430 | fn a_dsml_marker_mentioned_in_prose_stays_text() { |
| 1431 | let tools = vec![run_command_spec()]; |
| 1432 | for text in [ |
| 1433 | "The marker is `<｜DSML｜tool_calls>` and it never closes here.", |
| 1434 | "Quote `<｜DSML｜tool_calls>` then `<｜DSML｜/tool_calls>` in docs.", |
| 1435 | ] { |
| 1436 | assert_eq!( |
| 1437 | extract(text, &tools), |
| 1438 | Extracted { |
| 1439 | text: text.to_string(), |
| 1440 | ..Extracted::default() |
| 1441 | }, |
| 1442 | "{text}" |
| 1443 | ); |
| 1444 | } |
| 1445 | } |
| 1446 | |
| 1447 | #[test] |
| 1448 | fn scanner_recovers_dsml_tool_calls_split_across_chunk_boundaries_including_inside_the_marker() |
| 1449 | { |
| 1450 | let tools = vec![run_command_spec()]; |
| 1451 | let mut scanner = StreamScanner::new(&tools); |
| 1452 | let mut text = String::new(); |
| 1453 | let mut calls = Vec::new(); |
| 1454 | |
| 1455 | for chunk in [ |
| 1456 | "pre ", |
| 1457 | "<｜DSML", |
| 1458 | "｜tool_calls><｜DSML｜invoke name=\"run_command\"><｜DSML｜paramet", |
| 1459 | "er name=\"command\" string=\"true\">echo hi<｜DSML｜/parameter><｜DSML｜/invoke><｜DSML｜/tool_calls>", |
| 1460 | " post", |
| 1461 | ] { |
| 1462 | for event in scanner.push(chunk) { |
| 1463 | match event { |
| 1464 | ScanEvent::Text(t) => text.push_str(&t), |
| 1465 | ScanEvent::ToolCall(c) => calls.push(c), |
| 1466 | ScanEvent::Malformed(block) => panic!("unexpected malformed block: {block}"), |
| 1467 | } |
| 1468 | } |
| 1469 | } |
| 1470 | if let Some(ScanEvent::Text(rest)) = scanner.finish() { |
| 1471 | text.push_str(&rest); |
| 1472 | } |
| 1473 | |
| 1474 | assert_eq!(text, "pre post"); |
| 1475 | assert_eq!(calls.len(), 1); |
| 1476 | assert_eq!(calls[0].name, "run_command"); |
| 1477 | let args: serde_json::Value = serde_json::from_str(&calls[0].arguments).unwrap(); |
| 1478 | assert_eq!(args["command"], "echo hi"); |
| 1479 | } |
| 1480 | |
| 1481 | #[test] |
| 1482 | fn scanner_reports_an_unclosed_dsml_tag_at_stream_end_as_malformed() { |
| 1483 | let tools = vec![run_command_spec()]; |
| 1484 | let mut scanner = StreamScanner::new(&tools); |
| 1485 | let events = |
| 1486 | scanner.push("partial <｜DSML｜tool_calls><｜DSML｜invoke name=\"run_command\">"); |
| 1487 | assert_eq!(events, vec![ScanEvent::Text("partial ".to_string())]); |
| 1488 | assert_eq!( |
| 1489 | scanner.finish(), |
| 1490 | Some(ScanEvent::Malformed( |
| 1491 | "<｜DSML｜tool_calls><｜DSML｜invoke name=\"run_command\">".to_string() |
| 1492 | )) |
| 1493 | ); |
| 1494 | } |
| 1495 | } |