fix: stop Codex app-server transcript repeating older messages after idle

[dimavedenyapin]

Problem

In the Codex agent transcript, older messages get repeated after the last ongoing idle — every time a Codex turn completes and the thread goes idle, previously-shown messages are re-appended to the transcript, so the conversation keeps growing with duplicates.

Root cause

CodexAppServerAdapter.reconcileCompletedTurn() (added in #412) re-lists the entire thread via thread/items/list on every turn/completed, to make sure the final assistant message isn't lost. Re-buffering the whole history is only safe if it's idempotent — and it wasn't:

• The dedup guard bufferedThreadItemIds only registered an item when asString(item.id) was truthy.
• extractItemId() fell back to `item-${Date.now()}` when no id was found.

But many Codex thread item types have no top-level id — only a call_id, or nothing at all. Verified against real Codex rollout logs (~/.codex/sessions/.../rollout-*.jsonl) for the referenced task pf-web: analyse slow mongodb queries (thread 019f4a0d-e8c7-73f3-8592-6b5f312d1b5b, 12 turns):

item type	has id
function_call_output (168×)	❌ (only call_id)
custom_tool_call_output	❌
tool_search_output	❌
user/developer input message	❌

So on every reconcile those id-less items bypassed the dedup, got re-buffered, and were re-emitted to the renderer with a fresh item-${Date.now()} part id that neither seenPartIds nor the renderer's id-keyed dedup could collapse → the transcript repeated older messages after each idle.

Fix

Make reconcile buffering idempotent for every item:

• deriveItemIdentity() derives a stable identity from id / itemId / call_id.
• computeThreadItemKey() falls back to a deterministic content hash (turn + type + role + text) when there's no id-like field.
• bufferReconciledThreadItem() (shared by bufferThreadItems and bufferAllThreadTurns) skips already-buffered items by that key and forwards it as itemId, so the derived part id stays stable across passes.
• extractItemId() no longer uses Date.now() — it uses the same stable identity, then a deterministic content hash.

Reproduction / test

Added a regression test that fires three turn/completed reconciles returning the same id-less tool output and asserts it is buffered exactly once (was 3× before the fix).

• New + existing codex-app-server-adapter tests: 19 passing
• Full pre-commit suite: 1533 passing

🤖 Generated with Claude Code