@samitouri / QOSami_Agents_Zero / commits / 8a9de4a2

vision prototype, terminal sessions

Vision functionality, vision_load tool History.py re-work to support and optimize attachments token usage Code execution tool support for multiple sessions in parallel

frdel committed Apr 1, 2025 at 13:55 UTC 8a9de4a2ee334fc34d6a71673273c7272308480b

20 files changed +493 -210

agent.py

+42 -33

@@ -10,7 +10,12 @@ import models

10 from langchain_core.prompt_values import ChatPromptValue

11 from python.helpers import extract_tools, rate_limiter, files, errors, history, tokens

12 from python.helpers.print_style import PrintStyle

13 -from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder, HumanMessagePromptTemplate, StringPromptTemplate

13 +from langchain_core.prompts import (

14 + ChatPromptTemplate,

15 + MessagesPlaceholder,

16 + HumanMessagePromptTemplate,

17 + StringPromptTemplate,

18 +)

19 from langchain_core.prompts.image import ImagePromptTemplate

20 from langchain_core.messages import HumanMessage, SystemMessage, AIMessage, BaseMessage

21 from langchain_core.language_models.chat_models import BaseChatModel

@@ -91,7 +96,7 @@ class AgentContext:

96 else:

97 current_agent = self.agent0

94 - self.task =self.run_task(current_agent.monologue)

99 + self.task = self.run_task(current_agent.monologue)

100 return self.task

101

102 def communicate(self, msg: "UserMessage", broadcast_level: int = 1):

@@ -130,9 +135,9 @@ class AgentContext:

135 async def _process_chain(self, agent: "Agent", msg: "UserMessage|str", user=True):

136 try:

137 msg_template = (

133 - await agent.hist_add_user_message(msg) # type: ignore

138 + agent.hist_add_user_message(msg) # type: ignore

139 if user

135 - else await agent.hist_add_tool_result(

140 + else agent.hist_add_tool_result(

141 tool_name="call_subordinate", tool_result=msg # type: ignore

142 )

143 )

@@ -283,9 +288,6 @@ class Agent:

288 printer.stream(chunk)

289 self.log_from_stream(full, log)

290

286 - # store as last context window content

287 - self.set_data(Agent.DATA_NAME_CTX_WINDOW, prompt.format())

288 -

291 agent_response = await self.call_chat_model(

292 prompt, callback=stream_callback

293 )

@@ -296,10 +298,10 @@ class Agent:

298 self.loop_data.last_response == agent_response

299 ): # if assistant_response is the same as last message in history, let him know

300 # Append the assistant's response to the history

299 - await self.hist_add_ai_response(agent_response)

301 + self.hist_add_ai_response(agent_response)

302 # Append warning message to the history

303 warning_msg = self.read_prompt("fw.msg_repeat.md")

302 - await self.hist_add_warning(message=warning_msg)

304 + self.hist_add_warning(message=warning_msg)

305 PrintStyle(font_color="orange", padding=True).print(

306 warning_msg

307 )

@@ -307,7 +309,7 @@ class Agent:

309

310 else: # otherwise proceed with tool

311 # Append the assistant's response to the history

310 - await self.hist_add_ai_response(agent_response)

312 + self.hist_add_ai_response(agent_response)

313 # process tools requested in agent message

314 tools_result = await self.process_tools(agent_response)

315 if tools_result: # final response of message loop available

@@ -319,7 +321,7 @@ class Agent:

321 except RepairableException as e:

322 # Forward repairable errors to the LLM, maybe it can fix them

323 error_message = errors.format_error(e)

322 - await self.hist_add_warning(error_message)

324 + self.hist_add_warning(error_message)

325 PrintStyle(font_color="red", padding=True).print(error_message)

326 self.context.log.log(type="error", content=error_message)

327 except Exception as e:

@@ -358,23 +360,31 @@ class Agent:

360 extras += history.Message(False, content=extra).output()

361 loop_data.extras_temporary.clear()

362

361 - # combine history and extras

362 - history_combined: list[OutputMessage] = history.group_outputs_abab(loop_data.history_output + extras)

363 -

364 - # convert history to LLM format

365 - history_langchain: list[BaseMessage] = history.output_langchain(history_combined)

366 -

367 - PrintStyle(font_color="grey", background_color="black", bold=True, padding=True).print(

368 - f"History Langchain: {history_langchain}"

363 + # convert history + extras to LLM format

364 + history_langchain: list[BaseMessage] = history.output_langchain(

365 + loop_data.history_output + extras

366 )

367

368 # build chain from system prompt, message history and model

369 + system_text = "\n\n".join(loop_data.system)

370 prompt = ChatPromptTemplate.from_messages(

371 [

374 - SystemMessage(content="\n\n".join(loop_data.system)),

372 + SystemMessage(content=system_text),

373 *history_langchain,

374 ]

375 )

376 +

377 + # store as last context window content

378 + self.set_data(

379 + Agent.DATA_NAME_CTX_WINDOW,

380 + {

381 + "text": prompt.format(),

382 + "tokens": self.history.get_tokens()

383 + + tokens.approximate_tokens(system_text)

384 + + tokens.approximate_tokens(history.output_text(extras)),

385 + },

386 + )

387 +

388 return prompt

389

390 def handle_critical_exception(self, exception: Exception):

@@ -441,12 +451,12 @@ class Agent:

451 def set_data(self, field: str, value):

452 self.data[field] = value

453

444 - def hist_add_message(self, ai: bool, content: history.MessageContent):

445 - return self.history.add_message(ai=ai, content=content)

446 -

447 - async def hist_add_user_message(

448 - self, message: UserMessage, intervention: bool = False

454 + def hist_add_message(

455 + self, ai: bool, content: history.MessageContent, tokens: int = 0

456 ):

457 + return self.history.add_message(ai=ai, content=content, tokens=tokens)

458 +

459 + def hist_add_user_message(self, message: UserMessage, intervention: bool = False):

460 self.history.new_topic() # user message starts a new topic in history

461

462 # load message template based on intervention

@@ -476,19 +486,18 @@ class Agent:

