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

Prompt Engineering Artifact Management

Unified orchestrator for PE artifacts under the vision v15.4 default-full invocation contract and eight-parameter canonical option surface. This is the single review/optimization command — it absorbs strategic validation (formerly a separate strategic reviewer) as a conditional Phase 4.5 and holds review to the same guidance-quality bar the design prompts construct toward (design/review parity). Phase 0a resolves free-form input to canonical options, the parser validates the eight canonical parameters, breadth is derived, phases 1–8 execute (with Phase 4.5 inserted when --dim includes strategic dimensions), and the first-line Resolved invocation: log makes the actual execution observable before any side effects. --mode apply always materializes a plan then executes it through one execution engine.

v15.4 alignment. This prompt honors vision v15.4.0 § Plan output contract (a plan file is materialized on every mutating run — both --mode plan and --mode apply; apply additionally executes it through one execution engine — see pe-meta-plan-file-contract.md), § Plan execution modes (fresh / reconcile / trust off two booleans), § Model-routing seam (Phases 1–4 reasoning model, Phases 5–7 standard model), and § Iteration budget (overflow spillover plan — see pe-meta-iteration-budget.md). The full retired-flag → v14 destination map lives in the vision v15 changelog § Historical: v13 → v14 deprecated flag map and in the parser test inventory below.

v14 alignment (v2.1.0, 2026-05-29). This prompt was rebased from the v13 surface (--breadth, --since/--between, --area/--artifact/--consumer, --subject/--concern, --mode-review, catch-up, plain --incremental) onto the vision v14 contracts: seven canonical parameters, derived breadth, value-shape --scope parser, Phase 0a conversational pre-parser, per-artifact prompt invocation matrix, pipeline-phases / --skip mapping with rule #2, default-full invocation contract, and the first-line Resolved invocation: log.

Invocation options (canonical eight-parameter surface, vision v15.4)

The orchestrator accepts exactly eight canonical parameters. Any other --* flag is REJECTED with CF-05 and a deprecation notice pointing to its v14 destination. The full retired-flag → v14 destination map is in the migration table below and in the vision v15 changelog § Historical: v13 → v14 deprecated flag map.

--mode plan|apply (default: apply)

Execution modes (fresh / reconcile / trust). Whether the plan is regenerated or trusted is governed by two orthogonal booleans — (baseline available?) × (research runs?) — never by file existence. A baseline is a plan named by --plan-file OR a plan generated earlier in the same conversation. --skip research (existing parameter) decides regenerate-vs-trust:

In reconcile, human-authored rows (park-lot rulings, consent lines, scope tags, rationales) MUST NOT be silently overwritten — contradicting evidence escalates. See pe-meta-plan-file-contract.md § Execution modes / § Reconcile.

Re-running the same command does not raise confidence — it lowers it. A re-run of an identical --scope+--dim invocation is itself a signal the prior pass may be incomplete, so neither reconcile nor trust inherits the baseline's per-dimension verdict as coverage (the inherited verified/pass-weak is treated as never and re-exercised), and a baseline flagged shallow-sweep=suspected keeps that flag until the previously silent body groups are evidenced afresh.

--scope <artifact-type-token>|<path>[,<path>...] (default: all)

Value-shape parser per vision v14 § Option taxonomy. Exactly two unambiguous shapes:

Mixing the two shapes in one invocation is REJECTED with CF-05 (--scope must be EITHER a single artifact-type token OR a path set, not both). When --scope is a single file path, Research narrows to that artifact and its direct dependencies (from the dependency-tracking file — see 00.00-context-structure-index.md → Functional Categories in .copilot/context/00.00-prompt-engineering/).

--source <source-id>|<url>[,...] (default: all monitored sources)

Filters Phase 1 research and Tier-2 screening to the named sources (e.g., --source vscode-release-notes, --source vscode-release-notes,copilot-model-changelog). Each value carries one of two shapes (mirroring the two-shape --scope parser):

• A source-id shape — a token resolving to a configured source in pe-self-update.config.json → monitored_sources. Unknown source IDs are REJECTED with CF-05.
• A URL shape — a raw https://… release-notes URL for an ad-hoc source not in the monitored set (e.g. --source https://code.visualstudio.com/updates/v1_110). Phase 1 fetches it via fetch_webpage and treats it as a single-source ingestion for this run only; it is NOT persisted to the source ledger (no version_scheme, so version-shaped --start/--end is rejected against it — use a date window). This is the explicit external-platform reconciliation path that replaces the retired Release-monitor prompt: auto-fetch of the monitored set is the parameter-less default; a specific release URL is supplied via --source <url>.

--dim <group|D#> (default: full)

Dimension selection: full, freshness, quality, adherence, reliability, model, structural, strategic, context-full, context-health, or specific D1-metadata through D35-portability-boundary.

When --dim efficiency is active with --mode apply, the orchestrator delegates Phase 4 apply work to @meta-optimizer instead of per-type builders, and Phase 4 Research adopts an efficiency lens (token budgets, deduplication, structural improvements). When --dim adherence is active, the orchestrator routes Phase 4 via the per-artifact matrix to pe-meta-adherence for sampled consumer verification.

--start <date|version> / --end <date|version> (default: none)

Explicit research window. Either bound MAY be supplied alone (open-ended interval). Any non-empty value triggers derived breadth=bounded-delta per the resolution rules below.

