Operator direction (2026-07-15): « idéalement il faudrait que ça fonctionne avec Mistral ou bien Ornith et MLX en local » — ça being the harness loop proven this morning (builder subagent → independent cold verifier → merge, erp#38/PR erp#62).
model-fleet.md › new section "Harness portability": the atoms are multi-model by construction; the orchestration layer must be too. The protocol already carries everything in files + issues (AGENTS.md, Execution footers, in-issue acceptance tests, deterministic validators) — any runtime that reads the forge and runs a shell can attempt an issue. Evidence-gated migration ladder: (1) verifier roles first (re-run the already-passed erp#38 locate-test on Ornith and an MLX-served candidate, require verdict parity; then prefer cross-family verification — a different family refuting the builder is stronger evidence); (2) scoped builders benched (one small issue end-to-end on a Mistral-driven runtime, unchanged gates); (3) Claude default for multi-file/ambiguity/fiscal until the bench says otherwise.
agent-architecture.md › D2 row: records the direction (multi-runtime harness; Claude = escalation tier) and points at the ladder.
## What
Operator direction (2026-07-15): « idéalement il faudrait que ça fonctionne avec Mistral ou bien Ornith et MLX en local » — *ça* being the harness loop proven this morning (builder subagent → independent cold verifier → merge, erp#38/PR erp#62).
- **model-fleet.md › new section "Harness portability"**: the atoms are multi-model by construction; the orchestration layer must be too. The protocol already carries everything in files + issues (AGENTS.md, Execution footers, in-issue acceptance tests, deterministic validators) — any runtime that reads the forge and runs a shell can attempt an issue. Evidence-gated migration ladder: (1) **verifier roles first** (re-run the already-passed erp#38 locate-test on Ornith and an MLX-served candidate, require verdict parity; then prefer **cross-family verification** — a different family refuting the builder is stronger evidence); (2) **scoped builders benched** (one small issue end-to-end on a Mistral-driven runtime, unchanged gates); (3) Claude default for multi-file/ambiguity/fiscal **until the bench says otherwise**.
- **agent-architecture.md › D2 row**: records the direction (multi-runtime harness; Claude = escalation tier) and points at the ladder.
- Spike issue: erp#63.
QA: `scripts/prd_check.py` — NONE BROKEN.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Operator direction 2026-07-15: the orchestration layer itself (builder
sessions, cold verifiers) must run on Mistral or hermes+Ornith/MLX too.
The protocol already carries everything in files+issues; new model-fleet
section defines the evidence-gated ladder — verifier roles migrate
first (cross-family refutation is stronger verification), scoped
builders benched on unchanged acceptance gates, Claude default until
the bench says otherwise. D2 row records the direction; spike = erp#63.
Co-Authored-By: Claude Fable 5 <[email protected]>
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
What
Operator direction (2026-07-15): « idéalement il faudrait que ça fonctionne avec Mistral ou bien Ornith et MLX en local » — ça being the harness loop proven this morning (builder subagent → independent cold verifier → merge, erp#38/PR erp#62).
QA:
scripts/prd_check.py— NONE BROKEN.🤖 Generated with Claude Code