939 lines
36 KiB
Python
939 lines
36 KiB
Python
# ai_client.py
|
|
import tomllib
|
|
import json
|
|
import datetime
|
|
from pathlib import Path
|
|
import file_cache
|
|
import mcp_client
|
|
|
|
_provider: str = "gemini"
|
|
_model: str = "gemini-2.0-flash"
|
|
|
|
_gemini_client = None
|
|
_gemini_chat = None
|
|
|
|
_anthropic_client = None
|
|
_anthropic_history: list[dict] = []
|
|
|
|
# Injected by gui.py - called when AI wants to run a command.
|
|
# Signature: (script: str, base_dir: str) -> str | None
|
|
confirm_and_run_callback = None
|
|
|
|
# Injected by gui.py - called whenever a comms entry is appended.
|
|
# Signature: (entry: dict) -> None
|
|
comms_log_callback = None
|
|
|
|
# Injected by gui.py - called whenever a tool call completes.
|
|
# Signature: (script: str, result: str) -> None
|
|
tool_log_callback = None
|
|
|
|
# Increased to allow thorough code exploration before forcing a summary
|
|
MAX_TOOL_ROUNDS = 10
|
|
|
|
# Maximum characters per text chunk sent to Anthropic.
|
|
# Kept well under the ~200k token API limit.
|
|
_ANTHROPIC_CHUNK_SIZE = 120_000
|
|
|
|
_SYSTEM_PROMPT = (
|
|
"You are a helpful coding assistant with access to a PowerShell tool and MCP tools (file access: read_file, list_directory, search_files, get_file_summary, web access: web_search, fetch_url). "
|
|
"When asked to create or edit files, prefer targeted edits over full rewrites. "
|
|
"Always explain what you are doing before invoking the tool.\n\n"
|
|
"When writing or rewriting large files (especially those containing quotes, backticks, or special characters), "
|
|
"avoid python -c with inline strings. Instead: (1) write a .py helper script to disk using a PS here-string "
|
|
"(@'...'@ for literal content), (2) run it with `python <script>`, (3) delete the helper. "
|
|
"For small targeted edits, use PowerShell's (Get-Content) / .Replace() / Set-Content or Add-Content directly.\n\n"
|
|
"When making function calls using tools that accept array or object parameters "
|
|
"ensure those are structured using JSON. For example:\n"
|
|
"When you need to verify a change, rely on the exit code and stdout/stderr from the tool \u2014 "
|
|
"the user's context files are automatically refreshed after every tool call, so you do NOT "
|
|
"need to re-read files that are already provided in the <context> block."
|
|
)
|
|
|
|
|
|
_custom_system_prompt: str = ""
|
|
|
|
def set_custom_system_prompt(prompt: str):
|
|
global _custom_system_prompt
|
|
_custom_system_prompt = prompt
|
|
|
|
def _get_combined_system_prompt() -> str:
|
|
if _custom_system_prompt.strip():
|
|
return f"{_SYSTEM_PROMPT}\n\n[USER SYSTEM PROMPT]\n{_custom_system_prompt}"
|
|
return _SYSTEM_PROMPT
|
|
|
|
# ------------------------------------------------------------------ comms log
|
|
|
|
_comms_log: list[dict] = []
|
|
|
|
COMMS_CLAMP_CHARS = 300
|
|
|
|
|
|
def _append_comms(direction: str, kind: str, payload: dict):
|
|
entry = {
|
|
"ts": datetime.datetime.now().strftime("%H:%M:%S"),
|
|
"direction": direction,
|
|
"kind": kind,
|
|
"provider": _provider,
|
|
"model": _model,
|
|
"payload": payload,
|
|
}
|
|
_comms_log.append(entry)
|
|
if comms_log_callback is not None:
|
|
comms_log_callback(entry)
|
|
|
|
|
|
def get_comms_log() -> list[dict]:
|
|
return list(_comms_log)
|
|
|
|
|
|
def clear_comms_log():
|
|
_comms_log.clear()
|
|
|
|
|
|
def _load_credentials() -> dict:
|
|
with open("credentials.toml", "rb") as f:
|
|
return tomllib.load(f)
|
|
|
|
|
|
# ------------------------------------------------------------------ provider errors
|
|
|
|
class ProviderError(Exception):
|
|
def __init__(self, kind: str, provider: str, original: Exception):
|
|
self.kind = kind
|
|
self.provider = provider
|
|
self.original = original
|
|
super().__init__(str(original))
|
|
|
|
def ui_message(self) -> str:
|
|
labels = {
|
|
"quota": "QUOTA EXHAUSTED",
|
|
"rate_limit": "RATE LIMITED",
|
|
"auth": "AUTH / API KEY ERROR",
|
|
"balance": "BALANCE / BILLING ERROR",
|
|
"network": "NETWORK / CONNECTION ERROR",
|
|
"unknown": "API ERROR",
|
|
}
|
|
label = labels.get(self.kind, "API ERROR")
|
|
return f"[{self.provider.upper()} {label}]\n\n{self.original}"
|
|
|
|
|
|
def _classify_anthropic_error(exc: Exception) -> ProviderError:
|
|
try:
|
|
import anthropic
