Skip to content

About

Keep long, image-heavy Codex and Claude Code chats fast: earlier images become ID placeholders the model fetches back on demand, even after compaction. Local, cache-friendly, MIT.

Topics

Resources

Contributing

Security policy

Stars

814 stars

Watchers

1 watching

Forks

Repository files navigation

Context Attachment Manager icon

Context Attachment Manager

Keep long, image-heavy chats in Codex and Claude Code fast.
Earlier images become one-line placeholders with an ID. When the model needs one again, it fetches the original by its ID, also after the chat is compacted.

Release Platform: Windows, macOS, Linux License: MIT

English · 简体中文 · 日本語 · 한국어 · Español · Português (Brasil) · Français · Deutsch

A looping clip in three parts. One: every message sends all earlier images again; 24 keyframes pile up and the request grows to 15.6 MB, turning red. Two: with the plugin, earlier images fold into one-line ID placeholders such as IMG-004 Street, only the pinned character sheet still goes out, and the request drops to 1.3 MB. Three: the model fetches IMG-007 by its ID; the original is added at the end of this turn, its placeholder stays put, and everything above is read from the cache.

42.4 MB → 0.12 MB

One request in a Codex chat
with nine keyframes

20/20 vs 8/20

Questions about the images
after compaction (Sonnet 5.5)

About the same cost

Largest request: 41,859
vs 110,671 tokens

Get started in three steps

1. Install. Send this to your agent. It identifies its platform and installs the right setup:

Install Context Attachment Manager for me, following https://github.com/chipfighter/context-attachment-manager/blob/main/INSTALL.md

Then restart the agent (in Claude Desktop, start a new Code session). To do it by hand, see installation, updates and uninstall.

2. Work as usual. The plugin is on by default. Images from earlier turns go out as one-line placeholders; new images and the references you pin go out as they are. A placeholder looks like this:

[Image IMG-004 omitted | shot-1.png | generated image | turn 2 | 505×505]
Features: place: Street | Qinghe | medium shot, umbrella, lanterns
For the original, call cam_view_image with "IMG-004".

3. The model fetches by ID. When it needs to look at an earlier image again, it fetches the original by its ID, also after compaction. Nothing is deleted, and your images and chat history stay on your computer.

The panel: browse by character and place

The panel. Left: browsing by place, with six place cards such as Street, Railing and Courtyard. Right: the character Qinghe opened, her character sheet first, then her shots on the street. At the bottom, the next request is 1.3 MB instead of 15.6 MB, 91% smaller.

The plugin works without the panel; open it when you want control:

  • Browse by character and place: images are sorted into characters, places, screens and documents, with one card per name. Open a character to see their design images first, then the shots they appear in, place by place.
  • Find an image by a detail or a comment you remember, like "rainy night" or "best face".
  • Keep a reference in view: open an image and check Send every turn.
  • Choose images yourself: turn off Auto-select and pick what goes with your next message.

In Codex, open Context Assets in the task's side panel. In Claude Desktop's Code tab, the panel opens automatically when there are images; type /cam to bring it back.

Results

  • Codex: a chat with nine keyframes grew to 42.4 MB and could not be sent after a restart or a compaction; with the plugin the same requests were 0.12 MB.
  • Claude Code (Sonnet 5.5): in a 16-turn short-film task whose images exist only in the conversation, we compacted the chat and then asked about the images. With the plugin it answered 20 of 20 questions correctly; without it, 8 of 20. Its largest request was 41,859 tokens instead of 110,671, at about the same cost.

See the Codex comparison, the Claude Code compaction comparison and the Claude Code long-task comparison for the results and reproduction steps.

How it works: keep the cache working

Model services cache the start of a chat. When a request begins exactly like the one before, that part is read from the cache, which is faster and much cheaper. When something early in the chat changes, everything after the change is processed again. The plugin makes its change as late in the chat as it can:

Four ways to handle a long chat with images, each drawn as the next request. Sending everything keeps the cache but sends every image again. Shrinking the images or dropping early history makes nearly the whole chat be processed again. With the plugin's placeholders, only the rest of the previous turn is processed again.

  • A placeholder is written once: an image becomes a placeholder in the turn after it appeared, and stays the same from request to request.
  • Fetching changes nothing earlier: the original is added at the end of the current turn and its placeholder stays where it is, so earlier turns are still read from the cache.
  • The model gets the original back: not a smaller copy, and not a summary of it.
How fetching works, and what it costs

When the model fetches IMG-001, the original is attached at the end of the current turn and the placeholder in turn 1 stays as it is, so every earlier turn is read from the cache. In the next turn the fetched copy becomes a one-line note, and only what came after it is processed again.

