What happens
With deepseek/deepseek-v4-flash answering and mid tool-batch, picking zhipu/glm-5.3 from /model ends the turn with:
error: the request could not be sent as it was
Reproduces identically on plain dev @ 9be565861 with the same tmux script (lane B's drive on the Spark, ~/af-b-reports/), so it is not #993's. It is #935's cut-and-re-ask (the person's word wins at the request) meeting the new model's rejection of the previous model's tool-call transcript shape.
Replication a stranger can run
On the Spark, make build, then in tmux: bin/aforge chat --model deepseek/deepseek-v4-flash --yolo, send Run this exact shell command and then tell me what it printed: sleep 75 && echo done-waiting, wait for the bash call to be running, then /model zhipu/glm-5.3. The turn ends with the error above instead of the re-asked request answering.
Acceptance
End-to-end first: the same drive, and after the pick the turn continues on the new model (or the surface says in a person's words why it cannot carry this transcript across), with no error: row. Unit: a fake completer for the second model that refuses the first model's tool-call shape, and the cut-and-re-ask either repairs the transcript or refuses in words.
Found by lane B (#993) on 2026-09-12.
Drafted with CodeAF · reviewed and owned by the author
What happens
With
deepseek/deepseek-v4-flashanswering and mid tool-batch, pickingzhipu/glm-5.3from/modelends the turn with:Reproduces identically on plain
dev@9be565861with the same tmux script (lane B's drive on the Spark,~/af-b-reports/), so it is not #993's. It is #935's cut-and-re-ask (the person's word wins at the request) meeting the new model's rejection of the previous model's tool-call transcript shape.Replication a stranger can run
On the Spark,
make build, then in tmux:bin/aforge chat --model deepseek/deepseek-v4-flash --yolo, sendRun this exact shell command and then tell me what it printed: sleep 75 && echo done-waiting, wait for the bash call to be running, then/model zhipu/glm-5.3. The turn ends with the error above instead of the re-asked request answering.Acceptance
End-to-end first: the same drive, and after the pick the turn continues on the new model (or the surface says in a person's words why it cannot carry this transcript across), with no
error:row. Unit: a fake completer for the second model that refuses the first model's tool-call shape, and the cut-and-re-ask either repairs the transcript or refuses in words.Found by lane B (#993) on 2026-09-12.
Drafted with CodeAF · reviewed and owned by the author