Abstract
Starting from July to September 2026, numerous Codex users reported observable performance decay during long‑running interactive tasks. Session context would fill up, triggering automatic context compression, forgetting constraints, and abnormal token consumption. Many users initially attributed these symptoms to intrinsic model defects. However, publicly available GitHub issue threads, release notes, and official statements indicate three overlapping root causes: artificially capped effective context window for subscribed GPT‑5.6 Sol instances, flaws within the automatic compression workflow, and cumulative overhead from tool‑call outputs and background summarization jobs. This article reconstructs the three‑month incident timeline, explains hidden configuration constraints, breaks down compression‑loop mechanics, and provides actionable operational steps using /status inspection, config.toml tuning and new context‑management flags introduced in version 0.153. For engineering teams running multi‑model mixed workloads, an API gateway such as 4sapi can simplify endpoint switching and traffic governance when mixing different LLM backends in agent workflows.
1. Observable Phenomenon: Why Codex Appears to “Grow Sluggish” Over Time
Developers working inside Codex CLI or desktop clients on multi‑hour tasks frequently encounter a consistent set of symptoms. Users see the log message “Context compacted” appear on screen. After compression completes, previously applied constraints such as “only read from src directory” get partially dropped. Identical test failures trigger repeated re‑investigation because earlier reasoning records have been condensed and stripped of fine‑grained details. Token consumption metrics also accelerate noticeably on five‑hour usage windows.
These visible client‑side behaviors were widely discussed in technical community articles, with consensus pointing toward accumulated prompts, code snippets, tool returns and invocation logs. As token volume approaches the upper limit, model attention dilutes, and quality degrades before formal compression executes.
Nevertheless, user‑level observation cannot fully explain the sharp surge in community complaints starting in July 2026. Critical facts emerged from dozens of high‑vote issues filed within the openai/codex GitHub repository, exposing two less‑obvious root causes.
2. Core Fact One: Subscribed Deployments Are Limited to 258K Effective Tokens, Not 1 M
On July 9, a Pro‑plan user posted raw service‑side model metadata obtained via the Codex client. The configuration returned for GPT‑5.6 Sol showed context_window: 372000, max_context_window:372000, and effective_context_window_percent:95.
Calculated from these raw values, the theoretical effective window would be 353 400 tokens, matching the approximate “353K” figure shown on desktop UI. Automatic compression thresholds were set at 90 % of raw capacity, equivalent to 334 800 tokens. Public OpenAI documentation for GPT‑5.6 Sol advertises 1 M input tokens and 92.2 K output tokens.
Four days later on July 13, subsequent Codex CLI 0.144.1 updates altered the returned metadata fields. The real effective context window dropped to 258 400 tokens. Cross‑validation with codex debug models outputs and App Server usage event logs confirmed this represented a server‑side enforced limit rather than a local display bug. Local configuration attempting to override model_context_window to 1 M would be silently overwritten by remote service metadata.
This GitHub issue received 62 up‑votes before being closed. Comment threads highlighted major pain points: auto‑compaction triggers without advance user notification. Multiple related feature‑request tickets remained open: 168 requests demanded native 1 M context‑window support, 23 requested restoration of the original 372 K setting or toggle switches, and 112 wanted exposed configuration knobs for auto‑compression parameters.
This constitutes the primary layer of context‑overflow risk. Production subscribers receive only roughly one‑quarter of the context capacity advertised in API specifications. The Codex client respects server‑side metadata, so local tweaks cannot bypass this restriction. A second, more concealed factor further worsens resource exhaustion.
3. Core Fact Two: Compression Itself Consumes Context Budget
On August 22, Thibault Sottiaux, lead for Codex and ChatGPT teams, addressed public complaints about excessive token burn on X platform. He publicly confirmed three verified failure vectors for long‑session scenarios. First, image‑rich dialogues degraded compression efficiency. Second, Computer History functionality exhibited abnormal resource consumption for P95 heavy‑usage users. Third, title‑generation logic consumed unexpected token volume. Rollbacks were deployed globally during Pacific early‑morning hours on August 22 to reset user quotas and mitigate the bugs.
The mechanics of compression introduce subtle overhead. Codex compresses old conversation history to free capacity, yet prior releases retained embedded image payloads inside compressed segments. Even after shrinking text volume, payload size remained substantial, which could trigger another compression cycle shortly afterward. Each compression iteration itself expends context quota.
Version 0.150.1, released August 27, revised compression logic to count embedded images formally toward token budgets and evict image assets when necessary. Further fixes landed in 0.152.0 on September 1, addressing “Guardian” history overflow and implementing output‑token limits for long document excerpts.
