feat(matrix): land QA follow-ups and refresh docs

- harden Matrix onboarding/chat lifecycle after manual QA
- refresh README and Matrix docs to match current behavior
- add local ignores for runtime artifacts and include current planning/report docs

Closes #7
Closes #9
Closes #14
This commit is contained in:
Mikhail Putilovskij 2026-04-05 19:08:58 +03:00
parent 7fce4c9b3e
commit 6ced154124
35 changed files with 8380 additions and 67 deletions

635
bot-examples/llm_session.py Normal file
View file

@ -0,0 +1,635 @@
"""Claude CLI session manager.
Manages Claude Code CLI sessions per topic. Each topic gets a persistent
session ID so conversation context is maintained across messages.
Uses --output-format stream-json with asyncio subprocess to stream responses.
Falls back to claude-zai if primary claude fails.
Timeout: idle-based (resets on any output from Claude) + hard ceiling.
Status: streams tool_use/agent events via on_status callback.
Cancel: external cancel_event to stop processing.
"""
import asyncio
import json
import logging
import os
import shutil
import time
import uuid
from collections.abc import Callable
from pathlib import Path
from core.config import Config
logger = logging.getLogger(__name__)
def _session_path(data_dir: Path, topic_id: int | str, provider: str = "") -> Path:
"""Path to session ID file for a topic."""
suffix = f"_{provider}" if provider else ""
return data_dir / "topics" / str(topic_id) / f"session{suffix}.txt"
def load_session(data_dir: Path, topic_id: int | str, provider: str = "") -> str | None:
"""Load existing session ID for a topic, or None."""
path = _session_path(data_dir, topic_id, provider)
if path.exists():
return path.read_text().strip()
return None
def save_session(data_dir: Path, topic_id: int | str, session_id: str, provider: str = "") -> None:
"""Save session ID for a topic."""
path = _session_path(data_dir, topic_id, provider)
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(session_id)
async def send_message(
config: Config,
topic_id: int | str,
message: str,
on_chunk: Callable | None = None,
on_question: Callable | None = None,
on_status: Callable | None = None,
cancel_event: asyncio.Event | None = None,
idle_timeout_ref: list | None = None,
user_profile: str = "",
workspace_dir: Path | None = None,
) -> str:
"""Send a message to Claude CLI and return the response.
Args:
config: Application config.
topic_id: Topic ID (determines session and working directory).
message: User message text.
on_chunk: Optional async callback(text_so_far) for streaming updates.
on_question: Optional async callback(question) -> answer for ask-user tool.
on_status: Optional async callback(dict) for tool/agent status events.
cancel_event: Optional asyncio.Event set to cancel processing.
idle_timeout_ref: Optional mutable [int] current idle timeout in seconds.
Can be modified externally (e.g. user "more time" command).
user_profile: Optional user profile text (from user.md) to inject into system prompt.
workspace_dir: Optional per-user workspace directory path.
Returns:
Full response text.
Raises:
RuntimeError: If both primary and fallback CLI fail.
"""
# Try primary provider first
try:
return await _send_with_provider(config, topic_id, message, on_chunk, on_question,
on_status=on_status, cancel_event=cancel_event,
idle_timeout_ref=idle_timeout_ref,
provider="", user_profile=user_profile,
workspace_dir=workspace_dir)
except RuntimeError as e:
# Don't fallback if user cancelled
if cancel_event and cancel_event.is_set():
raise RuntimeError("Cancelled")
logger.warning("Primary claude failed (%s), trying fallback (claude-zai)", e)
# Fallback: claude-zai with separate session (using opus model)
try:
response = await _send_with_provider(
config, topic_id, message, on_chunk, on_question,
on_status=on_status, cancel_event=cancel_event,
idle_timeout_ref=idle_timeout_ref,
provider="zai", cmd_override="claude-zai", model_override="opus",
user_profile=user_profile, workspace_dir=workspace_dir,
)
# Add note that fallback provider was used
return response + "\n\n_[(via z.ai fallback)]_"
except RuntimeError:
raise RuntimeError("Both claude and claude-zai failed")
async def _watch_questions(topic_dir: Path, on_question: Callable) -> None:
"""Watch for ask-user.json and forward questions to the bot."""
