`ToolCallingAgent` silently skips inline-defined skill tools; extend `_build_tool_registry()` to support them #11222

New issue

Open

opened 2026-05-15 05:37:01 +00:00 by hurui200320 · 0 comments

hurui200320 commented

2026-05-15 05:37:01 +00:00

Member

Metadata

Commit Message: feat(actor-run): support inline skill tools in ToolCallingAgent tool registry
Branch: feature/m3-tool-calling-inline-tools

Background and Context

When a skill is resolved in ReactiveCleverAgentsApp._resolve_skills(), each
ResolvedToolEntry is either:

Named reference (is_inline=False): name is a registered tool name
such as "builtin/file-read", looked up in _builtin_registry.
Inline definition (is_inline=True): the full tool definition is embedded
in the inline_tool: SkillInlineTool field of the entry.

ToolCallingAgent._build_tool_registry() — introduced in #11211 — only handles
named references. When is_inline=True, the entry's name is a generated key
that does not exist in _builtin_registry, so get() returns None and the
tool is warned about and silently skipped:

WARNING: tool 'local/my-skill/custom-formatter' not found in builtin registry — skipping
(inline/MCP tools not yet supported in actor run tool-calling path)

This means any skill that defines its tools inline cannot be used with
agents actor run --skill, even though inline tools are entirely valid per
the spec.

Expected Behavior

When _build_tool_registry() encounters an entry with is_inline=True, it
should construct a ToolSpec directly from the inline_tool definition and
register it into the local registry — no _builtin_registry lookup needed.
The resulting tool must pass through the same SEC1 safety constraints that
govern inline tools elsewhere in the codebase (named-operation registry only;
raw code blocks rejected unconditionally).

Named-reference entries (is_inline=False) are unaffected.

Acceptance Criteria

A skill with an inline tool definition (is_inline=True, containing
name, description, input_schema, and a safe named operation) results
in a ToolSpec being registered in the per-run local ToolRegistry.
The LLM receives the inline tool's schema (via export_tool_schemas())
when --skill is attached to an actor that references a skill with inline tools.
Inline tools with a code block are rejected with a StreamRoutingError
consistent with the SEC1 policy already enforced by SimpleToolAgent.
Named-reference tools (is_inline=False) continue to work as before —
no regression.
All existing actor run BDD scenarios continue to pass.

Supporting Information

Relevant code locations:

src/cleveragents/reactive/tool_agent.py — ToolCallingAgent._build_tool_registry()
is where the inline branch needs to be added.
src/cleveragents/reactive/stream_router.py — SimpleToolAgent._SAFE_OPERATIONS
and the SEC1 code-block rejection are the reference for inline tool safety.
src/cleveragents/domain/models/core/skill.py — ResolvedToolEntry and
SkillInlineTool define the inline data model.
src/cleveragents/tool/runtime.py — ToolSpec is what needs to be constructed.

Introduced as follow-on to #11211.

Subtasks

Inspect SkillInlineTool schema to confirm available fields (name,
description, input_schema, operation, etc.).
In _build_tool_registry(), add an is_inline branch: when
entry.get("is_inline") is True, build a ToolSpec whose handler is
looked up from SimpleToolAgent._SAFE_OPERATIONS via the entry's operation
field; reject entries containing a code key with StreamRoutingError (SEC1).
Ensure inline entries use the name from the entry dict as ToolSpec.name
(no _builtin_registry lookup).
Deduplicate inline entries via the existing seen set.
Write BDD scenarios:
(A) inline tool with valid operation is registered and callable via the loop;
(B) inline tool with code block raises StreamRoutingError;
(C) skill mixing inline and named-reference tools resolves both correctly.
Verify coverage ≥ 96.5% via nox -s coverage_report.
Run nox (all default sessions), fix any errors.

Definition of Done

This issue is complete when:

All subtasks above are completed and checked off.
A Git commit is created where the first line of the commit message
matches the Commit Message in Metadata exactly, followed by a blank line,
then additional lines providing relevant details about the implementation.
The commit is pushed to the remote on the branch matching the Branch in
Metadata exactly.
The commit is submitted as a pull request to master, reviewed, and
merged before this issue is marked done.

