The thesis: the harness is the lever
Veyyon exists because the same model weights score wildly differently depending on the agent harness around them. The model is capable; scaffolding restricts or unlocks it.
The evidence
Benchmarks and production traces show tool shape, edit shape, context handling, and control flow can swing outcomes dramatically. Veyyon’s oh-my-pi lineage adds hashline editing, role-based models, compaction, and explicit modes (plan, goal, vibe) as harness levers, not only prompt text.
The two dominant levers
- Edit format / first-attempt edits. When the edit format is hard to emit, models burn turns on retries. Hashline and model-tuned edit prompts are the biggest swing in Veyyon.
- Control flow. Stop when verification passes; do not loop on repeated failures; budget context and subagent fan-out. Goal and plan modes encode some of this in the engine.
What this implies for the design
- Hashline and native edit tools as the primary write path in
packages/coding-agent(see engine docs underdocs/). - Per-model and per-role configuration via
modelRoles, thinking levels, and catalog selectors. - Engine-enforced modes (plan file + approval, goal continuation, tool approval tiers).
- Evidence discipline. Claims in this book must match tests and engine docs, or be labeled Spec — not shipped.
Harness improvements are runtime work
Dogfood traces, benchmark failures, and user corrections drive small runtime improvements: better tool hints, compaction, cache-stable prefixes, clearer progress, and bounded outputs. Name the lever each change moves.
Where to go next
- What makes Veyyon different
- Repair (note: the full repair cascade is Spec — not shipped)
- The hashline edit engine