486 self.last_user_message = msg

487 return msg

488

479 - async def hist_add_ai_response(self, message: str):

489 + def hist_add_ai_response(self, message: str):

490 self.loop_data.last_response = message

491 content = self.parse_prompt("fw.ai_response.md", message=message)

492 return self.hist_add_message(True, content=content)

493

484 - async def hist_add_warning(self, message: history.MessageContent):

494 + def hist_add_warning(self, message: history.MessageContent):

495 content = self.parse_prompt("fw.warning.md", message=message)

496 return self.hist_add_message(False, content=content)

497

488 - async def hist_add_tool_result(self, tool_name: str, tool_result: str, attachments: list[str] = []):

489 - attachments_str = json.dumps(attachments).replace("\n", "")

498 + def hist_add_tool_result(self, tool_name: str, tool_result: str):

499 content = self.parse_prompt(

491 - "fw.tool_result.md", tool_name=tool_name, tool_result=tool_result, attachments=attachments_str

500 + "fw.tool_result.md", tool_name=tool_name, tool_result=tool_result

501 )

502 return self.hist_add_message(False, content=content)

503

@@ -620,9 +629,9 @@ class Agent:

629 msg = self.intervention

630 self.intervention = None # reset the intervention message

631 if progress.strip():

623 - await self.hist_add_ai_response(progress)

632 + self.hist_add_ai_response(progress)

633 # append the intervention message

625 - await self.hist_add_user_message(msg, intervention=True)

634 + self.hist_add_user_message(msg, intervention=True)

635 raise InterventionException(msg)

636

637 async def wait_if_paused(self):

@@ -649,7 +658,7 @@ class Agent:

658 return response.message

659 else:

660 msg = self.read_prompt("fw.msg_misformat.md")

652 - await self.hist_add_warning(msg)

661 + self.hist_add_warning(msg)

662 PrintStyle(font_color="red", padding=True).print(msg)

663 self.context.log.log(

664 type="error", content=f"{self.agent_name}: Message misformat"

prompts/default/agent.system.instruments.md

+3 -2

@@ -1,4 +1,5 @@

1 # Instruments

2 -- following are instruments at disposal:

2 +- following are instruments at disposal

3 +- do not overly rely on them they might not be relevant

4 -{{instruments}}

\ No newline at end of file

5 +{{instruments}}

prompts/default/agent.system.memories.md

+2 -1

@@ -1,4 +1,5 @@

1 # Memories on the topic

2 -- following are your memories about current topic:

2 +- following are memories about current topic

3 +- do not overly rely on them they might not be relevant

5 {{memories}}

\ No newline at end of file

prompts/default/agent.system.solutions.md

+2 -1

@@ -1,4 +1,5 @@

1 # Solutions from the past

2 -- following are your memories about successful solutions of related problems:

2 +- following are memories about successful solutions of related problems

3 +- do not overly rely on them they might not be relevant

5 {{solutions}}

\ No newline at end of file

prompts/default/agent.system.tool.code_exe.md

+5 -2

@@ -3,10 +3,9 @@

3 execute terminal commands python nodejs code for computation or software tasks

4 place code in "code" arg; escape carefully and indent properly

5 select "runtime" arg: "terminal" "python" "nodejs" "output" "reset"

6 -for dialogues (Y/N etc.), use "terminal" runtime next step, send answer

6 +select "session" number, 0 default, others for multitasking

7 if code runs long, use "output" to wait, "reset" to kill process

8 use "pip" "npm" "apt-get" in "terminal" to install packages

9 -important: never use implicit print/output—it doesn't work!

9 to output, use print() or console.log()

10 if tool outputs error, adjust code before retrying; knowledge_tool can help

11 important: check code for placeholders or demo data; replace with real variables; don't reuse snippets

@@ -26,6 +25,7 @@ usage:

25 "tool_name": "code_execution_tool",

26 "tool_args": {

27 "runtime": "python",

28 + "session": 0,

29 "code": "import os\nprint(os.getcwd())",

30 }

31 }

@@ -41,6 +41,7 @@ usage:

41 "tool_name": "code_execution_tool",

42 "tool_args": {

43 "runtime": "terminal",

44 + "session": 0,

45 "code": "apt-get install zip",

46 }

47 }

@@ -55,6 +56,7 @@ usage:

56 "tool_name": "code_execution_tool",

57 "tool_args": {

58 "runtime": "output",

59 + "session": 0,

60 }

61 }

62 ~~~

@@ -68,6 +70,7 @@ usage:

70 "tool_name": "code_execution_tool",

71 "tool_args": {

72 "runtime": "reset",

73 + "session": 0,

74 }

75 }

76 ~~~

\ No newline at end of file

prompts/default/agent.system.tools_vision.md

+16 -1

@@ -1,3 +1,18 @@

1 ## "Multimodal (Vision) Agent Tools" available:

3 -None yet. In future, this section will contain vision-only tools

3 +### vision_load:

4 +load image data to LLM

5 +use paths arg for attachments

6 +

7 +**Example usage**:

8 +```json

9 +{

10 + "thoughts": [

11 + "I need to see the image...",

12 + ],

13 + "tool_name": "vision_load",

14 + "tool_args": {

15 + "paths": ["/path/to/image.png"],

16 + }

17 +}

18 +```

\ No newline at end of file

prompts/default/fw.tool_result.md

+3 -4

@@ -1,7 +1,6 @@

1 -~~~json

1 +```json

2 {

3 "tool_name": {{tool_name}},

4 - "tool_result": {{tool_result}},

5 - "attachments": {{attachments}}

4 + "tool_result": {{tool_result}}

5 }

7 -~~~

6 +```

python/api/ctx_window_get.py

+6 -2

@@ -9,6 +9,10 @@ class GetCtxWindow(ApiHandler):

9 context = self.get_context(ctxid)

10 agent = context.streaming_agent or context.agent0

11 window = agent.get_data(agent.DATA_NAME_CTX_WINDOW)

12 - size = tokens.approximate_tokens(window)

12 + if not window or not isinstance(window, dict):

13 + return {"content": "", "tokens": 0}

14 - return {"content": window, "tokens": size}

15 + text = window["text"]

