Skip to content
Documentation Automatic repair loop

Automatic repair loop

When proof fails, AgentDiff does not immediately ask the developer. It creates a small deterministic failure packet and sends it back to the same agent for a bounded repair attempt:

attempt 1
  → proof fails
  → failure packet: failed phases, failed tests, changed files, policy,
    allowed scope, risk evidence, patch digest
  → same agent runs a bounded repair attempt
attempt 2
  → proof passes
  → REPAIRED

agentdiff repair RUN_ID [--max-attempts 2] [--max-runtime 1800]

Failure packet

The packet is written to .agentdiff/repair/<run-id>/attempt-<n>-packet.json and contains exactly what the repair is allowed to see:

  • failed verification phases and test counts
  • changed files with their policy decisions
  • the immutable policy and the allowed write scope
  • immediate and future blast radius evidence
  • proof reasons, patch digest, base digest

Hard limits (non-negotiable)

  • max_attempts — no infinite retry loops
  • max_runtime_seconds — a monotonic budget (blocked when exceeded)
  • no silent scope expansion — any dependency/CI/config/security change, any review/deny policy action, or high future risk stops the loop
  • no permission escalation — repair runs under the same policy and sandbox
  • the agent can never approve its own permissions

Each repair attempt runs in a fresh workspace derived from the same trusted base, so proof always verifies a clean base plus the repaired patch.

Human Attention Router

HumanAttentionRouter classifies every outcome deterministically:

Kind Condition
AUTO normal source change, proof passes
RETRY proof fails, repair stays in scope
HUMAN dependency added, CI changed, review/deny action, high future risk

Repair loop outcomes: REPAIRED, FAILED, NEEDS_HUMAN, NEEDS_AGENT (no repair command configured — packet written), BLOCKED (budget exceeded).

Repair drivers

The default driver re-invokes the same agent CLI with the bounded repair prompt (prompt-taking CLIs such as codex exec, claude -p, gemini -p have their prompt argument replaced; other commands get the prompt appended). Programmatic callers can inject a custom repair_command_builder.