Skip to content

BUG-engine-crate-test-failures.md

Latest commit

 

History

History
94 lines (57 loc) · 4.88 KB

BUG-engine-crate-test-failures.md

File metadata and controls

94 lines (57 loc) · 4.88 KB

BUG: 5 pre-existing test failures in lunarwing_engine crate

Severity: Low (tests only, no production impact) Found: 2026-06-02 during v1.1.0 release validation Status: FIXED (verified 2026-06-07) — cargo test -p lunarwing_engine is green (271 passed, 0 failed). Original report retained below. Affects: cargo test -p lunarwing_engine

Summary

Five unit tests in lunarwing_engine fail consistently. All are test bugs caused by tests not being updated after production code refactors. No production code is broken.

Failure 1: executor::loop_engine::tests::action_then_text

Location: ic/crates/lunarwing_engine/src/executor/loop_engine.rs:621 Panic: assertion failed: exec.thread.messages.len() >= 3

Root cause: Test was written for the old Rust-driven execution loop. After the migration to the Python orchestrator, intermediate messages (action calls, action results) are stored in thread.internal_messages, not thread.messages. The visible thread.messages now only contains the system prompt + final response (2 messages), not >= 3.

Fix: Assert on internal_messages or lower the bound to >= 2.

Failure 2: executor::loop_engine::tests::codeact_multi_step

Location: ic/crates/lunarwing_engine/src/executor/loop_engine.rs:877 Panic: assertion failed: exec.thread.messages.iter().any(|m| m.content.contains("x = 30"))

Root cause: Same as failure 1. The code step output (x = 30) is part of the orchestrator's working transcript stored in thread.internal_messages, not the user-visible thread.messages.

Fix: Search exec.thread.internal_messages for "x = 30" instead of exec.thread.messages.

Failure 3: executor::loop_engine::tests::tool_intent_nudge_injected

Location: ic/crates/lunarwing_engine/src/executor/loop_engine.rs:739 Panic: assertion failed: exec.thread.messages.iter().any(|m| m.content.contains("did not include any tool calls"))

Root cause: Same as failures 1-2. The tool intent nudge is injected by the Python orchestrator into its working transcript (internal_messages), not the user-visible messages.

Fix: Search exec.thread.internal_messages for the nudge text. The wording may also differ between the old Rust implementation and the Python orchestrator.

Failure 4: executor::trace::tests::trace_serializes_approval_request_payload

Location: ic/crates/lunarwing_engine/src/executor/trace.rs:606 Panic: unexpected event kind: MessageAdded { role: "System", content_preview: "sys" }

Root cause: Thread::add_message() now emits a MessageAdded event for every message added. The test adds 2 messages (system + assistant) then manually pushes an ApprovalRequested event, then asserts trace.events[0] is ApprovalRequested. But the actual event order is:

  • [0] = MessageAdded { role: "System" }
  • [1] = MessageAdded { role: "Assistant" }
  • [2] = ApprovalRequested { ... }

Fix: Access trace.events[2] or find the event by type:

let event = trace.events.iter()
    .find(|e| matches!(e.kind, EventKind::ApprovalRequested { .. }))
    .expect("should have an ApprovalRequested event");

Failure 5: runtime::mission::tests::system_mission_requires_system_user_to_manage

Location: ic/crates/lunarwing_engine/src/runtime/mission.rs:2546 Panic: called Result::unwrap_err() on an Ok value: ()

Root cause: The test creates a mission owned by "system" and expects that user "alice" will be denied access when pausing it. However, "system" maps to OwnerId::Shared (it matches LEGACY_SHARED_OWNER_ID), and shared missions intentionally bypass per-user access checks at the engine level. The pause_mission guard:

if !mission.is_owned_by(user_id) && !mission.owner_id().is_shared() {
    return Err(EngineError::AccessDenied { ... });
}

...evaluates to false because is_shared() is true, so the method returns Ok(()) and the test's unwrap_err() panics.

Per the docstring: "For shared missions, the caller (web handler) must verify admin role before calling this. The engine only checks ownership."

Fix: Use a non-shared user_id (e.g., "admin-user") as the mission owner to test the ownership guard, or adjust the test expectation to match the documented shared-mission semantics.

Common Theme

Failures 1-3 share a single root cause: the execution loop tests were written against the pre-orchestrator Rust loop and never updated when ExecutionLoop::run() was refactored to delegate to the Python orchestrator via Monty. The orchestrator maintains its own working transcript in thread.internal_messages and only syncs the final response to thread.messages.

Impact

These failures do not indicate broken production behavior. The 266 passing engine tests cover the actual orchestrator logic, capability system, gate pipeline, memory retrieval, mission lifecycle, and type invariants.