The prior fix dropped numCtx to 8192 to guarantee full GPU offload,
but real multi-turn djAgentPick conversations peak around 7.5k tokens
-- leaving almost no headroom before llama.cpp's context-shift drops
the front of the prompt (system instructions + tool defs, including
`done`). Result: "agent stopped without calling done" recurring every
few picks. 11264 is the largest context that still offloads all 29
layers on this GPU (tested empirically -- 12288 fell 15MB short and
dropped to 28/29), leaving ~3.7k tokens of margin over the observed
peak.
Also: OLLAMA_CONTEXT_LENGTH had drifted onto the running container via
an earlier manual `docker run` and was never in this compose file, so
a prior git-redeploy silently kept the manual value instead of the
committed one. Committing it here closes that drift.
Claude-Session: https://claude.ai/code/session_01L7Rwa6guD5wK8F8tWQwcJX
poprhythm
b7bee3e02f