[Web UI] First user message bubble missing / messages misordered & duplicated in live rendering; server transcript is correct #3991

Activity

1. aircolorX commented on Sep 23, 2026

   

   aircolorX

   on Sep 23, 2026

   Author

   More actions

   Update (2026-09-23): additional findings across rounds 3–5

   • Round 3 — hook injection hypothesis invalidated. Reproduced on a fresh process (PID-verified) with all UserPromptSubmit hooks disabled and all browser extensions off (v0.39.1). The user bubble still disappears after the first assistant reply. The server-side transcript remains complete and correctly ordered, so this is a client-side rendering issue.
   • Round 4 — session list is a startup snapshot. Sessions created by another process (web ↔ TUI, same KIMI_CODE_HOME) don't appear until the receiving process is restarted; in-session refresh doesn't help. This suggests state is snapshotted at process start with no cross-process invalidation — a candidate mechanism for why "restart + refresh" is the only recovery. (Note: this alone doesn't explain the disappearing bubbles of the current session, so a pure rendering-layer fault is not excluded either.)
   • Round 5 — still reproduces on 2.0.2. After a full web-server restart on the new binary, the same "user bubbles gone after round 1" symptom appears in the browser. Affected range now 0.39.1 → 2.0.2.

   Also noted: the web UI never displays hook-injected content (only the TUI does), which is consistent with the round-3 result.
2. aircolorX commented on Sep 28, 2026

   

   aircolorX

   on Sep 28, 2026

   Author

   More actions

   Root cause found — two independent bugs stacking up. I traced this end-to-end (server event journal, WS frames, REST transcript paging, and the frontend projection path) and reproduced both symptom variants. Posting the analysis in case it helps an official fix.

   Layer 1 — transcript_since catch-up can never match → silent full-reset storms (hook-independent)

   In packages/kap-server, transcript frames were broadcast with the event-journal seq on the envelope, while clients resync with the ops-namespace seq carried in the payload. The two namespaces diverge, so transcript_since catch-up never matches and every WS reconnect falls back to a silent full reset. Additionally, when the transcript store detected an append gap, seeded clients were never told to resync — the server just diverged silently.

   This layer needs no hooks at all, which explains evidence point 5 above: the "nothing renders until server restart" variant still reproduces with UserPromptSubmit hooks fully disabled, because the per-session projection baseline is wrecked by reset storms and never heals within the session's lifetime.

   Fix (local): the envelope seq for transcript frames now uses the ops-namespace seq from the payload (transcriptService.ts, coreBinding.ts), and an append gap now broadcasts a transcript reset to seeded targets instead of silent divergence (sessionEventBroadcaster.ts).

   Layer 2 — hook_result injections open pseudo turns in the snapshot projection → every turn shifts by one (hook-dependent)

   In packages/transcript/src/history/groupTurns.ts, groupMessagesIntoSnapshot treated a user-role hook_result injection as a turn-opening user message (the hidden-user-origin set was missing hook_result). Snapshot turn ordinals shifted +1; then healTurnOps in kap-server's transcriptService.ts overwrote the live turn's correct origin with the misaligned snapshot origin. Net effect: user bubbles render one turn off and the first bubble is dropped — exactly variant 2 (first reply duplicated, first prompt bubble shifted under its own reply). With hooks enabled this compounds with Layer 1.

   Fix (local): project hook_result user injections as hook markers (matching the live path) instead of opening a turn (+4 lines in groupTurns.ts), and healTurnOps now prefers the live turn's origin over the snapshot header (+1 line in transcriptService.ts). Regression tests added in packages/transcript/test/layers.test.ts (+64 lines).

   Verification

   • packages/transcript: 127/127 tests pass; packages/kap-server: 133/133 pass.
   • End-to-end repro in an isolated environment confirms snapshot/live turn origins stay aligned.
   • Running a from-source build with both patches since yesterday across multiple new sessions: no missing bubbles, no duplication, no misordering, no server restarts needed.

   Happy to share the patches if useful.
3. aircolorX commented on Sep 28, 2026

   

   aircolorX

   on Sep 28, 2026

   Author

   More actions

   Root cause found — two independent bugs stacking up. I traced this end-to-end (server event journal, WS frames, REST transcript paging, and the frontend projection path) and reproduced both symptom variants. Posting the analysis in case it helps an official fix.

   Layer 1 — transcript_since catch-up can never match → silent full-reset storms (hook-independent)

   In packages/kap-server, transcript frames were broadcast with the event-journal seq on the envelope, while clients resync with the ops-namespace seq carried in the payload. The two namespaces diverge, so transcript_since catch-up never matches and every WS reconnect falls back to a silent full reset. Additionally, when the transcript store detected an append gap, seeded clients were never told to resync — the server just diverged silently.

   This layer needs no hooks at all, which explains evidence point 5 above: the "nothing renders until server restart" variant still reproduces with UserPromptSubmit hooks fully disabled, because the per-session projection baseline is wrecked by reset storms and never heals within the session's lifetime.

   Fix (local): the envelope seq for transcript frames now uses the ops-namespace seq from the payload (transcriptService.ts, coreBinding.ts), and an append gap now broadcasts a transcript reset to seeded targets instead of silent divergence (sessionEventBroadcaster.ts).

   Layer 2 — hook_result injections open pseudo turns in the snapshot projection → every turn shifts by one (hook-dependent)

   In packages/transcript/src/history/groupTurns.ts, groupMessagesIntoSnapshot treated a user-role hook_result injection as a turn-opening user message (the hidden-user-origin set was missing hook_result). Snapshot turn ordinals shifted +1; then healTurnOps in kap-server's transcriptService.ts overwrote the live turn's correct origin with the misaligned snapshot origin. Net effect: user bubbles render one turn off and the first bubble is dropped — exactly variant 2 (first reply duplicated, first prompt bubble shifted under its own reply). With hooks enabled this compounds with Layer 1.

   Fix (local): project hook_result user injections as hook markers (matching the live path) instead of opening a turn (+4 lines in groupTurns.ts), and healTurnOps now prefers the live turn's origin over the snapshot header (+1 line in transcriptService.ts). Regression tests added in packages/transcript/test/layers.test.ts (+64 lines).

   Verification

   • packages/transcript: 127/127 tests pass; packages/kap-server: 133/133 pass.
   • End-to-end repro in an isolated environment confirms snapshot/live turn origins stay aligned.
   • Running a from-source build with both patches since yesterday across multiple new sessions: no missing bubbles, no duplication, no misordering, no server restarts needed.

   Happy to share the patches if useful.