plat 525

← chat / reliability

PLAT-525 — Make chat history simple: one source of truth instead of merging the CLI's own transcript into the saved conversation

Field	Value
State	open
Priority	P2
Product	chat
Area	reliability
Summary	open (proposal): one source of truth (the event store) instead of merging the CLI's transcript; decision needed.

Coordination	Value
State	fixed on main (a3529fd9e), confirmed locally by the owner 2026-10-05 ("seems to work"); not deployed
Priority	P2
Date	2026-10-05
Owner	chat-reliability
Related	PLAT-518, PLAT-341, PLAT-178, PLAT-140, PLAT-324, PLAT-354

What exists today (verified 2026-10-05)

• The saved conversation file (.../conversation/.../session-<id>-conversation.json) is rebuilt from two sources: what the platform saved while the chat ran, and the coding CLI's own transcript (Codex rollout, Claude JSONL ...). claude_native_transcript_sync.go (refreshLatestBuilderConversationFromNativeTranscript, mergeBuilderConversationHistory, builderConversationMergeRefs) aligns them with a longest-common-subsequence step and appends what looks missing (log line [CHAT_HISTORY] Native transcript catch-up: merged N missing message(s)). It runs when a chat is loaded (sync_native_transcript=1) and after retained turns.
• It exists because a message sent into a running turn (live input) is saved at once but its reply cannot be, and because reconnects after a provider restart need the history back (commit dcac66a6e, 2026-09-17; native terminal mode 2026-09-29/30).
• The structured-events migration (PLAT-354) made COMPLETION structured; it did not remove this catch-up. The chat itself shows structured events from structured-chat-events-v2.sqlite.
• No decision, commit or open PR removes or replaces it.

Why it keeps hurting (same code, five tickets)

• PLAT-341: the alignment's cost saturated the RTS agent CPU (fixed by materialising keys once).
• PLAT-178 / PLAT-140 / PLAT-324: chat UI behind the terminal, a resume rebuilt from partial state, tabs losing their conversation.
• PLAT-518 (2026-10-05): two quick messages into one Codex turn gave you, you, reply saved versus you, reply, you, reply in Codex's file; a tie in the alignment appended the second message again, shown as a stuck, unanswered duplicate.

Proposal (not decided)

1. The event store is the one source of conversation history: user messages, replies, tool calls are saved from the structured events the chat already shows; the live-input path saves its reply itself.
2. Rebuild a chat after a restart or provider change from the event store, not from the CLI's transcript.
3. Delete the native transcript merge and the alignment; keep a read-only recovery tool for old conversations if needed.

Risks and what to check first

• The merge carries rules for old edge cases (replayed tails after provider restarts, stale tails, interleaved replies, structured tool entries); each needs an end-to-end check on every CLI (Claude, Codex, Cursor, Muse, Agy, Pi) with a restart and a provider switch.
• Conversations already saved may contain duplicates (PLAT-518 added none retroactively).
• It changes how history is restored; needs the owner's decision and a staged rollout (local first).

Decision (owner, 2026-10-05)

What the platform saved is the chat's history. What Claude/Codex save in their own transcripts is for debugging (and chat debugging) only, never merged back.

Done

• Found: since 2026-09-21 the platform already saves each reply itself, in order (structured_completion_persistence.go); the merge only ran as a "repair" on every chat open, resume, builder restore and after retained turns. In the PLAT-518 case both replies were in the platform's own events (8671, 8672).
• Removed the four triggers (polling.go sync_native_transcript, chat_history_routes.go resume, workflow_builder_session_routes.go restore and its Slack refresh limit, server.go retained-turn fallback and the startup recovery journal), the merge/LCS/stale-tail code, the PLAT-518 patch, and their tests (~1,900 lines).
• Kept: messages typed straight into the CLI terminal (native terminal observer) and the read-only proof that an uncertain send never arrived (canRetryUncertainChatSubmission).
• A retained-turn completion with no final reply is now logged ([CHAT_HISTORY] Retained turn completion carried no final reply to save), not repaired.

Left

• Done 2026-10-05: owner tested locally after restart; works.
• Frontend still sends sync_native_transcript=1; the server ignores it. Remove after the test.
• Check one Cursor and one Claude chat (send, reopen, nothing missing); watch for the "no final reply to save" log line.

Register notes

PLAT-525, open (proposal): one source of truth (the event store) instead of merging the CLI's transcript; decision needed.