|
|
if isinstance(exc, anthropic.RateLimitError):
|
|
return ProviderError("rate_limit", "anthropic", exc)
|
|
if isinstance(exc, anthropic.AuthenticationError):
|
|
return ProviderError("auth", "anthropic", exc)
|
|
if isinstance(exc, anthropic.PermissionDeniedError):
|
|
return ProviderError("auth", "anthropic", exc)
|
|
if isinstance(exc, anthropic.APIConnectionError):
|
|
return ProviderError("network", "anthropic", exc)
|
|
if isinstance(exc, anthropic.APIStatusError):
|
|
status = getattr(exc, "status_code", 0)
|
|
body = str(exc).lower()
|
|
if status == 429:
|
|
return ProviderError("rate_limit", "anthropic", exc)
|
|
if status in (401, 403):
|
|
return ProviderError("auth", "anthropic", exc)
|
|
if status == 402:
|
|
return ProviderError("balance", "anthropic", exc)
|
|
if "credit" in body or "balance" in body or "billing" in body:
|
|
return ProviderError("balance", "anthropic", exc)
|
|
if "quota" in body or "limit" in body or "exceeded" in body:
|
|
return ProviderError("quota", "anthropic", exc)
|
|
except ImportError:
|
|
pass
|
|
return ProviderError("unknown", "anthropic", exc)
|
|
|
|
|
|
def _classify_gemini_error(exc: Exception) -> ProviderError:
|
|
body = str(exc).lower()
|
|
try:
|
|
from google.api_core import exceptions as gac
|
|
if isinstance(exc, gac.ResourceExhausted):
|
|
return ProviderError("quota", "gemini", exc)
|
|
if isinstance(exc, gac.TooManyRequests):
|
|
return ProviderError("rate_limit", "gemini", exc)
|
|
if isinstance(exc, (gac.Unauthenticated, gac.PermissionDenied)):
|
|
return ProviderError("auth", "gemini", exc)
|
|
if isinstance(exc, gac.ServiceUnavailable):
|
|
return ProviderError("network", "gemini", exc)
|
|
except ImportError:
|
|
pass
|
|
if "429" in body or "quota" in body or "resource exhausted" in body:
|
|
return ProviderError("quota", "gemini", exc)
|
|
if "rate" in body and "limit" in body:
|
|
return ProviderError("rate_limit", "gemini", exc)
|
|
if "401" in body or "403" in body or "api key" in body or "unauthenticated" in body:
|
|
return ProviderError("auth", "gemini", exc)
|
|
if "402" in body or "billing" in body or "balance" in body or "payment" in body:
|
|
return ProviderError("balance", "gemini", exc)
|
|
if "connection" in body or "timeout" in body or "unreachable" in body:
|
|
return ProviderError("network", "gemini", exc)
|
|
return ProviderError("unknown", "gemini", exc)
|
|
|
|
|
|
# ------------------------------------------------------------------ provider setup
|
|
|
|
def set_provider(provider: str, model: str):
|
|
global _provider, _model
|
|
_provider = provider
|
|
_model = model
|
|
|
|
|
|
def reset_session():
|
|
global _gemini_client, _gemini_chat
|
|
global _anthropic_client, _anthropic_history
|
|
_gemini_client = None
|
|
_gemini_chat = None
|
|
_anthropic_client = None
|
|
_anthropic_history = []
|
|
file_cache.reset_client()
|
|
|
|
|
|
# ------------------------------------------------------------------ model listing
|
|
|
|
def list_models(provider: str) -> list[str]:
|
|
creds = _load_credentials()
|
|
if provider == "gemini":
|
|
return _list_gemini_models(creds["gemini"]["api_key"])
|
|
elif provider == "anthropic":
|
|
return _list_anthropic_models()
|
|
return []
|
|
|
|
|
|
def _list_gemini_models(api_key: str) -> list[str]:
|
|
from google import genai
|
|
try:
|
|
client = genai.Client(api_key=api_key)
|
|
models = []
|
|
for m in client.models.list():
|
|
name = m.name
|
|
if name.startswith("models/"):
|
|
name = name[len("models/"):]
|
|
if "gemini" in name.lower():
|
|
models.append(name)
|
|
return sorted(models)
|
|
except Exception as exc:
|
|
raise _classify_gemini_error(exc) from exc
|
|
|
|
|
|
def _list_anthropic_models() -> list[str]:
|
|
import anthropic
|
|
try:
|
|
creds = _load_credentials()
|
|
client = anthropic.Anthropic(api_key=creds["anthropic"]["api_key"])
|
|
models = []
|
|
for m in client.models.list():
|
|
models.append(m.id)
|
|
return sorted(models)
|
|
except Exception as exc:
|
|
raise _classify_anthropic_error(exc) from exc
|
|
|
|
|
|
# ------------------------------------------------------------------ tool definition
|
|
|
|
TOOL_NAME = "run_powershell"
|
|
|
|
def _build_anthropic_tools() -> list[dict]:
|
|
"""Build the full Anthropic tools list: run_powershell + MCP file tools."""
