fix(libsy): restate the routing instruction after windowed conversation

[gburachas]

Addresses the first of the two changes described in #279. The second, capping the payload size, will be a separate pull request.

What

When an llm_classifier route sets recent_turn_window, one sentence is appended as the last message sent to the judge:

Route the conversation above. Output ONLY the routing JSON object, nothing else.

Nothing changes when recent_turn_window is unset.

Why

The judge is asked to classify a request and reply with a routing JSON object. Its instructions lead the payload, which works while that payload is one short message.

With a window configured, the instructions are followed by assistant turns and tool output. The judge then sometimes answers the conversation rather than classifying it. That reply does not parse as a verdict, so the route falls back to its default target: the turn is served, and the only sign that routing did not happen is a warning in the log.

Example: a request of "check my account balance", with a window containing a long stack trace from a failed tool call. The verdict describes the stack trace.

Scope

Only the windowed path is affected. With recent_turn_window unset, task_messages sends the opening task and the latest user follow-up, with no assistant or tool content between the instruction and the end of the payload.

Tests

Two, in crates/libsy/src/algorithms/llm_class.rs:

• a_window_restates_the_routing_instruction_last asserts the instruction is the last message, not merely present. Position is the whole mechanism: moving the same message to the front makes this test fail.
• the_default_path_is_left_unchanged asserts the unwindowed request shape is untouched.

Naming

An earlier draft called this "reinforcement", which reads as reinforcement learning and is unrelated. The constant is TRAILING_ROUTING_INSTRUCTION and the issue has been retitled.

How tested

• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace
• uv run ruff check .
• Commit signed off per the DCO

One file, 66 lines, of which 45 are the two tests.

Note for anyone reading the CodeRabbit panel: the "Docstring Coverage" check measures Python docstrings and this diff is Rust only.

Summary by CodeRabbit

• Improvements

  • Windowed capability and custom classification requests now provide clearer routing guidance, resulting in JSON-only routing responses.
  • Requests without windowing retain their existing behavior and format.

• Tests

  • Added coverage to verify routing guidance appears in the correct position and that default requests remain unchanged.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 3b84ccf4-a5f4-4137-bc0b-9a57f44674b5

📥 Commits

Reviewing files that changed from the base of the PR and between 053a61e and 2913f51.

📒 Files selected for processing (1)

• crates/libsy/src/algorithms/llm_class.rs

Included review availability: Your plan provides up to 12 included reviews per hour; 10 remain after this review.

---

Walkthrough

Windowed judge inputs now append a final routing-only user instruction. The default no-window path remains unchanged. Tests verify the final message position and the default message shape.

Changes

Windowed routing behavior

Layer / File(s)	Summary
Append routing instruction to windowed messages crates/libsy/src/algorithms/llm_class.rs	TaskInput::build_messages defines and appends TRAILING_ROUTING_INSTRUCTION when recent_turn_window is configured. Tests verify windowed and no-window message sequences.

Estimated code review effort: 2 (Simple) | ~10 minutes

Merge Risk: ⚪ Minimal · up to 2913f

The PR adds a trailing routing instruction only for windowed conversations while preserving the default path; no actionable merge-blocking risk remains after normal checks and review.

Poem

I’m a rabbit with routes in my ear,
A final JSON hop makes the path clear.
Windowed turns line up in a row,
No-window messages stay as they go.
Hop, hop—clean prompts now flow!

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main routing-instruction change for windowed conversations.
Docstring Coverage	✅ Passed	Docstring coverage is 80.00% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 5 functions across 1 files.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

_{Comment @coderabbitai help to get the list of available commands.}