question_file = topic_dir / "ask-user.json"
fifo_file = topic_dir / "ask-user.fifo"
while True:
await asyncio.sleep(0.5)
if not question_file.exists():
continue
try:
data = json.loads(question_file.read_text())
question = data.get("question", "")
logger.info("Claude asks user: %s", question[:200])
answer = await on_question(question)
# Write answer to FIFO (unblocks ask-user script)
with open(fifo_file, "w") as f:
f.write(answer)
question_file.unlink(missing_ok=True)
except Exception as e:
logger.error("Error handling ask-user: %s", e)
question_file.unlink(missing_ok=True)
def _tool_preview(tool_name: str, raw_input: str) -> str:
"""Extract a human-readable preview from tool input JSON."""
try:
inp = json.loads(raw_input)
except (json.JSONDecodeError, TypeError):
return raw_input[:200]
if tool_name == "Bash":
return inp.get("command", "")[:500]
if tool_name in ("Read", "Write"):
return inp.get("file_path", "")[:300]
if tool_name == "Edit":
return inp.get("file_path", "")[:300]
if tool_name in ("Glob", "Grep"):
return inp.get("pattern", "")[:200]
if tool_name == "WebSearch":
return inp.get("query", "")[:200]
if tool_name == "WebFetch":
return inp.get("url", "")[:300]
if tool_name == "Agent":
desc = inp.get("description", "")
prompt = inp.get("prompt", "")
return desc[:200] if desc else prompt[:300]
if tool_name == "TodoWrite":
todos = inp.get("todos", [])
if todos:
items = [t.get("content", "")[:80] for t in todos[:3]]
return "; ".join(items)
# Generic: show first key=value
for k, v in inp.items():
return f"{k}={str(v)[:200]}"
return ""
def _load_conversation_log(data_dir: Path, topic_id: str, limit: int = 5) -> str:
"""Load recent conversation log for context.
Returns formatted summary of last N interactions from log.jsonl,
so Claude has context even after session resets or fallback switches.
"""
log_file = data_dir / "rooms" / str(topic_id) / "log.jsonl"
if not log_file.exists():
return ""
try:
with open(log_file) as f:
entries = [json.loads(line.strip()) for line in f if line.strip()]
except Exception:
return ""
if not entries:
return ""
recent = entries[-limit:]
parts = []
for e in recent:
ts = e.get("ts", "")[:16].replace("T", " ")
user = e.get("user", "")[:300]
bot = e.get("bot", "")[:500]
parts.append(f"[{ts}] User: {user}")
parts.append(f"[{ts}] Bot: {bot}")
return "\n".join(parts)
async def _send_with_provider(
config: Config,
topic_id: int | str,
message: str,
on_chunk: Callable | None,
on_question: Callable | None,
on_status: Callable | None = None,
cancel_event: asyncio.Event | None = None,
idle_timeout_ref: list | None = None,
provider: str = "",
cmd_override: str | None = None,
model_override: str | None = None,
user_profile: str = "",
workspace_dir: Path | None = None,
_retry_count: int = 0,
) -> str:
"""Send message using a specific provider."""
existing_session = load_session(config.data_dir, topic_id, provider)
topic_dir = config.data_dir / "topics" / str(topic_id)
topic_dir.mkdir(parents=True, exist_ok=True)
cmd = cmd_override or config.claude_cmd
# Build args: --resume for existing sessions, --session-id for new ones
if existing_session:
session_flag = ["--resume", existing_session]
else:
new_id = str(uuid.uuid4())
session_flag = ["--session-id", new_id]
# User profile: prefer explicit parameter, fallback to workspace user.md
user_context = ""
if user_profile:
user_context = f"\n\nUSER PROFILE:\n{user_profile}\n"
elif config.workspace_dir:
user_md = config.workspace_dir / "user.md"
if user_md.exists():
user_context = f"\n\nUSER PROFILE:\n{user_md.read_text().strip()}\n"