Value-shape (vision v15.3 — § Processing-state model). Each bound carries one of two shapes, resolved deterministically at Phase 0a (mirroring the two-shape --scope parser):

• A date shape — ISO-8601 (YYYY-MM-DD), now, or a relative offset (-7d, -90d).
• A source-version shape — a token matching the scoped source's version_scheme (e.g. 1.099 for a semver source), resolved to that version's publish timestamp via the source's version_scheme before phases 1–8 run.

A bounded-delta window overrides recorded coverage inside the window — every processing unit whose source-input falls in [--start, --end] is reprocessed even when its recorded status == pass (the re-baseline / distrust-recovery path). Two guard rails (enforced at Phase 0a): a version-shaped bound REQUIRES a singleton --source (rejected otherwise), and a source whose version_scheme is none rejects a version-shaped bound ("use a date window"). This adds no new parameter — it widens the value space of an existing one, so minimal-canonical-surface is preserved.

--deps none|direct|full|<N> (default: none)

Dependency-chain depth for Phase 4 per-artifact work. direct = first-level only; full = bounded recursive traversal (default depth 5); <N> = explicit numeric depth.

--skip <stage>[,<stage>...] (default: none)

--plan-file <path> (default: auto-name)

Plan-artifact location/identity only — the eighth canonical parameter (vision v15.4 § Option taxonomy). It does NOT change what is assessed or applied, and it never decides regenerate-vs-trust (that is governed by --skip research; see the Execution modes table under --mode).

• Omitted (default): the plan lands at the auto-name path <run-folder>/<NN>-<kebab-name>.plan.md; fallback .copilot/temp/pe-meta-state/plans/YYYYMMDD-HHMMSS-<kebab-name>.plan.md when no run folder applies.
• Supplied: --plan-file <path> overrides the location AND marks that path as a baseline for this run. With research running this resolves to reconcile; with --skip research it resolves to trust (drift-guarded). A same-conversation just-generated plan is an implicit baseline even without --plan-file.

See pe-meta-plan-file-contract.md for the full path algorithm, execution-ready-precision clause, and drift-guard rule.

Derived breadth (vision v14 § Default-full invocation contract)

breadth is a resolved attribute — never a flag. The orchestrator MUST derive it before any phase runs and MUST emit it in the first-line Resolved invocation: log.

Rejection. --breadth (any value) is REJECTED with CF-05: --breadth retired in v14; breadth is derived from caller-type and --start/--end — see vision v14 § Default-full invocation contract.

Retired-flag migration table (vision v14 — historical)

Every retired v13 flag is REJECTED with CF-05 using a uniform message template: <flag> retired in v14; use <v14-replacement> — see vision v15 changelog § Historical: v13 → v14 deprecated flag map. The parser MUST table-drive these rejections; do NOT hand-write per-flag prose.

Phase 0a — conversational pre-parser (vision v14 § Option taxonomy)

When it runs. Phase 0a runs BEFORE strict parsing whenever the raw invocation contains any token that is not one of the eight canonical parameters or a recognized value. It is LLM-mediated, so it is the ONLY place in the pipeline where free-form input is allowed to influence canonical resolution.

What it does.

1. Reads the raw invocation (including any natural-language hints like "recheck anything affected by the April VS Code release" or "focus on consumer-correctness for the adherence prompt").
2. Resolves free-form intent into the eight canonical parameters using these rules:
   • Subject keywords (e.g., "April VS Code release") → enumerate matching artifact paths and emit a comma-separated --scope enumeration.
   • Concern keywords (e.g., "consumer-correctness", "freshness") → set --dim to the matching group; if the keyword names a topic without a dimension, also emit a --scope enumeration.
   • Consumer chains (e.g., "the adherence prompt and what it depends on") → emit --scope <consumer-file> --deps full.
   • Temporal hints (e.g., "since April", "between April and May") → emit --start/--end.
3. Echoes the resolved canonical invocation back to the caller BEFORE Phase 1 runs (gives caller a chance to abort or correct).
4. Hands the canonical eight-parameter set to the strict parser. Phases 1–8 NEVER see free-form input.

Non-determinism risk. Phase 0a is LLM-mediated. Two callers passing the same prompt MAY get the same canonical resolution, but reproducibility is not guaranteed. Mitigation: every resolution is echoed back AND logged on the first line of the Phase 8 report (Resolved invocation: log). Downstream determinism is preserved because phases 1–8 only consume the canonical resolution.

Phase 0a value-shape --start/--end resolution (vision v15.3 § Processing-state model)

After free-form resolution settles and BEFORE the strict parser runs, Phase 0a resolves each supplied --start/--end bound to a timestamp so phases 1–8 only ever consume a resolved time window:

1. Detect the shape of each bound deterministically. ISO-8601 (YYYY-MM-DD), now, or a relative offset (-Nd) → date shape (used directly). A token matching the scoped source's version_scheme pattern (semver | dated | model-version) → source-version shape.
2. Resolve a version-shaped bound to a timestamp. Look up the scoped source's version_scheme in pe-self-update.config.json, find that version's publish timestamp (from the source ledger last_seen_* history or by fetching the source), and substitute the resolved timestamp into the window. version_scheme thus plays a dual role — ledger version recorder AND parser version-token recognizer.
3. Enforce the two guard rails (reject with CF-05):
   • A version-shaped bound with a non-singleton --source (or --source omitted, which resolves to all monitored sources) → version window requires a single --source.
   • A version-shaped bound against a source whose version_scheme is none → source <id> has no version scheme; use a date window.
4. Echo the resolved window on the first-line Resolved invocation: log as resolved timestamps (the version token is preserved in the audit narrative). No --incremental token is accepted at this stage.

