Crate newt_coder 

newt-coder — coder plugin closing failure mode T0b.

Background: see the authoritative knowledge card ~/workspaces/knowledge/board/drake/2026-05-29_newt-coder-failure-mode-taxonomy.md.

The default newt worker (newt-flat) sends the operator’s task prompt verbatim with no file contents, then asks the model to emit a unified diff. Local Ollama coder models reliably invent plausible-but-wrong context lines under that contract — apply_patch rejects the patch on context mismatch and the workspace diff stays empty (failure mode T0b).

newt-coder is the opinionated path: it scans the workspace for files the task mentions, injects them verbatim into the user message, and asks the model to emit each updated file in full (no fences, no prose, no diffs). The reply is parsed by normalize_emission and the changes are written to the workspace; the caller then captures a real diff via git diff.

Activation:

• env: NEWT_CODER=1, or
• per-session ACP param: { "coder": true } on new_session.

Both are checked by newt-acp-worker. The legacy newt-flat path is the default and remains unchanged.

Re-exports

pub use coder::Coder;

pub use coder::CoderRun;

pub use emission::normalize_emission;

pub use emission::Emission;

pub use error::CoderError;

pub use error::Result;

pub use prompt::build_prompt;

pub use prompt::CoderPrompt;

pub use prompt::DEFAULT_CONTEXT_CAP_CHARS;

pub use prompt::WHOLE_FILE_SYSTEM_PROMPT;

pub use workspace_scan::scan_workspace_for_files;

Modules

coder

The Coder orchestrator — prompt -> infer -> normalize -> apply.

emission

Parse the model’s raw reply into a structured Emission.

error

Error type for the newt-coder plugin.

prompt

Build the LLM prompt for a coder task.

workspace_scan

Scan the workspace for files relevant to the task.