Trigger token consolidation before prompt usage reaches the full context window so response tokens and tokenizer estimation drift still fit safely within the model budget.
Made-with: Cursor
- Add strip_think() to helpers.py as single source of truth
- Filter deltas in agent loop before dispatching to consumers
- Implement send_delta in TelegramChannel with progressive edit_message_text
- Remove duplicate think filtering from CLI stream.py and telegram.py
- Remove legacy fake streaming (send_message_draft) from Telegram
- Default Telegram streaming to true
- Update CHANNEL_PLUGIN_GUIDE.md with streaming documentation
Made-with: Cursor
Move ThinkingSpinner and StreamRenderer into a dedicated module to keep
commands.py focused on orchestration. Uses Rich Live with manual refresh
(auto_refresh=False) and ellipsis overflow for stable streaming output.
Made-with: Cursor
Preserve the provider and agent-loop streaming primitives plus the CLI experiment scaffolding so this work can be resumed later without blocking urgent bug fixes on main.
Made-with: Cursor
Resolve conflict in context.py: keep main's build_messages which already
merges runtime context into user message (achieving the same cache goal).
The real value-add from this PR is the second cache breakpoint in
litellm_provider.py.
Made-with: Cursor
estimate_prompt_tokens() only counted the `content` text field, completely
missing tool_calls JSON (~72% of actual payload), reasoning_content,
tool_call_id, name, and per-message framing overhead. This caused the
memory consolidator to never trigger for tool-heavy sessions (e.g. cron
jobs), leading to context window overflow errors from the LLM provider.
Also adds reasoning_content counting and proper per-message overhead to
estimate_message_tokens() for consistent boundary detection.
Made-with: Cursor
Merge process_direct() and process_direct_outbound() into a single
interface returning OutboundMessage | None. This eliminates the
dual-path detection logic in CLI single-message mode that relied on
inspect.iscoroutinefunction to distinguish between the two APIs.
Extract status rendering into a pure function build_status_content()
in utils/helpers.py, decoupling it from AgentLoop internals.
Made-with: Cursor
Only use process_direct_outbound when the agent loop actually exposes it as an async method, and otherwise fall back to the legacy process_direct path. This keeps the new CLI render-metadata flow without breaking existing test doubles or older direct-call implementations.
Made-with: Cursor
Keep status output responsive while estimating current context from session history, dropping low-value queue/subagent counters, and marking command-style replies for plain-text rendering in CLI. Also route direct CLI calls through outbound metadata so help/status formatting stays explicit instead of relying on content heuristics.
Made-with: Cursor
Handle /status at the run-loop level so it can return immediately while the agent is busy, and reset last-usage stats when providers omit usage data. Also keep Telegram help/menu coverage for /status without changing the existing final-response send path.
Made-with: Cursor
Only normalize nullable MCP tool schemas for OpenAI-compatible providers so optional params still work without collapsing unrelated unions. Also teach local validation to honor nullable flags and add regression coverage for nullable and non-nullable schemas.
Made-with: Cursor
Add native image content blocks for read_file and web_fetch, preserve the multimodal tool-result path through the agent loop, and keep session history compact with image placeholders. Also harden web_fetch against redirect-based SSRF bypasses and add regression coverage for image reads and blocked private redirects.
Keep multimodal tool outputs on the native content-block path while
restoring redirect SSRF checks for web_fetch image responses. Also share
image block construction, simplify persisted history sanitization, and
add regression tests for image reads and blocked private redirects.
Made-with: Cursor
Exclude openai_codex alongside github_copilot from generated config,
filter OAuth-only providers out of the onboarding wizard, and clarify in
README that OAuth login stores session state outside config. Also unify
the GitHub Copilot login command spelling and add regression tests.
Made-with: Cursor
Keep messages already collected in the current poll cycle when a stale
IMAP connection dies mid-fetch, so retrying once does not drop emails
that were already parsed and marked seen. Add a regression test covering
a mid-cycle disconnect after the first message succeeds.
Made-with: Cursor
Record run_at_ms, status, duration_ms and error for each execution,
keeping the last 20 entries per job in jobs.json. Adds CronRunRecord
dataclass, get_job() lookup, and four regression tests covering
success, error, trimming and persistence.
Closes#1837
Made-with: Cursor
- wire tools.exec.enable and deny_patterns into the current AgentLoop
- preserve the current WebSearchTool config-based registration path
- treat deny_patterns=[] as an explicit override instead of falling back
to the default blacklist
- add regression coverage for disabled exec registration and custom deny
patterns
Made-with: Cursor
- Mask sensitive fields (api_key/token/secret/password) in all display
surfaces, showing only the last 4 characters
- Replace all emoji with pure ASCII labels for consistent cross-platform
terminal rendering
- Extract _print_summary_panel helper, eliminating 5x duplicate table
construction in _show_summary
- Replace 3 one-line wrapper functions with declarative _SETTINGS_SECTIONS
dispatch tables and _MENU_DISPATCH in run_onboard
- Extract _handle_model_field / _handle_context_window_field into a
_FIELD_HANDLERS registry, shrinking _configure_pydantic_model
- Return FieldTypeInfo NamedTuple from _get_field_type_info for clarity
- Replace global mutable _PROVIDER_INFO / _CHANNEL_INFO with @lru_cache
- Use vars() instead of dir() in _get_channel_info for reliable config
class discovery
- Defer litellm import in model_info.py so non-wizard CLI paths stay fast
- Clarify README Quick Start wording (Add -> Configure)
Adjust the spawn tool description to keep the workspace-organizing hint while
avoiding language that sounds like the system automatically assigns a dedicated
working directory for subagents.
Made-with: Cursor
The root cause of "Pool timeout" errors is that long-polling (getUpdates)
and outbound API calls (send_message, send_photo, etc.) shared the same
HTTPXRequest pool — polling holds connections indefinitely, starving sends
under concurrent load (e.g. cron jobs + user chat).
- Split into two independent pools: API calls (default 32) and polling (4)
- Expose connection_pool_size / pool_timeout in TelegramConfig for tuning
- Add _call_with_retry() with exponential backoff (3 attempts) on TimedOut
- Apply retry to _send_text and remote media URL sends
Replace fire-and-forget consolidation with archive_messages(), which
retries until the raw-dump fallback triggers — making it effectively
infallible. /new now clears the session immediately and archives in
the background. Pending archive tasks are drained on shutdown via
close_mcp() so no data is lost on process exit.
Instead of adding a separate load_skill tool to bypass workspace restrictions,
extend ReadFileTool with extra_allowed_dirs so it can read builtin skill paths
while keeping write/edit tools locked to the workspace. Fixes the original issue
for both main agent and subagents.
Made-with: Cursor
Replace the static provider-level supports_vision check with a
reactive fallback: when a model returns an image-unsupported error,
strip image_url blocks from messages and retry once. This avoids
maintaining an inaccurate vision capability table and correctly
handles gateway/unknown model scenarios.
Also extract _safe_chat() to deduplicate try/except boilerplate
in chat_with_retry().
With custom_llm_provider kwarg handling routing, the openrouter/ prefix
caused model names like anthropic/claude-sonnet-4-6 to become
openrouter/anthropic/claude-sonnet-4-6, which OpenRouter API rejects.
- Add nanobot/utils/evaluator.py: lightweight LLM tool-call to decide notify/silent after background task execution
- Remove magic token injection from heartbeat and cron prompts
- Clean session history (no more <SILENT_OK> pollution)
- Add tests for evaluator and updated heartbeat three-phase flow
- Require both history_entry and memory_update, reject null/empty values
- Fallback to tool_choice=auto when provider rejects forced function call
- After 3 consecutive consolidation failures, raw-archive messages to
HISTORY.md without LLM summarization to prevent context window overflow
Some providers (e.g. Dashscope in thinking mode) reject object-style
tool_choice with "does not support being set to required or object".
Retry once with tool_choice="auto" instead of failing silently.
Made-with: Cursor
Xubin RenandXubin Ren