opencode-go/muse-spark-1.2-contributor 502 truncation on large tool-call tasks (mid tool call without terminal signal) - comprehensive evidence

Client or integration

Codex App

Area

Provider adapter

Summary

opencode-go/muse-spark-1.2-contributor via opencodex proxy 10100 consistently truncates on large-context + tool-call streams: 502 upstream_server_error upstream stream ended mid tool call without a terminal signal — possible truncation src/adapters/openai-chat.ts:1734. Small prompts succeed, the same large CodexBar sweep fails at all reasoning levels low/medium/high/xhigh even after reinstall, while opencode-go/deepseek-v4-flash on identical prompt succeeds 200. Codex then does 5/5 useless retries.

Reproduction

Environment: opencodex 2.27.0 bun ~/.hermes/node/lib/node_modules/@bitkyc08/opencodex , codex 0.147.0 openai_base_url http://127.0.0.1:10100/v1 , macOS 15.5 Darwin 24.x, muse-spark context 1048576 baseUrl https://opencode.ai/zen/go/v1, workspace CodexBar-bug:feat main a3457dafb
Minimal large reproduction (deterministic):

codex exec -m opencode-go/muse-spark-1.2-contributor -c model_reasoning_effort="low" --color never \
"Continue aligning the status of CodexBar-bug:feat — sweep branches, local changes and PR/CI. Just run bash tools \`git status\` and \`git branch -vv\` and summarize"
# also reproduces with medium/high/xhigh, same result

Observe Codex Reconnecting 1/5...5/5 then stream disconnected before completion: upstream stream ended mid tool call without a terminal signal — possible truncation, localhost:10100/#logs shows 502 with unreported tokens.
Control:

codex exec -m opencode-go/deepseek-v4-flash -c model_reasoning_effort="high" --color never "same prompt"
# → 200, git status + 41 branches returned
codex exec -m opencode-go/muse-spark-1.2-contributor "hi" # small → 200 even on xhigh

Version

2.27.0

Operating system

macOS 15.5 (Darwin 24.6.0, arm64)

Provider and model

opencode-go / muse-spark-1.2-contributor

Logs or error output

~/.opencodex/usage.jsonl truncated batch (high, still truncates):

ocx-mt1aawks-4f ts 1787216068972 eff xhigh/xhigh status 502 dur 3646 upstreamError mid tool call truncation
ocx-mt1b32u3-d ts 1787217383451 eff high/high status 502 dur 2654
ocx-mt1b3655-e ts 1787217387737 eff high/high status 502 dur 2433
ocx-mt1bdrba-3 ts 1787217881734 eff high/medium(after high->medium map) status 502 dur 3409
...
ocx-mt1bhd5b-h ts ... eff low/low status 502  # even low fails on large sweep

All share: adapter openai-chat status 502 errorCode upstream_server_error recoveryKinds [] usageStatus unreported

Small control (same model, small prompt, succeeds):

ocx-mt1b51yc-2 eff xhigh/xhigh status 200 totalTokens 1066 dur 2500
ocx-mt1bbmaw-2 eff high/medium status 200

DeepSeek control (same large sweep, succeeds):

ocx-mt1bns5q-3 model deepseek-v4-flash eff high status 200
ocx-mt1bo65y-5 model deepseek-v4-flash eff high status 200
codex exec deepseek → git status + 41 branches returned in <30s

Matrix tested:

Prompt size	Model	Effort	Result
small `hi` (8 tokens)	spark	xhigh/high/medium/low	200
synthetic 40k + bash tool	spark	medium/low	200 / high timeout
CodexBar large sweep	spark	low/medium/high/xhigh	502 (all)
CodexBar large sweep	deepseek	high	200
CodexBar large sweep	gpt-5.6-luna	low	200

~/.opencodex/service.log diagnostic:
debugProviderDiagnostic openai-chat stream-truncated {finishReason:null hadUsage:false pendingToolCalls:1} then yield error mid tool call

Screenshots and supporting files

localhost:10100/#logs 2026/08/20 17:16:39-58 5x muse-spark high 502 unreported
Codex UI Reconnecting 1/5...5/5 screenshots (large sweep) vs deepseek success screenshots available on request

Redacted configuration

// ~/.opencodex/config.json providers.opencode-go
{
  "modelReasoningEfforts": { "muse-spark-1.2-contributor": ["low","medium","high","xhigh","max"] },
  "modelDefaultReasoningEfforts": { "muse-spark-1.2-contributor": "high" },
  "modelContextWindows": { "muse-spark-1.2-contributor": 1048576 }
}
// ~/.codex/config.toml (reproduces on default)
model = "opencode-go/muse-spark-1.2-contributor"
model_reasoning_effort = "high"
openai_base_url = "http://127.0.0.1:10100/v1"

After reinstall (ocx uninstall → service install → sync --restart-codex) disabledModels reset and spark still reproduces, proving not local corruption. Mapping high->medium and modelMaxOutputTokens 65536 tested, still 502 on large sweep.