|
|
mcp_tools = []
|
|
for spec in mcp_client.MCP_TOOL_SPECS:
|
|
mcp_tools.append({
|
|
"name": spec["name"],
|
|
"description": spec["description"],
|
|
"input_schema": spec["parameters"],
|
|
})
|
|
powershell_tool = {
|
|
"name": TOOL_NAME,
|
|
"description": (
|
|
"Run a PowerShell script within the project base_dir. "
|
|
"Use this to create, edit, rename, or delete files and directories. "
|
|
"The working directory is set to base_dir automatically. "
|
|
"Always prefer targeted edits over full rewrites where possible. "
|
|
"stdout and stderr are returned to you as the result."
|
|
),
|
|
"input_schema": {
|
|
"type": "object",
|
|
"properties": {
|
|
"script": {
|
|
"type": "string",
|
|
"description": "The PowerShell script to execute."
|
|
}
|
|
},
|
|
"required": ["script"]
|
|
},
|
|
"cache_control": {"type": "ephemeral"},
|
|
}
|
|
return mcp_tools + [powershell_tool]
|
|
|
|
|
|
_ANTHROPIC_TOOLS = _build_anthropic_tools()
|
|
|
|
|
|
def _gemini_tool_declaration():
|
|
from google.genai import types
|
|
|
|
declarations = []
|
|
|
|
# MCP file tools
|
|
for spec in mcp_client.MCP_TOOL_SPECS:
|
|
props = {}
|
|
for pname, pdef in spec["parameters"].get("properties", {}).items():
|
|
props[pname] = types.Schema(
|
|
type=types.Type.STRING,
|
|
description=pdef.get("description", ""),
|
|
)
|
|
declarations.append(types.FunctionDeclaration(
|
|
name=spec["name"],
|
|
description=spec["description"],
|
|
parameters=types.Schema(
|
|
type=types.Type.OBJECT,
|
|
properties=props,
|
|
required=spec["parameters"].get("required", []),
|
|
),
|
|
))
|
|
|
|
# PowerShell tool
|
|
declarations.append(types.FunctionDeclaration(
|
|
name=TOOL_NAME,
|
|
description=(
|
|
"Run a PowerShell script within the project base_dir. "
|
|
"Use this to create, edit, rename, or delete files and directories. "
|
|
"The working directory is set to base_dir automatically. "
|
|
"stdout and stderr are returned to you as the result."
|
|
),
|
|
parameters=types.Schema(
|
|
type=types.Type.OBJECT,
|
|
properties={
|
|
"script": types.Schema(
|
|
type=types.Type.STRING,
|
|
description="The PowerShell script to execute."
|
|
)
|
|
},
|
|
required=["script"]
|
|
),
|
|
))
|
|
|
|
return types.Tool(function_declarations=declarations)
|
|
|
|
|
|
def _run_script(script: str, base_dir: str) -> str:
|
|
if confirm_and_run_callback is None:
|
|
return "ERROR: no confirmation handler registered"
|
|
result = confirm_and_run_callback(script, base_dir)
|
|
if result is None:
|
|
output = "USER REJECTED: command was not executed"
|
|
else:
|
|
output = result
|
|
if tool_log_callback is not None:
|
|
tool_log_callback(script, output)
|
|
return output
|
|
|
|
|
|
# ------------------------------------------------------------------ dynamic file context refresh
|
|
|
|
def _reread_file_items(file_items: list[dict]) -> list[dict]:
|
|
"""
|
|
Re-read every file in file_items from disk, returning a fresh list.
|
|
This is called after tool calls so the AI sees updated file contents.
|
|
"""
|
|
refreshed = []
|
|
for item in file_items:
|
|
path = item.get("path")
|
|
if path is None:
|
|
refreshed.append(item)
|
|
continue
|
|
from pathlib import Path as _P
|
|
p = _P(path) if not isinstance(path, _P) else path
|
|
try:
|
|
content = p.read_text(encoding="utf-8")
|
|
refreshed.append({**item, "content": content, "error": False})
|
|
except Exception as e:
|
|
refreshed.append({**item, "content": f"ERROR re-reading {p}: {e}", "error": True})
|
|
return refreshed
|
|
|
|
|
|
def _build_file_context_text(file_items: list[dict]) -> str:
|
|
"""
|
|
Build a compact text summary of all files from file_items, suitable for
|
|
injecting into a tool_result message so the AI sees current file contents.
