backward-sweep-auditor
Cold backward-sweep for a §1.7 Backward-check — given only a change's class/logic, enumerates every parallel surface where that class applies and reports GAP/CLEAN per surface; PR-blind, reporting-only.
backward-sweep-auditor
Status: shipped-beta (framework-side authoring tool) · Ships to: clone — excluded from consumer delivery (setup.d/20-agents.sh:28 case-skip) · Fires at: framework authoring sessions — dispatched by an author about to write a ### §1.7 Backward-check applied section (agents/backward-sweep-auditor.md:27)
What it is
Sub-agent prompt for the cold, PR-blind enumeration of every codebase surface where a given change-class applies, reporting GAP/CLEAN per surface for a §1.7 Backward-check (agents/backward-sweep-auditor.md:11-13). Reporting-only — no fix, edit, or commit; classification is operator-only (authoring-only), not shipped to consumers (agents/backward-sweep-auditor.md:29).
How it works
- Cold mechanism. The §1.7 Backward-check fails as restatement (T21) when the author's context is saturated with the PR's own narrative; the syntactic CI gate cannot tell restatement from sweep. This agent defeats it structurally — "you never saw the PR, so you cannot restate it" (
agents/backward-sweep-auditor.md:42-43). - Input contract. Only the change's class/logic — a content predicate abstracted from where it was applied. If dispatched with a diff and no explicit class, derive the class first, then sweep — never sweep the diff's file list (
agents/backward-sweep-auditor.md:64-66). - Method. State the class as a one-sentence predicate → enumerate the complete surface set with real Grep/Glob/Bash evidence (population count stated before verdicting, per T10) → per-surface verdict
SWEPT-CLEAN(quoted guardfile:line) orGAP-FOUND(quoted site + reachability proof) → list the surfaces NOT touched by the originating change — "the deliverable's whole point" (agents/backward-sweep-auditor.md:68,:79-81). - Output. Per-surface rows plus overall token GO/REVISE/STOP: any
GAP-FOUND→ REVISE; incomplete population or too-vague class → STOP; complete population + zero GAP-FOUND → GO (agents/backward-sweep-auditor.md:88-90).
Satellites & companions
- USES: the §1.7 backward-check discipline (SSOT: phase-research-coverage.md §1.7) and the T21 trap definition — pointed at, never re-described (
agents/backward-sweep-auditor.md:7). - PEER: shares the reviewer-discipline clauses and the GO/REVISE/STOP verdict vocab sourced from dispatch-input-checker.md §Output grammar (
agents/backward-sweep-auditor.md:21); same authoring-only skip-loop family as the other cold auditors. - ADDS: the incident-grounded cold-seat construction — PR #857 shipped a restatement backward-check; the real parallel gap was caught only by operator challenge (
agents/backward-sweep-auditor.md:38-40).
Anchors
agents/backward-sweep-auditor.md:2—name: backward-sweep-auditoragents/backward-sweep-auditor.md:3—description: Cold backward-sweep for a §1.7 Backward-check. Given ONLY a change's class/logic (never the PR diff or narrative) …setup.d/20-agents.sh:28—backward-sweep-auditor.md) continue ;; # authoring-only tool (§1.7 backward-check cold-sweep, T21)agents/backward-sweep-auditor.md:11—> **Authoritative for:** the `backward-sweep-auditor` sub-agent prompt — the cold, PR-blindagents/backward-sweep-auditor.md:29—commit. **Classification — operator-only (authoring-only), not shipped to consumers.**agents/backward-sweep-auditor.md:43—load-bearing property — **you never saw the PR**, so you _cannot_ restate it.agents/backward-sweep-auditor.md:68—## Method (no prose-only findings — per T3)agents/backward-sweep-auditor.md:86—## Output format
All anchors at framework pin b069c59328aa8e08671d3a19fd25efa618dfc1c5 (fetched 2026-09-11 from artyhoo/getff staging; ancestry of census pin a1337cb301 verified).
adapter-jig-reviewer
Cold adversarial multi-dimension review of an ecosystem-adapter wiring diff against the eight §3 conformance groups, one structured verdict per group; PR-blind, reporting-only, never CI-invoked.
dispatch-input-checker
Cold dispatch-input reality-check at the aif-dispatch station boundary — judges whether a dispatch input is fit for an executor to burn tokens on, with a machine-consumed DISPATCH-INPUT verdict recorded as a calibration-ledger row.