# Load recent conversation log — provides context after session resets,
# fallback switches, or timeouts. Always included so Claude knows what happened.
conv_log = _load_conversation_log(config.data_dir, str(topic_id))
conv_context = ""
if conv_log:
conv_context = (
"\n\nRECENT CONVERSATION LOG (from bot's perspective, "
"may overlap with your session memory — use to fill gaps "
"after timeouts or session switches):\n" + conv_log + "\n"
)
# Per-user workspace context
workspace_context = ""
if workspace_dir and workspace_dir.is_dir():
ws_md = workspace_dir / "WORKSPACE.md"
if ws_md.exists():
workspace_context = (
f"\n\nUSER WORKSPACE ({workspace_dir}):\n"
f"{ws_md.read_text().strip()}\n"
f"\nYour working directory is the topic dir ({topic_dir}). "
f"Use it for scratch work (scripts, downloads, temp files). "
f"Save important/refined results to the workspace at {workspace_dir}. "
f"The workspace is a git repo — your changes will be committed automatically.\n"
)
# Paths Claude should know about
room_dir = config.data_dir / "rooms" / str(topic_id)
log_file = room_dir / "log.jsonl"
history_file = room_dir / "history.jsonl"
# System prompt with topic context
system_extra = (
f"Topic/room ID: {topic_id}. Data dir: {topic_dir}. "
f"After responding, update {config.data_dir / 'topic-map.yml'} "
f"with this topic's ID, path, and a short label. "
f"The bot renames the topic from the label. "
f"CONVERSATION HISTORY: Full conversation log is at {log_file} (JSONL, "
f"fields: ts, user, bot — every interaction with timestamps). "
f"Detailed message history with sender info: {history_file}. "
f"If you lose context (after timeout, session switch, or restart), "
f"READ these files to recover the full conversation. "
f"Entries ending with '[timed out]' or '[idle timeout]' mean your previous "
f"response was cut short — check what you were doing and continue. "
f"FORMATTING: User reads on mobile (Telegram/Matrix Element). "
f"NEVER use markdown tables — they render as broken text on mobile. "
f"Prefer bullet lists, bold headers, numbered lists to structure data. "
f"Small tables (2-4 cols, few rows): use monospace code block with aligned columns. "
f"Large/complex tables: generate HTML, convert to PDF via "
f"`html-to-pdf input.html output.pdf`, send via send-to-user. "
f"Do NOT use wkhtmltopdf — its PDFs are broken on iOS. "
f"SCREENSHOTS: `screenshot-page <url-or-file> output.png [--width 1280] [--height 900] "
f"[--wait 3] [--full-page] [--stealth]`. Works with URLs and local HTML files (folium maps etc). "
f"IMAGE SEARCH: `search-images \"query\" -o dir/ -n 4 -p prefix [--size large] "
f"[--orient horizontal]`. Uses Yandex Image Search API. Downloads images automatically. "
f"Add --no-download to just list URLs. "
f"WEB SEARCH: `search-web \"query\" [-n 10] [--lang ru]`. Yandex web search — "
f"best for Russian-language queries. Returns titles, URLs, snippets. "
f"Use for research, reviews, travel tips, local info. Lang: ru (default), en, tr. "
f"SENDING FILES: To send files to the user, use: `send-to-user <path> [caption]`. "
f"It is in PATH. The file will be delivered after your response. "
f"ASKING USER: To ask the user a question and wait for their reply, use: "
f"`ask-user \"your question\"`. It blocks until the user responds via the chat. "
f"IMAGE GENERATION: Use `generate-image` (NanoBanana/Gemini 3 Pro). "
f"It supports multi-turn chat for iterative refinement of images. "
