Steering during an in-flight skill-context turn can reuse turn_index and overwrite the persisted user message

Environment

• Copilot CLI 1.0.83-4
• Package build commit: c05c8e5
• Native session-store tracking implementation

Summary

A steering user.message delivered while an injected skill-context turn is pending can receive the same turn_index as that skill-context message. The conflicting SQLite upsert replaces the earlier skill-context text, while append-only usage events from both interactions retain the shared index.

turn_index is intended to be monotonically increasing and unique for each persisted user.message. This is not expected steering behavior.

Observed sequence

1. Human input A receives index 536.
2. Skill-context message 1 receives 537.
3. Skill-context message 2 receives 538.
4. Model-usage events for skill context 2 are recorded with 538.
5. Before the pending turn is flushed, steering input B arrives with a new interactionId.
6. B incorrectly also receives 538.
7. Later B usage is also recorded with 538.
8. The next human input C receives 539.
9. The final turns row for 538 contains B's text; skill-context message 2 has been overwritten.

Expected

turns

turn_index	Message
536	A
537	Skill context 1
538	Skill context 2
539	B
540	C

Usage

• Pre-B usage: 538
• Post-B usage: 539

Actual

turns

turn_index	Message
536	A
537	Skill context 1
538	B
539	C

Usage

• Pre-B usage: 538
• Post-B usage: 538

Relevant implementation

Current main locations:

• src/runtime/src/session/store_tracking.rs:510-526
  • initial_tracking_state and initial_tracking_state_result seed turn_counter from persisted max_turn_index + 1.
• src/runtime/src/session/store_tracking.rs:544-554
  • event_operations flushes the previous pending turn and stores the new user.message in pending_user_message.
• src/runtime/src/session/store_tracking.rs:869-894
  • flush_pending_turn writes the pending message at the current turn_counter, then increments the counter.
• src/runtime/src/session_bindings/api_store_tracking.rs:61-78
  • session_store_tracking_init_session_state replaces session.store_tracking_state.
• src/runtime/src/session_bindings/api_store.rs:295-303
  • assistant.usage reads the mutable tracking_turn_counter and writes an assistant_usage_events row with that index.
• src/runtime/src/session_bindings/api_store.rs:1033-1047
  • tracking_turn_counter reads the counter from the current session tracking state.
• src/runtime/src/session/store.rs:971-1002
  • insert_turn_conn uses ON CONFLICT(session_id, turn_index) DO UPDATE, with user_message = COALESCE(excluded.user_message, user_message).
• src/runtime/src/session/store.rs:1244-1307
  • insert_assistant_usage_event_conn appends usage rows without a uniqueness constraint on (session_id, turn_index).
• src/cli/core/localSessionManager.ts:256-283
  • initSessionStoreTracking has a waitFor gate intended to flush a previous tracker before resetting shared tracking state and reading getMaxTurnIndex.
• src/runtime/src/skills/core.rs:2484-2594
  • format_skill_content constructs the injected <skill-context> message.
• src/runtime/src/tools/session_tool_invoker.rs
  • Tool results carry injected follow-up content through newMessages.

Relevant exported symbols:

• sessionStoreTrackingInitSessionState
• SessionStoreHandle.getMaxTurnIndex
• SessionStoreHandle.handleTrackingEventForSession
• SessionStoreHandle.flushTrackingForSession
• SessionStoreHandle.insertTurn
• SessionStoreHandle.insertAssistantUsageEvent

Likely race

1. The original tracker has index 538 only in its in-memory pending_user_message.
2. A steering/session lifecycle path initializes or replaces another tracker state.
3. It seeds its counter from persisted MAX(turn_index)=537.
4. It cannot see the other tracker's pending index 538.
5. It independently allocates 538 to B.

The low-level serialized handler does not reproduce the collision when exercised directly. The bug appears to require the integrated steering/session lifecycle path and overlapping tracker initialization or replacement.

Current main already attempts to gate tracker replacement with waitFor, so the reproducing path may be bypassing that gate or creating an additional lifecycle transition not covered by it.

Destructive upsert

insertTurn uses:

ON CONFLICT(session_id, turn_index) DO UPDATE SET
    user_message = COALESCE(excluded.user_message, user_message),
    assistant_response = COALESCE(excluded.assistant_response, assistant_response)

That behavior is appropriate for combining the user and assistant halves of one turn, but silently corrupts data when two different non-null user messages collide.

Impact

• Loss of a persisted user message.
• Usage generated before B existed appears associated with B.
• No reliable usage-to-interaction correlation remains because usage rows do not persist interactionId.
• Consumers receive corrupted per-turn AIU, token, latency, and model-loop distributions.
• Winsorization and percentile metrics can change even when aggregate AIU is unchanged.
• If the colliding interactions begin under different experiment assignments, A/B arm attribution can also be wrong.
• Overwritten skill-context markers can prevent consumers from recognizing and excluding skill-loading usage.

Missing regression coverage

• Two distinct non-null user messages colliding at one index.
• Multiple skill injections followed by immediate steering.
• Tracker initialization while another tracker has pending_user_message.
• Preservation of usage correlation after an upsert conflict.

Requested regression test

1. Seed the store so the next index is 536.
2. Submit human A.
3. Return two skill-context newMessages.
4. Pause after usage is emitted for the second skill context but before tracker flush.
5. Submit immediate steering B through the real steering admission path with a new interactionId.
6. Emit B usage, release the paused turn, flush, and submit C.
7. Assert the expected indexes and usage assignments above.
8. Assert that insertTurn is never called with two unequal, non-null user_message values for one (session_id, turn_index).

Acceptance criteria

• Index allocation is serialized or atomically reserved across tracker lifecycle transitions.
• A pending in-memory index cannot be duplicated by a tracker initialized from persisted MAX(turn_index).
• Conflicting non-null user-message upserts are rejected or surfaced rather than silently overwriting data.
• The integrated steering-plus-skill regression test passes.
• Ideally, usage events also persist interaction_id and a model/request identifier so consumers do not depend exclusively on turn_index.

Bắt đầu bằng cách lần theo việc thay thế tracker và xử lý các tin nhắn đang chờ trong src/runtime/src/session/store_tracking.rs và src/cli/core/localSessionManager.ts, sau đó kiểm tra insertTurn và việc duy trì dữ liệu sử dụng trong src/runtime/src/session/store.rs và api_store.rs. Tái hiện vòng đời steering-plus-skill tích hợp được mô tả trong issue, đồng thời bổ sung coverage cho các assertion được yêu cầu về index, bảo toàn tin nhắn và gán mức sử dụng. Hoàn tất có nghĩa là regression vượt qua mà không có index trùng lặp hoặc việc ghi đè im lặng các tin nhắn xung đột.