A third under‑documented source of context inflation arrived with release 0.151.0: automatic recap triggers. After at least three dialogue turns or three minutes of user inactivity, a background thread launches structural summarization jobs against full conversation history. Community‑raised issues attracted 41 up‑votes within two days, prompting OpenAI to add the auto_recap toggle in version 0.153.
Combined together, the full failure chain becomes clear. Actual usable window sits at 258 400 tokens. Tool outputs and screenshot assets push usage close to the 90 % threshold. Auto‑compression activates, but image data are not purged completely. Minutes later another compression pass runs, interleaved with background recap jobs. Token budget and model attention degrade cumulatively.
4. Practical Operations: Three‑Step Workflow to Stabilize Context Consumption
All procedures below are validated for Codex CLI version 0.153; desktop configuration files and command interfaces behave identically.
Step 1: Inspect Real Effective Context Limits
Inside active Codex sessions, execute /status. The field model_context_window reports currently active effective capacity. The value N% context left shows remaining budget.
To fetch original server‑side metadata:
If output returns context_window: 258400, users encounter the documented production subscription cap. Manual local configuration edits cannot raise this hard remote limit, so task planning must respect this constraint.
Step 2: Tune config.toml to Constrain Compression and Tool Output Volume
Modify configuration file located at ~/.codex/config.toml. Key configurable parameters reference official documentation:
Setting tool_output_token_limit = 8000 truncates lengthy test outputs and grep results before they flood conversation history. For individual MCP tools generating large payloads, further override output_token_limit separately under mcp_servers.tools. This parameter is among the most effective controls against runaway context usage.
Step 3: Trigger Manual Compression or Reset Session
Run /compact at natural task breakpoints, explicitly specifying what information should persist: for instance “Summarize current requirements, preserve modified file list, discard failed test logs”. Manual compression yields far more predictable outcomes than automatic background routines.
When a subtask finishes completely, start a fresh session via /new. Transfer only brief key takeaways under 300 tokens to the new session.
Users with ChatGPT Plus, Pro or Pro‑Lite accounts can enable experimental context‑management mode available starting in 0.153.
When enabled, Codex no longer condenses full history into single monolithic summaries. Instead it maintains searchable notes across isolated context partitions. Legacy conversation records remain retrievable. New context_notes objects store condensed takeaways. This experimental function does not support API‑key self‑hosted providers or custom temporary structured pipelines, and is disabled by default.
Engineering teams handling mixed‑provider agent workloads frequently adopt supplementary traffic‑splitting patterns. Offload exploratory, disposable one‑shot subtasks to cheaper secondary models and pass only key conclusions back to the primary Codex session. 4sapi provides OpenAI‑compatible API interfaces, allowing teams to configure base‑url endpoints within a single profile to implement such traffic diversion. Teams with steady monthly model consumption can also evaluate enterprise token‑plan offerings for multi‑model mixed deployments.
5. Context‑Management Roadmap under GPT‑6 Astra
OpenAI’s official response to Codex context‑limitation issues is not simply lifting window capacity to 1 M tokens, but fundamentally changing context representation mechanisms. As announced with the September 3 release of GPT‑6 Astra, high‑capacity storage shifts toward note‑based retrieval‑oriented memory, reducing reliance on raw token window expansion. Codex will adopt comparable note‑oriented memory patterns for long‑duration agent tasks. Context windows will still fill, yet the system differentiates which information must stay resident and which can be off‑loaded into searchable note storage.
For existing Codex users, three major user requests remain unfulfilled as of September 2026: native 1 M context window rollout, official disable switches for automatic compression, and granular visual indicators for real‑time context utilization. Note‑based memory features demonstrated in GPT‑6 Astra have not yet been ported to mainstream Codex clients.
All analysis draws from openai/codex repository issues and release logs, official Codex configuration documentation, and public statements published by Thibault Sottiaux on X platform. Codex receives frequent iterative updates; exact behaviors should always be cross‑checked against local /status output and version‑specific reference documents.
6. Conclusion
The performance decline observed during extended Codex sessions is not purely caused by model‑inherent reasoning decay. It arises from a combination of server‑side context‑window caps, poorly‑tuned automatic compression workflows, unconstrained tool‑output payloads and background summarization overhead.
Users can mitigate most pain points today through inspection workflows, explicit config.toml tuning, manual compression and session reset strategies. Looking ahead, note‑oriented memory architectures represented by GPT‑6 Astra point toward the industry direction for long‑running agent systems, moving away from endlessly expanding raw token windows.
International access: https://4sapi.com
Domestic access: https://4sapi.cn