f"First generation: `generate-image \"prompt\" output.png --chat history.json [-a 16:9]`. "
f"Refinement (edits the PREVIOUS image): `generate-image --chat history.json --refine \"change X to Y\" output2.png`. "
f"The --chat flag saves conversation context so the model remembers what it generated. "
f"ALWAYS use --chat with a history file in the current dir so you can refine later. "
f"The model can modify its own previous output when you use --refine — "
f"it does NOT generate from scratch, it edits the existing image. "
f"You can also pass reference images (up to 14): `generate-image \"prompt\" out.png --chat h.json --ref photo.jpg --ref photo2.jpg`. "
f"Aspect ratios: 9:16, 16:9, 1:1, 4:3, 3:4. Sizes: 1K, 2K, 4K (default). "
f"THREAD VISIBILITY: Your response is posted in a Matrix thread. "
f"The user sees ONLY the final message at a glance — intermediate tool output "
f"and thread messages are hidden unless expanded. "
f"All text the user needs to read MUST be in your response message, not only in files. "
f"Writing to files for persistence is fine, but the conversation text — "
f"analysis, notes, discussion points — must appear in the response itself. "
f"The user is chatting with you, not reading files. "
f"IMAGES IN CONTEXT: When conversation history contains entries like "
f"'[image: /path/to/file.png]', these are actual image files on disk. "
f"Use the Read tool to view them — they contain photos, screenshots, or book pages "
f"that the user shared. Always review referenced images before responding about them. "
f"TOOL DISCOVERY: Before installing packages or writing scripts, check what tools "
f"are already available. Common tools in PATH: transcribe-audio, send-to-user, "
f"ask-user, search-web, search-images, screenshot-page, generate-image, html-to-pdf, browser. "
f"BROWSER: If BROWSER_CDP_URL is set, you have access to a real Chrome browser via "
f"`browser <command>`. Commands: navigate <url>, screenshot [file], click <selector>, "
f"type <selector> <text>, read [selector], eval <js>, tabs, new [url], close. "
f"Use this for web interaction, authenticated sites, downloads, form filling. "
f"Run `ls /opt/agent-core/common-tools/` to see all. "
f"Prefer existing tools over writing new code."
f"{user_context}"
f"{workspace_context}"
f"{conv_context}"
)
claude_args = [
cmd,
*session_flag,
"-p",
"--verbose",
"--output-format", "stream-json",
"--append-system-prompt", system_extra,
"--allowedTools", ",".join(config.allowed_tools),
"--max-turns", "50",
]
if model_override:
claude_args.extend(["--model", model_override])
claude_args.append(message)
# Wrap with bwrap if available
bwrap_path = Path(__file__).resolve().parent.parent / "bwrap-claude"
if bwrap_path.exists() and shutil.which("bwrap"):
args = [str(bwrap_path)] + claude_args
else:
args = claude_args
# Build clean environment for Claude subprocess
_strip_prefixes = ("CLAUDECODE", "CLAUDE_CODE")
_strip_keys = {
"BOT_TOKEN", "MATRIX_ACCESS_TOKEN", "MATRIX_HOMESERVER",
"MATRIX_USER_ID", "MATRIX_OWNER_MXID", "MATRIX_DEVICE_ID",
}
# Auth env vars that must pass through to Claude CLI
_passthrough_keys = {"CLAUDE_CODE_OAUTH_TOKEN"}
env = {
k: v for k, v in os.environ.items()
if k in _passthrough_keys
or (not any(k.startswith(p) for p in _strip_prefixes) and k not in _strip_keys)
}
# Add common-tools to PATH so Claude can use send-to-user, generate-image, etc.
common_tools = str(Path(__file__).resolve().parent.parent / "common-tools")
env["PATH"] = common_tools + ":" + env.get("PATH", "")
