ADR 0007 — Auto-approval policy rules (D3, v2 mechanism)
Date: 2026-06-11 · Status: accepted
Context
ADR 0002 made applying the patch the single Queen-side approval, performed by a
human via review --approve. The north-star workloads (U1: the dependency/audit
bot) need that approval to happen without a human when it is safe to — but
"safe" has to mean something checkable, and every grant of autonomy has to be
auditable after the fact.
Decision
Policy gains the power to grant autonomy, not only to deny actions:
declarative AutoApprovalRules (skep/supervisor/policy.py) that auto-apply a
worker's patch when all enabled conditions hold:
require_verification_passed— the worker reportedpassed;require_reverified— the supervisor's own re-verification confirmed it (G10, ADR 0006) — autonomy is gated on independent evidence, not the worker's word;forbid_risk_flags— the result carried norisk_flags;diff_scope— every changed file matches an allowed glob (e.g. lockfiles only);max_changed_files— an optional ceiling.
The first matching rule fires: the patch is applied on skep/<task_id> through
the same shared apply_patch_on_branch the human path uses, and an approval row
is recorded resolved_by = auto:<rule> with a note naming the rule and the
conditions it matched. This sits entirely on top of patch-as-approval — zero
contract change.
Built in v2, active in v3 — dormant by default
SupervisorConfig.auto_approval_rules defaults to empty: with no rules, nothing
is ever auto-approved and the human loop is byte-for-byte unchanged (the golden
smoke and reliability gate run with no rules). The mechanism is fully built and
tested now; it is switched on for real recurring workloads in v3 (U1), which is
also when the inputs it reasons about (network allowlist, scheduling) arrive.
Consequences
- The G10 ↔ D3 seam is the safety property: a worker that claims
completedbut fails re-verification is never auto-approved (proven by aMODE:liarend-to-end test). Re-verification is not decoration — it is the gate autonomy depends on. - Escalation over silence: a rule that matches but whose patch fails to apply is left as a pending approval for a human, not dropped.
- Every autonomous action is as auditable as a human one — same approval row,
same
Approved-bycommit trailer, withauto:<rule>instead of a username.