streaming: server-authoritative message identity + idempotent client reducer

[TirmanSidhu]

Summary

Re-architects the daemon→macOS streaming pipeline so the same-message-rendered-twice mid-stream bug is structurally unrepresentable instead of chased through ad-hoc dedupe.

Before: daemon emitted anonymous assistant_text_delta events; client used a lazy "find the bubble for currentAssistantMessageId, else create a new one" path. Three failure modes (double subscriber window during loop restart, stale-id forking after handoff, reconnect replay) all routed through that "else create new" branch.

After:

• Daemon assigns messageId (UUIDv7) at message open, not at persist. Every event carries (messageId, blockIndex, seq). Events are persisted to a durable SQLite log; SSE supports Last-Event-Id replay on reconnect.
• Client runs a single MessageStreamReducer keyed by messageId, idempotent via per-(messageId, blockIndex) seq watermarks. Renderer reads from MessageStore.
• EventStreamClient fan-out collapsed into 11 typed per-domain dispatchers with bounded buffers. startMessageLoop/messageLoopTask/messageLoopGeneration removed.

PRs merged into this branch

• streaming: pre-assign message_id, add block_index and seq #32440 — streaming: pre-assign message_id, add block_index and seq to all events
• streaming: add MessageStreamReducer and MessageStore #32445 — streaming: add MessageStreamReducer and MessageStore (unused by UI)
• streaming: durable event log + Last-Event-Id replay #32447 — streaming: durable event log + Last-Event-Id replay
• streaming: render chat from MessageStore #32454 — streaming: render chat from MessageStore, remove implicit current-message id
• streaming: collapse EventStreamClient fan-out #32459 — streaming: collapse EventStreamClient fan-out into typed routing
• streaming: retire legacy vocabulary #32462 — streaming: retire legacy vocabulary (narrowed scope — see below)

Where the fix actually lands

PR #32454. The renderer now reads from MessageStore; renderedMessages drops legacy lazy-created bubbles that lack a daemonMessageId. The MessageStore snapshot is authoritative.

• Failure mode feat: initialize Next.js app in /web directory #1 (double subscriber window during loop restart) — killed by PR streaming: collapse EventStreamClient fan-out #32459 (no more cancel-and-resubscribe).
• Failure mode feat: add platform terraform for GKE deployment #2 (stale-id forking after handoff) — killed by PR streaming: render chat from MessageStore #32454 (no "current message" pointer; every event carries its own messageId).
• Failure mode ci: add terraform apply workflow on platform changes #3 (reconnect replay) — killed by PR streaming: durable event log + Last-Event-Id replay #32447 (Last-Event-Id) + PR streaming: add MessageStreamReducer and MessageStore #32445 (idempotent reducer).

Scope notes

• PR streaming: retire legacy vocabulary #32462 narrowed: it removed the TranscriptProjector dedupe-by-id and updated wire-contract docs, but did NOT delete currentAssistantMessageId, streamingDeltaBuffer, partialOutputBuffer, clearCurrentTurnTracking, appendTextToCurrentMessage, or the legacy messages array. The agent judged those entanglements (~160 call sites; load-bearing semantics beyond text streaming — tool result anchoring, surface widget routing, send-coordination, watchdog recovery) too risky for a single PR. Daemon-side message_complete emission was also kept because CLI/voice/Slack consumers still need it.
• The structural fix is in place because the renderer no longer sources from those legacy helpers — they are dead-write paths from the view's perspective. Removing them is straightforward follow-up.

Test plan

• Send several messages in quick succession in a fresh conversation, especially the 2nd message right after the 1st — the previously-reproducible duplicate-bubble-mid-stream should be gone.
• Reload an in-flight conversation mid-stream — message identity should be preserved (no re-creation as new bubbles).
• Reconnect the SSE stream (kill daemon, restart) mid-turn — replay should be exact, no duplicates.
• History reload of a long conversation should render identically to streaming.
• Multi-step skill turns with generation_handoff — second turn's text should not fork into a new bubble.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 36a80a28f1

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Filter reducer events to the active conversation

Because subscribeChatEvents() now delivers the global chat category, this reducer applies assistant lifecycle/delta events from every conversation without the belongsToConversation guard that the legacy ChatActionHandler still uses. When another conversation streams (for example a pop-out window, background task, or another active VM), its message_open/delta events are inserted into this VM's messageStore, and renderedMessages will append those snapshots into the wrong transcript.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Do not drop buffered streaming deltas

Using .bufferingNewest(256) silently discards older chat events whenever the main-actor consumer falls behind a long or fast assistant/tool stream. Chat messages use incremental assistant_text_delta and tool_input_delta payloads rather than snapshots, so dropping an older buffered event permanently removes part of the transcript (and no reconnect/replay is triggered because the SSE connection itself stayed healthy).

Useful? React with 👍 / 👎.