# Load per-user workspace .env (Readest keys, Linkwarden keys, etc.)
if workspace_dir:
ws_env = workspace_dir / ".env"
if ws_env.exists():
for line in ws_env.read_text().splitlines():
line = line.strip()
if line and not line.startswith("#") and "=" in line:
key, _, val = line.partition("=")
env[key.strip()] = val.strip().strip("'\"") # handle KEY="value" and KEY='value'
session_label = existing_session[:8] if existing_session else f"new:{new_id[:8]}"
logger.info("Claude CLI: topic=%s session=%s cmd=%s", topic_id, session_label, cmd)
proc = await asyncio.create_subprocess_exec(
*args,
stdout=asyncio.subprocess.PIPE,
stderr=asyncio.subprocess.PIPE,
cwd=str(topic_dir),
env=env,
limit=10 * 1024 * 1024, # 10MB — stream-json lines can be huge (base64 images)
)
response_parts: list[str] = []
full_text = ""
result_text = "" # clean final response from result event
result_session_id = None
timeout_reason = None
# Tool tracking for status events
block_tools: dict[str, str] = {} # tool_use_id -> tool name
# Idle timeout state — mutable so watchdog can read, user can extend
idle_timeout = idle_timeout_ref if idle_timeout_ref is not None else [config.claude_idle_timeout]
last_activity = [time.monotonic()]
start_time = time.monotonic()
# Start question watcher if callback provided
question_task = None
if on_question:
question_task = asyncio.create_task(_watch_questions(topic_dir, on_question))
# Watchdog: checks idle timeout, hard timeout, and cancel
async def _watchdog():
nonlocal timeout_reason
while True:
await asyncio.sleep(2)
now = time.monotonic()
if cancel_event and cancel_event.is_set():
timeout_reason = "cancelled"
proc.kill()
return
idle = now - last_activity[0]
if idle > idle_timeout[0]:
timeout_reason = "idle"
proc.kill()
return
elapsed = now - start_time
if elapsed > config.claude_max_timeout:
timeout_reason = "max"
proc.kill()
return
watchdog_task = asyncio.create_task(_watchdog())
# Stream log — save all events from Claude CLI for debugging/replay
stream_log_path = topic_dir / "stream.jsonl"
stream_log = open(stream_log_path, "a")
try:
async for line in proc.stdout:
last_activity[0] = time.monotonic() # reset idle timer on ANY output
line = line.decode("utf-8", errors="replace").strip()
if not line:
continue
# Log raw event to stream.jsonl
stream_log.write(line + "\n")
stream_log.flush()
try:
event = json.loads(line)
except json.JSONDecodeError:
logger.debug("Non-JSON stdout: %s", line[:200])
continue
etype = event.get("type")
# Capture session_id from init or result events
if etype == "system" and event.get("session_id"):
result_session_id = event["session_id"]
elif etype == "result" and event.get("session_id"):
result_session_id = event["session_id"]
# Handle result events — this has the clean final response
if etype == "result":
if event.get("is_error"):
errors = event.get("errors", [])
logger.error("Claude CLI error: %s", "; ".join(errors))
if event.get("result"):
result_text = event["result"]
# --- Status events from stream-json ---
# Claude CLI emits full "assistant" snapshots (with tool_use blocks)
# followed by "user" events (with tool_result).
if etype == "assistant":
content = event.get("message", {}).get("content", [])
has_tools = any(b.get("type") == "tool_use" for b in content)
for block in content:
if block.get("type") == "tool_use" and on_status:
tool_name = block.get("name", "")
tool_id = block.get("id", "")
inp = block.get("input", {})
preview = _tool_preview(tool_name, json.dumps(inp, ensure_ascii=False))
if tool_id:
block_tools[tool_id] = tool_name
if tool_name == "Agent":
desc = inp.get("description", "")
bg = inp.get("run_in_background", False)
await on_status({
"event": "agent_start",
"description": desc,
"background": bg,
})
else:
await on_status({
"event": "tool_start",
"tool": tool_name,
"input_preview": preview,
})
