Fix issue #1823: Memory consolidation does not inherit agent temperature
and maxTokens configuration.
The agent's configured generation parameters were not being passed through
to the memory consolidation call, causing it to fall back to default values.
This resulted in the consolidation response being truncated before the
save_memory tool call was emitted.
- Pass temperature, max_tokens, reasoning_effort from AgentLoop to
MemoryConsolidator and then to MemoryStore.consolidate()
- Forward these parameters to the provider.chat_with_retry() call
Fixes#1823
Share assistant message construction between the main agent and subagents, and add a regression test to keep reasoning_content and thinking_blocks in follow-up tool rounds.
Move consolidation policy into MemoryConsolidator, keep backward compatibility for legacy config, and compress history by token budget instead of message count.
Subagent's _run_subagent() was dropping reasoning_content and
thinking_blocks when building assistant messages for the conversation
history. Providers like Deepseek Reasoner require reasoning_content on
every assistant message when thinking mode is active, causing a 400
BadRequestError on the second LLM round-trip.
Align with the main AgentLoop which already preserves these fields via
ContextBuilder.add_assistant_message().
Closes#1834
MCP SDK's anyio cancel scopes can leak CancelledError on timeout or
failure paths. Since CancelledError is a BaseException (not Exception),
it escapes both MCPToolWrapper.execute() and ToolRegistry.execute(),
crashing the agent loop.
Now catches CancelledError and returns a graceful error to the LLM,
while still re-raising genuine task cancellations from /stop.
Also catches general Exception for other MCP failures (connection
drops, invalid responses, etc.).
Related: #1055
Major changes:
- Replace message-count-based memory window with token-budget-based compression
- Add max_tokens_input, compression_start_ratio, compression_target_ratio config
- Implement _maybe_compress_history() that triggers based on prompt token usage
- Use _build_compressed_history_view() to provide compressed history to LLM
- Refactor MemoryStore.consolidate() -> consolidate_chunk() for chunk-based compression
- Remove last_consolidated from Session, use _compressed_until metadata instead
- Add background compression scheduling to avoid blocking message processing
Key improvements:
- Compression now based on actual token usage, not arbitrary message counts
- Better handling of long conversations with large context windows
- Non-destructive compression: old messages remain in session, but excluded from prompt
- Automatic compression when history exceeds configured token thresholds
Some LLM providers return tool_calls[0].arguments as a list instead of
dict or str. Add handling to extract the first dict element from the list.
Fixes /new command warning: 'unexpected arguments type list'
Feishu downloads images with incorrect extensions (e.g. .jpg for PNG files).
mimetypes.guess_type() relies on the file extension, causing a MIME mismatch
that Anthropic rejects with 'image was specified using image/jpeg but appears
to be image/png'.
Fix: read the first bytes of the image data and detect the real MIME type via
magic bytes (PNG: 0x89PNG, JPEG: 0xFFD8FF, GIF: GIF87a/GIF89a, WEBP: RIFF+WEBP).
Fall back to mimetypes.guess_type() only when magic bytes are inconclusive.
provider.chat() had no retry logic — a transient 429 rate limit,
502 gateway error, or network timeout would permanently fail the
entire message. For a system running cron jobs and heartbeats 24/7,
even a brief provider blip causes lost tasks.
Adds _chat_with_retry() that:
- Retries up to 3 times with 1s/2s/4s exponential backoff
- Only retries transient errors (429, 5xx, timeout, connection)
- Returns immediately on permanent errors (400, 401, etc.)
- Falls through to the final attempt if all retries exhaust
ReadFileTool had no file size check — reading a multi-GB file would
load everything into memory and crash the process. Now:
- Rejects files over ~512KB at the byte level (fast stat check)
- Truncates at 128K chars with a notice if content is too long
- Guides the agent to use exec with head/tail/grep for large files
This matches the protection already in ExecTool (10KB) and
WebFetchTool (50KB).
datetime.fromisoformat(at) raises ValueError for malformed strings,
which propagated uncaught and crashed the tool execution. Now catches
ValueError and returns a user-friendly error message instead.
Re-bin
WhalerO
ethanclaw
lailoo
suger-m
Alfredo ArenasandGitHub
VITOHJL
Barry WangandGitHub
coldxiangyu
Sergio Sánchez Vallés
Xubin RenandGitHub
Kiplangatkorir