|
|
"""
|
|
if not file_items:
|
|
return ""
|
|
parts = []
|
|
for item in file_items:
|
|
path = item.get("path") or item.get("entry", "unknown")
|
|
suffix = str(path).rsplit(".", 1)[-1] if "." in str(path) else "text"
|
|
content = item.get("content", "")
|
|
parts.append(f"### `{path}`\n\n```{suffix}\n{content}\n```")
|
|
return "\n\n---\n\n".join(parts)
|
|
|
|
|
|
# ------------------------------------------------------------------ content block serialisation
|
|
|
|
def _content_block_to_dict(block) -> dict:
|
|
"""
|
|
Convert an Anthropic SDK content block object to a plain dict.
|
|
This ensures history entries are always JSON-serialisable dicts,
|
|
not opaque SDK objects that may fail on re-serialisation.
|
|
"""
|
|
if isinstance(block, dict):
|
|
return block
|
|
if hasattr(block, "model_dump"):
|
|
return block.model_dump()
|
|
if hasattr(block, "to_dict"):
|
|
return block.to_dict()
|
|
# Fallback: manually construct based on type
|
|
block_type = getattr(block, "type", None)
|
|
if block_type == "text":
|
|
return {"type": "text", "text": block.text}
|
|
if block_type == "tool_use":
|
|
return {"type": "tool_use", "id": block.id, "name": block.name, "input": block.input}
|
|
return {"type": "text", "text": str(block)}
|
|
|
|
|
|
# ------------------------------------------------------------------ gemini
|
|
|
|
def _ensure_gemini_client():
|
|
global _gemini_client
|
|
if _gemini_client is None:
|
|
from google import genai
|
|
creds = _load_credentials()
|
|
_gemini_client = genai.Client(api_key=creds["gemini"]["api_key"])
|
|
|
|
|
|
def _send_gemini(md_content: str, user_message: str, base_dir: str, file_items: list[dict] | None = None) -> str:
|
|
global _gemini_chat
|
|
from google import genai
|
|
from google.genai import types
|
|
|
|
try:
|
|
_ensure_gemini_client()
|
|
mcp_client.configure(file_items or [], [base_dir])
|
|
|
|
system_text = _get_combined_system_prompt() + f"\n\n<context>\n{md_content}\n</context>"
|
|
if _gemini_chat is None:
|
|
_gemini_chat = _gemini_client.chats.create(
|
|
model=_model,
|
|
config=types.GenerateContentConfig(
|
|
system_instruction=system_text,
|
|
tools=[_gemini_tool_declaration()]
|
|
)
|
|
)
|
|
else:
|
|
_gemini_chat = _gemini_client.chats.create(
|
|
model=_model,
|
|
config=types.GenerateContentConfig(
|
|
system_instruction=system_text,
|
|
tools=[_gemini_tool_declaration()]
|
|
),
|
|
history=_gemini_chat.get_history()
|
|
)
|
|
|
|
payload_to_send = user_message
|
|
|
|
_append_comms("OUT", "request", {
|
|
"message": f"[context {len(md_content)} chars + user message {len(user_message)} chars]",
|
|
})
|
|
|
|
all_text_parts = []
|
|
|
|
# We allow MAX_TOOL_ROUNDS, plus 1 final loop to get the text synthesis
|
|
for round_idx in range(MAX_TOOL_ROUNDS + 2):
|
|
response = _gemini_chat.send_message(payload_to_send)
|
|
|
|
text_parts_raw = [
|
|
part.text
|
|
for candidate in response.candidates
|
|
for part in candidate.content.parts
|
|
if hasattr(part, "text") and part.text
|
|
]
|
|
if text_parts_raw:
|
|
all_text_parts.append("\n".join(text_parts_raw))
|
|
|
|
tool_calls = [
|
|
part.function_call
|
|
for candidate in response.candidates
|
|
for part in candidate.content.parts
|
|
if hasattr(part, "function_call") and part.function_call is not None
|
|
]
|
|
|
|
usage_dict = {}
|
|
if hasattr(response, "usage_metadata") and response.usage_metadata:
|
|
meta = response.usage_metadata
|
|
if hasattr(meta, "prompt_token_count") and meta.prompt_token_count is not None:
|
|
usage_dict["input_tokens"] = meta.prompt_token_count
|
|
if hasattr(meta, "candidates_token_count") and meta.candidates_token_count is not None:
|
|
usage_dict["output_tokens"] = meta.candidates_token_count
|
|
if hasattr(meta, "cached_content_token_count") and meta.cached_content_token_count:
|
|
usage_dict["cache_read_input_tokens"] = meta.cached_content_token_count
|
|
|
|
stop_reason = ""
|
|
if response.candidates and hasattr(response.candidates[0], "finish_reason"):
|
|
fr = response.candidates[0].finish_reason
|
|
stop_reason = str(fr.name) if hasattr(fr, "name") else str(fr)
|
|
|
|
_append_comms("IN", "response", {
|
|
"round": round_idx,
|
|
"stop_reason": stop_reason,
|
|
"text": "\n".join(text_parts_raw),
|
|
"tool_calls": [{"name": fc.name, "args": dict(fc.args)} for fc in tool_calls],
|
|
"usage": usage_dict,
|
|
})
|
|
|
|
if not tool_calls:
|
|
break
|
|
|
|
if round_idx > MAX_TOOL_ROUNDS:
