fix: strengthen default Apply prompt for local models

[shanevcantwell]

Summary

• Local/open-weight models respond to the default Apply prompt with prose summaries of changes instead of actual code, because defaultApplyPrompt returns a single-sentence plain string with no code-only constraint
• Change defaultApplyPrompt to return ChatMessage[] with labeled ORIGINAL CODE / SUGGESTED EDIT sections, explicit "Output ONLY code" instructions, and an assistant prefill opening a code block
• This routes through streamChat (instead of streamComplete), and the assistant prefill forces the model to continue with code rather than prose
• The existing filterCodeBlockLines filter strips the closing fence from the output

Context

defaultApplyPrompt is the fallback Apply prompt used when no model-specific promptTemplates.apply is configured. It applies when a model calls edit_existing_file and the Apply step merges the suggested code into the original file.

The previous prompt was:

{original_code}\n\nThe following code was suggested as an edit:\n```\n{new_code}\n```\nPlease apply it to the previous code.

Local models (tested with gpt-oss-20b via LM Studio) interpreted this as a request to describe changes rather than output code. The new prompt with explicit rules and assistant prefill produces clean code diffs.

What this doesn't change

• Claude Sonnet lazy-apply path (separate code path in core/edit/lazy/)
• GPT/Claude edit prompts (gptEditPrompt, not defaultApplyPrompt)
• Users with custom promptTemplates.apply in config (custom overrides defaultApplyPrompt)

Test plan

• cd core && npm run vitest — 1654 tests pass, 0 failures
• Tested with gpt-oss-20b (local model via LM Studio) — Apply produces code diffs instead of prose summaries
• Verify cloud model Apply still works (they typically use model-specific prompts, not defaultApplyPrompt)

🤖 Generated with Claude Code

---

Summary by cubic

Make the default Apply prompt produce code (not prose) for local/open‑weight models by switching to a chat-based prompt with code-only instructions and an assistant prefill. Apply now routes through streamChat and returns the full modified file.

• Bug Fixes
  • defaultApplyPrompt now returns ChatMessage[] with ORIGINAL CODE / SUGGESTED EDIT blocks and “output ONLY code” instructions.
  • Assistant prefill opens a code block to force code continuation; existing filter removes the closing fence.
  • No changes to model-specific prompts, lazy-apply path, or user overrides.

^{Written for commit 25026d7. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 1 file

[RomneyDa]

@shanevcantwell just to double check any reason the new apply prompt is now messages instead of the raw text?

[shanevcantwell]

The messages format lets us include an assistant prefill: the role: "assistant" message that opens with a code fence. Without it, local models (Ollama, LM Studio, etc.) tend to respond to the raw text prompt with prose like "Sure! Here's the modified code:" before the actual code, which breaks the diff parser.

The assistant prefill locks the model into continuing inside a code fence, so the output is code-only. This pattern already exists in the lazy apply prompt (core/edit/lazy/prompts.ts) for the same reason.

The fallback path in renderPromptTemplate handles providers that don't support prefill. It compiles the messages down to a completion prompt instead (canPutWordsInModelsMouth check in core/llm/index.ts).

[RomneyDa]

@shanevcantwell appreciate the explanation! Looks good to me.