self-healing-policies

Self-Healing Prompt Policies

FlowDeck can learn from its own mistakes. When the same type of failure recurs, a new policy is added to ~/.fd-plan/<slug>/.codebase/POLICIES.json to prevent it from happening again.

How Policies Work

A policy is a trigger → rule pair:

• trigger: the pattern that precedes the mistake (e.g., "editing auth middleware")
• rule: what the agent must do/avoid when the trigger matches (e.g., "always run auth integration tests before committing")

Policy Lifecycle

1. Detection: A failure is recorded in FAILURES.json with recurrence_count ≥ 2
2. Proposal: The agent proposes a new policy to prevent the recurrence
3. Review: Human reviews and approves/rejects via /policy-update
4. Active: Policy is marked active=true and checked before relevant edits
5. Violation tracking: Every time a policy is nearly violated, record it
6. Pruning: Policies with 0 violations in 90 days are candidates for deactivation

Policy Sources

Source	Meaning
manual	Added by a human directly
learned	Proposed by the agent after failure pattern detected

Workflow

Adding a Learned Policy

After recording a recurrent failure:

1. Check if recurrence_count ≥ 2 for any failure entry
2. If yes, propose a policy:

   { "action": "add", "policy": {
       "id": "no-auth-edit-without-tests",
       "name": "Auth edits require test run",
       "trigger": "editing files matching src/auth/ or src/middleware/",
       "rule": "Always run auth test suite before committing. If no auth tests exist, create at least one regression test.",
       "source": "learned",
       "failure_count": 3
   }}
   

3. Notify the human for approval

Checking Policies Before Edit

Before editing a file:

1. Query active policies: { "action": "query", "query": { "active_only": true } }
2. Check if any policy trigger matches the current file path or change type
3. If match: apply the rule before proceeding