|
|
# The model ignored the MAX ROUNDS warning and kept calling tools.
|
|
# Force abort to prevent infinite loop.
|
|
break
|
|
|
|
function_responses = []
|
|
sent_results_log = []
|
|
|
|
for i, fc in enumerate(tool_calls):
|
|
fc_name = fc.name
|
|
fc_args = dict(fc.args)
|
|
|
|
if fc_name in mcp_client.TOOL_NAMES:
|
|
_append_comms("OUT", "tool_call", {"name": fc_name, "args": fc_args})
|
|
output = mcp_client.dispatch(fc_name, fc_args)
|
|
_append_comms("IN", "tool_result", {"name": fc_name, "output": output})
|
|
elif fc_name == TOOL_NAME:
|
|
script = fc_args.get("script", "")
|
|
_append_comms("OUT", "tool_call", {"name": TOOL_NAME, "script": script})
|
|
output = _run_script(script, base_dir)
|
|
_append_comms("IN", "tool_result", {"name": TOOL_NAME, "output": output})
|
|
else:
|
|
output = f"ERROR: unknown tool '{fc_name}'"
|
|
|
|
# Inject dynamic updates directly into the LAST tool's output string.
|
|
# Gemini strictly expects function_responses only, so we piggyback on the string.
|
|
if i == len(tool_calls) - 1:
|
|
if file_items:
|
|
file_items = _reread_file_items(file_items)
|
|
refreshed_ctx = _build_file_context_text(file_items)
|
|
if refreshed_ctx:
|
|
output += f"\n\n[SYSTEM: FILES UPDATED — current contents below. Do NOT re-read these files.]\n\n{refreshed_ctx}"
|
|
|
|
if round_idx == MAX_TOOL_ROUNDS:
|
|
output += "\n\n[SYSTEM WARNING: MAX TOOL ROUNDS REACHED. YOU MUST PROVIDE YOUR FINAL ANSWER NOW WITHOUT CALLING ANY MORE TOOLS.]"
|
|
|
|
function_responses.append(
|
|
types.Part.from_function_response(name=fc_name, response={"output": output})
|
|
)
|
|
sent_results_log.append({"tool_use_id": fc_name, "content": output})
|
|
|
|
_append_comms("OUT", "tool_result_send", {"results": sent_results_log})
|
|
payload_to_send = function_responses
|
|
|
|
final_text = "\n\n".join(all_text_parts)
|
|
return final_text if final_text.strip() else "(No text returned by the model)"
|
|
|
|
except ProviderError:
|
|
raise
|
|
except Exception as exc:
|
|
raise _classify_gemini_error(exc) from exc
|
|
|
|
|
|
|
|
# ------------------------------------------------------------------ anthropic history management
|
|
|
|
# Rough chars-per-token ratio. Anthropic tokeniser averages ~3.5-4 chars/token.
|
|
# We use 3.5 to be conservative (overestimate token count = safer).
|
|
_CHARS_PER_TOKEN = 3.5
|
|
|
|
# Maximum token budget for the entire prompt (system + tools + messages).
|
|
# Anthropic's limit is 200k. We leave headroom for the response + tool schemas.
|
|
_ANTHROPIC_MAX_PROMPT_TOKENS = 180_000
|
|
|
|
# Marker prefix used to identify stale file-refresh injections in history
|
|
_FILE_REFRESH_MARKER = "[FILES UPDATED"
|
|
|
|
|
|
def _estimate_message_tokens(msg: dict) -> int:
|
|
"""Rough token estimate for a single Anthropic message dict."""
|
|
total_chars = 0
|
|
content = msg.get("content", "")
|
|
if isinstance(content, str):
|
|
total_chars += len(content)
|
|
elif isinstance(content, list):
|
|
for block in content:
|
|
if isinstance(block, dict):
|
|
text = block.get("text", "") or block.get("content", "")
|
|
if isinstance(text, str):
|
|
total_chars += len(text)
|
|
# tool_use input
|
|
inp = block.get("input")
|
|
if isinstance(inp, dict):
|
|
import json as _json
|
|
total_chars += len(_json.dumps(inp, ensure_ascii=False))
|
|
elif isinstance(block, str):
|
|
total_chars += len(block)
|
|
return max(1, int(total_chars / _CHARS_PER_TOKEN))
|
|
|
|
|
|
def _estimate_prompt_tokens(system_blocks: list[dict], history: list[dict]) -> int:
|
|
"""Estimate total prompt tokens: system + tools + all history messages."""
