fix: change default Ollama model to qwen2.5:7b (coder variant breaks tool calls)

[fubak]

Problem

The default Ollama model qwen2.5-coder:7b does not use Ollama's structured tool_calls API. When SmallHarness sends tool definitions with a request, the coder variant emits tool invocations as raw JSON text inside the content field:

{"name": "shell", "arguments": {"command": "ls"}}

SmallHarness reads tool calls from the tool_calls array in the API response, not from response text. These invocations are never executed — the session displays them as prose and stalls, with no error message to explain why.

This is the first thing a new Ollama user hits, and there's no indication of what's wrong.

Fix

Change the default from qwen2.5-coder:7b to qwen2.5:7b.

The base model uses the <tool_call> XML template that Ollama correctly translates into structured tool_calls responses. Tool calls from the base model are received and executed as expected in both interactive and one-shot (--print) modes.

Users who specifically want the coder fine-tune can set modelOverride in agent.config.json. The coder variant is a fine coding model — it just cannot reliably invoke tools, which is a prerequisite for an interactive coding agent.

Test

Added defaults_ollama_to_base_qwen_not_coder_variant which asserts the default is qwen2.5:7b and explicitly rejects qwen2.5-coder:7b with a comment explaining why.

All 384 tests pass.

Verified with

Ollama 0.30.8, both qwen2.5:7b and qwen2.5-coder:7b pulled locally. The base model calls tools correctly; the coder variant emits raw JSON text.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Updates Ollama’s default model selection to prefer the Qwen2.5 base model due to structured tool call support, and adds a regression test to lock in the behavior.

Changes:

• Switch default Ollama model from qwen2.5-coder:7b to qwen2.5:7b.
• Add a unit test asserting Ollama defaults to the base variant (and not the coder variant).
• Document the rationale for the default change in code comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[bennewell35]

I reviewed the diff and ran the upstream-style local checks in a disposable Docker Rust environment.

Validation run:

docker run --rm -v /tmp/smallharness-pr9:/work -w /work rust:latest bash -c 'rustup component add rustfmt clippy >/dev/null && rustc --version && cargo --version && cargo fmt --all -- --check && cargo clippy --all-targets -- -D warnings && cargo test'\n```\n\nResult:\n- `rustc 1.96.0` / `cargo 1.96.0`\n- `cargo fmt --all -- --check` passed\n- `cargo clippy --all-targets -- -D warnings` passed\n- `cargo test` passed: 381 passed, 0 failed\n- The new `defaults_ollama_to_base_qwen_not_coder_variant` test passed as part of the suite\n\nI did not live-test Ollama/model tool-call behavior on my machine, so this is code/test validation rather than independent model-runtime validation. The patch is narrow and the regression test covers the intended default-model contract.