Repository navigation
bug: Hermes Agent — gpt-5.6-luna via opencode-go provider returns finish_reason:null (no [DONE]) #40420
Description
Description
When using opencode-go as a provider in Hermes Agent, every gpt-5.6-luna response fails. The OpenCode Go gateway (https://opencode.ai/zen/go/v1/chat/completions) never sends a terminal finish_reason for this model — streamed or non-streamed: non-streaming returns "finish_reason":null on a complete response, and streaming ends with data: {"choices":[],"cost":"0"} and no [DONE] marker. Hermes requires a terminal event and retries, then errors with "Response remained truncated after 4 continuation attempts" — gpt-5.6-luna is unusable via this provider. All other Go models tested (deepseek-v4-flash, deepseek-v4-pro, kimi-k2.7-code, glm-5.2) return finish_reason: "stop" correctly. Related to but distinct from #36766 (truncated tool args on native OpenAI Responses) — this is the Go gateway omitting the terminal event.
Plugins
N/A — no OpenCode plugins. Consumer is Hermes Agent v0.19.1 (docker, Linux), which fails on every gpt-5.6-luna response.
OpenCode version
N/A — issue is against the hosted Go gateway endpoint, not a local OpenCode install.
Steps to reproduce
POST /chat/completions,model: gpt-5.6-luna,stream: false→ HTTP 200, content correct,choices[0].finish_reason: null- Same with
stream: true→ content chunk arrives withfinish_reason: null, then final chunk{"choices":[],"cost":"0"}— no[DONE] - Control (same calls,
deepseek-v4-flash/kimi-k2.7-code/glm-5.2/deepseek-v4-pro) →finish_reason: "stop"and a proper[DONE]chunk
Screenshot and/or share link
No UI screenshot applicable (API-level bug). Verbatim captures from a fresh repro run:
Non-streaming:
{"id":"gen-1785843293-273noTConb4mUERLVDI2","object":"chat.completion","created":1785843294,"model":"gpt-5.6-luna","choices":[{"index":0,"message":{"role":"assistant","content":"OK"},"finish_reason":null}],"usage":{"prompt_tokens":11,"completion_tokens":5,"total_tokens":16}}Streaming (all data: lines):
data: {"id":"gen-...","object":"chat.completion.chunk","choices":[]}
data: {"id":"","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":"OK"},"finish_reason":null}]}
data: {"id":"gen-...","choices":[],"usage":{"prompt_tokens":11,"completion_tokens":5,"total_tokens":16}}
data: {"choices":[],"cost":"0"}
Operating System
Linux (Docker container)
Terminal
N/A — no terminal UI involved. Reproduced with plain curl from a non-interactive container shell (TERM=dumb, no TTY).
Activity
github-actions commentedon Aug 4, 2026
This issue might be a duplicate of existing issues. Please check:
gpt-5.6-luna) on OpenCode Go with streaming issues — stream degradation, mid-stream cuts, and continuation loops. The root cause may overlap with thefinish_reason:null/ missing[DONE]reported here.If your issue is distinct (specifically the missing terminal event /
[DONE]marker rather than stream quality), please clarify how it differs from #39881 so maintainers can track them separately.jcmpgit commentedon Aug 7, 2026
Workaround available (verified working 2026-08-07).
Confirmed root cause matches #39881: the OpenCode Go chat/completions path never sends a terminal
finish_reason/[DONE]forgpt-5.6-luna(andkimi-k3), while other models on the same endpoint returnfinish_reason:"stop"+[DONE]. Hermes deliberately refuses to fabricate a terminal marker on a text-only stream (guards a real truncation class — NousResearch/hermes-agent#32086), so the response is treated as a mid-stream drop and every turn fails with "Response remained truncated after 4 continuation attempts".Since this is an upstream endpoint defect, we shipped a narrow, model-scoped client-side workaround rather than waiting: a patch script that lets Hermes treat luna streams as complete (text IS fully delivered — only the terminal marker is missing):
https://gist.github.com/jcmpgit/60087ac45da47c58ffe654781bada8fc
gpt-5.6-lunaonly (extendable tokimi-k3)revertrestores itFor maintainers: the fix belongs upstream — either emit
finish_reason:"stop"/[DONE]on the chat/completions path for these models, or route luna to the/zen/go/v1/responsespath where termination is already correct (see #39881). Happy to run additional probes if useful.antoniosena-code commentedon Aug 8, 2026
Confirming reproduction from a second, independent Hermes Agent deployment (v0.20.0, Docker/Linux), fresh probe as of 2026-08-08:
Still broken 4+ days after the original report — every Hermes turn on
gpt-5.6-lunavia this endpoint fails with "Response remained truncated after 4 continuation attempts".One additional data point isolating the defect: on the same endpoint,
deepseek-v4-flashhandled a 369,114-prompt-token request (needle retrieval over ~1.2 M chars) in 6 s and returnedfinish_reason: "stop"with correct usage accounting. The gateway and its large-context path are healthy; the bug is confined to the terminal-event handling of the luna shim (and reportedlykimi-k3) on the chat/completions path, consistent with the routing analysis in #39881.github-actions commentedon Oct 9, 2026
To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.