16 + tokens = window["tokens"]

17 +

18 + return {"content": text, "tokens": tokens}

python/api/history_get.py

+2 -2

@@ -8,8 +8,8 @@ class GetHistory(ApiHandler):

8 ctxid = input.get("context", [])

9 context = self.get_context(ctxid)

10 agent = context.streaming_agent or context.agent0

11 - history = agent.history.output()

12 - size = tokens.approximate_tokens(agent.history.output_text())

11 + history = agent.history.output_text()

12 + size = agent.history.get_tokens()

14 return {

15 "history": history,

python/helpers/files.py

+36

@@ -1,6 +1,7 @@

1 from fnmatch import fnmatch

2 import json

3 import os, re

4 +import base64

6 import re

7 import shutil

@@ -45,6 +46,32 @@ def read_file(_relative_path, _backup_dirs=None, _encoding="utf-8", **kwargs):

46 return content

49 +def read_file_bin(_relative_path, _backup_dirs=None):

50 + # init backup dirs

51 + if _backup_dirs is None:

52 + _backup_dirs = []

53 +

54 + # get absolute path

55 + absolute_path = find_file_in_dirs(_relative_path, _backup_dirs)

56 +

57 + # read binary content

58 + with open(absolute_path, "rb") as f:

59 + return f.read()

60 +

61 +

62 +def read_file_base64(_relative_path, _backup_dirs=None):

63 + # init backup dirs

64 + if _backup_dirs is None:

65 + _backup_dirs = []

66 +

67 + # get absolute path

68 + absolute_path = find_file_in_dirs(_relative_path, _backup_dirs)

69 +

70 + # read binary content and encode to base64

71 + with open(absolute_path, "rb") as f:

72 + return base64.b64encode(f.read()).decode('utf-8')

73 +

74 +

75 def replace_placeholders_text(_content: str, **kwargs):

76 # Replace placeholders with values from kwargs

77 for key, value in kwargs.items():

@@ -175,6 +202,15 @@ def write_file_bin(relative_path: str, content: bytes):

202 f.write(content)

203

204

205 +def write_file_base64(relative_path: str, content: str):

206 + # decode base64 string to bytes

207 + data = base64.b64decode(content)

208 + abs_path = get_abs_path(relative_path)

209 + os.makedirs(os.path.dirname(abs_path), exist_ok=True)

210 + with open(abs_path, "wb") as f:

211 + f.write(data)

212 +

213 +

214 def delete_file(relative_path: str):

215 abs_path = get_abs_path(relative_path)

216 if os.path.exists(abs_path):

python/helpers/history.py

+162 -123

@@ -1,16 +1,14 @@

1 from abc import abstractmethod

2 import asyncio

3 from collections import OrderedDict

4 +from collections.abc import Mapping

5 import json

6 import math

6 -import os

7 -from typing import Coroutine, Literal, TypedDict, cast

7 +from typing import Coroutine, Literal, TypedDict, cast, Union, Dict, List, Any, override

8 from python.helpers import messages, tokens, settings, call_llm

9 from enum import Enum

10 -from langchain_core.messages import HumanMessage, SystemMessage, AIMessage, BaseMessage

11 -from python.helpers.print_style import PrintStyle

12 -from langchain_core.prompts import HumanMessagePromptTemplate

13 -from typing import Any

10 +from langchain_core.messages import BaseMessage, HumanMessage, SystemMessage, AIMessage

11 +

12 BULK_MERGE_COUNT = 3

13 TOPICS_KEEP_COUNT = 3

14 CURRENT_TOPIC_RATIO = 0.5

@@ -18,14 +16,22 @@ HISTORY_TOPIC_RATIO = 0.3

16 HISTORY_BULK_RATIO = 0.2

17 TOPIC_COMPRESS_RATIO = 0.65

18 LARGE_MESSAGE_TO_TOPIC_RATIO = 0.25

19 +RAW_MESSAGE_OUTPUT_TEXT_TRIM = 100

20 +

21 +

22 +class RawMessage(TypedDict):

23 + raw_content: "MessageContent"

24 + preview: str | None

25 +

22 -MessageContent = (

23 - list["MessageContent"]

24 - | OrderedDict[str, "MessageContent"]

25 - | list[OrderedDict[str, "MessageContent"]]

26 - | str

27 - | list[str]

28 -)

27 +MessageContent = Union[

28 + List["MessageContent"],

29 + Dict[str, "MessageContent"],

30 + List[Dict[str, "MessageContent"]],

31 + str,

32 + List[str],

33 + RawMessage,

34 +]

37 class OutputMessage(TypedDict):

@@ -37,9 +43,9 @@ class Record:

43 def __init__(self):

44 pass

46 + @abstractmethod

47 def get_tokens(self) -> int:

41 - out = self.output_text()

42 - return tokens.approximate_tokens(out)

48 + pass

50 @abstractmethod

51 async def compress(self) -> bool:

@@ -70,10 +76,25 @@ class Record:

78 class Message(Record):

73 - def __init__(self, ai: bool, content: MessageContent):

79 + def __init__(self, ai: bool, content: MessageContent, tokens: int = 0):

80 self.ai = ai

81 self.content = content

76 - self.summary: MessageContent = ""

82 + self.summary: str = ""

83 + self.tokens: int = tokens or self.calculate_tokens()

84 +

85 + @override

86 + def get_tokens(self) -> int:

87 + if not self.tokens:

88 + self.tokens = self.calculate_tokens()

89 + return self.tokens

90 +

91 + def calculate_tokens(self):

92 + text = self.output_text()

93 + return tokens.approximate_tokens(text)

94 +

95 + def set_summary(self, summary: str):

96 + self.summary = summary

97 + self.tokens = self.calculate_tokens()

99 async def compress(self):

100 return False

@@ -93,12 +114,15 @@ class Message(Record):

114 "ai": self.ai,

115 "content": self.content,

116 "summary": self.summary,

117 + "tokens": self.tokens,

118 }

119

120 @staticmethod

121 def from_dict(data: dict, history: "History"):

