schneewolflabs/egirl-delegation-dpo

egirl-DPO

558 pairs teaching one lesson — hand the job to the code agent, or do the small thing yourself, but stop exploring.

Built for egirl, where a local model acts as an operator: it has small tools (read_file, edit_file, glob_files, execute_command, git_*) and can delegate to code_agent — a full coding agent (Claude Code / Codex) that explores a codebase, makes multi-file edits, and runs tests. The operator is supposed to delegate real engineering work and handle only trivia itself.

Local models do not do this.

The problem

Benchmarked on ten operator requests in egirl's own <tool_call> dialect, with egirl's real system prompt — which states "Delegate coding work to the code agent by default — you're the human-in-the-loop, not a code generator":

model	delegated
Qwen/Qwen3.6-27B	0/3
nbeerbower/Qwen3.6-27B-TIES	0/3
huihui-ai/Huihui-ThinkingCap-Qwen3.6-27B-abliterated	0/3
nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451	0/3
two local TIES merges	0/3
gpt-5.6-sol	1/3

Instead of delegating, they narrate — "Let me start by understanding the current structure" — and then call glob_files. Swapping a generic prompt for egirl's real one moved overall tool accuracy 6/10 → 7/10 and changed delegation not at all, so this is model behaviour, not prompt wording. Even a frontier model only manages it about a third of the time under the same single-turn conditions.

What the data teaches

A three-way policy, not "always delegate":

situation	correct action	share
read a file, check status, show a diff or log	the specific direct tool	~11%
change exactly one line in one named file	edit_file	~24%
anything else that touches real code	code_agent, with a specified task	~65%

Each row also carries workspace_root, so you can filter to a single layout if you prefer.

field	content
system	a deliberately minimal operator prompt: identity, workspace, and the <tools> definitions — nothing else. It carries no delegation policy and no "call the tool immediately" instruction, so a model trained on it has to learn when to delegate as behaviour rather than learning to obey a sentence that happens to be in context. At inference egirl's real prompt does state the policy, and the two compose. Every row carries the tool list, because the target is a tool call.
prompt	a developer request naming a real project path
chosen	the tool call a strong model actually makes under egirl's prompt
rejected	what the local model does instead — sampled from Qwen3.6-27B-TIES
tier	trivial / borderline / complex
chosen_tool, rejected_tools	for filtering and analysis

What the local model reaches for instead, across all 186 rejections:

	count
glob_files	71
read_file	56
execute_command	49
other	10

Two distinct failures. On code work it explores (glob, shell) instead of delegating. On one-line edits it reads the file and stops instead of editing. Both are in the set.

Construction

1. Requests — a strong model (gpt-5.6-sol) generated realistic developer requests across three difficulty tiers, each naming an unambiguous project path.
2. chosen — each request was run through egirl's real system prompt and the resulting tool call kept. Behaviour, not stated judgement: asked to explain whether a symbol rename or an added test should be delegated, the model says "direct — small and localized," while its actions delegate both. egirl's prompt sides with the actions, so building from explanations would have trained the wrong policy.
3. rejected — the same prompts run through Qwen3.6-27B-TIES, on-policy for the model you would actually train. Pairs where the local model already matched the target were dropped; there is nothing to learn from a pair whose sides agree (69 of 255 dropped).

Limitations

• Paths are augmented, not diverse by nature. The 186 original requests all named one developer's workspace. Each is reproduced under several roots (/home/alice/code, /srv/repos, /Users/dev/work, /opt/projects, /home/rin/dev) with projects renamed, so no single root or project name predicts the answer — but the underlying tasks are still the 186 originals seen about three times each.
• Single-turn. Each pair is one request and the immediately following tool call. It does not teach what to do with the results, when to follow up, or how to recover from a failed call.
• The borderline tier is 100% delegate. That reflects egirl's stated policy, but it is a choice: a different operator design might want renames and single-test additions handled directly. Filter on tier if you disagree.
• chosen is synthetic. These are a strong model's tool calls, not human-authored gold. They were verified to parse and to name real tools, not to be optimal.
• Not decontaminated against any benchmark.

Related

• egirl — the agent this targets
• schneewolflabs/Kadenz-DPO — same single-axis discipline, prose rhythm