---
name: deft-directive-spec-reconstruct
description: >-
  Brownfield specification reconstruction + advise|shadow|enforce drift guard
  (#1589 / #5350). Draft-only synthesis from completed xBRIEFs + #1595
  codebase MAP; never auto-promotes. Use when reconstructing a failed/stale
  spec, recovering requirements, or checking semantic spec drift. Do NOT
  trigger on ordinary spec:render / spec:validate / verify:spec-prd-fresh.
triggers:
  - reconstruct spec
  - recover spec
  - rebuild spec
  - spec is stale
  - spec is failed
  - spec:reconstruct
  - verify:spec-drift
  - spec drift
---
<!-- AUTO-GENERATED by task packs:render -- DO NOT EDIT MANUALLY -->
<!-- Purpose: rendered skill -->
<!-- Source of truth: packs/skills/skills-pack-0.1.json -->
<!-- Regenerate with: task packs:render -->
<!-- Edit the source, not this file. Slice instead of loading every SKILL.md: task packs:slice skills by-trigger --trigger <kw> (or list) -->

# Spec Reconstruct + Drift Guard (#1589 / #5350)

Recover a trustworthy project specification from completed xBRIEFs, then audit drift under advise|shadow|enforce.

Legend (from RFC2119): !=MUST, ~=SHOULD, ≉=SHOULD NOT, ⊗=MUST NOT, ?=MAY.

## When to Use

- Spec baseline has failed or drifted after RFC/ADR/feature landings
- Operator says "reconstruct spec", "recover spec", or "spec is stale/failed"
- Auditing semantic drift via `verify:spec-drift`
- ⊗ Ordinary `spec:render` / `spec:validate` / `verify:spec-prd-fresh` freshness work

## Live surfaces (re-authored)

- Policy: `packages/core/src/policy/spec-guard.ts` via `plan["x-directive/policy"].specGuard`
- Reconstruct: `packages/core/src/spec-reconstruct/` + `task spec:reconstruct`
- Drift: `packages/core/src/verify-source/spec-drift.ts` + `task verify:spec-drift`
- Promote: `task policy:set-spec-guard-enforcement -- --set shadow|enforce --confirm`
- Override hatch: `deft authz:grant -- --template spec-drift-override --target <baseline> --plan-ref <scopeId> --story-ids <ids> --confirm`
- Sync hook: extend `syncSpecificationAfterScopeMove` (#2566) + pre-move gate in `scope:complete`
- Code oracle: shipped `#1595` codebase MAP + `verify:codebase-map-fresh`
- Corpus: `xbrief/completed/`
- Installer skill deposit under `.deft/core/skills/`

## C1 — draft-only reconstruct

```bash
task spec:reconstruct -- --project-root .
# consumer: task deft:spec:reconstruct
```

- ! Emit a **draft candidate only** under `xbrief/.audit/spec-reconstruct-draft.json`
- ! Tag provenance + confidence; keep observed behavior, intended requirement, and unresolved conflict as separate fields
- ! Fold with supersession awareness; defer remainder past adjudication budget to `pending-human-decisions`
- ! Sufficiency feeds `resolveSpecAuthority` (one greenfield predicate) — count is discovery heuristic only
- ! Consume `#1595` codebase MAP as code oracle; do not build a parallel agent-read oracle
- ⊗ Auto-promote draft to `xbrief/specification.xbrief.json`
- ⊗ Treat registry write / projection-fresh alone as proof requirements match reality

## C2 — advise drift guard

- ! Resolve `plan.policy.specGuard` via `readPlanPolicy` (`policy:show --field=specGuard`)
- ! Declare completion impact as `plan.items[].x-directive/specImpact` (`none|delta|new`) — never a bare key
- ! Wire advise trigger by **extending** `syncSpecificationAfterScopeMove` (#2566)
- ! Audit with `task verify:spec-drift` (0 clean / 1 drift / 2 unassessable)
- ! Discharge only against baseline revision + per-completion requirements/delta coverage
- ⊗ Clear unresolved drift with registry/render-only or unrelated spec touches
- ⊗ Splice advisory validate into `task check` in v1
- ⊗ Enforce-default on the framework tree without shadow promote

## C3 — enforce capability (#5350)

- ! Enforcement ladder `advise|shadow|enforce` (default advise). Promote via `policy:set-spec-guard-enforcement` — refuse advise→enforce skip without recorded shadow attestation
- ! Shadow: same evaluator + durable warnings; MUST NOT refuse `scope:complete` or fail CI solely for shadow hits
- ! Enforce named consumers: pre-move `scope:complete` + `tasks/verify.yml` passthrough; exit 1 and 2 hard-fail; not bare `deft check`
- ! Under shadow/enforce: per-item `x-directive/specImpact` (nested delta|new MUST NOT clear uncovered siblings); completion-scoped rewrite proof on durable SPECIFICATION via `resolveSpecArtifactPath`
- ! Operator mint hatch: `authz:grant --template spec-drift-override` binds scopeId + item ids + baseline/revision; Wave-1 grant store is discharge oracle
- ! Seed/reseed: `verify:spec-drift --seed|--reseed`; empty advise ledger is not proven under enforce
- ! Ship `sqaPass` schema only; engine is a documented no-op in v1
- ⊗ Agent-forged override via completing-brief edit; ⊗ one-shot advise→enforce without shadow

## Hardening

- ~ After draft: run interview/probe with a **bounded adjudication budget**
- ! Remainder stays in pending-human-decisions until human resolve
- ! Only after hardening may the operator promote a specification artifact

## Anti-Patterns

- ⊗ Retired Python policy/spec-render script paths as implementation targets (Prefer-A Bound re-author onto TypeScript surfaces)
- ⊗ Bare `plan.policy.specGuard` / bare `specImpact` under #1650
- ⊗ Parallel code-oracle beside #1595
- ⊗ Second greenfield predicate that can disagree with `resolveSpecAuthority`
- ⊗ Building the C3 SQA judgment engine in v1
- ⊗ Splicing enforce into bare `deft check` in v1

## EXIT

deft-directive-spec-reconstruct complete — draft written and/or drift audited; chain to interview/probe for hardening, then operator-promoted `spec:render` only after explicit promote.