This widens an existing parameter's value space exactly as the value-shape --scope parser does; it introduces no new parameter and keeps breadth a derived attribute.

Phase 0a precondition — artifact-type/path consistency check (CF-05)

Spec SoT: 04.05-pe-meta-invocation-gates.md § Phase 0a CF-05. The table, rejection message format, and applicability rules below MUST stay byte-equivalent (modulo wording) to the context file. Per-artifact prompts pin their single expected root locally and cite the context file for the canonical mapping.

When it runs. At the END of Phase 0a, after free-form resolution has settled but BEFORE Phase 0b. Runs on every invocation of a per-artifact prompt (e.g. /pe-meta-context-review, /pe-meta-prompt-review). Orchestrator-level prompts (/pe-meta-review, /pe-meta-review, /pe-meta-create-update, /pe-meta-design, /pe-meta-adherence, /pe-meta-scheduled-review) are artifact-type-agnostic by design and SKIP this check.

What it does. Compares the artifact-type ROOT implied by the prompt name against the artifact-type ROOT resolved from the positional <file-path> or --scope value. Mismatch is REJECTED with CF-05 before Phase 0b runs.

Prompt-name-prefix → expected-root table.

CF-05 rejection message format.

CF-05: artifact-type/path mismatch. Invoked prompt /<invoked-prompt> expects positional paths under <expected-root>; supplied path resolves under <actual-root>. Canonical replacement: /<canonical-prompt> '<supplied-path>' <other-args>.

Operates on ROOT, not domain. This check is deterministic from path. It does NOT read frontmatter and is NOT a domain check. Domain coherence is a separate concern handled by Phase 0b (which reads domain: frontmatter from each in-scope file).

Per-artifact prompt invocation matrix (vision v14 § Per-artifact prompt invocation matrix)

After parsing, the orchestrator routes Phase 4 work via this matrix from (--scope-resolved-artifact-type, --dim) → per-artifact prompt. The orchestrator MUST NOT hand-write per-artifact-type branches.

Path-shape --scope. When --scope is a path set, the orchestrator infers the artifact type from the path prefix (e.g., .github/prompts/ → prompts) and applies the matching matrix row.

Per-artifact prompt invocation matrix — Phase 0b inheritance. When a per-artifact prompt is invoked DIRECTLY by the user with a positional <file-path>, it runs its own minimal Phase 0b stub before delegating to the orchestrator. The stub MUST handle the positional-path scope-extraction step per vision v15 § Domain detection § per-invocation-type matrix and MUST read each in-scope file's declared domain: frontmatter to compute the footprint (the seed file's path does NOT constrain consumer domains when --deps full traverses the closure). The algorithm SoT is 04.05-pe-meta-invocation-gates.md — per-artifact prompts cite that file (not this prompt) so sibling-to-sibling coupling is avoided.

Phase 0b — Domain coherence check (vision v15 § Domain-coherent batching)

Algorithm SoT: 04.05-pe-meta-invocation-gates.md. The spec below is this prompt's implementation of the shared contract — algorithm, decision matrix, dispatch table, bundle=… closed set, and bundle=accept consent semantics MUST stay byte-equivalent (modulo wording) to the context file. When the contract changes, update the context file FIRST and re-validate this section with pe-meta-adherence.

Goal. Prevent silent heterogeneous batching by computing the semantic-domain footprint of the resolved scope (seed AND --deps-closure) BEFORE Phase 1 runs. When the footprint exceeds one domain, gate --mode apply until the caller either splits the run per-domain or appends explicit bundle=accept consent. Honest per-file metadata is the input; deterministic dispatch is the output.

Inputs.

• The canonical eight-parameter invocation resolved by Phase 0a (and validated against the Phase 0a precondition CF-05 check).
• The metadata-first 3-tier domain resolution algorithm defined in vision v15 § Domain detection:
  • Tier 1 (authoritative): Read each in-scope file's domain: frontmatter value. Tier 1 NEVER loses to a lower tier when present.
  • Tier 2 (optional heuristic): If pe-domain-map.yaml exists at repo root, apply path-slug heuristic for files without declared domain:. Map entries are flagged per-file in the Phase 8 report so authors can migrate to declared metadata.
  • Tier 3 (fallback): Files unresolved by Tier 1 or Tier 2 receive domain: unknown. unknown is a reserved domain-id; repos MUST NOT declare domain: unknown in any file (the orchestrator emits a Phase 8 warning when a declared value collides with the reserved id).

Algorithm (deterministic, 5 steps).

1. Resolve scope to file set. Expand --scope (artifact-type token → all files under root; path → enumerated files; path-set → union). If --deps direct|full|<N> is present, traverse the dependency-tracking closure to depth N and union into the file set. Tag each file as role=seed (named by the caller) or role=dep (added by closure traversal).

2. Resolve domain per file. For each file, apply Tier 1 → Tier 2 → Tier 3 and record domain-source (declared | path-heuristic | unknown). --dim is NEVER consulted here — --dim is a dimension-group selector per v14 (filters which audit dimensions Phase 2–4 exercises); it never affects domain resolution.

