Prompt file imported from diginsight/tools (.github/prompts/00.09-pe-meta/pe-meta-instruction-review.prompt.md). Copyright stays with the author.

Instruction Review

v15.4 alignment. This prompt honors the vision v15.4 apply = plan + execute contract: every --mode apply run first materializes/reconciles a plan, then executes it; --mode plan materializes the same plan and stops (see § Plan output contract and pe-meta-plan-file-contract.md). The eighth canonical parameter --plan-file sets plan location/identity only; the fresh / reconcile / trust execution modes follow from (baseline-available? × research-runs?), and the § Iteration budget checkpoint (see pe-meta-iteration-budget.md) emits a plan with a trust-mode resume when a run hits the per-cycle change cap. The first-line Resolved invocation: log echoes plan-file=<path-or-none> and spillover=<path-or-none> markers.

Phase 0a CF-05 + Phase 0b — Invocation gates

This prompt enforces the Phase 0a CF-05 artifact-type/path consistency check AND the Phase 0b domain coherence gate defined in 04.05-pe-meta-invocation-gates.md (upstream authority: vision v15 § Domain detection, § Pipeline phases).

Locally true for this prompt:

• CF-05 expected root. Any resolved target path (positional <file-path>, --scope, or design output target) MUST resolve under .github/instructions/. Paths outside this root are REJECTED before Phase 0b runs; the rejection message MUST suggest the canonical replacement prompt name from the SoT § Per-prompt-class applicability matrix.
• Phase 0b scope. Resolved file set = the target path (+ closure under --deps full when this prompt's argument-hint exposes --deps); degenerate single-file scope is single-domain by construction.
• Algorithm. 3-tier metadata-first per the SoT: Tier 1 = each in-scope file's declared domain: frontmatter; Tier 2 = optional pe-domain-map.yaml; Tier 3 = unknown. The seed path does NOT constrain consumer domains when --deps full traverses the closure.
• Gate timing. Runs BEFORE delegating to handoffs declared in this prompt's frontmatter.
• Consent tokens. bundle=accept is recognized AND propagated when delegating to the orchestrator so it does not re-prompt; --skip domain-coherence is REJECTED with CF-05.
• When delegated from an orchestrator. Phase 0a CF-05 is verified by the dispatcher and Phase 0b has already run on the single-domain resolved scope — this section's gate is a no-op in that path.

Process

1. Parse --dim, --deps, --scope, and --skip flags
2. Resolve the applicable-dimension SET from the 05.07-pe-meta-dimension-catalog.md applicability matrix for the instruction type — this is the <applicable> denominator (the full type-applicable set, NOT the 05.08 enumerated subset); then load 05.08-pe-meta-type-checklists.md → instruction section for the SUB-CHECKS of those applicable dimensions that declare rows
3. Audit applyTo pattern — count matched files, flag over-broad patterns
4. Check minimization — flag unnecessary examples, prose, or commentary
5. Run selected dimensions via @pe-meta-validator, recording dim_evidence[] (one anchored {dim, status, evidence_ref} per applicable dimension — passes included) per the § Assess-phase evidence coverage contract
6. If --deps full: run full dependency-aware review (bounded recursion, default depth 2, scope-filtered) and check for overlap with other instruction files
7. If --deps direct: run first-level-only dependency checks (scope-filtered) and check direct overlap risks
8. Run the independent Coverage Audit (@pe-meta-validator, Coverage Audit mode) and reconcile pu-evidence/subcheck-coverage/shallow-sweep; apply the evidence-depth hard-fails before any clean health score
9. Generate severity-ranked report

Phase ordering and option behavior

1. Phase ordering: parse inputs first, execute the type-specific workflow second, then validate and report.
2. Default mode is --mode apply — assess and implement non-breaking improvements autonomously. Use --mode plan to opt into assessment-only output.
3. --dim selects which dimension groups run; default (omitted) = full — the full 05.07 type-applicable set. --dim is subtractive: it may NARROW the evaluated set but the default is never a silent subset. --deps controls dependency traversal: none (per-artifact only), direct (first-level deps), full (bounded recursive). Default: none.
4. --scope filters which dependency types to focus on during --deps traversal (e.g., --scope context focuses on context file dependencies only). When omitted, traverse all dependency types.
5. --skip research|external suppresses external source fetching during review.
6. Guidance-first behavior is handled through /pe-meta-adherence.
7. When --mode apply is active and findings require multi-file changes, delegate to @pe-con-builder.

Risk Classification

#file:.github/prompt-snippets/pe-meta-risk-classification.md

Assess-phase evidence coverage (entry-point depth parity)

A direct /pe-meta-instruction-review call MUST reach the same evidence depth as the /pe-meta-review orchestrator — depth is a property of the work, not the entry path. This prompt therefore includes the first technique module (engine Assess phase):

#file:.github/prompt-snippets/pe-meta-evidence-coverage.md

dim_evidence[] (MANDATORY). For EVERY applicable dimension — passes included — record one {dim, status, evidence_ref} object with a non-empty, anchored evidence_ref (path:line + verbatim quote). A status: pass with an empty evidence_ref does NOT count as covered. The SET of applicable dimensions is the 05.07 applicability-matrix set for the instruction type (NOT the 05.08 subset); an applicable dimension with no 05.08 rows still requires one anchored evidence_ref. Each dimension's evidence_ref set MUST discharge every sub-check declared for the instruction type in 05.08-pe-meta-type-checklists.md.

Independent Coverage Audit (before any clean health score). Hand the run's outcome log to the existing Validate handoff — @pe-meta-validator in Coverage Audit mode (read-only, separate context) — which independently re-derives pu-evidence/subcheck-coverage/shallow-sweep per the shared evidence-bound coverage contract § Independent audit. Divergence is a hard-fail — reconciled, NOT self-attested.

Evidence-depth hard-fail (single-file, mode-independent). Before emitting the report:

• pu-evidence <evidenced> < <applicable> → hard failure on BOTH --mode plan and --mode apply (an evidence-free PASS is indistinguishable from a dimension never exercised).
• any dimension subcheck-coverage <evaluated>/<declared> < 1 → graded partial, BLOCKS a clean health score (a declared sub-check never ran).
• shallow-sweep=suspected → BLOCKS clean until body-level evidence_ref or an explicit acknowledgment is recorded.

The Layer-A anchor checks (resolvability, literal-containment, distinctness) are Evaluation: hook:.github/hooks/scripts/pe-check-evidence-anchors.ps1 — deterministic, zero-LLM, every PU.

Output contract (coverage + plan-file + spillover markers)

The report MUST open with a first-line Resolved invocation: log echoing the plan-file=, spillover=, and the three evidence-coverage markers:

Resolved invocation: --mode=<plan|apply> … | plan-file=<path-or-none> | spillover=<path-or-none> | pu-evidence=<evidenced>/<applicable> | subcheck-coverage=<fully-covered-dims>/<applicable-dims> | shallow-sweep=<clean|suspected>

• pu-evidence / subcheck-coverage / shallow-sweep: computed per the § Assess-phase evidence coverage contract above and reconciled with the independent Coverage Audit. Each carries the evidence-depth hard-fail described there.

• --mode plan (assessment-only): emit an actionable plan file at the canonical plan-mode path per pe-meta-plan-file-contract.md and record plan-file=<path>. When --mode apply, record plan-file=none.

• --mode apply overflow: if the per-cycle change cap is hit with validated findings remaining, emit a spillover plan at <run-folder>/<NN>-<kebab-name>-spillover.plan.md per pe-meta-iteration-budget.md and record spillover=<path>; otherwise record spillover=none.