Name each Codex conversation, and draw every one that is working

[Splinter0000]

Name each Codex conversation, and draw every one that is working

Every Codex row said "Codex". The monitor drew one row at most — the newest
rollout or the newest desktop thread, whichever moved last — while the Claude
rows beside it named every session. Two conversations running at once drew as
one, and nothing on the notch said what either was about.

Codex does name its conversations, in the same threads row the rollout path
was already being read from. Independent of #321; the two touch no file in
common.

What a row is called

1. Codex's own threads.name — every VS Code thread on the machine this was
   built on has one, a moment after its first turn.
2. Otherwise the request itself, first line, capped at 48 characters. codex exec never names a thread, so this is permanent there. When Codex has put
   attached files or in-app browser state ahead of the request, it marks
   where the request begins with a ## My request line; the label is taken
   from under that marker (or from after a leading <tag>…</tag> block), so a
   row never reads "# Files mentioned by the user:". Numeric character
   references ( ) are decoded. A request that opens with a #hashtag of
   the user's own is left alone — only Codex's own markers are taken off.
3. Then the folder, as a Claude session falls back to.
4. Then "Codex", as before.

One row per conversation — not per thread

Sub-agents are folded into the conversation they work for. Codex can run half a
dozen for one request, each with a rollout of its own, and the parent, waiting
on them, may write nothing for minutes. The row keeps its root's name and spins
for as long as any of it is moving.

That includes guardians, the helpers auto-review runs to vet each action.
Their threads.source is {"subagent":{"other":"guardian"}} — no parent at
all — and taken for a conversation, each review drew a row of its own, named
after the guardian prompt, that chimed "Complete" while the real request was
mid-turn (six false chimes in forty minutes in one real session). The parent is
in the opening session_meta line of the guardian's rollout, so it is read
from there, once per change of the store and only for helpers the store has no
parent for. A helper whose conversation cannot be found is not drawn at all.

A row's state is only ever its root's. A helper writing task_complete is not
the request finishing.

Also

• The desktop app's copy of a conversation. local_thread_catalog shares
  thread_id with threads.id when the same conversation is open in both —
  ten of forty-three here. The old newest-of-two rule hid that; it is now drawn
  once.
• since means what AgentSession says: when the row entered its state.
  A rollout's last write moves every second while a conversation works; as
  since it made two busy rows swap places every few seconds and made every
  tick look like a change. The monitor now keeps the first sighting of a row in
  a state until the state changes (settled(_:entered:)).
• A rollout is parsed again only when its (mtime, size) moves. The parse
  walks up to a megabyte backwards on the main actor, and it now runs per live
  conversation rather than once per tick.
• Older stores. Columns are asked for only when pragma_table_info says the
  table has them, so a store from before name existed reads exactly as before.
• Ids in queries. A parent id comes out of JSON another program writes; it
  is checked (isPlainID) before it is put in a query.
• A debug line on change, as the Claude monitor has, so what the notch thinks is
  running can be seen with log stream.

Notes for review

• CodexStore.newestRollout and CodexStoreCache.newestRollout no longer have
  a caller in the app. I left them and their tests in place rather than delete
  your tests in the same change — happy to remove them here if you prefer.
• A new codex exec thread can read as its folder for a second or two before
  Codex writes its preview; it then takes the request's first line.
• I did not pre-filter the per-tick stats by updated_at_ms. It would cut
  ~30 stats a tick to a handful, but it makes liveness depend on Codex updating
  that column during a long turn, and a stat is microseconds.

Checked against a real ~/.codex

194 threads, 133 of them helpers: all 133 folded into one of 61 conversations,
none dropped for want of a parent, no row labelled with injected context. Live,
with the app running: two conversations working at once drew as two named rows,
and the one that finished went busy → success.

Tests: 25 new in CodexConversationTests — naming and its fallbacks, the
request under Codex's markers, one row per conversation, sub-agents and
guardians folded, a helper finishing not finishing the row, an orphaned helper
not drawn, the desktop copy drawn once, an older schema, the stable order and
clock, and a rollout not parsed again while unchanged. Full suite 1759 / 0
failures on this branch alone.

---

🤖 Generated with Claude Code