100 - msg = Message(ai=data["ai"], content=data.get("content", "Content lost"))

122 + content = data.get("content", "Content lost")

123 + msg = Message(ai=data["ai"], content=content)

124 msg.summary = data.get("summary", "")

125 + msg.tokens = data.get("tokens", 0)

126 return msg

127

128

@@ -108,8 +132,16 @@ class Topic(Record):

132 self.summary: str = ""

133 self.messages: list[Message] = []

134

111 - def add_message(self, ai: bool, content: MessageContent):

112 - msg = Message(ai=ai, content=content)

135 + def get_tokens(self):

136 + if self.summary:

137 + return tokens.approximate_tokens(self.summary)

138 + else:

139 + return sum(msg.get_tokens() for msg in self.messages)

140 +

141 + def add_message(

142 + self, ai: bool, content: MessageContent, tokens: int = 0

143 + ) -> Message:

144 + msg = Message(ai=ai, content=content, tokens=tokens)

145 self.messages.append(msg)

146 return msg

147

@@ -118,7 +150,7 @@ class Topic(Record):

150 return [OutputMessage(ai=False, content=self.summary)]

151 else:

152 msgs = [m for r in self.messages for m in r.output()]

121 - return group_outputs_abab(msgs)

153 + return msgs

154

155 async def summarize(self):

156 self.summary = await self.summarize_messages(self.messages)

@@ -134,22 +166,31 @@ class Topic(Record):

166 )

167 large_msgs = []

168 for m in (m for m in self.messages if not m.summary):

169 + # TODO refactor this

170 out = m.output()

171 text = output_text(out)

139 - tok = tokens.approximate_tokens(text)

172 + tok = m.get_tokens()

173 leng = len(text)

174 if tok > msg_max_size:

175 large_msgs.append((m, tok, leng, out))

176 large_msgs.sort(key=lambda x: x[1], reverse=True)

177 for msg, tok, leng, out in large_msgs:

178 trim_to_chars = leng * (msg_max_size / tok)

146 - trunc = messages.truncate_dict_by_ratio(

147 - self.history.agent,

148 - out[0]["content"],

149 - trim_to_chars * 1.15,

150 - trim_to_chars * 0.85,

151 - )

152 - msg.summary = trunc

179 + # raw messages will be replaced as a whole, they would become invalid when truncated

180 + if _is_raw_message(out[0]["content"]):

181 + msg.set_summary(

182 + "Message content replaced to save space in context window"

183 + )

184 +

185 + # regular messages will be truncated

186 + else:

187 + trunc = messages.truncate_dict_by_ratio(

188 + self.history.agent,

189 + out[0]["content"],

190 + trim_to_chars * 1.15,

191 + trim_to_chars * 0.85,

192 + )

193 + msg.set_summary(_json_dumps(trunc))

194

195 return True

196 return False

@@ -175,6 +216,7 @@ class Topic(Record):

216 return False

217

218 async def summarize_messages(self, messages: list[Message]):

219 + # FIXME: vision bytes are sent to utility LLM, send summary instead

220 msg_txt = [m.output_text() for m in messages]

221 summary = await self.history.agent.call_utility_model(

222 system=self.history.agent.read_prompt("fw.topic_summary.sys.md"),

@@ -194,9 +236,9 @@ class Topic(Record):

236 @staticmethod

237 def from_dict(data: dict, history: "History"):

238 topic = Topic(history=history)

197 - topic.summary = data["summary"]

239 + topic.summary = data.get("summary", "")

240 topic.messages = [

199 - Message.from_dict(m, history=history) for m in data["messages"]

241 + Message.from_dict(m, history=history) for m in data.get("messages", [])

242 ]

243 return topic

244

@@ -214,7 +256,7 @@ class Bulk(Record):

256 return [OutputMessage(ai=False, content=self.summary)]

257 else:

258 msgs = [m for r in self.records for m in r.output()]

217 - return group_outputs_abab(msgs)

259 + return msgs

260

261 async def compress(self):

262 return False

@@ -253,8 +295,15 @@ class History(Record):

295 self.current = Topic(history=self)

296 self.agent: Agent = agent

297

298 + def get_tokens(self) -> int:

299 + return (

300 + self.get_bulks_tokens()

301 + + self.get_topics_tokens()

302 + + self.get_current_topic_tokens()

303 + )

304 +

305 def is_over_limit(self):

257 - limit = get_ctx_size_for_history()

306 + limit = _get_ctx_size_for_history()

307 total = self.get_tokens()

308 return total > limit

309

@@ -267,15 +316,10 @@ class History(Record):

316 def get_current_topic_tokens(self) -> int:

317 return self.current.get_tokens()

318

270 - def get_tokens(self) -> int:

271 - return (

272 - self.get_bulks_tokens()

273 - + self.get_topics_tokens()

274 - + self.get_current_topic_tokens()

275 - )

276 -

277 - def add_message(self, ai: bool, content: MessageContent):

278 - return self.current.add_message(ai, content=content)

319 + def add_message(

320 + self, ai: bool, content: MessageContent, tokens: int = 0

321 + ) -> Message:

322 + return self.current.add_message(ai, content=content, tokens=tokens)

323

324 def new_topic(self):

325 if self.current.messages:

@@ -287,7 +331,6 @@ class History(Record):

331 result += [m for b in self.bulks for m in b.output()]

332 result += [m for t in self.topics for m in t.output()]

333 result += self.current.output()

290 - result = group_outputs_abab(result)

334 return result

335

336 @staticmethod

@@ -307,7 +350,7 @@ class History(Record):

350

351 def serialize(self):

352 data = self.to_dict()

310 - return json.dumps(data)

353 + return _json_dumps(data)

354

355 async def compress(self):

356 compressed = False

@@ -317,7 +360,7 @@ class History(Record):

360 self.get_topics_tokens(),

361 self.get_bulks_tokens(),

362 )

320 - total = get_ctx_size_for_history()

363 + total = _get_ctx_size_for_history()

364 ratios = [

365 (curr, CURRENT_TOPIC_RATIO, "current_topic"),

366 (hist, HISTORY_TOPIC_RATIO, "history_topic"),

@@ -392,25 +435,46 @@ class History(Record):

