Bedrock Runtime: inline image history still overflows after successful compaction #46247

What version of Codex CLI is running?

codex-cli 0.154.0 (latest stable release checked on 2026-09-18).

What subscription do you have?

Amazon Bedrock API, pay-as-you-go. This is the built-in amazon-bedrock-runtime provider, not a ChatGPT subscription.

Which model were you using?

global.openai.gpt-6-astra. Similar failures were observed in a separate global.openai.gpt-5.6-sol session. The controlled reproduction below keeps the model, AWS identity, and endpoint Region unchanged.

What platform is your computer?

Linux 6.17.0-1020-aws, x86_64; Ubuntu 24.04.

What terminal emulator and version are you using?

CLI through tmux 3.4. Also reproduced through the installed Codex app-server's stdio API, so this is not limited to terminal rendering.

What issue are you seeing?

Image-heavy conversations can fail with the following Bedrock Runtime response while still well below Codex's configured token-based compaction threshold:

{
  "error": {
    "message": "Input is too long.",
    "type": "invalid_request_error",
    "param": null,
    "code": "validation_error"
  }
}

Attempting compaction can fail with the same error:

Error running remote compact task: {"error":{"message":"Input is too long.","type":"invalid_request_error","param":null,"code":"validation_error"}}

Reducing the token threshold alone did not resolve this. With model_auto_compact_token_limit = 50000, a fresh thread accepted two PNGs per turn. Turns 6 through 10 each completed a remote compaction, but submitting images 21–22 failed with the same error. The last successful request used 74,965 input tokens. Completed compacted records continued retaining the historical inline images.

Controlled observations

A/B measurements were collected on 2026-09-17.

A separate direct Responses API test used the same 23 images, model, Region, and image dimensions. The request only asked for a short acknowledgement, with store=false. The image set is private and is not attached.

The lossless conversion was checked by comparing decoded RGBA pixel buffers. These observations point to an encoded-payload constraint that token accounting alone does not protect against. They do not establish an exact or universal Bedrock request-size limit.

In affected full conversations, the latest active history contained approximately 20–23 inline images and 30–36 MB of encoded image content. One affected conversation's last successful full-history request used only 153,997 input tokens.

What steps can reproduce the bug?

1. Use Codex 0.154.0 with the built-in amazon-bedrock-runtime provider and a supported multimodal model.
2. Keep the model, credentials, and endpoint Region unchanged throughout the test.
3. Start a fresh thread and submit two reasonably large PNG screenshots per turn, asking for a short acknowledgement without tool calls. Alternatively, accumulate screenshots through image-producing tools.
4. Continue until the active inline image payload grows into the tens of megabytes.
5. Observe a Bedrock validation_error even though token usage has not reached the configured context limit.
6. Try /compact; it can fail with the same oversized input.
7. Repeat in a fresh thread with the auto-compaction threshold lowered to 50,000 tokens. In the observed run, remote compaction completed five times but retained enough inline images for the 22-image request to fail anyway.

For the controlled API comparison, use POST https://bedrock-runtime.us-east-1.amazonaws.com/openai/v1/responses, a user content array of input_image data URLs, the same detail values, and the same short text instruction. Compare original PNG data URLs against losslessly re-encoded copies of those images.

What is the expected behavior?

• Track encoded image/request bytes separately from model token usage.
• Trigger compaction before the provider's encoded-payload constraint is exhausted.
• Ensure that the retained post-compaction image payload fits a bounded budget; a successful compaction should make continued progress possible.
• Provide a recovery path for already oversized histories without requiring users to edit private rollout files or encrypted reasoning content.
• Keep image content, text, and associated history metadata consistent during any transformation.

Technical observations

On the rust-v0.154.0 source:

• Image token estimation intentionally discounts raw Base64 bytes. This is appropriate for vision-token accounting but does not enforce a separate transport budget.
• core/src/compact_remote_v2.rs uses a RETAINED_MESSAGE_TOKEN_BUDGET of 64,000. An image set can fit that token budget while its encoded representation remains too large for the provider.
• Bedrock returns validation_error for the observed failure, rather than context_length_exceeded.

Configuration used in the lower-threshold reproduction:

model_provider = "amazon-bedrock-runtime"
model = "global.openai.gpt-6-astra"
model_context_window = 872000
model_auto_compact_token_limit = 50000
model_auto_compact_token_limit_scope = "total"
model_reasoning_effort = "xhigh"

[model_providers.amazon-bedrock-runtime.aws]
region = "us-east-1"

Codex reported an effective model context window of 828,400 tokens. No custom provider proxy was used.

Related reports

• Unbounded image history payload can make threads unrecoverable with 413 #30441: unbounded image history payloads can make threads unrecoverable with HTTP 413.
• Desktop compaction repeatedly embeds full image base64 in compacted checkpoints #23257: compacted checkpoints repeatedly retain full inline image Base64.
• Codex remote compaction deadlocks when input_image payloads remain in compacted replacement history #24388: retained images can make remote compaction itself fail.
• Account for file images in context budgets and Guardian reviews #46072 addresses file-image references; this report concerns inline Base64 images on Bedrock Runtime.

This report adds a current stable Bedrock Runtime reproduction, the validation_error response shape, a same-pixel encoding comparison, and a negative control showing that earlier token-based compaction can still retain an oversized image payload.

Privacy and validation limits

Only sanitized structural measurements are included. No screenshots, private prompts, account identifiers, credentials, or raw rollouts are attached. The exact provider byte limit and behavior on other providers have not been established.

A local experimental implementation allowed the 23-image batch followed by another 32 images to continue, with automatic compaction, manual compaction, and a restart/recall check. That implementation is still under validation and is not being presented as a completed upstream fix.