Stream error: “Your input exceeds the context window” when Codex streams large tool output (e.g., 300+ tests)

What feature would you like to see?

Summary

• When Codex CLI streams large command output back to the model (e.g., running >300

tests), the session repeatedly fails with “Your input exceeds the context window” and
disconnects after retries.

Steps To Reproduce

1. Open Codex in a repo with a large test suite (e.g., >300 tests).
2. Ask Codex to run tests, for example:

• make test or
• bun x vitest run

1. Observe Codex CLI logs during the run.

Expected Behavior

• Codex should handle large tool outputs gracefully:
• Chunk or truncate output.
• Summarize long output before sending to the model.
• Continue the session without disconnecting.

Actual Behavior

• Codex retries several times and then fails with a context window error.

Observed Logs

⚠️ stream error: stream disconnected before completion: Your input exceeds the context
window of this model. Please adjust your input and try again.; retrying 1/5 in 202ms…

⚠️ stream error: stream disconnected before completion: Your input exceeds the context
window of this model. Please adjust your input and try again.; retrying 2/5 in 418ms…

⚠️ stream error: stream disconnected before completion: Your input exceeds the context
window of this model. Please adjust your input and try again.; retrying 3/5 in 810ms…

⚠️ stream error: stream disconnected before completion: Your input exceeds the context
window of this model. Please adjust your input and try again.; retrying 4/5 in 1.621s…

⚠️ stream error: stream disconnected before completion: Your input exceeds the context
window of this model. Please adjust your input and try again.; retrying 5/5 in 3.28s…

■ stream disconnected before completion: Your input exceeds the context window of this
model. Please adjust your input and try again.

Minimal Repro (alternative)

• Any large output triggers it, not only tests:

1. Start Codex in any folder.
2. Ask it to run a command that prints a lot, e.g.:

• yes "x" | head -c 3000000

1. Codex attempts to stream the entire output to the model and hits the context

limit.

Environment

• Codex CLI version: 0.42.0
• OS: macOS 15.6.1
• Node version: 24.2.0
• Bun version (if applicable): 1.2.2
• Model: gpt-5-high
• Terminal/shell: zsh in ghostty

Impact

• Prevents using Codex for CI-like tasks (tests, lint, builds) where output can exceed

the model’s token budget.

• Causes session interruptions and lost context.

Workarounds Tried

• Use less verbose reporters: bun x vitest run --reporter=dot or --silent.
• Pipe output to a file and ask Codex to summarize the file:
• bun x vitest run --reporter=json > test-output.json and then “summarize test-

output.json”.

• Tail only the last N lines: bun x vitest run | tail -n 400.

Hypothesis

• Tool result payloads are forwarded to the model without enforcing a token/byte budget

or auto-summarization. Large outputs push the prompt beyond the model’s context window,
causing the stream to disconnect.

Suggested Fixes

• Add a hard cap on tool-result size sent to the model with:
• Automatic chunking + iterative summarization.
• Configurable truncation strategy (head/tail/middle) and max bytes/tokens.
• Provide a setting to store full tool outputs as artifacts/files and send only

summaries + links/paths into the chat context.

• Consider auto-dropping or summarizing older conversation turns when approaching

context limits, with a visible notice.

Additional Context

• Reproduces consistently on large test runs (300+ tests) in our Next.js monorepo.
• Happy to provide a minimal reproduction repo or additional debug logs if helpful.

If you’d like, I can open this issue for you using the GitHub CLI. Otherwise, paste
the above into a new issue at https://github.com/openai/codex/issues/new and add the
missing environment details.

Are you interested in implementing this feature?

_No response_

Additional information

_No response_