Documentation menu
On this page

ADR 0006 — Supervisor-side re-verification (G10, v2)

Date: 2026-06-11 · Status: accepted

Context

v1 recorded a deliberate trust gap (G10): the run record's verification outcome was whatever the worker reported. A buggy or dishonest worker could claim completed / passed and the supervisor would believe it. The patch-as-approval model (ADR 0002) means a human reviews evidence before applying — but "review evidence, not promises" is hollow if the central piece of evidence (did verification actually pass?) is itself an unchecked promise.

Decision

After a run the worker reports completed, the supervisor re-verifies it independently (skep/supervisor/reverify.py):

  1. Create a fresh, clean git worktree at the same baseline (repo@ref).
  2. git apply the worker's patch artifact to it.
  3. Re-run the worker's own recorded verification command(s) under the Seatbelt sandbox (deny-all network, writes confined to the throwaway worktree).
  4. Compare exit codes: all zero ⇒ passed; any non-zero ⇒ failed; command not found (127) ⇒ unavailable (the supervisor lacks the toolchain — it cannot honestly confirm or deny).

The worker surfaces the verification command(s) in the verify.result event payload (commands: list[str]). The contract Event keeps unknown payload keys verbatim, so this is additive — not a contract change, and no golden fixture regenerates.

The result is stored in its own reverifications row alongside the worker's claim, with confirmed = (worker said passed AND re-run passed). The canonical run state stays exactly what the worker reported; re-verification is a parallel, equally-durable supervisor judgment. When they disagree, status and review say so loudly ("NOT CONFIRMED — DO NOT TRUST"), and (ADR 0007) auto-approval refuses to fire.

Why not a "validator" agent

The decision record was explicit: the validator is a ~plain function, not a third agent. Re-verification trusts only the exit code of a command the worker already recorded — there is nothing for an LLM to decide. Adding an agent would re-introduce the very trust problem it exists to close.

Consequences

  • The G10 gap is closed for completed claims with a patch and a recorded command. A lying worker (claims passed, patch fails the command) is caught — proven by a MODE:liar fake worker and by the real worker re-running pytest on a clean worktree.
  • Re-verification reuses the Stage A sandbox, so the supervisor's own re-run is itself network-denied and write-confined.
  • Honest limits: re-verification fidelity assumes the supervisor shares the worker's toolchain (true in v2 — co-located on one Mac, G3); a missing tool is reported as unavailable, never a false failed. It does not re-verify failed/pending_approval/no-patch runs (nothing was claimed verified).