fix(libsy): keep classifier history from orphaning a tool result

[eric-liu-nvidia]

Issue

LlmClassifier trims conversation history by message count, so the window can begin on a tool result and leave out the assistant tool call that introduced its id. The classifier subrequest then carries an orphan tool result, which violates the tool-call protocol and can be rejected by the classifier backend or scored on structurally incomplete context.

Solution

The window start now moves back to the nearest index that keeps every tool call and its result together. If no earlier start can pair a result, the counted start is kept, so an unanswerable result cannot widen the window to the whole conversation.

Summary by CodeRabbit

• Bug Fixes
  • Conversation history trimming now preserves complete tool-call and tool-result pairs when preparing input for classification.
  • Plain conversation history continues to use the requested trailing window without changes.

[coderabbitai]

Walkthrough

The classifier now preserves complete tool-call and tool-result pairs when trimming judge conversation history. The requested trailing window remains unchanged for plain conversations. Tests cover paired tools, orphan results, and standard message windows.

Changes

LLM history trimming

Layer / File(s)	Summary
Pair-aware trimming and validation crates/libsy/src/algorithms/llm_class.rs	trim_messages adjusts the window start when a retained tool result requires its originating tool call. New fixtures and tests verify paired-tool preservation and unchanged plain-conversation trimming.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Poem

I’m a rabbit with tools in my pack,
Keeping each call with its result intact.
The window moves back,
No orphan hops the track,
Plain messages stay on their exact stack.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
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.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: preventing classifier history trimming from orphaning tool results.

✨ Finishing Touches 💡 1

⚔️ Resolve merge conflicts 💡

• Resolve merge conflict in branch eric-liu/libsy-classifier-tool-pair-trim

---

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

[coderabbitai]

🧹 Nitpick comments (1)

crates/libsy/src/algorithms/llm_class.rs (1)

1518-1538: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Assert the retained tool pair directly.

Line 1537 passes if a regression removes ToolResult("call-1") from the retained window. Assert that kept contains both ToolCall("call-1") and ToolResult("call-1"). Add a fallback test where the call is before the first user message and assert that trimming keeps the original counted boundary.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@crates/libsy/src/algorithms/llm_class.rs` around lines 1518 - 1538,
Strengthen the test trimming_keeps_the_call_that_introduced_a_kept_tool_result
by directly asserting that kept contains both ToolCall("call-1") and
ToolResult("call-1"), rather than only checking orphan_tool_results. Add a
fallback test covering a tool call before the first user message, and verify
trimming preserves the original counted boundary.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@crates/libsy/src/algorithms/llm_class.rs`:
- Around line 1518-1538: Strengthen the test
trimming_keeps_the_call_that_introduced_a_kept_tool_result by directly asserting
that kept contains both ToolCall("call-1") and ToolResult("call-1"), rather than
only checking orphan_tool_results. Add a fallback test covering a tool call
before the first user message, and verify trimming preserves the original
counted boundary.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 1404a642-abe6-4a9e-bf34-010973bf12e3

📥 Commits

Reviewing files that changed from the base of the PR and between 54d4b66 and 9b0036d.

📒 Files selected for processing (1)

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

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://NVIDIA-NeMo.github.io/Switchyard/pr-preview/pr-309/
Built to branch gh-pages at 2026-08-05 20:18 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[elyasmnvidian]

The current scan revisits a long classifier history for every candidate start, and it accepts a tool result when a matching call appears later. The regression test also stays green if trimming drops the pair. I ran cargo fmt --all --check, workspace clippy with warnings denied, and cargo test -p switchyard-libsy; all passed.

[elyasmnvidian]

please fix merge conflicts