Checks

I searched existing issues and documentation.
I removed secrets, tokens, account details, request credentials, and personal data.

⚠️ Possible Duplicate Issue(s)

opencode-go/muse-spark-1.2: frequent 'upstream stream ended mid tool call without a terminal signal — possible truncation' in streaming tool calls (Codex App via ocx, Pi direct OK) #2156

🔗 Related PRs

#2165 - fix(openai-chat): treat a non-string repeat as padding once that field is known [merged]
#2162 - fix(anthropic): frame the opening turn so AgentRouter stops blocking non-English [merged]
#2186 - [WRONG BRANCH] release: promote dev to main for 2.28.0 [closed]
#2187 - [WRONG BRANCH] release: sync dev into preview for 2.28.0 [closed]
#2160 - fix(providers): identify opencode-free with the client User-Agent it claims [merged]

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

Create Plan

🧪 Issue enrichment is currently in open beta.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

리뷰 · 우선순위 69 / 80

#2156이랑 같은 구멍임. 제목에 contributor 붙은 모델만 다른데 에러 문자열이 똑같음. src/adapters/openai-chat.ts:1825가 finishReason 없고 pendingToolCalls.length > 0이면 upstream stream ended mid tool call without a terminal signal — possible truncation 냄. 이슈에 적은 stream-truncated {finishReason:null hadUsage:false pendingToolCalls:1}이 그 로그 그대로임.

라인 번호는 2.27.0 기준 1734가 아님. 지금 dev HEAD 96f288d59에선 그 자리는 툴콜 arguments 검증이고, 잘림 분기는 1825임. #2180이 하트비트 넣으면서 밀렸음. 그 PR 주석도 이거 고친 거 아니라고 적혀 있음 (:1752-1755). 스톨 워치독이랑 다른 프레임임.

작은 hi는 200, CodexBar 큰 스윕은 low/medium/high/xhigh 전부 502, deepseek랑 luna는 같은 스윕이 200. 업스트림이 툴콜 중간에 소켓을 닫거나, spark가 finish_reason/[DONE] 없이 끊는 쪽에 가깝음. ocx가 임의로 자른 증거는 아직 없음. recoveryKinds []라서 ocx 재시도가 아님. Codex App 5/5는 클라 재접속임. 같은 잘림을 다섯 번 더 보는 거임.

#2193은 이미 닫혔음. 이 이슈랑 #2156을 둘 다 열어두면 중복임. 증거는 여기가 더 빵빵하지만 SSE 원본은 여전히 없음. 캡처 없이 flush-on-EOF 때리면 잘린 툴콜을 성공으로 포장함.

해결방안: #2156을 메인으로 두고 이 매트릭스/로그를 거기 코멘트로 옮긴 다음 #2194는 duplicate로 닫음. 고치려면 zen/go SSE 원본이 필요함. finish_reason 없이 끊기는지, 우리가 먼저 끊는지. 캡처 오기 전엔 머지하지 말 것. #2180 재리베이스도 의미 없음.

이 댓글은 grok-bot이 작성했습니다

Thanks for the detailed matrix. I moved the additional evidence into #2156, which owns the identical openai-chat EOF-with-pending-tool-call failure.

The exact error and diagnostic state are the same, and keeping both open would split the raw-SSE investigation. The current guard must remain fail-closed until we can distinguish an upstream terminal-frame omission from a proxy-side truncation; flushing partial tool JSON would be unsafe.

Closing this as a duplicate of #2156.

Navigation Menu

Repository navigation

opencode-go/muse-spark-1.2-contributor 502 truncation on large tool-call tasks (mid tool call without terminal signal) - comprehensive evidence #2194

Client or integration

Area

Summary

Reproduction

Version

Operating system

Provider and model

Logs or error output

Screenshots and supporting files

Redacted configuration

Checks

리뷰 · 우선순위 69 / 80

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

opencode-go/muse-spark-1.2-contributor 502 truncation on large tool-call tasks (mid tool call without terminal signal) - comprehensive evidence #2194

Description

Client or integration

Area

Summary

Reproduction

Version

Operating system

Provider and model

Logs or error output

Screenshots and supporting files

Redacted configuration

Checks

Activity

coderabbitai commented on Aug 20, 2026

lidge-jun commented on Aug 20, 2026

리뷰 · 우선순위 69 / 80

Ingwannu commented on Aug 20, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions