Drop the PluginToolLoader.get_tool() override that returned a raw
ComponentManifest, so every loader's get_tool() now returns a uniform
resource_tool.LLMTool (PluginToolLoader.get_tools() already did this
conversion). This removes the only source of tool-shape heterogeneity.
- ToolManager.get_tool_schema(): drop the ComponentManifest-vs-LLMTool branch
- ToolManager.get_tool_detail(): new host-level shape {name, description,
human_desc, parameters}
- handler.py GET_TOOL_DETAIL: call tool_mgr.get_tool_detail(); delete the
handler-local _build_tool_detail + _i18n_to_dict/_i18n_to_text adapters and
the litellm TODO
- ToolLookupResult is now just LLMTool
The dropped label/spec fields were not consumed by any runner (local-agent
build_llm_tool and external harnesses use only name/description/parameters).
Extract the AgentRunner Protocol v1 host-side surface from the giant
RuntimeConnectionHandler.__init__ into sibling modules using a registration-
function pattern (behavior-preserving; @h.action == @self.action):
- agent_run_support.py: shared constants + authorization/scope/projection helpers
- agent_pull_actions.py: register(h) for history/event pull APIs
- agent_runner_actions.py: register(h) for run/runtime/stats/claim lifecycle
- agent_state_actions.py: register(h) for steering/state APIs
__init__ now calls the three register(self) functions. handler.py keeps the
pre-existing plugin/llm/vector/knowledge handlers, get_prompt/call_tool/
get_tool_detail (coupled to retained helpers), shared helpers, and outbound
methods; it re-imports _validate_agent_run_session so external imports keep
working. handler.py: 4066 -> 1871 lines.
test_state_api_auth.py: repoint get_session_registry patch targets to
agent_run_support (the lookup moved modules). 385 agent unit tests pass; ruff clean.
- nsjail: full create→exec→register→activate→exec-from-activated-path chain
returns exit 0; activated mount runs scripts/use.py (reads data/input.json)
and writes activated_writeback.txt through to the host skill store.
- docker: same chain now passes after langbot-plugin-sdk#87 (recreate sandbox
container when extra_mounts change). Corrected #2271 root cause from
'docker masks nested bind mount' to container-reuse: extra_mounts was not in
the box session compatibility check, so docker reused a running container and
could not append the activated skill's bind mount.
- Exit criterion 3 (real end-to-end skill use) now DONE; all 5 criteria met.
- Documents the nsjail stale-docker-artifact environment gotcha.
Prior matrix recorded acp as blocked needing langbot-assets-gateway-public-url
(PROBEDONE 0 0 / timeout). That was an environment artifact: a duplicate
LangBot-master/ backend contending on box ws-control-port 5410 plus a wedged
plugin runtime (host emit_event / list_agent_runners timing out). On a clean
single-instance runtime acp discovers skills via the SDK SSH reverse tunnel
with no public-url: PROBEDONE 1 17 (8-24s), parity with claude-code (1 15).
- claude-code-agent (new pipeline, remote-ssh->101): langbot_list_assets returns
skills=1 tools=15 in 24s -> all-tool 'skills' asset class is discoverable
end-to-end by an external harness on the unmodified branch
- document the runner transport difference: claude-code uses a stdio bridge
(works on remote-ssh out of the box), acp uses an HTTP proxy (needs
langbot-assets-gateway-public-url on remote-ssh). This is a runner-plugin
detail, not a host all-tool-branch issue
- references/skill-all-tool-acceptance.md: acceptance matrix for the skill
all-tool model (runner x lifecycle x backend), case status, exit criteria,
and the #2271 known issue (pre-existing box nested-mount, not this branch)
- cases/skill-discovery-via-mcp-gateway.yaml: schema-valid case proving an
external harness discovers skills via langbot_list_assets (the new 'skills'
asset class); marked blocked-env until remote claude-code is responsive
Expose skill tools (activate/register_skill/native exec) like native tools
instead of gating them behind the skill_authoring capability:
- toolmgr.get_all_tools drops include_skill_authoring; SkillToolLoader
self-gates on sandbox + skill_mgr
- preproc drops the include_skill_authoring branch; pipeline-bound skills
and the skills resource gate on skill_mgr presence
Persist activated skills into host.activated_skills conversation state so
they survive across runs (host writes at activate; last-write-wins); drop
the dead restore_activated_skills helper.
Prefill ToolResource.parameters host-side (tool_mgr.get_tool_schema) so
runners build LLM tools without per-tool get_tool_detail round-trips.
Align agent-runner-pluginization design docs to the all-tool model.
Move the Lark SDK synchronous connection URL lookup off the main asyncio loop, serialize reconnects, and cover the incident with a non-blocking regression test.
* fix(provider): strip think tags for MiniMax-M3 and other OpenAI-compatible models
MiniMax-M3 (and other OpenAI-compatible providers) emit chain-of-thought
reasoning directly in the content field wrapped in tags, instead
of using a separate reasoning_content field or the legacy CRETIRE_REASONING
markers. The existing remove_think logic only handled CRETIRE_* tags, so
think blocks leaked into user-visible output even when remove_think was enabled.
- Add _ThinkStripState: a stateful filter that correctly handles tags
split across streaming chunk boundaries.
- Add _strip_think classmethod with regex patterns for both and
CRETIRE_* tags.
- Wire think_state into invoke_llm_stream so deltas are filtered before
reaching the accumulator.
- Add remove_think safety net in _StreamAccumulator so the final message
from tool-call rounds also gets stripped.
- Fix remove_think resolution to use defensive nested .get() so
pipelines missing output.misc don't raise AttributeError.
* fix(litellmchat): add missing _CLOSE_TAG class attribute on _ThinkStripState
* fix(provider): handle think stripping across LiteLLM paths
---------
Co-authored-by: WangCham <651122857@qq.com>
huanghuoguoguo