Change one edit tool, watch a weak model jump from 6.7% to 68.3%
Swap the edit tool for hashline edits to lift weak models. Harness tinkerers; users of oh-my-pi (OMP).
what to steal
If a cheap model fails at edits, fix the edit format before you upgrade the model.
Our takeaway from the source, not a quote.
can-boluk-omp-hashline.toml
# Swap the edit tool for hashline edits to lift weak models # source: https://blog.can.ac/2026/02/12/the-harness-problem/ [rig] who = "Harness tinkerers; users of oh-my-pi (OMP)" author = "Can Bölük" date = "2026-02-12" harness = ["OMP (oh-my-pi)"] models = ["Grok Code Fast 1 (and others benchmarked)"] [keep] - Hashline edit format (line-hash-anchored edits) replacing the default edit tool [cut] - Default edit tool
⚠ Grok Code Fast 1: older model(s), named by the source on 2026-02-12. See current models.
| thing | status | why (per source) |
|---|---|---|
| Default edit tool | replaced | Grok Code Fast 1 went from 6.7% to 68.3% in his benchmark; ~+15 pts on average; ~$300 spent benchmarking |
Source: https://blog.can.ac/2026/02/12/the-harness-problem/ · https://github.com/can1357/oh-my-pi · checked 2026-10-14
// similar rigs
~/rigs/merlin-claude-manager-codex-….toml𝕏 · 2026-10-14
Claude manages, Codex grinds
steal → Use one model to plan architecture and another to churn through build steps.
Claude CodeCodex Claude Fable as manageGPT-6 Astra on Codex
by @Merlin412188240
~/rigs/mikeholp-route-claude-codex.toml2026-10-11
Route by task type and make each agent review the other
steal → Have the agent that didn't write the change review it.
Claude CodeCodexCursor Claude Opus 5.5Claude Sonnet 5.5 git push / commit in C
by Mike Holp
~/rigs/joekarlsson-opencode-litellm.toml2026-10
Broke up with Claude Code: open-weight models at about two cents a task
steal → Put every model call behind one gateway with per-app budgets, and force a test run before 'done'.
OpenCodeClaude Code DeepSeek V4.1 FlashClaude Claude Code as default
by Joe Karlsson