This is not free. After a turn with images, the rest of that turn is processed again once, and the model writes a short line for each new image. Changing an old image in the panel (unchecking it, checking it again, Send every turn) changes the chat at that image, so what follows it is processed again once. In our long-task comparison in Claude Code, 87% of the input was read from the cache with the plugin and 91% without it.

An agent's own compaction, and clearing old tool results on the API side (Anthropic's context editing, for example), remove early content so that what follows is processed again: the third row of the first figure.

Works with

Codex and Claude Code are tested, and Codex and Claude Desktop also have the panel. The other agents below are supported through adapters written from their code and docs, but we have not tested them live yet. If you use one of them, please tell us how it went, whether it worked or not.

Agent What it is Support
Codex OpenAI's coding agent, desktop and CLI Auto-select + panel, tested
Claude Code Anthropic's coding agent, desktop and CLI Auto-select + panel (desktop), tested
WorkBuddy Tencent's desktop AI work assistant, with CodeBuddy inside Auto-select, not tested yet
OpenCode Open-source terminal coding agent for any model provider Auto-select, not tested yet
Gemini CLI Google's open-source terminal agent Auto-select, not tested yet
pi Minimal, extensible terminal coding agent Auto-select (pi extension), not tested yet
DeepSeek Harness DeepSeek's agent harness where everything is a plugin Auto-select (dsh plugin), not tested yet
Qwen Code Qwen's open-source terminal agent Auto-select, not tested yet
Cline Open-source coding agent for VS Code, JetBrains and the terminal Auto-select, not tested yet
Kilo Code Open-source coding agent for VS Code, JetBrains and the terminal Auto-select, not tested yet
goose Open-source agent for desktop and terminal, extended with MCP Auto-select, not tested yet
Cherry Studio Open-source desktop chat client for many models Auto-select, not tested yet

Setup for these agents: INSTALL.md. To connect another agent, see ADAPTERS.md.

FAQ

Why do Codex and Claude Code get slow in a long chat with many images?

The agent sends every earlier image again with each message, so a long chat with many images gets slow and can stop going through. In Codex, a chat with nine keyframes grew to 42.4 MB and could not be sent after a restart or a compaction; with Context Attachment Manager the same requests were 0.12 MB.

Does Context Attachment Manager break the prompt cache?

Only a little. A placeholder is written once and then stays the same, and a fetched image is added at the end of the current turn, so earlier turns are still read from the cache. After a turn with images, the rest of that turn is processed again once. In our long-task comparison in Claude Code, 87% of the input was read from the cache with the plugin and 91% without it.

How is it different from compaction, or from shrinking the images?

Compaction and dropping early history remove the early part of the chat, and shrinking the images changes it. Either way nearly the whole chat is processed again, and the model is left with a summary or a smaller copy. With Context Attachment Manager, only the rest of the previous turn is processed again, and the model gets the original image back by its ID.

Does the model still know the earlier images after compaction?

Yes. Every image has an ID such as IMG-001, and the model notes one line on each: its category, a name, and the characters in it. These stay in the placeholders after compaction, and the model fetches the original by its ID when it needs to look again.

Does it handle the many screenshots of computer use and browser use?

Yes. When tools return many screenshots in one turn, the older ones become placeholders in batches and the newest few go out as they are. The model can note a run of similar screenshots in one line and fetch any of them back by its ID.

Are my images deleted, or sent anywhere else?

No. The plugin runs on your computer. Your original images and chat history stay as they are, and requests go to the service your agent already uses.

Which agents does it work with, and is it free?

Codex and Claude Code are tested. WorkBuddy, OpenCode, Gemini CLI, pi, DeepSeek Harness, Qwen Code, Cline, Kilo Code, goose and Cherry Studio are supported through adapters that we have not tested live yet. It is free and open source under the MIT license.

How do I install it? Do I need a terminal?

No. Send your agent the one sentence in Get started above: it identifies its platform and runs the setup itself, and you only restart the agent afterwards (in Claude Desktop, start a new Code session). In Codex you can also install it from the Plugins page with a few clicks. If you prefer to do it by hand, it takes one command in Codex and two claude commands in Claude Desktop; see INSTALL.md.

Changelog · Contributing · MIT License

About

Keep long, image-heavy Codex and Claude Code chats fast: earlier images become ID placeholders the model fetches back on demand, even after compaction. Local, cache-friendly, MIT.

Topics

Resources

Contributing

Security policy

Stars

814 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages