maintainer edit: remove the _last_compacted_at maintenance state, gate idle compaction on whether a session still has a removable tail, and sort WebUI sessions by the latest visible transcript activity.
MCP server URLs can carry secrets in userinfo
(`https://user:token@host/sse`) or a query string (`?token=...`). A few
connect/validate paths logged the raw `cfg.url` / `request.url`, so those
secrets could land in log files that are often shared or aggregated.
Add a small `_redact_url()` helper that keeps only scheme/host/port/path
and use it at the four sites that log a server or request URL. Logging
only; no other behavior changes.
Replace the tuple(list[dict], int) return of
Session.retain_recent_legal_suffix with a named RetentionResult
dataclass that exposes retained, dropped,
already_consolidated_count, and new_last_consolidated fields.
The tuple return was easy to misuse because the second value only
made sense relative to the first and the old last_consolidated
cursor. The named fields make the archive-skip semantics explicit
at every call site.
No behavior change. All existing tests pass unchanged in semantics.
Refs #4136
Signed-off-by: axelray-dev <110029405+axelray-dev@users.noreply.github.com>
Treat max_messages as a last-resort replay guard now that consolidation and idle auto-compact own normal history reduction. Raising the default avoids frequent sliding-window prefix churn in moderate conversations without adding a new cache-policy knob.
Combine malformed tool-call handling with placeholder filtering and a
no-tools fallback so a relay that returns tool_use blocks with null
id/name/input can no longer crash a turn or permanently wedge a session.
Adapted to the ContextGovernor architecture (context governance now lives
in nanobot/agent/context_governance.py, not runner.py):
- ToolCallRequest.has_valid_name(): single source of truth for "usable
name" (non-empty string).
- tool_hints.format_tool_hints(): skip tool calls with a non-string/empty
name instead of raising AttributeError on the whole turn.
- ContextGovernor.strip_placeholder_assistant_messages() and
strip_malformed_tool_calls() (plus the _tool_call_name_is_valid helper):
history-cleaning staticmethods invoked at the START of
prepare_for_model() — strip_placeholder, then strip_malformed, then the
existing drop_orphan/backfill chain. Both only repair the model-facing
copy and leave persisted history untouched (return a copy, or the same
list when nothing changes). Also wired into runner's minimal-repair path.
- AgentRunner._drop_malformed_tool_calls(): returns
(dropped, all_dropped, original_finish_reason); clears finish_reason to
"stop" when all calls are dropped.
- AgentRunner._malformed_tool_call_retry_messages() + _request_model
malformed_retry flag: when an all-dropped tool_calls response comes back,
retry once with a corrective note; if the retry STILL comes back
all-dropped, fall back to _request_no_tools for graceful text degradation.
Tests for the history-cleaning methods live with ContextGovernor in
tests/agent/test_runner_governance.py; response-layer and tool-hint tests
stay on AgentRunner / tool_hints.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
MCPToolWrapper.execute only handled TextContent; every other block was
rendered with str(block). An MCP ImageContent block therefore became a
large base64 string embedded in the tool result, which (a) was truncated
by max_tool_result_chars, corrupting the data, and (b) could never reach a
channel because it was plain text, not an image artifact.
Decode ImageContent (and EmbeddedResource blobs with an image/* MIME type)
and persist them via store_generated_image_artifact, returning the same
compact {artifacts, next_step} JSON the built-in image_generation tool
produces. The base64 stays out of the model context; the model delivers the
saved file via the message tool's media parameter.
- Removes match_text regex that stripped # comments before pattern matching
(broke on quoted # inside strings)
- allow_patterns now run re.fullmatch against the full lowercased command
- deny_patterns search the original lowercased command
- Replaces comment-stripping test with comment-tail bypass regression
(touch canary # echo allowlisted must be blocked)
- Adds Re-bin regression for quoted hash + blocked command
(echo "#" followed by blocked command must be caught)
- All 10 tests pass
Signed-off-by: axelray-dev <110029405+axelray-dev@users.noreply.github.com>
Extract model-facing context governance from AgentRunner.
Only compact in-flight tool results when the model request is over budget, keep compacted IDs stable within a turn, and allow the newest result to be compacted as a last resort when it is the remaining source of overflow.
maintainer edit: add a regression test for the public ExecTool.execute path so omitted login stays non-login by default, and update the Unix environment docstring to match the new explicit login behavior.
The exec tool defaults login=True for bash/zsh, which causes the shell
to source ~/.bash_profile and similar startup files. This reintroduces
secrets from shell startup files into the exec environment, even though
_build_env() intentionally starts with a curated environment.
Change the default to login=False in both _prepare_command() and _spawn(),
and update the schema default accordingly.
Add fail_on_tool_error to AgentDefaults and wire it through
AgentLoop -> SubagentManager -> AgentRunSpec.
Previously hardcoded to True in SubagentManager._run_subagent.
Now configurable via config.json with default True for backward
compatibility. When set to False, subagents can retry on minor
tool errors instead of immediately failing.
Changes:
- nanobot/config/schema.py: add fail_on_tool_error field (default True)
- nanobot/agent/subagent.py: accept and forward fail_on_tool_error
- nanobot/agent/loop.py: pass config through to SubagentManager
- tests/agent/test_subagent.py: add regression test
Signed-off-by: axelray-dev <110029405+axelray-dev@users.noreply.github.com>
When dream.enabled is false, the Dream cron job never runs, so the
dream cursor (.dream_cursor) stays at its initial value (0). This
causes read_recent_history_for_prompt() to treat every history entry
as unprocessed, injecting the full chat history into every system
prompt and growing without bound.