# All assistant text goes to thread as narration.
# Only result.result is the final clean response.
if block.get("type") == "text" and block.get("text"):
text = block["text"]
if on_status:
await on_status({
"event": "thinking",
"text": text,
})
# Also accumulate for on_chunk (Telegram streaming)
response_parts.append(text)
full_text = "".join(response_parts)
if on_chunk:
await on_chunk(full_text)
# Tool results mark tool completion
if etype == "user" and on_status:
content = event.get("message", {}).get("content", [])
if isinstance(content, list):
for block in content:
if isinstance(block, dict) and block.get("type") == "tool_result":
tool_id = block.get("tool_use_id", "")
tool_name = block_tools.pop(tool_id, "tool")
await on_status({"event": "tool_end", "tool": tool_name})
# Check if watchdog killed the process
if watchdog_task.done():
break
await proc.wait()
except Exception:
if not watchdog_task.done():
watchdog_task.cancel()
raise
finally:
stream_log.close()
if not watchdog_task.done():
watchdog_task.cancel()
try:
await watchdog_task
except asyncio.CancelledError:
pass
if question_task:
question_task.cancel()
try:
await question_task
except asyncio.CancelledError:
pass
elapsed = int(time.monotonic() - start_time)
# Handle timeout/cancel
if timeout_reason:
await proc.wait()
if timeout_reason == "cancelled":
logger.info("Claude CLI cancelled by user after %ds", elapsed)
suffix = "\n\n[cancelled by user]"
elif timeout_reason == "idle":
logger.warning("Claude CLI idle timeout after %ds (idle limit: %ds)", elapsed, idle_timeout[0])
suffix = f"\n\n[idle timeout — no output for {idle_timeout[0]}s]"
else:
logger.error("Claude CLI hard timeout after %ds (max: %ds)", elapsed, config.claude_max_timeout)
suffix = f"\n\n[timeout — {elapsed}s elapsed]"
# Save session even on timeout — don't lose conversation history
if result_session_id:
save_session(config.data_dir, topic_id, result_session_id, provider)
# On timeout: prefer result_text (clean), fall back to full_text (has thinking)
response = result_text or full_text
error_patterns = ["Failed to authenticate", "API Error:", "authentication_error", "401"]
if response and not any(p in response for p in error_patterns):
return response + suffix
raise RuntimeError(f"Claude CLI {timeout_reason} after {elapsed}s (error response: {full_text[:100]})")
# Save session ID for future resume
if result_session_id:
save_session(config.data_dir, topic_id, result_session_id, provider)
# Check for error responses (auth failures, API errors) - these should trigger fallback
error_patterns = ["Failed to authenticate", "API Error:", "authentication_error", "401"]
is_error_response = any(p in full_text for p in error_patterns)
if proc.returncode != 0 or is_error_response:
stderr = await proc.stderr.read()
stderr_text = stderr.decode("utf-8", errors="replace").strip()
logger.error("Claude CLI failed (rc=%d): %s", proc.returncode, stderr_text[:500])
if is_error_response:
raise RuntimeError(f"Claude CLI returned error: {full_text[:200]}")
response = result_text or full_text
if response:
return response
# Non-auth failure with no output — raise to trigger fallback
# but preserve session file (conversation history is valuable)
raise RuntimeError(f"Claude CLI exited with code {proc.returncode}")
response = result_text or full_text
if not response and _retry_count < 1:
logger.warning("Claude CLI returned empty response, retrying (attempt %d)", _retry_count + 1)
return await _send_with_provider(
config, topic_id, message, on_chunk, on_question,
on_status=on_status, cancel_event=cancel_event,
idle_timeout_ref=idle_timeout_ref,
provider=provider, cmd_override=cmd_override, model_override=model_override,
user_profile=user_profile, workspace_dir=workspace_dir,
_retry_count=_retry_count + 1,
)
return response or "(no response)"
def _extract_text(event: dict) -> str | None:
"""Extract text content from a stream-json event."""
etype = event.get("type")
if etype == "assistant":
content = event.get("message", {}).get("content", [])
texts = []
for block in content:
if block.get("type") == "text":
texts.append(block.get("text", ""))
return "".join(texts) if texts else None
if etype == "content_block_delta":
delta = event.get("delta", {})
if delta.get("type") == "text_delta":
return delta.get("text", "")
# Don't extract from "result" — it duplicates what was already
# streamed via "assistant" events. The caller uses it as fallback
# only if full_text is empty after processing all events.
return None