What it's for
Turning a failing adversarial test into a concrete piece of engineering work, rather than a red badge nobody actions.
When you use it
After an evidence run with failures — especially when a failure has made a control contested.
What's on the screen
Expandable cards, each with the title, affected control ID, agent, a status dropdown (draft / PR opened / done), a View PR link where applicable, and the full plan rendered inline.
How it works
Fix plan on a failing test generates a document containing what happened, category-specific implementation guidance, and an acceptance-criteria checklist. Guidance is drawn from a per-category playbook — prompt injection gets instruction-hierarchy and input-delimiting advice; tool abuse gets allowlists, argument validation, and spend caps — and is sharpened by the LLM when a key is configured.
With GitHub connected, the plan is committed to a new branch as docs/veriflow/<slug>.md and opened as a pull request automatically. Without it, the plan is saved as a draft here.
Gotchas
- Creating a plan doesn't fix anything. The control stays contested until the test passes.
- The status dropdown is for your tracking — it doesn't verify the fix. Only a passing test does that.
- Scale plan only.
Related
Fix failures and escalate hard questions · Work through your control checklist