fix(ollama): honor max_tokens and stop defaulting to the model-max context

[RasulOs]

Problem (Discord report)

Why is setting max_tokens not doing anything? … the context is 262144?

Two independent bugs in the Ollama path:

1. max_tokens was silently dropped. It is not a field on llama-index's Ollama class, so pydantic discarded it — no error, no effect. Verified: load_llm("Ollama", ..., max_tokens=16) produced an 890-char response with additional_kwargs: {}.
2. The context defaulted to the model's maximum. llama-index's context_window=-1 default resolves to the model's max context via a hidden client.show() call, allocating the full KV cache (256K-context models → ~19 GB, CPU spill). Because mobilerun sends num_ctx per request, this overrides every Ollama-side setting (OLLAMA_CONTEXT_LENGTH, Modelfile, /set parameter) — users could not fix it from the Ollama side, which is exactly the reporter's confusion.

Fix

A pure _prepare_ollama_kwargs() helper in load_llm's Ollama branch — the single chokepoint covering CLI, SDK, and config-profile paths:

• max_tokens → additional_kwargs.num_predict; explicit num_predict wins (one-line warning on conflict); non-integer values warn and skip (today they're a silent no-op, so warn+skip strictly improves); a model_fields guard auto-disables the shim if llama-index adds native support.
• context_window defaults to 32768 when neither it nor additional_kwargs.num_ctx is set (in-repo precedent: the DeepSeek branch's context_window default). An explicit num_ctx is mirrored into context_window so they stay aligned and the hidden show() network call is never triggered. context_window: -1 remains the documented escape hatch for model-max on big-GPU machines.
• Other unknown Ollama kwargs log a deduped warning instead of vanishing (warn-only, never popped).

Plus: wizard-created Ollama profiles write an explicit context_window: 32768 (self-documenting configs; the runtime injection is the fallback for hand-written ones), an Ollama-gated "Context window" prompt in mobilerun configure advanced settings, the stale loader docstring path, and a commented Ollama example in config_example.yaml.

Back-compat: no config migration — runtime translation is the compatibility layer; existing additional_kwargs.num_predict/num_ctx workaround configs behave byte-for-byte identically under the precedence rules. Two visible changes: wizard-written max_tokens now actually caps output, and unconfigured Ollama profiles drop from model-max to 32K context (one line restores either).

Testing

• 169/169 unit tests, 16 new in tests/test_llm_picker.py: translation, both conflict directions, numeric-string coercion, invalid-value warn+skip (string/bool/None), default injection, explicit/-1 preservation, num_ctx mirroring (incl. non-numeric fallback), deduped unknown-kwarg warning, the model_fields future-guard, end-to-end load_llm("Ollama"), and the wizard default.
• Live A/B vs released 0.6.5 (real Ollama server, gemma4 = 131072-context model):

	0.6.5	this branch
ollama ps CONTEXT	131072 (model max)	32768
max_tokens=16	ignored → 890-char response	num_predict: 16 → output capped
hidden show() call	yes	no

• Live emulator agent run on Ollama (-p Ollama -m gemma4:latest): completed the screen-title task with the 32K context active at 100% GPU.

🤖 Generated with Claude Code

[mintlify]

Docs PR opened: https://github.com/droidrun/mobilerun-docs/pull/12

Documented Ollama profile kwargs in the configuration guide, covering max_tokens, the new 32K context_window default, and -1 override.