File size: 10,618 Bytes
48eb149 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 | """Terminal chat: live reasoning stream plus SmolTalk tool-call rounds.
Tercet-R assistant turns may contain:
- a zero-loss `<|think|>` / `<|no_think|>` control prefix
- a `<think>…</think>` reasoning block
- one or more SmolTalk JSON `<tool_call>` blocks (NVIDIA XML is also parsed)
This module colours those regions as tokens arrive and, after a completed
turn, collects tool observations (prefixed with `<|tool_response|>`) so the
model can continue the same conversation.
"""
from __future__ import annotations
import json
import sys
from collections.abc import Callable, Sequence
from dataclasses import dataclass, field
from typing import Any, Literal, TextIO
from tiny_gdn.smoltalk_chat import (
SMOLTALK_WEB_SEARCH_TOOL,
TOOL_RESPONSE_TOKEN,
format_smoltalk_tool_call,
wrap_smoltalk_tool_result,
)
from tiny_gdn.tools import ParsedToolCall, parse_tool_calls
SegmentKind = Literal["answer", "think", "tool_call"]
MarkerKind = Literal[
"think_open",
"think_close",
"think_control",
"no_think",
"tool_open",
"tool_close",
]
THINK_OPEN = "<think>"
THINK_CLOSE = "</think>"
TOOL_CALL_OPEN = "<tool_call>"
TOOL_CALL_CLOSE = "</tool_call>"
THINK_CONTROL = "<|think|>"
NO_THINK_CONTROL = "<|no_think|>"
MARKERS: tuple[tuple[str, MarkerKind], ...] = (
(THINK_CLOSE, "think_close"),
(TOOL_CALL_CLOSE, "tool_close"),
(THINK_OPEN, "think_open"),
(TOOL_CALL_OPEN, "tool_open"),
(NO_THINK_CONTROL, "no_think"),
(THINK_CONTROL, "think_control"),
)
ANSI = {
"think": "\033[2;33m",
"tool_call": "\033[36m",
"answer": "\033[0m",
"reset": "\033[0m",
}
DEFAULT_MAX_TOOL_ROUNDS = 8
@dataclass(frozen=True)
class ContentSegment:
kind: SegmentKind
text: str
open: bool = False
@dataclass(frozen=True)
class GeneratedTurn:
text: str
token_count: int
stop_reason: str
tool_calls: tuple[ParsedToolCall, ...]
def first_marker(text: str) -> tuple[int, str, MarkerKind] | None:
best: tuple[int, str, MarkerKind] | None = None
for marker, kind in MARKERS:
at = text.find(marker)
if at < 0:
continue
if (
best is None
or at < best[0]
or (at == best[0] and len(marker) > len(best[1]))
):
best = (at, marker, kind)
return best
def holdback_prefix_length(text: str) -> int:
if not text:
return 0
keep = 0
for marker, _kind in MARKERS:
limit = min(len(marker) - 1, len(text))
for size in range(1, limit + 1):
if marker.startswith(text[-size:]):
keep = max(keep, size)
return keep
def mode_after_marker(kind: MarkerKind, current: SegmentKind) -> SegmentKind:
if kind in {"think_open", "think_control"}:
return "think"
if kind in {"think_close", "no_think", "tool_close"}:
return "answer"
if kind == "tool_open":
return "tool_call"
return current
def split_assistant_segments(source: str) -> list[ContentSegment]:
"""Split a completed (or in-progress) assistant turn for tests / replay."""
segments: list[ContentSegment] = []
mode: SegmentKind = "answer"
cursor = 0
while cursor < len(source):
found = first_marker(source[cursor:])
if found is None:
tail = source[cursor:]
if tail:
segments.append(ContentSegment(kind=mode, text=tail, open=True))
break
at, marker, kind = found
at += cursor
if at > cursor:
segments.append(
ContentSegment(kind=mode, text=source[cursor:at], open=False)
)
mode = mode_after_marker(kind, mode)
cursor = at + len(marker)
return [segment for segment in segments if segment.text]
class LiveReasoningStreamer:
"""Colour reasoning and tool-call regions as decoded text grows."""