3. Compute seed footprint and dependency footprint SEPARATELY. Seed footprint = {distinct domains across files where role=seed}. Dependency footprint = {distinct domains across files where role=dep AND domain ∉ seed footprint}. Recording the two sets separately is what enables the cross-domain-deps discrimination in step 4.

4. Dispatch from the seed-vs-deps decision matrix.

   Seed footprint	Additional dep-domains	Disposition (bundle=…)	Phase 1 action
   1	0	single-domain	Proceed without prompt
   1	≥ 1	cross-domain-deps	Proceed as ONE review with per-dep-domain specialized analysis lenses in Phase 2–4 (no split — splitting a single-seed cross-domain-deps invocation produces incomplete reviews because the consumer artifact needs all declared deps present to be evaluated)
   ≥ 2	n/a	multi-domain-gated (--mode apply) or multi-domain-advisory (--mode plan)	apply: emit numbered split proposal and BLOCK Phase 1 until user selects a split or appends bundle=accept (→ bundle=accepted-bundle); plan: include split proposal in Phase 8 report and proceed

5. Emit the first-line Resolved invocation: log marker. Append | bundle=<disposition> to the canonical first-line log so the disposition is observable before any side effects occur.

Outputs.

• Domain footprint table in the Phase 8 report with one row per in-scope file and columns path | role (seed/dep) | domain | domain-source (declared/path-heuristic/unknown). Files resolved by Tier 2 or Tier 3 are flagged in the report so authors can backfill declared metadata.

• Numbered split proposal (only when disposition is multi-domain-gated or multi-domain-advisory): one numbered line per detected domain plus a final [N] line for "proceed as one atomic bundle (equivalent to appending bundle=accept)". Example shape:

  Phase 0b detected 3 distinct domains in the resolved scope:
  
  [1] prompt-engineering  (12 files)
  [2] article-writing     (4 files)
  [3] learning-hub        (2 files)
  [4] proceed as one atomic bundle (bundle=accept)
  
  Select a split number (1-3) to run that domain only, run all three sequentially (all), or [4] to proceed as one atomic bundle.
  
  Note: 2 files were resolved by path-heuristic (Tier 2) and 2 files by unknown fallback (Tier 3). Consider adding `domain:` to those files' YAML frontmatter; see the Phase 8 report for the exact file list.

Gate behavior summary.

bundle=accept consent token. Single-keystroke bypass for the multi-domain gate. Appended as a trailing positional token (e.g. /pe-meta-review --mode apply --scope context bundle=accept). It is the ONLY accepted consent value — bundle=skip, bundle=yes, bundle=true, etc. are REJECTED with CF-05. Consent is recorded on the first-line Resolved invocation: log as bundle=accepted-bundle (distinct from the un-consented bundle=multi-domain-gated marker so the audit trail preserves the discrimination).

CF-05 rejection — --skip domain-coherence. Phase 0b is NOT in the skippable set. --skip domain-coherence is REJECTED with CF-05:

CF-05: --skip domain-coherence is rejected; Phase 0b is not skippable per vision v15 § Domain-coherent batching. To bypass the gate on a multi-domain scope, append bundle=accept to the invocation.

--dim orthogonality. --dim is a dimension-group selector per v14 (filters which audit dimensions Phase 2–4 exercises). It is NEVER a domain override. The domain footprint is always computed from per-file declared domain: metadata (with Tier 2/3 fallback) regardless of the --dim value. Parser test P0b-13 pins this invariant.

Determinism. The domain map is computed ONCE at Phase 0b entry and frozen for the rest of the run. The 3-tier algorithm runs in < 1 second on the largest current workspace (51 files, 3 domains); each file's YAML frontmatter is read once and cached. No LLM call required.

Pipeline phases and --skip mapping (vision v14 § Pipeline phases and --skip mapping)

Rule #1. --skip external retires internet/URL fetching across ALL phases that would otherwise use fetch_webpage.

Rule #2 (CRITICAL — vision v15.4). --skip research is INCOMPATIBLE with derived breadth=full because a full sweep without source research is structurally meaningless under the default-full invocation contract — EXCEPT when --plan-file references a validated baseline plan, because the plan IS the prior research product (trust mode). Without a baseline the parser MUST REJECT the combination with CF-05: --skip research is incompatible with derived breadth=full unless --plan-file references a validated baseline plan; either drop --skip research, supply --plan-file, or narrow the window with --start/--end (which derives breadth=bounded-delta). In trust mode the cross-run drift guard (target-section hashing) is REQUIRED: a hash mismatch escalates the drifted row rather than applying a stale edit.

Rule #3 (CRITICAL — execution discipline; no batch-marking). The audit phases (1, 1.5, 2, 3, 4) MUST be executed and status-marked one at a time — each phase is marked in-progress when its work begins and completed only after its work is actually performed and its outputs (findings, outcome-log entries, coverage markers) exist. The orchestrator MUST NOT batch-mark multiple audit phases complete in a single step, and MUST NOT mark a phase complete on the basis of a cheap proxy (e.g., a frontmatter grep standing in for the mandated per-artifact body review). Batch-marking is the behavioral root of the 2026-06-06 full-processing collapse, where audit phases were marked done without execution; the Phase 8 full-coverage linter is the runtime backstop, but this rule is the first line of defense.

