Codex Desktop Ctrl+K task search can miss known tasks by visible numeric/content terms

Open 💬 3 comments Opened Jul 25, 2026 by claell
💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of the Codex App are you using (From "About Codex" dialog)?

Current Windows merged ChatGPT/Codex desktop app, observed on 2026-07-25.

Exact version was not captured yet.

What subscription do you have?

ChatGPT account with Codex access.

What platform is your computer?

Windows desktop app.

What issue are you seeing?

Codex Desktop's Ctrl+K task search can fail to find a known task even when searching for simple terms that are visibly present in that task.

The recent example was a Codex task/report about the usage-limit meter off-by-one behavior. The task contained repeated visible references to:

  • 99
  • Codex
  • usage
  • 5-hour

However, opening Ctrl+K and searching for 99, or combining the numeric query with Codex, returned no matching tasks.

This is surprising because the task itself exists and the searched terms are short, exact, and distinctive. In previous Codex Desktop builds, task search often found this kind of task reliably, so this may be a regression in the merged desktop app's task-search index, query path, or hydration state.

What steps can reproduce the bug?

  1. Open or create a Codex task whose visible conversation content contains a distinctive numeric term such as 99, together with another common term such as Codex.
  2. Open the desktop task switcher/search with Ctrl+K.
  3. Search for the numeric term alone, for example 99.
  4. Search for a combined query, for example 99 Codex.
  5. Observe whether the known task appears in the matching task list.

Observed result:

  • The task search reported no matching tasks for the numeric/content query, even though the current/known task contained the searched terms.

Expected result:

  • Ctrl+K task search should find existing Codex tasks by visible conversation content, including short numeric terms and exact terms in recent messages.
  • At minimum, search should reliably include the current task and recent local tasks, and it should make clear if it only searches titles/previews rather than message content.

What is the expected behavior?

Task search should be a reliable recovery/discovery path for known Codex work.

If a task exists locally and contains a visible term, searching that term should either return the task or clearly explain the search scope. A no-results state is misleading when the task exists and the query term is visibly present in the task content.

Additional information

I searched existing openai/codex issues before filing this and did not find an exact report for Ctrl+K returning no matches for a known task by numeric/content terms.

Related but not exact matches:

  • #25463: project threads disappear from project views/search while local session JSONL remains readable. This report is closely related, but the symptom here is specifically Ctrl+K search missing a known task by visible numeric/content terms in the current merged desktop app.
  • #25761: the Desktop sidebar appears limited to around 50 conversations and older project chats disappear. The current issue may share the same search/list pagination or indexing layer.
  • #32182: a recent Codex task was missing from history and not searchable by exact ID after the new ChatGPT desktop update.
  • #32614: an agent-created top-level task was hidden from desktop search and Codex Mobile Remote.
  • #33579: pinned/default task listings can diverge from locally readable tasks.
  • #34183: when /wham/tasks/list failed, the merged Windows desktop app replaced healthy local Codex task state with the generic ChatGPT shell, showing that local task discovery can be affected by merged-shell hydration and remote task-list state.

This report is intentionally scoped to the user-visible search failure: Ctrl+K can report no matching tasks for simple exact terms that are present in a known Codex task.

View original on GitHub ↗

3 Comments

github-actions[bot] contributor · 1 month ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

  • #34754

Powered by Codex Action

claell · 1 month ago

Additional sanitized reproduction detail from the same Windows merged desktop app behavior.

The failure is not limited to one missing task. It appears to involve the query semantics of the Ctrl+K task search itself:

  • Searching for a single word from a task title can return the expected task.
  • Searching for two words in the same order as they appear in the title can return the expected task.
  • Searching for the same two words in the opposite order can return no results.
  • Searching for a short numeric term that is visibly present in the task, such as 99, can return no results.

This suggests the search may be using an order-sensitive substring/prefix match or a stale projection rather than tokenized, order-independent matching across task title/preview/content. It also suggests numeric terms may be ignored, filtered out, or not indexed.

Expected behavior:

  • A query containing multiple title/content tokens should match regardless of token order.
  • Exact numeric tokens that are visible in a task should be searchable.
  • If Ctrl+K intentionally searches only a limited field, such as title prefix or current loaded sidebar rows, the UI should make that scope clear instead of presenting it as general task search.

This comment is intentionally keeping the affected task details generic. The useful reproduction signal is the query behavior: word1 word2 can match while word2 word1 does not, and numeric-only queries can fail even when the number is visibly present in the task.

prcris · 6 days ago

Corroborating Windows report from Codex Desktop 26.818.31338:

Task search behaved as though it matched only the auto-generated title. A known task created through iOS Remote could be found when searching its generated title, but terms present in the task request/messages did not retrieve it.

This is especially problematic because generated titles can be truncated or only loosely represent the task. When the same task was also missing from the Recent list due to a synchronization/index issue, the only recovery path was to look up the generated title on iOS and enter that exact title on Windows.

Expected search scope:

  • generated title
  • user prompts and assistant messages
  • task summary/preview
  • project/repository
  • status and origin (active/completed, local/remote)

At minimum, the UI should clearly state the indexed fields if full conversation-content search is not available. Full-text or semantic retrieval would make task search a reliable recovery path rather than a title lookup.