435 def deserialize_history(json_data: str, agent) -> History:

436 history = History(agent=agent)

437 if json_data:

395 - data = json.loads(json_data)

438 + data = _json_loads(json_data)

439 history = History.from_dict(data, history=history)

440 return history

441

442

400 -def get_ctx_size_for_history() -> int:

443 +def _get_ctx_size_for_history() -> int:

444 set = settings.get_settings()

445 return int(set["chat_model_ctx_length"] * set["chat_model_ctx_history"])

446

447

405 -def serialize_output(output: OutputMessage, ai_label="ai", human_label="human"):

406 - return f'{ai_label if output["ai"] else human_label}: {serialize_content(output["content"])}'

448 +def _stringify_output(output: OutputMessage, ai_label="ai", human_label="human"):

449 + return f'{ai_label if output["ai"] else human_label}: {_stringify_content(output["content"])}'

450

451

409 -def serialize_content(content: MessageContent) -> str:

452 +def _stringify_content(content: MessageContent) -> str:

453 + # already a string

454 if isinstance(content, str):

455 return content

456 +

457 + # raw messages return preview or trimmed json

458 + if _is_raw_message(content):

459 + preview: str = content.get("preview", "") # type: ignore

460 + if preview:

461 + return preview

462 + text = _json_dumps(content)

463 + if len(text) > RAW_MESSAGE_OUTPUT_TEXT_TRIM:

464 + return text[:RAW_MESSAGE_OUTPUT_TEXT_TRIM] + "... TRIMMED"

465 + return text

466 +

467 + # regular messages of non-string are dumped as json

468 + return _json_dumps(content)

469 +

470 +

471 +def _output_content_langchain(content: MessageContent):

472 + if isinstance(content, str):

473 + return content

474 + if _is_raw_message(content):

475 + return content["raw_content"] # type: ignore

476 try:

413 - return json.dumps(content)

477 + return _json_dumps(content)

478 except Exception as e:

479 raise e

480

@@ -421,98 +485,73 @@ def group_outputs_abab(outputs: list[OutputMessage]) -> list[OutputMessage]:

485 if result and result[-1]["ai"] == out["ai"]:

486 result[-1] = OutputMessage(

487 ai=result[-1]["ai"],

424 - content=merge_outputs(result[-1]["content"], out["content"]),

488 + content=_merge_outputs(result[-1]["content"], out["content"]),

489 )

490 else:

491 result.append(out)

492 return result

493

494

431 -def output_langchain(messages: list[OutputMessage]) -> list[BaseMessage]:

495 +def group_messages_abab(messages: list[BaseMessage]) -> list[BaseMessage]:

496 result = []

433 - for m in messages:

434 - if m["ai"]:

435 - result.append(AIMessage(content=serialize_content(m["content"])))

497 + for msg in messages:

498 + if result and isinstance(result[-1], type(msg)):

499 + # create new instance of the same type with merged content

500 + result[-1] = type(result[-1])(content=_merge_outputs(result[-1].content, msg.content)) # type: ignore

501 else:

437 - contents = m["content"]

438 -

439 - # sometimes content is a list sometimes not

440 - if not isinstance(contents, list):

441 - contents = [contents]

502 + result.append(msg)

503 + return result

504

443 - PrintStyle(font_color="grey", background_color="black", bold=True, padding=True).print(

444 - f"Contents: {json.dumps(contents, indent=2)}"

445 - )

505

447 - template: list[dict[str, str]] = [] # type: ignore

448 - message = ""

449 - images = {}

450 - for _, content in enumerate(contents):

451 - if message:

452 - # the first message is the user message, then the memory and solutions

453 - message += "\n\n--- Memory & Solutions Section: ---\n\n"

454 -

455 - message += serialize_content(content)

456 -

457 - if isinstance(content, dict) and "attachments" in content:

458 - attachments: list[str] = cast(list[str], content["attachments"])

459 - for attachment in attachments:

460 - if not os.path.exists(str(attachment)):

461 - continue

462 - if attachment not in images:

463 - import base64

464 - from mimetypes import guess_type

465 - mime_type, _ = guess_type(str(attachment))

466 - if mime_type.startswith("image/"):

467 - # Read and encode the image file

468 - with open(str(attachment), "rb") as image_file:

469 - base64_encoded_data = base64.b64encode(image_file.read()).decode('utf-8')

470 - # Construct the data URL

471 - images[attachment] = f"data:{mime_type};base64,{base64_encoded_data}"

472 -

473 - if message:

474 - template.append({"type": "text", "text": message})

475 - if images:

476 - for _, image in images.items():

477 - template.append({"type": "image_url", "image_url": image})

478 - if template:

479 - # only jinja2 is safe for json, both mustache({{...}}) and f-string({...}) are not

480 - result.append(HumanMessagePromptTemplate.from_template(template=template, partial_variables={}, template_format="jinja2")) # type: ignore

481 -

482 - PrintStyle(font_color="grey", background_color="black", bold=True, padding=True).print(

483 - f"Result: {result}"

484 - )

506 +def output_langchain(messages: list[OutputMessage]):

507 + result = []

508 + for m in messages:

509 + if m["ai"]:

510 + # result.append(AIMessage(content=serialize_content(m["content"])))

511 + result.append(AIMessage(_output_content_langchain(content=m["content"]))) # type: ignore

512 + else:

513 + # result.append(HumanMessage(content=serialize_content(m["content"])))

514 + result.append(HumanMessage(_output_content_langchain(content=m["content"]))) # type: ignore

515 + # ensure message type alternation

516 + result = group_messages_abab(result)

517 return result

518

519

520 def output_text(messages: list[OutputMessage], ai_label="ai", human_label="human"):

489 - return "\n".join(serialize_output(o, ai_label, human_label) for o in messages)

521 + return "\n".join(_stringify_output(o, ai_label, human_label) for o in messages)

522 +

523

524 +def _merge_outputs(a: MessageContent, b: MessageContent) -> MessageContent:

525 + if isinstance(a, str) and isinstance(b, str):