Examples

• /pe-meta-review — Parameter-less manual invocation. Derives breadth=full. Runs Phases 0 → 1 → 1.5 → 2 → 3 → 4 → 5 → 6 → 7 → 8 against the entire monitored-sources catalog and the entire PE artifact tree. (default-full-investigation contract.)
• /pe-meta-review --mode plan — Same as above but stops after Phase 4 and produces an assessment-only report.
• /pe-meta-review --scope context — Manual, --scope is the artifact-type token context. Derived breadth=full, but Phase 1.5 Organizational Pass STILL runs because the resolved scope (all context files) is broader than a single file.
• /pe-meta-review --scope .copilot/context/00.00-prompt-engineering/01.07-critical-rules-priority-matrix.md — Single-file path scope. Derived breadth=full, but Phase 1.5 is SKIPPED (single-file gate).
• /pe-meta-review --start 2026-04-01 — Manual + window. Derives breadth=bounded-delta. Phase 1 produces a bounded-window digest between 2026-04-01 and "now".
• /pe-meta-review --start 2026-04-01 --end 2026-04-30 — Same as above with both endpoints.
• /pe-meta-review --source vscode-release-notes — Manual, source-filtered. Derived breadth=full, but Phase 1 consults only the vscode-release-notes source.
• /pe-meta-review --mode apply --dim efficiency --skip structure,consistency --scope context — Efficiency-lens review of context files only.
• /pe-meta-review --mode apply --dim adherence --scope .copilot/context/00.00-prompt-engineering/01.07-critical-rules-priority-matrix.md --deps full — Adherence sampling for a specific guidance file across its full dependency chain.
• /pe-meta-review --incremental (trigger-fired only) — Single-window deprecation alias. Resolves to derived breadth=incremental. REJECTED if invoked manually.
• /pe-meta-review --mode apply --scope .github/prompts/00.09-pe-meta/pe-meta-update.prompt.md --deps full — Manual + single-seed-file + --deps full. Phase 0b expands the dependency closure (vision contexts, dependency-tracking files, instruction files). Seed footprint = {prompt-engineering} (1 domain); dep closure adds {article-writing, learning-hub}. Disposition = bundle=cross-domain-deps — ONE review runs with per-dep-domain specialized lenses applied in Phase 2–4; the review is NOT split because the consumer artifact needs all declared deps present to be evaluated.
• /pe-meta-review --mode apply --scope context — Manual + artifact-type token. Phase 0b enumerates all .copilot/context/ files; seed footprint spans {prompt-engineering, article-writing, learning-hub} (3 domains). Disposition = bundle=multi-domain-gated — Phase 1 is BLOCKED until the user selects a numbered split (1/2/3 per domain) OR appends bundle=accept to convert to bundle=accepted-bundle.
• /pe-meta-review --mode apply --scope context bundle=accept — Same scope as above but with explicit consent. Phase 0b records bundle=accepted-bundle on the first-line log and proceeds without gating Phase 1. Phase 8 report still emits the domain-footprint table.

Rejected examples (CF-05).

• /pe-meta-review --breadth full → --breadth retired in v14; breadth is derived from caller-type and --start/--end — see vision v14 § Default-full invocation contract.
• /pe-meta-review --since 2026-04-01 → --since retired in v14; use --start <YYYY-MM-DD> — see vision v15 changelog § Historical: v13 → v14 deprecated flag map.
• /pe-meta-review --area context → --area retired in v14; use --scope <artifact-type-token> — see vision v15 changelog § Historical: v13 → v14 deprecated flag map.
• /pe-meta-review --subject "April VS Code release" → Phase 0a resolution required; if invoked directly without Phase 0a route, CF-05: --subject retired in v14; free-form intent is resolved by Phase 0a — see vision v14 § Option taxonomy.
• /pe-meta-review --mode-review guidance-first → CF-05: --mode-review retired in v14; use --dim adherence — see vision v15 changelog § Historical: v13 → v14 deprecated flag map.
• /pe-meta-review --incremental (manual caller) → CF-05: --incremental is accepted only for trigger-fired callers; manual invocations are contracted to derived breadth=full per default-full-investigation.
• /pe-meta-review --skip research (manual, no window) → CF-05: --skip research is incompatible with derived breadth=full; either drop --skip research or narrow the window with --start/--end (which derives breadth=bounded-delta).
• /pe-meta-review healthcheck → CF-05: "healthcheck" is no longer a supported preset. Use the eight canonical parameters per vision v15.4 — see § Invocation options.
• /pe-meta-review --skip domain-coherence → CF-05: --skip domain-coherence is rejected; Phase 0b is not skippable per vision v15 § Domain-coherent batching. To bypass the gate on a multi-domain scope, append bundle=accept to the invocation.
• /pe-meta-review --scope context bundle=skip → CF-05: bundle=skip is not a valid consent token; the closed set is {accept}. Use bundle=accept to bypass the multi-domain gate.
• /pe-meta-review --scope context bundle=yes → CF-05: bundle=yes is not a valid consent token; the closed set is {accept}. Use bundle=accept to bypass the multi-domain gate.
• /pe-meta-context-review .github/prompts/00.09-pe-meta/pe-meta-update.prompt.md → CF-05 (artifact-type/path mismatch from the Phase 0a precondition): Invoked prompt /pe-meta-context-review expects positional paths under .copilot/context/; supplied path resolves under .github/prompts/. Canonical replacement: /pe-meta-prompt-review '.github/prompts/00.09-pe-meta/pe-meta-update.prompt.md'.