|
|
total = 0
|
|
# System blocks
|
|
for block in system_blocks:
|
|
text = block.get("text", "")
|
|
total += max(1, int(len(text) / _CHARS_PER_TOKEN))
|
|
# Tool definitions (rough fixed estimate — they're ~2k tokens for our set)
|
|
total += 2500
|
|
# History messages
|
|
for msg in history:
|
|
total += _estimate_message_tokens(msg)
|
|
return total
|
|
|
|
|
|
def _strip_stale_file_refreshes(history: list[dict]):
|
|
"""
|
|
Remove [FILES UPDATED ...] text blocks from all history turns EXCEPT
|
|
the very last user message. These are stale snapshots from previous
|
|
tool rounds that bloat the context without providing value.
|
|
"""
|
|
if len(history) < 2:
|
|
return
|
|
# Find the index of the last user message — we keep its file refresh intact
|
|
last_user_idx = -1
|
|
for i in range(len(history) - 1, -1, -1):
|
|
if history[i].get("role") == "user":
|
|
last_user_idx = i
|
|
break
|
|
for i, msg in enumerate(history):
|
|
if msg.get("role") != "user" or i == last_user_idx:
|
|
continue
|
|
content = msg.get("content")
|
|
if not isinstance(content, list):
|
|
continue
|
|
cleaned = []
|
|
for block in content:
|
|
if isinstance(block, dict) and block.get("type") == "text":
|
|
text = block.get("text", "")
|
|
if text.startswith(_FILE_REFRESH_MARKER):
|
|
continue # drop this stale file refresh block
|
|
cleaned.append(block)
|
|
if len(cleaned) < len(content):
|
|
msg["content"] = cleaned
|
|
|
|
|
|
def _trim_anthropic_history(system_blocks: list[dict], history: list[dict]):
|
|
"""
|
|
Trim the Anthropic history to fit within the token budget.
|
|
Strategy:
|
|
1. Strip stale file-refresh injections from old turns.
|
|
2. If still over budget, drop oldest turn pairs (user + assistant).
|
|
Returns the number of messages dropped.
|
|
"""
|
|
# Phase 1: strip stale file refreshes
|
|
_strip_stale_file_refreshes(history)
|
|
|
|
est = _estimate_prompt_tokens(system_blocks, history)
|
|
if est <= _ANTHROPIC_MAX_PROMPT_TOKENS:
|
|
return 0
|
|
|
|
# Phase 2: drop oldest turn pairs until within budget
|
|
dropped = 0
|
|
while len(history) > 2 and est > _ANTHROPIC_MAX_PROMPT_TOKENS:
|
|
# Always drop from the front in pairs (user, assistant) to maintain alternation
|
|
# But be careful: the first message might be user, followed by assistant
|
|
if history[0].get("role") == "user" and len(history) > 1 and history[1].get("role") == "assistant":
|
|
removed_user = history.pop(0)
|
|
removed_asst = history.pop(0)
|
|
dropped += 2
|
|
est -= _estimate_message_tokens(removed_user)
|
|
est -= _estimate_message_tokens(removed_asst)
|
|
# If the next message is a user tool_result that belonged to the dropped assistant,
|
|
# we need to drop it too to avoid dangling tool_results
|
|
while history and history[0].get("role") == "user":
|
|
content = history[0].get("content", [])
|
|
if isinstance(content, list) and content and isinstance(content[0], dict) and content[0].get("type") == "tool_result":
|
|
removed_tr = history.pop(0)
|
|
dropped += 1
|
|
est -= _estimate_message_tokens(removed_tr)
|
|
# And the assistant reply that followed it
|
|
if history and history[0].get("role") == "assistant":
|
|
removed_a2 = history.pop(0)
|
|
dropped += 1
|
|
est -= _estimate_message_tokens(removed_a2)
|
|
else:
|
|
break
|
|
else:
|
|
# Edge case: history starts with something unexpected. Drop one message.
|
|
removed = history.pop(0)
|
|
dropped += 1
|
|
est -= _estimate_message_tokens(removed)
|
|
|
|
return dropped
|
|
|
|
|
|
# ------------------------------------------------------------------ anthropic
|
|
|
|
def _ensure_anthropic_client():
|
|
global _anthropic_client
|
|
if _anthropic_client is None:
|
|
import anthropic
|
|
creds = _load_credentials()
|
|
_anthropic_client = anthropic.Anthropic(api_key=creds["anthropic"]["api_key"])
|
|
|
|
|
|
def _chunk_text(text: str, chunk_size: int) -> list[str]:
|
|
return [text[i:i + chunk_size] for i in range(0, len(text), chunk_size)]
|
|
|
|
|
|
def _build_chunked_context_blocks(md_content: str) -> list[dict]:
|
|
"""
|
|
Split md_content into <=_ANTHROPIC_CHUNK_SIZE char chunks.
|
|
cache_control:ephemeral is placed only on the LAST block so the whole
|
|
prefix is cached as one unit.