526 + return a + b

527

492 -def merge_outputs(a: MessageContent, b: MessageContent) -> MessageContent:

528 if not isinstance(a, list):

529 a = [a]

530 if not isinstance(b, list):

531 b = [b]

497 - return a + b # type: ignore

498 - # return merge_properties(a, b)

532 +

533 + return cast(MessageContent, a + b)

534

535

501 -def merge_properties(a: MessageContent, b: MessageContent) -> MessageContent:

502 - if isinstance(a, list):

503 - if isinstance(b, list):

504 - return a + b # type: ignore

536 +def _merge_properties(

537 + a: Dict[str, MessageContent], b: Dict[str, MessageContent]

538 +) -> Dict[str, MessageContent]:

539 + result = a.copy()

540 + for k, v in b.items():

541 + if k in result:

542 + result[k] = _merge_outputs(result[k], v)

543 else:

506 - return a + [b]

507 - elif isinstance(b, list):

508 - return [a] + b # type: ignore

509 - elif isinstance(a, dict) and isinstance(b, dict):

510 - for key, value in b.items():

511 - if key in a:

512 - a[key] = merge_properties(a[key], value)

513 - else:

514 - a[key] = value

515 - return a

516 - elif isinstance(a, str) and isinstance(b, str):

517 - return a + b

518 - raise ValueError(f"Cannot merge {a} and {b}")

544 + result[k] = v

545 + return result

546 +

547 +

548 +def _is_raw_message(obj: object) -> bool:

549 + return isinstance(obj, Mapping) and "raw_content" in obj

550 +

551 +

552 +def _json_dumps(obj):

553 + return json.dumps(obj, ensure_ascii=False)

554 +

555 +

556 +def _json_loads(obj):

557 + return json.loads(obj)

python/helpers/images.py new

+35

@@ -0,0 +1,35 @@

1 +from PIL import Image

2 +import io

3 +import math

4 +

5 +

6 +def compress_image(image_data: bytes, *, max_pixels: int = 256_000, quality: int = 50) -> bytes:

7 + """Compress an image by scaling it down and converting to JPEG with quality settings.

8 +

9 + Args:

10 + image_data: Raw image bytes

11 + max_pixels: Maximum number of pixels in the output image (width * height)

12 + quality: JPEG quality setting (1-100)

13 +

14 + Returns:

15 + Compressed image as bytes

16 + """

17 + # load image from bytes

18 + img = Image.open(io.BytesIO(image_data))

19 +

20 + # calculate scaling factor to get to max_pixels

21 + current_pixels = img.width * img.height

22 + if current_pixels > max_pixels:

23 + scale = math.sqrt(max_pixels / current_pixels)

24 + new_width = int(img.width * scale)

25 + new_height = int(img.height * scale)

26 + img = img.resize((new_width, new_height), Image.Resampling.LANCZOS)

27 +

28 + # convert to RGB if needed (for JPEG)

29 + if img.mode in ('RGBA', 'P'):

30 + img = img.convert('RGB')

31 +

32 + # save as JPEG with compression

33 + output = io.BytesIO()

34 + img.save(output, format='JPEG', quality=quality, optimize=True)

35 + return output.getvalue()

python/helpers/runtime.py

+22

@@ -2,6 +2,9 @@ import argparse

2 import inspect

3 from typing import TypeVar, Callable, Awaitable, Union, overload, cast

4 from python.helpers import dotenv, rfc, settings

5 +import asyncio

6 +import threading

7 +import queue

9 T = TypeVar('T')

10 R = TypeVar('R')

@@ -102,3 +105,22 @@ def _get_rfc_url() -> str:

105 url = url+":"+str(set["rfc_port_http"])

106 url += "/rfc"

107 return url

108 +

109 +

110 +def call_development_function_sync(func: Union[Callable[..., T], Callable[..., Awaitable[T]]], *args, **kwargs) -> T:

111 + # run async function in sync manner

112 + result_queue = queue.Queue()

113 +

114 + def run_in_thread():

115 + result = asyncio.run(call_development_function(func, *args, **kwargs))

116 + result_queue.put(result)

117 +

118 + thread = threading.Thread(target=run_in_thread)

119 + thread.start()

120 + thread.join(timeout=30) # wait for thread with timeout

121 +

122 + if thread.is_alive():

123 + raise TimeoutError("Function call timed out after 30 seconds")

124 +

125 + result = result_queue.get_nowait()

126 + return cast(T, result)

python/helpers/tool.py

+3 -4

@@ -1,5 +1,6 @@

1 from abc import abstractmethod

2 -from dataclasses import dataclass, field

2 +from dataclasses import dataclass

3 +

4 from agent import Agent

5 from python.helpers.print_style import PrintStyle

@@ -8,8 +9,6 @@ from python.helpers.print_style import PrintStyle

9 class Response:

10 message:str

11 break_loop: bool

11 - attachments: list[str] = field(default_factory=list[str])

12 -

13 class Tool:

@@ -34,7 +33,7 @@ class Tool:

34 async def after_execution(self, response: Response, **kwargs):

35 text = response.message.strip()

37 - await self.agent.hist_add_tool_result(self.name, text, response.attachments)

36 + self.agent.hist_add_tool_result(self.name, text)

37 PrintStyle(font_color="#1B4F72", background_color="white", padding=True, bold=True).print(f"{self.agent.agent_name}: Response from tool '{self.name}'")

38 PrintStyle(font_color="#85C1E9").print(response.message)

39 self.log.update(content=response.message)

python/helpers/whisper.py

+1 -1

@@ -30,7 +30,7 @@ async def _preload(model_name:str):

30 is_updating_model = True

31 if not _model or _model_name != model_name:

32 PrintStyle.standard(f"Loading Whisper model: {model_name}")

33 - _model = whisper.load_model(model_name)

33 + _model = whisper.load_model(name=model_name) # type: ignore

34 _model_name = model_name

35 finally:

36 is_updating_model = False

python/tools/call_subordinate.py

+1 -1

@@ -18,7 +18,7 @@ class Delegation(Tool):

19 # add user message to subordinate agent

20 subordinate: Agent = self.agent.get_data(Agent.DATA_NAME_SUBORDINATE)