CRITICAL BOUNDARIES

Always Do

• Parse mode, scope, flags FIRST
• Run Phase 0a precondition (artifact-type/path consistency check, CF-05) on every per-artifact prompt invocation: compare prompt-name prefix against positional-path root; reject with CF-05 and suggest the canonically-correct prompt name on mismatch. Orchestrator-level prompts skip this check (artifact-type-agnostic by design).
• Run Phase 0b (domain coherence check) on every invocation BEFORE Phase 1: resolve scope, apply the metadata-first 3-tier algorithm, compute seed and dependency footprints separately, dispatch from the decision matrix, emit bundle=… on the first-line Resolved invocation: log.
• Honor bundle=accept as the ONLY consent token for the multi-domain gate; record consented runs as bundle=accepted-bundle (distinct from un-consented bundle=multi-domain-gated) so the audit trail preserves the discrimination.
• Load dependency map (the dependency-tracking files — see 00.00-context-structure-index.md → Functional Categories in .copilot/context/00.00-prompt-engineering/)
• Use three-tier classification for every proposed change (see Classification Protocol below)
• In each audit phase Research substep: challenge current state, propose 2+ alternative approaches per finding, compare on effectiveness/reliability/efficiency
• Risk-ordered execution: When multiple findings are produced, execute in this order:
  1. Non-regressive changes first (autonomous-eligible: LOW severity, deterministic classification)
  2. Potentially-regressive changes next (require human approval: MEDIUM/HIGH severity)
  3. Optimization-only changes last (separate cycle if budget allows)
  4. No step is blocked by a higher-risk independent step — execute what you can, escalate what you must
• Propagation-aware priority: Before presenting findings, check the dependency-tracking file (see 00.00-context-structure-index.md → Functional Categories in .copilot/context/00.00-prompt-engineering/) for each affected artifact's dependent count. Sort findings by: severity (primary) × dependent count (secondary). A HIGH finding in a Tier 1 file with 15 dependents takes priority over a HIGH finding in a Tier 5 file with 2.
• Present consolidated plan to user BEFORE applying changes
• Max 3 files between validation checkpoints
• Produce final report regardless of mode
• After every applied change: bump version:, update last_updated:, verify scope.covers topics match content
• Autonomous execution for LOW-severity changes (Phase 1 rollout artifacts only: templates, prompt snippets, hooks JSON):
  • When ALL of these conditions are met, apply without human approval: (1) severity is LOW, (2) classification is deterministic (Tier 1 or Tier 2), (3) pre-change guard passes, (4) post-change validation passes
  • For all other changes: present plan and wait for approval
• Structured change logging (MANDATORY after every applied change):
  • Append a structured entry to 05.04-meta-review-log.md with: artifact path, classification (breaking/non-breaking), confidence (deterministic/LLM-assisted), autonomy level (autonomous/approved), validations passed, outcome (success/pending)

Ask First

• High-impact files (6+ dependents)
• Uncertain source reliability
• 5+ CRITICAL findings in any audit phase
• Any change above LOW severity — present plan before applying

Never Do

• NEVER apply MEDIUM/HIGH/CRITICAL changes without user approval
• NEVER skip validation after changes
• NEVER remove capabilities — only extend, refine, or deprecate
• NEVER classify a change without checking N-1 block labels first (when available)
• NEVER skip Phase 0b — --skip domain-coherence is REJECTED with CF-05; the only legitimate bypass is bundle=accept (which surfaces the disposition on the audit log rather than hiding it)
• NEVER treat --dim as a domain override — --dim is a dimension-group selector per v14; domain footprint is always computed from per-file domain: metadata regardless of --dim value (parser test P0b-13 pins this invariant)
• NEVER override declared domain: frontmatter with a Tier 2 heuristic match — Tier 1 always wins when present; Tier 2 only applies to files without declared domain: and emits a Phase 8 report flag suggesting metadata backfill

Classification Protocol (Three-Tier)

For every proposed change, classify using these tiers in order. Stop at the first tier that produces a confident result.

Tier 1: Deterministic Structural (metadata)

Check without LLM judgment:

• Required YAML fields present (goal:, scope:, version:, last_updated:)
• version: will be bumped after change
• scope.covers: topics not removed from content (compare topic strings against section headings)

If any check fails → CRITICAL finding. No LLM needed.

Tier 2: Deterministic Content (N-1 block labels)

For context files and instruction files with **Rule**:/**Rationale**:/**Example**: labels:

• If the diff touches a **Rule**: block → BREAKING CANDIDATE → requires full validation
• If the diff touches only **Rationale**: or **Example**: blocks → NON-BREAKING → eligible for streamlined processing
• If the section has no N-1 labels → fall through to Tier 3 + flag for N-1 adoption

This is deterministic — parse block labels, classify the diff. No LLM judgment needed.

Tier 3: LLM-Assisted Semantic (metadata as reference)

When Tiers 1-2 don't resolve, use LLM judgment WITH metadata as reference:

• Compare proposed change against goal: — does it contradict the stated purpose?
• Compare against boundaries: — does it violate any declared constraint?
• Compare against rationales: — does it invalidate a design decision?

This is LLM judgment, but significantly better than judgment without any metadata reference point.

