Context
When a model returns empty replies (no text, no tool call) repeatedly, borg stops with an actionable message naming the likely cause: the model can't make structured tool calls (internal/agent/loop.go, emptyReplyMsg). It's good, but it doesn't point the user anywhere concrete.
The task
Extend that message to link the README's tested-models table and suggest a known-good model (e.g. qwen2.5-coder). Keep it model- and vendor-agnostic — describe the missing capability and point at the doc, don't hardcode a single "correct" model in the harness logic.
Notes / constraints
- Per
CONTRIBUTING.md, harness text must not overfit to one model. A doc link + example is fine; a hardcoded "use model X or fail" is not.
Acceptance
- The error references the tested-models guidance.
- Existing empty-reply tests updated; behavior otherwise unchanged. Coverage ≥90%.
Context
When a model returns empty replies (no text, no tool call) repeatedly, borg stops with an actionable message naming the likely cause: the model can't make structured tool calls (
internal/agent/loop.go,emptyReplyMsg). It's good, but it doesn't point the user anywhere concrete.The task
Extend that message to link the README's tested-models table and suggest a known-good model (e.g.
qwen2.5-coder). Keep it model- and vendor-agnostic — describe the missing capability and point at the doc, don't hardcode a single "correct" model in the harness logic.Notes / constraints
CONTRIBUTING.md, harness text must not overfit to one model. A doc link + example is fine; a hardcoded "use model X or fail" is not.Acceptance