[Provider compatibility] Claude Code /compact gets stuck on 502 retries with native GPT-6.1 Sol #6504

Client or integration

Claude Code CLI 2.1.288 on macOS, launched through the claude-ocx script included below. The gateway runs remotely. I haven't tried reproducing this with a fresh stock ocx claude launch yet.

Provider or upstream service

OpenAI through my ChatGPT/Codex subscription, using gpt-6.1-sol. Claude Code selects it as ocx-claude-native--gpt-6.1-sol.

OpenCodex version

@bitkyc08/opencodex 2.76.0.

Endpoint or capability

/v1/messages to native OpenAI Responses, especially input-limit errors and Claude Code's /compact command.

Current behaviour

My Claude Code session got stuck after the context grew to roughly 660k tokens. The next turn returned API Error 502 and kept retrying. I tried /compact, but that failed too, so I couldn't continue the session.

At the time, my launcher used:

export CLAUDE_CODE_MAX_CONTEXT_TOKENS=922000
export CLAUDE_CODE_AUTO_COMPACT_WINDOW=829800

Both DISABLE_COMPACT and DISABLE_AUTO_COMPACT were unset. I had copied the context values from GPT-5.6 guidance. After checking the source, I changed the maximum to 872000 for GPT-6.1 and moved compaction earlier, to 600000. The attached launcher has those corrected values.

The last successful request before the failure accepted 658,820 input tokens. I don't know the exact upstream cutoff or what caused the following request to be rejected.

I managed to recover a fork by running /compact on the same model with tools and MCPs disabled:

• Claude's recorded context went from 664,375 to 26,517 tokens.
• The native request used 632,140 input / 18,094 output tokens and returned HTTP 200.
• Compaction took about 13 minutes.
• After restoring normal tools and MCPs, /context succeeded at 85.7k / 600k.

The original history is intact. I haven't reached the new automatic compaction threshold yet.

Expected behaviour

I'd expect an input-limit rejection to reach Claude Code as an error it understands, so it stops retrying the same request. /compact should either work or explain how to recover if the compaction request also exceeds the limit.

I'd also like a supported long-context setting that leaves enough room for tools and the compaction request itself.

Minimal redacted request or reproduction

I hit this in an existing session with normal tools and MCPs enabled. I haven't reduced it to a standalone payload yet.

1. Launch Claude Code through the script below, using the original 922000 / 829800 values shown above, and select ocx-claude-native--gpt-6.1-sol.
2. In my session, the next turn failed after input grew to roughly 660k tokens.
3. Run /compact in that session. It also fails with the input rejection and Claude retries with 502 errors.

For recovery, I used the corrected launcher, forked the session, and compacted with tools/MCPs disabled. That changed the request size. I can't yet say whether the earlier compaction setting will prevent this from happening again.

Actual response or error

Claude showed API Error: 502 and retried. The associated gateway record showed status 400, invalid_request_error, terminal_sse, payload-rejected, and protocol-prelude, with this message:

The provider rejected this turn because its input exceeds the provider size or context limit. Reduce the current input or compact the conversation before retrying.

I haven't captured a complete redacted Claude-facing wire exchange, so I can't pinpoint where this turns into a retrying 502.

Upstream documentation

I couldn't find a public specification for the exact input limit on this subscription endpoint. These are the OpenCodex references I checked:

• Claude Code alias and context-window handling.
• GPT-6/GPT-6.1 context metadata: 272k default, 872k extended ceiling/max input.
• GPT-5.6 subscription context measurements: 922k ceiling.

Suggested mapping or implementation notes

Could you check how the native input-limit error reaches Claude Code through the Messages translation? #3170 and #4112 cover similar failures in the Responses/Codex paths. I haven't confirmed whether this has the same cause.

I'd also like to select 1M context for native ChatGPT-subscription models in Claude Code, with manual and automatic compaction working. Please support that wherever the upstream model/account allows it, and show the supported maximum and compaction setting for each model. I'm currently setting both values by hand in the launcher.

I've read #3958 about the existing extended-context control and #4950 about a 1M option. GPT-6.1 currently advertises 872k, and my recovery doesn't establish that it accepts a full 1M. Could you clarify what's supported for this route and whether 1M can be added?

Additional context and attachments

Here's my current claude-ocx launcher. It loads credentials from a local claude-ocx.env file; I've left that file out. The script contains variable references for the gateway URL and token, with no embedded credential values.

#!/usr/bin/env bash
set -euo pipefail

ENV_FILE="${XDG_CONFIG_HOME:-$HOME/.config}/claude-ocx.env"

if [[ -f "$ENV_FILE" ]]; then
    set -a
    # shellcheck disable=SC1090
    source "$ENV_FILE"
    set +a
fi

: "${OPENCODEX_BASE_URL:?Set OPENCODEX_BASE_URL in $ENV_FILE}"
: "${OPENCODEX_API_AUTH_TOKEN:?Set OPENCODEX_API_AUTH_TOKEN in $ENV_FILE}"

# Don't let a local Anthropic API key bypass OpenCodex.
unset ANTHROPIC_API_KEY

export ANTHROPIC_BASE_URL="$OPENCODEX_BASE_URL"
export ANTHROPIC_AUTH_TOKEN="$OPENCODEX_API_AUTH_TOKEN"

# Tell Claude Code that authentication/model access is managed by the gateway.
export CLAUDE_CODE_PROVIDER_MANAGED_BY_HOST=1
export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1
# Summarize before the observed upstream rejection near 661k tokens.
export CLAUDE_CODE_AUTO_COMPACT_WINDOW=600000
export ANTHROPIC_DEFAULT_HAIKU_MODEL=google-antigravity/gemini-3.8-flash[1m]
export CLAUDE_CODE_AUTO_MODE_SERVER=0
export EDITOR=edit
# Native GPT-6 / GPT-6.1 input ceiling advertised by OpenCodex.
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=872000

exec claude "$@"

I checked open and closed issues before filing. Related reports are #1218 (the earlier 200k fallback), #3958 (extended native context), #3170 / #4112 (input-limit and compaction errors), and #4950 (the 1M request). I didn't find an exact match for this Claude Code/GPT-6.1 failure. #4950 was closed by the template checker.

Checks

• I searched existing provider and compatibility issues.
• The request and response were redacted.
• The expected behaviour is based on an upstream specification or a concrete client requirement.