def __init__(
self,
writer: TextIO | None = None,
*,
color: bool | None = None,
) -> None:
self.writer = writer if writer is not None else sys.stdout
if color is None:
color = bool(getattr(self.writer, "isatty", lambda: False)())
self.color = color
self._seen = ""
self._hold = ""
self._mode: SegmentKind = "answer"
self._style: SegmentKind | None = None
self._emitted_think_label = False
self._emitted_tool_label = False
def update(self, decoded: str) -> None:
if decoded.startswith(self._seen):
delta = decoded[len(self._seen) :]
else:
delta = decoded
self._seen = decoded
if delta:
self._consume(delta, final=False)
def finish(self) -> str:
if self._hold:
self._emit(self._hold)
self._hold = ""
self._set_style(None)
self.writer.write("\n")
self.writer.flush()
return self._seen
def _consume(self, delta: str, *, final: bool) -> None:
buffer = self._hold + delta
self._hold = ""
while buffer:
found = first_marker(buffer)
if found is None:
keep = 0 if final else holdback_prefix_length(buffer)
if keep:
self._emit(buffer[:-keep])
self._hold = buffer[-keep:]
else:
self._emit(buffer)
return
at, marker, kind = found
if at:
self._emit(buffer[:at])
self._switch(kind)
buffer = buffer[at + len(marker) :]
if final:
return
def _switch(self, kind: MarkerKind) -> None:
nxt = mode_after_marker(kind, self._mode)
if nxt != self._mode and nxt == "answer":
self._emit_plain("\n")
self._mode = nxt
if nxt == "think" and not self._emitted_think_label:
self._emit_plain("\n")
self._set_style("think")
self._emit_plain("reasoning ")
self._emitted_think_label = True
elif nxt == "tool_call" and not self._emitted_tool_label:
self._emit_plain("\n")
self._set_style("tool_call")
self._emit_plain("tool_call ")
self._emitted_tool_label = True
elif nxt == "answer":
self._set_style("answer")
def _emit(self, text: str) -> None:
if not text:
return
self._set_style(self._mode)
self.writer.write(text)
self.writer.flush()
def _emit_plain(self, text: str) -> None:
if not text:
return
self._set_style(None)
self.writer.write(text)
self.writer.flush()
def _set_style(self, kind: SegmentKind | None) -> None:
if not self.color:
self._style = kind
return
if kind == self._style:
return
self.writer.write(ANSI["reset"])
if kind in {"think", "tool_call"}:
self.writer.write(ANSI[kind])
self._style = kind
def prompt_tool_results(
calls: Sequence[ParsedToolCall],
*,
read_line: Callable[[str], str],
writer: TextIO | None = None,
) -> list[str]:
out = writer if writer is not None else sys.stdout
results: list[str] = []
out.write(
f"\n{len(calls)} tool call(s). Paste each observation; "
f"it is sent as {TOOL_RESPONSE_TOKEN}.\n"
)
out.flush()
for index, call in enumerate(calls, start=1):
out.write(f"\n[{index}/{len(calls)}] {format_smoltalk_tool_call(call.name, call.arguments)}\n")
out.flush()
raw = read_line(f"result[{call.name}]> ")
results.append(raw)
return results
def append_tool_round(
messages: list[dict[str, Any]],
assistant_text: str,
raw_results: Sequence[str],
) -> None:
messages.append({"role": "assistant", "content": assistant_text})
for raw in raw_results:
wrapped = wrap_smoltalk_tool_result(raw)
if not wrapped:
raise ValueError("Tool result cannot be empty")
messages.append({"role": "tool", "content": raw})
def resolve_cli_tools(spec: str | None, tools_json: str | None) -> list[dict[str, Any]] | None:
tools: list[dict[str, Any]] = []
if spec:
for name in spec.split(","):
key = name.strip().lower()
if not key or key in {"none", "off"}:
continue
if key == "web_search":
tools.append(SMOLTALK_WEB_SEARCH_TOOL)
continue
raise ValueError(
f"Unknown built-in tool {name!r}. Use web_search or --tools-json."
)
if tools_json:
payload = json.loads(tools_json)
if isinstance(payload, dict):
tools.append(payload)
elif isinstance(payload, list):
tools.extend(payload)
else:
raise ValueError("tools JSON must be an object or array")
return tools or None
@dataclass
class ChatLoopState:
messages: list[dict[str, Any]] = field(default_factory=list)
system: str = ""
enable_thinking: bool = True
tools: list[dict[str, Any]] | None = None
def reset(self) -> None:
self.messages = []
if self.system.strip():
self.messages.append({"role": "system", "content": self.system.strip()})
def add_user(self, text: str) -> None:
self.messages.append({"role": "user", "content": text})
def apply_slash_command(state: ChatLoopState, text: str) -> str | None:
"""Return a status string if `text` is a slash command, else None."""
command = text.strip()
lowered = command.lower()
if lowered in {"/exit", "/quit"}:
return "exit"
if lowered == "/reset":
state.reset()
return "history cleared"
if lowered == "/think":
if state.messages:
return "thinking can only be changed on a fresh conversation (/reset first)"
state.enable_thinking = True
return "thinking on — next assistant turn is prefixed with <|think|>"
if lowered in {"/no_think", "/nothink"}:
if state.messages:
return "thinking can only be changed on a fresh conversation (/reset first)"
state.enable_thinking = False
return "thinking off — next assistant turn is prefixed with <|no_think|>"
if lowered.startswith("/system"):
rest = command[len("/system") :].strip()
state.system = rest
state.reset()
return "system prompt updated" if rest else "system prompt cleared"
return None
|