21 - await subordinate.hist_add_user_message(UserMessage(message=message, attachments=[]))

21 + subordinate.hist_add_user_message(UserMessage(message=message, attachments=[]))

22 # run subordinate monologue

23 result = await subordinate.monologue()

24 # result

python/tools/code_execution_tool.py

+69 -30

@@ -12,7 +12,7 @@ from python.helpers.docker import DockerContainerManager

13 @dataclass

14 class State:

15 - shell: LocalInteractiveSession | SSHInteractiveSession

15 + shells: dict[int, LocalInteractiveSession | SSHInteractiveSession]

16 docker: DockerContainerManager | None

@@ -27,19 +27,26 @@ class CodeExecution(Tool):

27 # os.chdir(files.get_abs_path("./work_dir")) #change CWD to work_dir

29 runtime = self.args.get("runtime", "").lower().strip()

30 + session = int(self.args.get("session", 0))

32 if runtime == "python":

32 - response = await self.execute_python_code(self.args["code"])

33 + response = await self.execute_python_code(

34 + code=self.args["code"], session=session

35 + )

36 elif runtime == "nodejs":

34 - response = await self.execute_nodejs_code(self.args["code"])

37 + response = await self.execute_nodejs_code(

38 + code=self.args["code"], session=session

39 + )

40 elif runtime == "terminal":

36 - response = await self.execute_terminal_command(self.args["code"])

41 + response = await self.execute_terminal_command(

42 + command=self.args["code"], session=session

43 + )

44 elif runtime == "output":

45 response = await self.get_terminal_output(

39 - wait_with_output=5, wait_without_output=60

46 + session=session, wait_with_output=5, wait_without_output=60

47 )

48 elif runtime == "reset":

42 - response = await self.reset_terminal()

49 + response = await self.reset_terminal(session=session)

50 else:

51 response = self.agent.read_prompt(

52 "fw.code_runtime_wrong.md", runtime=runtime

@@ -72,11 +79,15 @@ class CodeExecution(Tool):

79 # PrintStyle().print()

81 def get_log_object(self):

75 - return self.agent.context.log.log(type="code_exe", heading=f"{self.agent.agent_name}: Using tool '{self.name}'", content="", kvps=self.args)

76 -

82 + return self.agent.context.log.log(

83 + type="code_exe",

84 + heading=f"{self.agent.agent_name}: Using tool '{self.name}'",

85 + content="",

86 + kvps=self.args,

87 + )

89 async def after_execution(self, response, **kwargs):

79 - await self.agent.hist_add_tool_result(self.name, response.message)

90 + self.agent.hist_add_tool_result(self.name, response.message)

92 async def prepare_state(self, reset=False):

93 self.state = self.agent.get_data("_cot_state")

@@ -97,7 +108,11 @@ class CodeExecution(Tool):

108

109 # initialize local or remote interactive shell insterface

110 if self.agent.config.code_exec_ssh_enabled:

100 - pswd = self.agent.config.code_exec_ssh_pass if self.agent.config.code_exec_ssh_pass else await rfc_exchange.get_root_password()

111 + pswd = (

112 + self.agent.config.code_exec_ssh_pass

113 + if self.agent.config.code_exec_ssh_pass

114 + else await rfc_exchange.get_root_password()

115 + )

116 shell = SSHInteractiveSession(

117 self.agent.context.log,

118 self.agent.config.code_exec_ssh_addr,

@@ -108,42 +123,63 @@ class CodeExecution(Tool):

123 else:

124 shell = LocalInteractiveSession()

125

111 - self.state = State(shell=shell, docker=docker)

126 + self.state = State(shells={0: shell}, docker=docker)

127 await shell.connect()

128 self.agent.set_data("_cot_state", self.state)

129

115 - async def execute_python_code(self, code: str, reset: bool = False):

130 + async def execute_python_code(self, session: int, code: str, reset: bool = False):

131 escaped_code = shlex.quote(code)

132 command = f"ipython -c {escaped_code}"

118 - return await self.terminal_session(command, reset)

133 + return await self.terminal_session(session, command, reset)

134

120 - async def execute_nodejs_code(self, code: str, reset: bool = False):

135 + async def execute_nodejs_code(self, session: int, code: str, reset: bool = False):

136 escaped_code = shlex.quote(code)

137 command = f"node /exe/node_eval.js {escaped_code}"

123 - return await self.terminal_session(command, reset)

138 + return await self.terminal_session(session, command, reset)

139

125 - async def execute_terminal_command(self, command: str, reset: bool = False):

126 - return await self.terminal_session(command, reset)

140 + async def execute_terminal_command(

141 + self, session: int, command: str, reset: bool = False

142 + ):

143 + return await self.terminal_session(session, command, reset)

144

128 - async def terminal_session(self, command: str, reset: bool = False):

145 + async def terminal_session(self, session: int, command: str, reset: bool = False):

146

147 await self.agent.handle_intervention() # wait for intervention and handle it, if paused

148 # try again on lost connection

149 for i in range(2):

150 try:

134 -

151 +

152 if reset:

153 await self.reset_terminal()

154

138 - self.state.shell.send_command(command)

139 -

140 - PrintStyle(background_color="white", font_color="#1B4F72", bold=True).print(

141 - f"{self.agent.agent_name} code execution output"

142 - )

143 - return await self.get_terminal_output()

155 + if session not in self.state.shells:

156 + if self.agent.config.code_exec_ssh_enabled:

157 + pswd = (

158 + self.agent.config.code_exec_ssh_pass

159 + if self.agent.config.code_exec_ssh_pass

160 + else await rfc_exchange.get_root_password()

161 + )

162 + shell = SSHInteractiveSession(

163 + self.agent.context.log,

164 + self.agent.config.code_exec_ssh_addr,

165 + self.agent.config.code_exec_ssh_port,

166 + self.agent.config.code_exec_ssh_user,

167 + pswd,

168 + )

169 + else:

170 + shell = LocalInteractiveSession()

171 + self.state.shells[session] = shell

172 + await shell.connect()