## Metadata ``` Commit Message: feat(actor-run): support inline skill tools in ToolCallingAgent tool registry Branch: feature/m3-tool-calling-inline-tools ``` ## Background and Context When a skill is resolved in `ReactiveCleverAgentsApp._resolve_skills()`, each `ResolvedToolEntry` is either: - **Named reference** (`is_inline=False`): `name` is a registered tool name such as `"builtin/file-read"`, looked up in `_builtin_registry`. - **Inline definition** (`is_inline=True`): the full tool definition is embedded in the `inline_tool: SkillInlineTool` field of the entry. `ToolCallingAgent._build_tool_registry()` — introduced in #11211 — only handles named references. When `is_inline=True`, the entry's `name` is a generated key that does not exist in `_builtin_registry`, so `get()` returns `None` and the tool is warned about and silently skipped: ``` WARNING: tool 'local/my-skill/custom-formatter' not found in builtin registry — skipping (inline/MCP tools not yet supported in actor run tool-calling path) ``` This means any skill that defines its tools inline cannot be used with `agents actor run --skill`, even though inline tools are entirely valid per the spec. ## Expected Behavior When `_build_tool_registry()` encounters an entry with `is_inline=True`, it should construct a `ToolSpec` directly from the `inline_tool` definition and register it into the local registry — no `_builtin_registry` lookup needed. The resulting tool must pass through the same SEC1 safety constraints that govern inline tools elsewhere in the codebase (named-operation registry only; raw `code` blocks rejected unconditionally). Named-reference entries (`is_inline=False`) are unaffected. ## Acceptance Criteria - [ ] A skill with an inline tool definition (`is_inline=True`, containing `name`, `description`, `input_schema`, and a safe named `operation`) results in a `ToolSpec` being registered in the per-run local `ToolRegistry`. - [ ] The LLM receives the inline tool's schema (via `export_tool_schemas()`) when `--skill` is attached to an actor that references a skill with inline tools. - [ ] Inline tools with a `code` block are rejected with a `StreamRoutingError` consistent with the SEC1 policy already enforced by `SimpleToolAgent`. - [ ] Named-reference tools (`is_inline=False`) continue to work as before — no regression. - [ ] All existing `actor run` BDD scenarios continue to pass. ## Supporting Information Relevant code locations: - `src/cleveragents/reactive/tool_agent.py` — `ToolCallingAgent._build_tool_registry()` is where the inline branch needs to be added. - `src/cleveragents/reactive/stream_router.py` — `SimpleToolAgent._SAFE_OPERATIONS` and the SEC1 `code`-block rejection are the reference for inline tool safety. - `src/cleveragents/domain/models/core/skill.py` — `ResolvedToolEntry` and `SkillInlineTool` define the inline data model. - `src/cleveragents/tool/runtime.py` — `ToolSpec` is what needs to be constructed. Introduced as follow-on to #11211. ## Subtasks - [ ] Inspect `SkillInlineTool` schema to confirm available fields (`name`, `description`, `input_schema`, `operation`, etc.). - [ ] In `_build_tool_registry()`, add an `is_inline` branch: when `entry.get("is_inline")` is `True`, build a `ToolSpec` whose handler is looked up from `SimpleToolAgent._SAFE_OPERATIONS` via the entry's `operation` field; reject entries containing a `code` key with `StreamRoutingError` (SEC1). - [ ] Ensure inline entries use the `name` from the entry dict as `ToolSpec.name` (no `_builtin_registry` lookup). - [ ] Deduplicate inline entries via the existing `seen` set. - [ ] Write BDD scenarios: (A) inline tool with valid `operation` is registered and callable via the loop; (B) inline tool with `code` block raises `StreamRoutingError`; (C) skill mixing inline and named-reference tools resolves both correctly. - [ ] Verify coverage ≥ 96.5% via `nox -s coverage_report`. - [ ] Run `nox` (all default sessions), fix any errors. ## Definition of Done This issue is complete when: - All subtasks above are completed and checked off. - A Git commit is created where the **first line** of the commit message matches the Commit Message in Metadata exactly, followed by a blank line, then additional lines providing relevant details about the implementation. - The commit is pushed to the remote on the branch matching the **Branch** in Metadata exactly. - The commit is submitted as a **pull request** to `master`, reviewed, and **merged** before this issue is marked done.

hurui200320 added the

labels

2026-05-15 05:37:22 +00:00