openai / openai/codex

[Desktop][macOS] 7-day usage remaining jumped from 48% to 23% after context/memory feature change

Open
#38,191 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

app bug context memory rate-limits
Dominant language
Rust
Stars
125k
Forks
19.5k
PR merge metrics
PR metrics pending

Description

What issue are you seeing?

The Codex Desktop 7-day usage indicator appeared to drop instantly from 48% remaining to 23% remaining, despite very little intentional use that day.

Local session telemetry shows this is the primary 7-day rate limit, not the model context-window indicator:

  • Earlier record: used_percent: 52.0 (48% remaining)
  • Current record: used_percent: 77.0 (23% remaining)
  • window_minutes: 10080
  • Model context window is separately reported as 258,400 tokens

The first token_count event in a new diagnostic task already reported used_percent: 77.0, before its diagnostic tool calls. Therefore the apparent 25-point drop was not caused solely by that task. It looks like stale UI state being refreshed all at once, delayed server-side accounting, or incorrect attribution.

A context/memory/window-capture-related option had been enabled the previous day. Relevant observations:

  • Later session metadata records show memory_mode: enabled.
  • The currently persisted auto-context setting reports disabled.
  • The affected session record does not contain actual window-context, screen-capture, OCR, or image input items.
  • New turns do contain a large automatically injected payload, including plugin metadata, project instructions, skill metadata, memory guidance, and tool schemas.
  • In one short diagnostic session, cumulative token events progressed from approximately 29k to 61k to 104k within about 32 seconds, with a large cached-input component. The weekly meter stayed at 77% during those calls.

Please investigate:

  1. Whether the remaining-usage UI can remain stale and then jump when a new task starts.
  2. Whether memory/context/window-capture features change quota accounting or cause repeated system/tool context injection.
  3. Whether cached or automatically injected input is attributed correctly.
  4. Whether the UI can distinguish weekly quota remaining from model context-window remaining and expose per-task attribution.
Steps to reproduce
  1. Enable the context/memory/window-capture-related feature.
  2. Leave Codex mostly idle or perform only a small amount of intentional work.
  3. Start or resume a project task.
  4. Observe whether the weekly remaining-usage percentage refreshes abruptly.
  5. Compare the UI with local JSONL token_count events.

The issue may be intermittent because it appears to involve delayed refresh or accounting.

Expected behavior
  • Weekly usage should update promptly enough that a 25-point change is attributable to visible activity.
  • Enabling memory/context support should not unexpectedly consume a large portion of the weekly quota.
  • Automatically injected and cached context should have transparent, correct quota treatment.
  • The UI should clearly label weekly quota versus current context-window capacity.
Environment
  • Product: Codex Desktop on macOS
  • macOS: 26.5
  • Architecture: arm64
  • Codex backend/CLI version recorded in session metadata: 0.147.0-alpha.6.5

No local paths, account/plan details, session identifiers, private project content, or full logs are included.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by comparing the Desktop weekly usage indicator with local JSONL token_count events, including used_percent, window_minutes, and cached-input values. Review the session metadata for memory_mode and auto-context settings, then define done as identifying whether refresh timing, injected context, or attribution explains the jump and documenting a reproducible check.

Written by the indexing model from the issue text.

Assessment

Tech stack
macos, rust
Domain
backend, desktop
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Quiet
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.