173 +

174 + self.state.shells[session].send_command(command)

175 +

176 + PrintStyle(

177 + background_color="white", font_color="#1B4F72", bold=True

178 + ).print(f"{self.agent.agent_name} code execution output")

179 + return await self.get_terminal_output(session)

180

181 except Exception as e:

146 - if i==1:

182 + if i == 1:

183 # try again on lost connection

184 PrintStyle.error(str(e))

185 await self.prepare_state(reset=True)

@@ -153,6 +189,7 @@ class CodeExecution(Tool):

189

190 async def get_terminal_output(

191 self,

192 + session=0,

193 reset_full_output=True,

194 wait_with_output=3,

195 wait_without_output=10,

@@ -165,10 +202,10 @@ class CodeExecution(Tool):

202

203 while max_exec_time <= 0 or time.time() - start_time < max_exec_time:

204 await asyncio.sleep(SLEEP_TIME) # Wait for some output to be generated

168 - full_output, partial_output = await self.state.shell.read_output(

205 + full_output, partial_output = await self.state.shells[session].read_output(

206 timeout=max_exec_time, reset_full_output=reset_full_output

207 )

171 - reset_full_output = False # only reset once

208 + reset_full_output = False # only reset once

209

210 await self.agent.handle_intervention() # wait for intervention and handle it, if paused

211

@@ -184,8 +221,10 @@ class CodeExecution(Tool):

221 break

222 return full_output

223

187 - async def reset_terminal(self):

188 - self.state.shell.close()

224 + async def reset_terminal(self, session=0):

225 + if session in self.state.shells:

226 + self.state.shells[session].close()

227 + del self.state.shells[session]

228 await self.prepare_state(reset=True)

229 response = self.agent.read_prompt("fw.code_reset.md")

230 self.log.update(content=response)

python/tools/input.py

+1 -1

@@ -20,4 +20,4 @@ class Input(Tool):

20 return self.agent.context.log.log(type="code_exe", heading=f"{self.agent.agent_name}: Using tool '{self.name}'", content="", kvps=self.args)

22 async def after_execution(self, response, **kwargs):

23 - await self.agent.hist_add_tool_result(self.name, response.message)

\ No newline at end of file

23 + self.agent.hist_add_tool_result(self.name, response.message)

\ No newline at end of file

python/tools/vision_load.py new

+79

@@ -0,0 +1,79 @@

1 +import base64

2 +from python.helpers.print_style import PrintStyle

3 +from python.helpers.tool import Tool, Response

4 +from python.helpers import runtime, files, images

5 +from mimetypes import guess_type

6 +from python.helpers import history

7 +

8 +# image optimization and token estimation for context window

9 +MAX_PIXELS = 768_000

10 +QUALITY = 75

11 +TOKENS_ESTIMATE = 1500

12 +

13 +

14 +class VisionLoad(Tool):

15 + async def execute(self, paths: list[str] = [], **kwargs) -> Response:

16 +

17 + self.images_dict = {}

18 + template: list[dict[str, str]] = [] # type: ignore

19 +

20 + for path in paths:

21 + if not await runtime.call_development_function(files.exists, str(path)):

22 + continue

23 +

24 + if path not in self.images_dict:

25 + mime_type, _ = guess_type(str(path))

26 + if mime_type and mime_type.startswith("image/"):

27 + # Read binary file

28 + file_content = await runtime.call_development_function(

29 + files.read_file_base64, str(path)

30 + )

31 + file_content = base64.b64decode(file_content)

32 + # Compress and convert to JPEG

33 + compressed = images.compress_image(

34 + file_content, max_pixels=MAX_PIXELS, quality=QUALITY

35 + )

36 + # Encode as base64

37 + file_content_b64 = base64.b64encode(compressed).decode("utf-8")

38 +

39 + # DEBUG: Save compressed image

40 + # await runtime.call_development_function(

41 + # files.write_file_base64, str(path), file_content_b64

42 + # )

43 +

44 + # Construct the data URL (always JPEG after compression)

45 + self.images_dict[path] = file_content_b64

46 +

47 + return Response(message="dummy", break_loop=False)

48 +

49 + async def after_execution(self, response: Response, **kwargs):

50 +

51 + # build image data messages for LLMs, or error message

52 + content = []

53 + if self.images_dict:

54 + for _, image in self.images_dict.items():

55 + content.append(

56 + {

57 + "type": "image_url",

58 + "image_url": {"url": f"data:image/jpeg;base64,{image}"},

59 + }

60 + )

61 + # append as raw message content for LLMs with vision tokens estimate

62 + msg = history.RawMessage(raw_content=content, preview="<Base64 encoded image data>")

63 + self.agent.hist_add_message(

64 + False, content=msg, tokens=TOKENS_ESTIMATE * len(content)

65 + )

66 + else:

67 + self.agent.hist_add_tool_result(self.name, "No images processed")

68 +

69 + # print and log short version

70 + message = (

71 + "No images processed"

72 + if not self.images_dict

73 + else f"{len(self.images_dict)} images processed"

74 + )

75 + PrintStyle(

76 + font_color="#1B4F72", background_color="white", padding=True, bold=True

77 + ).print(f"{self.agent.agent_name}: Response from tool '{self.name}'")

78 + PrintStyle(font_color="#85C1E9").print(message)

79 + self.log.update(result=message)

webui/js/history.js

+3 -2

@@ -3,9 +3,10 @@ import { getContext } from "../index.js";

3 export async function openHistoryModal() {

4 try {

5 const hist = await window.sendJsonData("/history_get", { context: getContext() });

6 - const data = JSON.stringify(hist.history, null, 4);

6 + // const data = JSON.stringify(hist.history, null, 4);

7 + const data = hist.history

8 const size = hist.tokens

8 - await showEditorModal(data, "json", `History ~${size} tokens`, "Conversation history visible to the LLM. History is compressed to fit into the context window over time.");

9 + await showEditorModal(data, "markdown", `History ~${size} tokens`, "Conversation history visible to the LLM. History is compressed to fit into the context window over time.");

10 } catch (e) {

11 window.toastFetchError("Error fetching history", e)

12 return