Per-model repair posture

Status: Partial — repair shipped; per-model knobs are an early cut via harness profiles.

Shipped

• Repair threads the active model through the repair hook, so per-model posture can vary.
• Harness profiles: harness.profiles in config.yml or harness-profiles.yml in the agent dir. Keys match provider/model-id or provider/* wildcards.

harness:
  profiles:
    "anthropic/claude-sonnet-4-20250514":
      repair: true
      tools: ["read", "edit", "grep", "bash"]
      promptSectionOrder: ["activeRepo", "system", "project"]
    "google/*":
      repair: false

• repair: false skips schema repair for that model.
• tools: [...] filters the initial tool allowlist (MVP hint surface).
• promptSectionOrder: [...] reorders the top-level system-prompt blocks for that model. The three addressable blocks are system (the rendered system-prompt.md/custom prompt), project (the environment/workspace footer), and activeRepo (nested-repo notice, when applicable). Named blocks move to the front in the order given; unlisted blocks keep their default relative order after the named ones. This reorders whole blocks only — it does not reorder content within system-prompt.md (e.g. moving “Delegation” before “Tool Policy”), because that template renders as one piece, not a list of named sections.

Disable all repair: VEYYON_REPAIR_DISABLE=1.

Spec — not shipped

• Per-(model,tool,shape) telemetry counters
• Strictness tables (refuse earlier vs coerce more) beyond the ambiguity guard
• Full backends.toml three-backend tuning
• Reordering within system-prompt.md (Skills vs. Tool Policy vs. Delegation, etc.) — only the three coarse top-level blocks above are addressable today; the template itself is monolithic.

See Why repair exists and Models.