Handoff Data Contracts

📖 Researcher output format: .github/templates/00.00-prompt-engineering/output-researcher-report.template.md

Summarization Protocol

Trigger: Before EVERY handoff, estimate accumulated context. If >10,000 tokens: MUST summarize all prior phases to their "Summarize to" format. This is CRITICAL for pe-meta-review which has 8 phases — without summarization, accuracy drops to ~30% at 32K tokens.

📖 Full strategies: .copilot/context/00.00-prompt-engineering/02.02-context-window-and-token-optimization.md

Context Checkpoint Protocol

Between EVERY phase, run this deterministic context size check:

1. Count: List all files read in the completed phase
2. Estimate: files_read × avg_file_lines × 6.67 = tokens_this_phase (avg context file ~150 lines → ~1,000 tokens; avg agent/prompt ~300 lines → ~2,000 tokens)
3. Running total: prior_phases_total + tokens_this_phase
4. Act on thresholds (📖 see 01.06-system-parameters.md → Meta-Pipeline Context Thresholds):
   • ≤10,000 tokens: Continue normally
   • >10,000 tokens: MUST summarize all prior phases per the Summarization Protocol table above
   • >20,000 tokens: MUST write state to .copilot/temp/pe-meta-state/ and recommend new session

File-based isolation boundary: Phases 5–8 MUST use file-based isolation. At the end of Phase 4 (or at Phase 5 entry), write the consolidated approved change list to .copilot/temp/pe-meta-state/phase-5-changelist.md. Phases 6–8 read from this file — they do NOT depend on conversation history from Phases 1–4.

Processing-state model (vision v15.3 § Processing-state model)

Processing state is recorded on two independent axes — never collapsed to one scalar timestamp.

Source ledger (input axis). One file per monitored source at <state.path>/triggers/<source-id>.json (path from pe-self-update.config.json → state.path). It is a rebuildable cache of what the world last looked like:

{
  "source_id": "vscode-release-notes",
  "version_scheme": "semver",
  "last_seen_version": "1.107",
  "last_seen_timestamp": "2026-05-29T00:00:00Z",
  "last_digest_hash": "<sha256>"
}

Artifact coverage (processing axis). The durable single source of truth, recorded per artifact in the bottom validation-metadata block — one cell per (artifact × applicable-dimension):

coverage:
  <dimension-id>:
    source_versions: { "<source-id>": "<version-or-timestamp>", ... }
    depth: research | screening | deep
    status: pass | fail | partial | never
    last_run: "<YYYY-MM-DDTHH:MM:SSZ>"

A processing unit (PU) is one (artifact × applicable-dimension) — applicability follows active-dimensions-follow-evidence. A PU is stale when its status == never, OR status != pass, OR any dependency source's ledger last_seen_version / last_seen_timestamp is newer than the PU's recorded source_versions[dep]. Every writing prompt MUST emit the coverage block in this identical shape (defined here as the canonical block).

Incremental filter (trigger-fired derived breadth=incremental only)

When derived breadth=incremental (trigger-fired caller, no --start/--end), build the PU work set BEFORE Phase 1 from the source ledger and the per-artifact coverage metadata:

1. Read the source ledgers — for each source in resolved --source (or all monitored sources when omitted), read <state.path>/triggers/<source-id>.json (schema above). Version-keyed sources compare last_seen_version; time-keyed sources compare last_seen_timestamp.
2. Compute the PU work set — for each in-scope artifact, for each applicable dimension (per active-dimensions-follow-evidence), include the PU iff it is stale: its recorded source_versions[dep] is older than the dependency's ledger watermark, OR status != pass, OR status == never. Never-covered and non-pass PUs are ALWAYS included (at-least-once guarantee). PUs whose recorded source_versions equal the ledger latest AND status == pass are skipped (no-redundant).
3. Build scope — only artifacts owning at least one stale PU are in scope. Phase 1 produces a change digest (not a snapshot) per the researcher output contract.
4. Apply filter — all audit phases (2, 3, 4) process ONLY the stale PUs.
5. Persist new state — at end of Phase 8, write each processed PU's coverage cell and advance the per-source ledger last_seen_*.

If a source ledger file is missing on a trigger-fired call, warn No prior ledger for source <id>; treating all its PUs as never-covered (full sweep for this source) and treat that source's PUs as stale while leaving the global derived breadth=incremental intact for unaffected sources.

Manual callers MUST NOT reach this branch. --incremental is rejected for manual callers per CF-05 (default-full-investigation). Manual + window callers reach the bounded-delta branch instead, where the explicit --start bound is the watermark and recorded pass coverage is overridden inside [--start, --end] (every PU in the window is reprocessed — the re-baseline / distrust-recovery path).

Phase 1: Source Research (skip with --skip research — only when derived breadth ≠ full)

External knowledge gathering. Feeds findings into Phases 1.5, 2, 3, 4 to inform audit research substeps.

