delegate-coder
Let Claude think. Let a cheaper agent type.
A Claude Code skill that delegates execution-heavy coding work — bulk codebase reading, implementation, refactoring, first-pass review — to a second CLI coding agent (MiMo Code, Aider, Codex CLI, Gemini CLI, Qwen Code, OpenCode, or any headless agent), while Claude keeps the planning, architecture, and final judgment.
Why: Claude Code usage is precious. Most of what burns it isn't thinking — it's reading 50 files and typing routine code. Free-tier worker agents can do that part. This skill makes the handoff automatic and safe.
What makes this different
This is not just a CLI wrapper. The skill encodes a trust framework for delegation:
- Precise spec handoff — Claude writes exact scope, constraints, and verification criteria before delegating. Vague handoffs are the #1 cause of wasted cycles.
- Git branch isolation — every exec task runs on a
delegate/* branch. Nothing touches your working state.
- Cheap, deterministic verification — results are verified by
git diff --stat + your test suite, not by Claude re-reading code. Diffs can't lie; summaries can.
- Read-only modes where supported — analysis tasks use the worker's read-only mode (MiMo
plan, Codex read-only sandbox) for zero-risk delegation.
- Escalation rule — two failures on the same task and Claude stops delegating it. No silent retry loops eating your savings.
- Safe permissions guidance — granular allow/deny config instead of blanket
--dangerously-skip-permissions / --yolo.
Experimental (v2)
v2 adds several new config options: model selection, enable/scope guards, strict/graceful fallback, and a path allowlist. These all default to off or prior behavior, so existing setups are unaffected. However, none of these features have been benchmarked or unit-tested yet — the benchmark results below reflect v1 behavior only. Treat them as opt-in and report issues.
Supported worker agents
| Agent | exec | read-only | sessions |
|---|
| MiMo Code | ✅ | ✅ (plan mode) | ✅ |
| Aider | ✅ | ◐ (--dry-run) | — |
| Codex CLI | ✅ | ✅ (sandbox) | ✅ |
| Gemini CLI | ✅ | ◐ | — |
| Qwen Code | ✅ | ◐ | — |
| OpenCode | ✅ | ◐ | ◐ |
| Anything else | via command_override — any CLI with a headless mode works | | |
Install
# Personal (all projects)
git clone https://github.com/pctan88/delegate-coder.git
cp -r delegate-coder/skills/delegate-coder ~/.claude/skills/
# Or per-project
cp -r delegate-coder/skills/delegate-coder YOUR_PROJECT/.claude/skills/
Or use the install script:
curl -fsSL https://raw.githubusercontent.com/pctan88/delegate-coder/main/install.sh | bash
Quick start
-
Install a worker agent (e.g. MiMo: curl -fsSL https://mimo.xiaomi.com/install | bash)
-
Open Claude Code in your project and say something like:
"Use the delegate worker to summarize the architecture of this repo"
or just work normally — the skill triggers when delegation fits the task.
-
First run, Claude detects installed workers, asks which to use, and saves the choice to .claude/delegate-coder.json.
Configuration details: skills/delegate-coder/references/setup.md
Does it actually save credits?
This repo includes a benchmark harness that measures it properly: paired A/B runs (with/without the skill), real cost numbers from Claude Code's own JSON output, and objective pass/fail verification per task.
What we actually measured (mimo worker, commander.js, Claude Sonnet 4.6, n=3 — see full results): installing the skill gave a ~11% reduction in cost per successful task ($0.493 → $0.440), but the result is uneven and should be read as directional, not conclusive. The savings come from heavy implementation/refactor tasks when delegation actually fires (e.g. a cross-file feature −37%). Bulk codebase reading was a loss with this worker — ~20% costlier and ~3× slower — so despite the intuition, offloading large reads to mimo did not pay off. Review was inconclusive. Run the benchmark on your own repo before trusting any number, including ours.
Note on triggering: with natural, unmodified prompts the skill delegates selectively — readily for heavy reads, reluctantly for small implementation/refactor/review tasks (overall it fired in ~42% of eligible runs), and it correctly declines trivial one-line tasks. Getting it to trigger at all required leading the skill description with task types and stating that installing it implies intent to delegate; see RESULTS.md for the before/after wording. Triggering reliability on execution tasks is the main open item for v2.
Measured results (directional)
