`ToolCallingLLMCaller` does not render Jinja2 template variables in actor system prompts #11221

New issue

Open

opened 2026-05-15 05:36:22 +00:00 by hurui200320 · 0 comments

hurui200320 commented

2026-05-15 05:36:22 +00:00

Member

Metadata

Commit Message: feat(actor-run): render Jinja2 system prompt templates in ToolCallingLLMCaller
Branch: feature/m3-tool-calling-jinja2-context

Background and Context

SimpleLLMAgent.process() renders the actor's system_prompt: config through
Jinja2 before handing it to the LLM, using the graph's global context dict
(e.g. {"writing_stage": "intro", "paper_details": {...}}). This lets system
prompts use template expressions:

system_prompt: |
  You are a {{ writing_stage }} writer.
  Topic: {{ paper_details.topic }}

ToolCallingAgent.process() — introduced in #11211 — receives the same
context dict, but never forwards it to ToolCallingLLMCaller.invoke().
The system prompt is therefore used as a raw string; any {{ ... }} expressions
appear verbatim in the SystemMessage sent to the LLM.

ToolCallingLLMCaller already has _render_prompt() and self._template_env
(a jinja2.sandbox.SandboxedEnvironment), but they are never called during
invoke().

Expected Behavior

When ToolCallingLLMCaller.invoke() builds the SystemMessage, it should call
_render_prompt(system_prompt, context) so that template expressions in the
actor's system_prompt: are expanded against the graph's global context —
exactly as SimpleLLMAgent does.

Acceptance Criteria

An actor config with system_prompt: "You are a {{ stage }} expert." and
a context dict {"stage": "code-review"} produces a SystemMessage with
content "You are a code-review expert." when run via ToolCallingAgent.
An actor config with no template expressions in system_prompt is
unaffected (render is a no-op for plain strings).
If context is None or empty, rendering falls back gracefully (no
exception; unexpanded variables remain as-is per Jinja2 default).
Existing actor run BDD scenarios continue to pass.

Supporting Information

Relevant code locations:

src/cleveragents/reactive/tool_agent.py — ToolCallingAgent.process()
receives context but does not forward it.
src/cleveragents/reactive/tool_caller.py — ToolCallingLLMCaller.invoke()
needs to accept and use a render_context dict for Jinja2 rendering.
src/cleveragents/reactive/stream_router.py — SimpleLLMAgent._render_prompt()
is the reference implementation to follow.

Introduced as follow-on to #11211.

Subtasks

Add optional render_context: dict[str, Any] | None = None parameter to
ToolCallingLLMCaller.__init__() so the caller can hold the context for the
lifetime of the run.
In ToolCallingLLMCaller.invoke(), call
self._render_prompt(system_prompt, self._render_context or {}) before
building the SystemMessage (same pattern as SimpleLLMAgent.process()).
In ToolCallingAgent.process(), pass render_context=context when
constructing ToolCallingLLMCaller(self._actor_config, render_context=context).
Write BDD scenarios:
(A) system prompt with {{ var }} is expanded correctly;
(B) plain system prompt is unchanged;
(C) context=None does not raise.
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): render Jinja2 system prompt templates in ToolCallingLLMCaller Branch: feature/m3-tool-calling-jinja2-context ``` ## Background and Context `SimpleLLMAgent.process()` renders the actor's `system_prompt:` config through Jinja2 before handing it to the LLM, using the graph's global context dict (e.g. `{"writing_stage": "intro", "paper_details": {...}}`). This lets system prompts use template expressions: ```yaml system_prompt: | You are a {{ writing_stage }} writer. Topic: {{ paper_details.topic }} ``` `ToolCallingAgent.process()` — introduced in #11211 — receives the same `context` dict, but never forwards it to `ToolCallingLLMCaller.invoke()`. The system prompt is therefore used as a raw string; any `{{ ... }}` expressions appear verbatim in the `SystemMessage` sent to the LLM. `ToolCallingLLMCaller` already has `_render_prompt()` and `self._template_env` (a `jinja2.sandbox.SandboxedEnvironment`), but they are never called during `invoke()`. ## Expected Behavior When `ToolCallingLLMCaller.invoke()` builds the `SystemMessage`, it should call `_render_prompt(system_prompt, context)` so that template expressions in the actor's `system_prompt:` are expanded against the graph's global context — exactly as `SimpleLLMAgent` does. ## Acceptance Criteria - [ ] An actor config with `system_prompt: "You are a {{ stage }} expert."` and a context dict `{"stage": "code-review"}` produces a `SystemMessage` with content `"You are a code-review expert."` when run via `ToolCallingAgent`. - [ ] An actor config with no template expressions in `system_prompt` is unaffected (render is a no-op for plain strings). - [ ] If `context` is `None` or empty, rendering falls back gracefully (no exception; unexpanded variables remain as-is per Jinja2 default). - [ ] Existing `actor run` BDD scenarios continue to pass. ## Supporting Information Relevant code locations: - `src/cleveragents/reactive/tool_agent.py` — `ToolCallingAgent.process()` receives `context` but does not forward it. - `src/cleveragents/reactive/tool_caller.py` — `ToolCallingLLMCaller.invoke()` needs to accept and use a `render_context` dict for Jinja2 rendering. - `src/cleveragents/reactive/stream_router.py` — `SimpleLLMAgent._render_prompt()` is the reference implementation to follow. Introduced as follow-on to #11211. ## Subtasks - [ ] Add optional `render_context: dict[str, Any] | None = None` parameter to `ToolCallingLLMCaller.__init__()` so the caller can hold the context for the lifetime of the run. - [ ] In `ToolCallingLLMCaller.invoke()`, call `self._render_prompt(system_prompt, self._render_context or {})` before building the `SystemMessage` (same pattern as `SimpleLLMAgent.process()`). - [ ] In `ToolCallingAgent.process()`, pass `render_context=context` when constructing `ToolCallingLLMCaller(self._actor_config, render_context=context)`. - [ ] Write BDD scenarios: (A) system prompt with `{{ var }}` is expanded correctly; (B) plain system prompt is unchanged; (C) `context=None` does not raise. - [ ] 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 05:37:21 +00:00