reasoning-replay=keep-last mandatory with pi/gpt-oss-20b but lead to request crash

[NicolasJungers]

pi/qwen3.6-35b-a3b loses reasoning without reasoning-replay=keep-last, but its use leads to an internal forge crash:

Error: "can only concatenate list (not \"str\") to list"

Forge version: 0.9.3 (standalone installer)
Backend: --backend openai
Model: gpt-oss-20b
Harness: pi v0.84.2
Flags: --serialize --reasoning-replay keep-last

Traceback

File "forge/clients/llamafile.py", line 736, in _send_native
      merged = _merge_consecutive(messages)
File "forge/clients/llamafile.py", line 250, in _merge_consecutive
      "content": target.get("content", "") + "\n\n" + m.get("content", ""),
TypeError: can only concatenate list (not "str") to list

Assisted by Claude Sonnet 5

[antoinezambelli]

Thanks—we’ve isolated this to multipart/list-style message.content reaching Forge’s native message merger. It isn’t caused by reasoning replay itself; keep-last is exposing the affected message shape.

We’re working on a broader fix at the message-primitives level rather than adding a one-off concatenation patch, and hope to have that ready within a few days.

Could you confirm which server is behind --backend openai? If it’s Ollama, switching to Forge’s dedicated --backend ollama adapter and pointing it at the Ollama daemon root should work around this in the meantime, since that path already normalizes multipart content.

[NicolasJungers]

On 24/08/2026 17:12, Antoine wrote: *antoinezambelli* left a comment (antoinezambelli/forge#151) <#151 (comment)> Thanks—we’ve isolated this to multipart/list-style |message.content| reaching Forge’s native message merger. It isn’t caused by reasoning replay itself; |keep-last| is exposing the affected message shape. We’re working on a broader fix at the message-primitives level rather than adding a one-off concatenation patch, and hope to have that ready within a few days. Could you confirm which server is behind |--backend openai|? If it’s Ollama, switching to Forge’s dedicated |--backend ollama| adapter and pointing it at the Ollama daemon root should work around this in the meantime, since that path already normalizes multipart content.

It's a chain: {kimi-code,aider,pi} -> haproxy for tls -> forge ->  llama-swap -> llama.cpp -> {qwen3.6-35b-a3b,gpt-oss-20b, qwen3-coder-30b-a3b-instruct} What else would you know? Regards, Nicolas

…

— Reply to this email directly, view it on GitHub <#151?email_source=notifications&email_token=AHYELCPM3PWXN2H3RBXU3D35LRLN7A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMZZG4ZDIOJRHE2KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-5397249194>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHYELCK43IEXHKXYZRNV3TL5LRLN7AVCNFSNUABGKJSXA33TNF2G64TZHMYTCNJZGM4DCNZXGQ5US43TOVSTWNJSGM2TANRVGI3TBILWAI>. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/AHYELCNZIWREUE3OFJ4HEUD5LRLN7A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMZZG4ZDIOJRHE2KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KUZTPN52GK4S7NFXXG> and Android <https://github.com/notifications/mobile/android/AHYELCJORG2FKROILZNODAT5LRLN7A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMZZG4ZDIOJRHE2KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2K4ZTPN52GK4S7MFXGI4TPNFSA>. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***>

[antoinezambelli]

This is helpful, thanks!

Sadly, yeah - we have an incompatibility right now due to how the Message objects within forge were built, some of the multi part structures don't play well right now.

I have an upgrade mapped that will cover this and normalize everything more cleanly, but it'll take me a few days to land - it shouldn't have any external surface impact.

At the moment, I think Ollama is the only forge client that normalizes the message shape and would be a workaround, so forking and duplicating that approach should unblock things in the short term.

I'll get cracking on the upgrade in the meantime.