|
|
"""
|
|
chunks = _chunk_text(md_content, _ANTHROPIC_CHUNK_SIZE)
|
|
blocks = []
|
|
for i, chunk in enumerate(chunks):
|
|
block: dict = {"type": "text", "text": chunk}
|
|
if i == len(chunks) - 1:
|
|
block["cache_control"] = {"type": "ephemeral"}
|
|
blocks.append(block)
|
|
return blocks
|
|
|
|
|
|
def _strip_cache_controls(history: list[dict]):
|
|
"""
|
|
Remove cache_control from all content blocks in message history.
|
|
Anthropic allows max 4 cache_control blocks total across system + tools +
|
|
messages. We reserve those slots for the stable system/tools prefix and
|
|
the current turn's context block, so all older history entries must be clean.
|
|
"""
|
|
for msg in history:
|
|
content = msg.get("content")
|
|
if isinstance(content, list):
|
|
for block in content:
|
|
if isinstance(block, dict):
|
|
block.pop("cache_control", None)
|
|
|
|
def _repair_anthropic_history(history: list[dict]):
|
|
"""
|
|
If history ends with an assistant message that contains tool_use blocks
|
|
without a following user tool_result message, append a synthetic tool_result
|
|
message so the history is valid before the next request.
|
|
"""
|
|
if not history:
|
|
return
|
|
last = history[-1]
|
|
if last.get("role") != "assistant":
|
|
return
|
|
content = last.get("content", [])
|
|
tool_use_ids = []
|
|
for block in content:
|
|
if isinstance(block, dict):
|
|
if block.get("type") == "tool_use":
|
|
tool_use_ids.append(block["id"])
|
|
if not tool_use_ids:
|
|
return
|
|
history.append({
|
|
"role": "user",
|
|
"content": [
|
|
{
|
|
"type": "tool_result",
|
|
"tool_use_id": tid,
|
|
"content": "Tool call was not completed (session interrupted).",
|
|
}
|
|
for tid in tool_use_ids
|
|
],
|
|
})
|
|
|
|
|
|
def _send_anthropic(md_content: str, user_message: str, base_dir: str, file_items: list[dict] | None = None) -> str:
|
|
try:
|
|
_ensure_anthropic_client()
|
|
mcp_client.configure(file_items or [], [base_dir])
|
|
|
|
system_text = _get_combined_system_prompt() + f"\n\n<context>\n{md_content}\n</context>"
|
|
system_blocks = _build_chunked_context_blocks(system_text)
|
|
|
|
user_content = [{"type": "text", "text": user_message}]
|
|
|
|
_strip_cache_controls(_anthropic_history)
|
|
_repair_anthropic_history(_anthropic_history)
|
|
_anthropic_history.append({"role": "user", "content": user_content})
|
|
|
|
n_chunks = len(system_blocks)
|
|
_append_comms("OUT", "request", {
|
|
"message": (
|
|
f"[system {n_chunks} chunk(s), {len(md_content)} chars context] "
|
|
f"{user_message[:200]}{'...' if len(user_message) > 200 else ''}"
|
|
),
|
|
})
|
|
|
|
all_text_parts = []
|
|
|
|
# We allow MAX_TOOL_ROUNDS, plus 1 final loop to get the text synthesis
|
|
for round_idx in range(MAX_TOOL_ROUNDS + 2):
|
|
# Trim history to fit within token budget before each API call
|
|
dropped = _trim_anthropic_history(system_blocks, _anthropic_history)
|
|
if dropped > 0:
|
|
est_tokens = _estimate_prompt_tokens(system_blocks, _anthropic_history)
|
|
_append_comms("OUT", "request", {
|
|
"message": (
|
|
f"[HISTORY TRIMMED: dropped {dropped} old messages to fit token budget. "
|
|
f"Estimated {est_tokens} tokens remaining. {len(_anthropic_history)} messages in history.]"
|
|
),
|
|
})
|
|
|
|
response = _anthropic_client.messages.create(
|
|
model=_model,
|
|
max_tokens=16384,
|
|
system=system_blocks,
|
|
tools=_build_anthropic_tools(),
|
|
messages=_anthropic_history,
|
|
)
|
|
|
|
# Convert SDK content block objects to plain dicts before storing in history
|
|
serialised_content = [_content_block_to_dict(b) for b in response.content]
|
|
|
|
_anthropic_history.append({
|
|
"role": "assistant",
|
|
"content": serialised_content,
|
|
})
|
|
|
|
text_blocks = [b.text for b in response.content if hasattr(b, "text") and b.text]
|
|
if text_blocks:
|
|
all_text_parts.append("\n".join(text_blocks))
|
|
|
|
tool_use_blocks = [
|
|
{"id": b.id, "name": b.name, "input": b.input}
|
|
for b in response.content
|
|
if getattr(b, "type", None) == "tool_use"
|
|
]
|
|
|
|
usage_dict: dict = {}
|
|
if response.usage:
|
|
usage_dict["input_tokens"] = response.usage.input_tokens
|
|
usage_dict["output_tokens"] = response.usage.output_tokens
|
|
cache_creation = getattr(response.usage, "cache_creation_input_tokens", None)
|
|
cache_read = getattr(response.usage, "cache_read_input_tokens", None)
|
|
if cache_creation is not None:
|
|
usage_dict["cache_creation_input_tokens"] = cache_creation
|
|
if cache_read is not None:
|
|
usage_dict["cache_read_input_tokens"] = cache_read
|
|
|
|
_append_comms("IN", "response", {
|
|
"round": round_idx,
|
|
"stop_reason": response.stop_reason,
|
|
"text": "\n".join(text_blocks),
|
|
"tool_calls": tool_use_blocks,
|
|
"usage": usage_dict,
|
|
})
|
|
|
|
if response.stop_reason != "tool_use" or not tool_use_blocks:
|
|
break
|
|
|
|
if round_idx > MAX_TOOL_ROUNDS:
|
|
# The model ignored the MAX ROUNDS warning and kept calling tools.
|
|
# Force abort to prevent infinite loop.
|
|
break
|
|
|
|
tool_results = []
|
|
for block in response.content:
|
|
if getattr(block, "type", None) != "tool_use":
|
|
continue
|
|
b_name = getattr(block, "name", None)
|
|
b_id = getattr(block, "id", "")
|
|
b_input = getattr(block, "input", {})
|
|
if b_name in mcp_client.TOOL_NAMES:
|
|
_append_comms("OUT", "tool_call", {"name": b_name, "id": b_id, "args": b_input})
|
|
output = mcp_client.dispatch(b_name, b_input)
|
|
_append_comms("IN", "tool_result", {"name": b_name, "id": b_id, "output": output})
|
|
tool_results.append({
|
|
"type": "tool_result",
|
|
"tool_use_id": b_id,
|
|
"content": output,
|
|
})
|
|
elif b_name == TOOL_NAME:
|
|
script = b_input.get("script", "")
|
|
_append_comms("OUT", "tool_call", {
|
|
"name": TOOL_NAME,
|
|
"id": b_id,
|
|
"script": script,
|
|
})
|
|
output = _run_script(script, base_dir)
|
|
_append_comms("IN", "tool_result", {
|
|
"name": TOOL_NAME,
|
|
"id": b_id,
|
|
"output": output,
|
|
})
|
|
tool_results.append({
|
|
"type": "tool_result",
|
|
"tool_use_id": b_id,
|
|
"content": output,
|
|
})
|
|
|
|
# Refresh file context after tool calls and inject into tool result message
|
|
if file_items:
|
|
file_items = _reread_file_items(file_items)
|
|
refreshed_ctx = _build_file_context_text(file_items)
|
|
if refreshed_ctx:
|
|
tool_results.append({
|
|
"type": "text",
|
|
"text": (
|
|
"[FILES UPDATED — current contents below. "
|
|
"Do NOT re-read these files with PowerShell.]\n\n"
|
|
+ refreshed_ctx
|
|
),
|
|
})
|
|
|
|
if round_idx == MAX_TOOL_ROUNDS:
|
|
tool_results.append({
|
|
"type": "text",
|
|
"text": "SYSTEM WARNING: MAX TOOL ROUNDS REACHED. YOU MUST PROVIDE YOUR FINAL ANSWER NOW WITHOUT CALLING ANY MORE TOOLS."
|
|
})
|
|
|
|
_anthropic_history.append({
|
|
"role": "user",
|
|
"content": tool_results,
|
|
})
|
|
|
|
_append_comms("OUT", "tool_result_send", {
|
|
"results": [
|
|
{"tool_use_id": r["tool_use_id"], "content": r["content"]}
|
|
for r in tool_results if r.get("type") == "tool_result"
|
|
],
|
|
})
|
|
|
|
final_text = "\n\n".join(all_text_parts)
|
|
return final_text if final_text.strip() else "(No text returned by the model)"
|
|
|
|
except ProviderError:
|
|
raise
|
|
except Exception as exc:
|
|
raise _classify_anthropic_error(exc) from exc
|
|
|
|
|
|
# ------------------------------------------------------------------ unified send
|
|
|
|
def send(
|
|
md_content: str,
|
|
user_message: str,
|
|
base_dir: str = ".",
|
|
file_items: list[dict] | None = None,
|
|
) -> str:
|
|
"""
|
|
Send a message to the active provider.
|
|
|
|
md_content : aggregated markdown string from aggregate.run()
|
|
user_message: the user question / instruction
|
|
base_dir : project base directory (for PowerShell tool calls)
|
|
file_items : list of file dicts from aggregate.build_file_items() for
|
|
dynamic context refresh after tool calls
|
|
"""
|
|
if _provider == "gemini":
|
|
return _send_gemini(md_content, user_message, base_dir, file_items)
|
|
elif _provider == "anthropic":
|
|
return _send_anthropic(md_content, user_message, base_dir, file_items)
|
|
raise ValueError(f"unknown provider: {_provider}")
|
|
|