fix(domain): collapse system messages into a single one (#2894)

[ImBIOS]

Summary

Context::set_system_messages previously inserted each entry in content
as a separate ContextMessage::system(...), producing N system messages
in the resulting chat. Chat templates such as Qwen3.5/3.6 (used by
llama.cpp and vLLM) only allow a single system message at messages[0]
and raise:

raise_exception('System message must be at the beginning.')

for any additional system message, breaking those providers entirely with
a 400/500 error from the upstream inference server.

Fix: join all entries into a single system message (separated by \n\n)
and insert it once at position 0. Empty payloads are filtered out, and an
all-empty content vector leaves the context untouched.

Changes

• crates/forge_domain/src/context.rs
  • Rewrite Context::set_system_messages to:
    1. Drop every existing system message.
    2. Combine the new entries into a single \n\n-joined string (filtering
       out empty entries).
    3. Insert exactly one ContextMessage::system(combined) at position 0.
  • Added a Rustdoc block describing the chat-template constraint that
    motivates the change.
• Two regression tests in the same file:
  • test_set_system_messages_collapses_into_single_message — asserts the
    context contains exactly one system message and that its content is the
    \n\n-joined payload.
  • test_set_system_messages_replaces_existing_single_system_message —
    asserts a pre-existing system message is replaced, not retained or
    duplicated.

Why this is safe

• Single-function, single-crate change. The public API of
  Context::set_system_messages is unchanged.
• Behaviour is a strict superset of the old implementation for callers that
  pass a Vec of length 1 (still produces a single system message at
  position 0).
• The downstream caller in forge_app/src/system_prompt.rs:136 already
  passes exactly two entries (static_block, non_static_block), so this
  is the path actually exercised in production.

Verification

• cargo check -p forge_domain --tests: clean
• cargo check -p forge_app --tests: clean
• cargo clippy -p forge_domain --tests --no-deps: clean
• cargo test -p forge_domain --lib: 596 passed, 0 failed (incl. 2 new
  regression tests)
• cargo test -p forge_app --lib system_prompt: 7 passed, 0 failed

Related issues

• Fixes [Bug]: Multiple system messages break models with strict chat templates (e.g. Qwen3.5) #2894
• Fixes [Bug]: template error with qwen3.5-27b (llama.cpp) #2586 (same System message must be at the beginning. Jinja error on Qwen3.5-27B / llama.cpp) - it's a duplicate issue

How to test locally

1. cargo build -p forge_domain (or run the two new tests directly).
2. After the fix is merged, a Qwen3.5 / Qwen3.6 model served via llama.cpp
   or vLLM should no longer reject requests with the
   System message must be at the beginning. Jinja exception.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Updates Context::set_system_messages to ensure compatibility with chat templates that require exactly one system message at messages[0].

Changes:

• Replace any existing system messages and collapse all provided entries into a single system message block.
• Filter out empty system-message entries and avoid inserting a system message when the combined payload is empty.
• Add regression tests covering multi-entry collapse and replacement of pre-existing system messages.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.