Research output contract (vision v14). The researcher MUST emit ONE of three template-bound output shapes, selected by derived breadth:

Derived breadth	Output shape	Template
full	Current-state snapshot across the entire monitored-sources catalog	pe-meta-research-snapshot.template.md
incremental	Change digest for the stale PUs (per § Incremental filter)	pe-meta-research-digest.template.md
bounded-delta	Bounded-window digest between explicit --start/--end (date or source-version)	pe-meta-research-window-digest.template.md (includes window.start/window.end fields)

1a: Full Research (default)

Delegate to @meta-researcher with the resolved canonical invocation (the agent reads --source, derived breadth, --start/--end). The agent loads pe-self-update.config.json for the monitored-sources catalog and state location. Produce a self-contained report shaped by the contract above with:

• Evidence from authoritative sources (distilled context files always, 05.02 reference articles always, internet research always unless --skip external)
• User-provided authoritative sources (URLs, files) analyzed when supplied
• Critical validation of internet findings — each external finding must be evaluated for whether integrating it would improve artifact reliability, effectiveness, or efficiency; findings that are unuseful, unverifiable, or potentially misleading are flagged and excluded from recommendations
• Improvement opportunities mapped to affected artifacts and quality dimensions
• PE structure assessment (inventory, symmetry, orphans, context coverage)
• For each opportunity: 2+ alternative approaches compared on effectiveness, reliability, efficiency
• Prioritized recommendations classified by audit phase (structure / consistency / content)

When scope is a specific file path: Research focuses on that artifact and its dependency chain only.

1b: Direct application (--scope <file.md> + --mode apply --skip research, derived breadth ≠ full)

When the caller scopes to a single file path AND --skip research AND derived breadth ≠ full (i.e., bounded-delta):

1. Read the target artifact completely
2. Load relevant instruction file for the artifact type
3. Load the dependency map for consumer impact
4. Skip to Phase 5 (user approval) with user's change description as the change spec

This branch is unreachable from manual + no-window callers (rule #2 rejection prevents it).

Phase 1.5: Organizational Pass (skip with --skip organizational)

Gate. Runs ONLY when derived breadth=full AND the resolved --scope is broader than a single file (either an artifact-type token resolving to ≥2 files, or a path set with ≥2 files). Skipped on single-file scopes and on incremental/bounded-delta breadths.

Goal. Before the per-artifact audits (Phases 2–4) drill into individual files, the organizational pass verifies cross-artifact organizational concerns that only surface at the catalog level:

1. Inventory completeness — every PE artifact type has at least one researcher/builder/validator pairing where the architecture requires symmetry.
2. Orphan detection — every artifact has at least one declared consumer in the dependency map (orphan flag if not).
3. Layer cross-cuts — context-layer rules referenced by ≥2 instruction layers are flagged for canonical-source verification.
4. Naming and locator parity — file paths, frontmatter name: fields, and STRUCTURE-README table rows reconcile.
5. Series and sequence integrity — Order in group: markers, sequence labels, and duplicate H1s are detected.
6. Cross-domain consumer chains — agents/prompts that depend on context files outside their declared category are surfaced.

Findings from this phase scope Phase 2 (structure) — Phase 2 narrows from "everything" to "everything plus the organizational concerns Phase 1.5 surfaced".

Execution. Orchestrator runs this directly (using list_dir, file_search, grep_search, read_file). Do NOT delegate to a per-type validator. Output: organizational findings report consumed by Phase 2's screening step.

Phase 2: Structure Audit (skip with --skip structure)

Goal: Validate the PE artifact ecosystem's structural integrity — what files exist, where they are, what role and rules each contains, whether the layout follows conventions.

When --mode apply (and --skip structure is not set): Runs all three substeps (Research, Build, Validate). When --mode plan: Runs Research substep only — Build and Validate substeps are skipped because plan mode is assessment-only. Findings feed into Phase 8 report.

2-Research

The orchestrator performs structural inventory directly (using list_dir, file_search, grep_search, read_file). This is orchestrator-owned work — do NOT delegate to meta-validator for this step.

Checks:

1. Artifact inventory by type across all PE locations
2. Location compliance, role clarity (YAML frontmatter), rules presence (boundaries, tool alignment)
3. Builder/validator symmetry, orphan detection (via dependency map)
4. STRUCTURE-README alignment, dependency map accuracy

When Phase 1 produced findings, incorporate validated external best practices. When --skip external, compare against internal conventions and 05.02 reference articles.

Challenge step: For each issue, propose 2+ resolution options. Compare on effectiveness, reliability, efficiency. Recommend with rationale.

Output: Structural findings report with severity-scored issues and ranked options.

2-Build (--mode apply only)

Delegate to @meta-designer: Transform structural findings into change specifications. Each spec independently executable by a type-specific builder. Include: rationale, alternatives considered, why recommended option wins, layer-ordered execution sequence.

2-Validate (--mode apply only)

Delegate to @meta-validator (Design Validation mode): Verify structural changes won't break capabilities, respect dependency order, preserve builder/validator symmetry. Verdict: SAFE / FIX / UNSAFE.

Phase 3: Consistency Audit (skip with --skip consistency)

Goal: Validate cross-artifact consistency — goal alignment across related artifacts, non-ambiguity, non-redundancy, non-contradiction between files.

When --mode apply (and --skip consistency is not set): Runs all three substeps (Research, Build, Validate). When --mode plan: Runs Research substep only — Build and Validate substeps are skipped because plan mode is assessment-only. Findings feed into Phase 8 report.

3-Research

Delegate to @meta-validator (Ecosystem Audit mode, dimensions: coherence+rules+references). The valida

*Truncated - read the full file at https://github.com/diginsight/smartdocs/blob/6675c3ffb9137858a564427b3a5ea28173114f14/.github/prompts/00.09-pe-meta/pe-meta-review.prompt.m