Fix: fast-forward the dream cursor to the latest history entry at
gateway startup when Dream is disabled.
The onboard wizard dispatched field handlers by bare field name, so
WebSearchConfig.provider was hijacked by the LLM-provider handler and
showed LLM providers instead of search engines. Keenable was also never
wired into the CLI wizard when it landed in the WebUI.
- Add a single source of truth for selectable search providers
(SEARCH_PROVIDER_OPTIONS in web.py); WebUI settings now import it.
- Add a WebSearchConfig-aware search-provider picker to the wizard and
resolve handlers by (model type, field name) so the LLM and search
provider fields no longer collide.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The enabledTools allowlist was only enforced for MCP tools returned by
session.list_tools(). Resources and prompts from session.list_resources()
and session.list_prompts() were registered unconditionally, allowing a
deny-all or restrictive enabledTools config to leak resource and prompt
capabilities to the model.
Now resources and prompts are only registered when allow_all_tools is
true (default ["*"] wildcard). Any explicit tool restriction — including
enabledTools: [] (deny-all) or a list of specific tool names — also
blocks resource and prompt registration from that server.
Fixes#4435
The DuckDuckGo search provider instantiated DDGS(timeout=10) without
passing the configured proxy, making web_search unusable in environments
that require a proxy (e.g. behind GFW). DDGS supports a proxy parameter
and the proxy value is already available as self.proxy — it was simply
not forwarded.
Add a test verifying the proxy kwarg is forwarded to DDGS.
maintainer edit: native reasoning streams can split <thinking> wrapper tags across chunks. Buffer the stream and emit only cleaned incremental reasoning so raw partial tags do not reach WebUI.
maintainer edit: move duplicate tool_use history repair out of AgentRunner and into Anthropic message conversion, reusing the OpenAI-compatible queue-mapping approach locally without broadening the shared runner path.
maintainer edit: remap duplicate tool_use/tool_call ids instead of dropping later calls, so Anthropic-compatible providers that reuse ids for distinct parallel tool calls keep all requested work while still sending unique ids.
Anthropic rejects any request where two tool_use blocks share an id
("messages.N.content.M: tool_use ids must be unique"). A mis-assembled
stream could surface the same tool_use block twice in one assistant turn;
the runner persisted it verbatim, so the malformed message was re-sent on
every subsequent turn and permanently bricked the session — the agent
silently stopped replying.
Fix at two layers:
- AnthropicProvider._parse_response: drop duplicate tool_use ids (keep
first) as the response enters nanobot, so corruption is never persisted.
- AgentRunner._dedup_tool_calls: a new context-governance pass that dedupes
assistant tool_calls and tool results by id before each send, healing any
history that was already corrupted.
Add regression tests covering both the dedup and the no-op fast path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Document that SDK callers always go through run() → process_direct(ephemeral=False),
so per_call hooks are intentionally excluded from ephemeral turns per chengyongru
review feedback.
The test_run_populates_tools_used_across_iterations,
test_run_populates_final_messages, and
test_run_user_hooks_still_fire_alongside_capture tests were reading
bot._loop._extra_hooks directly in their fake_process_direct mocks,
but the contextvar change moved per-call hooks to _per_call_hooks.
Also fixes the ephemeral guard: per_call hooks should not be used in
ephemeral mode (the original code incorrectly applied them regardless).
Nanobot.run() previously mutated the shared self._loop._extra_hooks
attribute under a try/finally. When two run() calls with different
session_keys execute concurrently, they overwrite each other's hook
lists — the second call saves the first call's hooks as 'prev', and
the finally block restores stale state.
Use a contextvars.ContextVar instead, which is per-task in asyncio.
This gives each concurrent run() call its own hook list without
changing any function signatures. Falls back to self._extra_hooks
when no per-call hooks are set (non-SDK usage).
Match the _BOCHA_SEARCH_API_URL / _VOLCENGINE_SEARCH_API_URL convention
instead of hardcoding the URL inline.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Keenable's public endpoint serves the free tier (1000 req/hour) without
auth. Route to /v1/search/public with the X-Keenable-Title header when no
key is configured, instead of falling back to DuckDuckGo; keep the
authenticated /v1/search path when an apiKey or KEENABLE_API_KEY is set.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
timeout=None causes the httpx client to wait indefinitely when connecting
to MCP servers via streamableHttp transport, blocking the entire event loop.
This prevents Telegram polling and other async operations from running.
Set timeout=httpx.Timeout(30.0, connect=10.0) to match the pattern used
in the SSE transport's httpx_client_factory.
Manual testing against the live API showed the Keenable REST endpoint
(/v1/search) returns 401 without a key — the keyless "free tier" applies
only to the CLI, not the HTTP API. Treat Keenable like every other
key-based provider: fall back to DuckDuckGo when no key is configured,
and drop the now-inaccurate free-tier wording from the docs.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
chengyongru
chengyongru
Xubin Ren
xiaweiwei67-stack
axelray-dev
wangjunweiandClaude Opus 4.8
codedragon
Ilya GusevandClaude Opus 4.8
michaelxer
hyoukadev
Zhou
Teddy Yan
Teddy YanandClaude Opus 4.8
yu-xin-c
Luc Nguyen