root/fusion
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(FN-4180): compact room transcript context to reduce token usage

Browse Source 
Adds room transcript context compaction (FN-4180) to the dashboard chat layer — widens the fetch window then compacts older messages into a summary window to reduce token overhead while keeping full context recent. Includes a new 127-line test suite covering the compaction logic, a small fix to reus

Fusion-Task-Id: FN-4180

Fusion-Task-Lineage: 9e7e3a32-f2da-4c86-871b-37891f747561
This commit is contained in:
Fusion
2026-05-12 13:22:28 -07:00
committed by gsxdsm
parent efdc6c4625
commit f27ac8bf59
8 changed files with 292 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/dashboard-guide.md
View File

@@ -119,7 +119,7 @@ Chat Rooms are project-scoped group conversations for multiple agents. They are
- The dashboard backend now orchestrates room responders on that POST: mentioned members are routed as direct responders, additional ambient members may reply (up to the room ambient responder cap), and each assistant reply is persisted with `senderAgentId` via `chatStore.addRoomMessage(...)`.
- If room replies cannot be generated (for example no resolvable responders or all responders fail), the POST fails with an API error (HTTP 502) instead of silently returning only the user message.
- If room responders cannot be resolved or all room-reply generations fail, the POST now returns an error instead of silently succeeding with only the user message, so failures are surfaced deterministically.
- Room responder prompt construction now includes a bounded recent room transcript (role, sender label, timestamp, content) plus an explicit latest-user-message marker, so replies stay thread-aware without unbounded prompt growth.
- Room responder prompt construction now keeps the most recent room messages verbatim and, when the room runs long, prepends a compacted summary of older history (span, participants, and key highlights) plus an explicit latest-user-message marker so replies stay thread-aware without unbounded prompt growth.
- On send failure, `useChatRooms` rolls back/reconciles optimistic state and rethrows; `ChatView` catches once, preserves composer text for retry/edit, and surfaces a single error toast (no duplicate hook+view notifications).
- After each send attempt, the room transcript still re-fetches authoritative messages so persisted user/assistant replies remain visible even when SSE delivery is delayed, and `chat:room:message:*` SSE updates continue live fan-out.
- Relationship summary: direct Chat runs one target (agent or model) per session; rooms are shared threads with multiple agent members and now use the same message contract as direct Chat; Quick Chat stays a floating single-target panel and does not host rooms.