User Guide

• When to use: Memory-bank governance audit — checks whether T0 laws/current state, T1 supporting context, T2 execution mirrors, adapter freshness, and T3 archive indexes align. The audit is read-only by default.
• Inputs: Command arguments: /constitute-check [optional: 'full' for verbose report, or a specific principle number]; project artifacts referenced below; user decisions and approvals before writes.
• Outputs: A T0-T3 health report plus live and recorded adapter freshness. Write only the approved adapter-state record described below.
• Memory-bank writes: At most memory_bank/t2_execution/adapter_state.yaml, and only after showing the exact YAML draft and receiving explicit approval. No other file may be written.
• Next steps: Follow the workflow hand-off or next-step guidance below; recommendations do not auto-run and require explicit user command/approval.

Purpose before Governance Audit

Confirm asset ownership before instruction purpose. Apply standards/skill-purpose.md after establishing actual CDD framework provenance, ownership and delivery role for the complete package and its required support. Classify purpose only within that confirmed CDD asset collection. Project-owned skills, agents/roles, tools, configuration and documents follow their project's rules, including development capabilities; development use, native instruction fields and skill-directory placement do not make them CDD assets. Renaming or tailoring a CDD template does not itself change its ownership. Unresolved ownership stays pending only for the affected action.

Confirmed CDD developer templates, full platform projections and exclusive support remain local-only in a Target, excluded from Git tracking/staging/commits/delivery/ publication. Within that CDD collection, inspect complete bodies, conditional branches and direct/indirect support; split mixed CDD/project content before submission. Native developer_instructions fields and standalone roles receive the same ownership review even without a skill dependency. Minimal adaptations and independent configuration must not carry or reconstruct CDD framework instruction bodies or exclusive support; a native loader alone does not qualify a copied CDD role as independent configuration. This CDD-related restriction does not exclude independently established project-owned development skills or roles. Names/fields such as test or release do not decide ownership or purpose.

Git exclusion does not disable authorized local retention, discovery, loading, effect or use. Preserve complete required local CDD inputs at usable paths; an input allowlist does not grant them Git eligibility. Missing/unread required originals leave the affected CDD workflow pending. A fresh Target clone may need those inputs provisioned separately. Review Git candidates separately from local inputs and update-write scope. Ordinary template use must not depend on a local maintenance toolkit, reader, lock, live Source or Source ADR reader; optional authorized local maintenance operations remain allowed.

Do not submit a general CDD source mirror, template-maintenance tools/annotations, tool-only metadata, maintenance locks/caches or actual template-maintenance output. The designated CDD Source owns general CDD definitions independently of local storage. Local copies alone do not establish Source authority or authorize moving, deleting or changing log paths.

Identify the actual project and whether the request uses its delivered functions or develops/governs it before domain/audit routing. This CDD governance audit is a local-only Target development template; complete platform copies, bespoke prompts and exclusive framework overlays do not change its CDD ownership. Source delivers its templates/projections as framework product assets. Independent product-use tasks use their real capabilities, without requiring a Memory Bank. Ownership/ purpose review and Git eligibility remain separate from freshness and authorized local use.

Phase 0: Domain Routing

Detect the project domain before checking constitution compliance:

• design/cdd/game-concept.md -> [Game] verify principles against player fantasy, core loop, game pillars, playtest evidence, and game-specific CDDs.
• design/cdd/product-concept.md -> [Product] verify principles against user promise, JTBD, target workflows, product modules, API/CLI contracts, and product-specific CDDs.
• If neither exists, check only the constitution and report that domain-specific validation is pending.

Do not remove game constitution examples. Product checks are an added validation path.

Constitution Check — T0-T3 Memory Health Audit

This skill is read-only by default. Its only write exception is an explicitly approved refresh of memory_bank/t2_execution/adapter_state.yaml; every other project file remains read-only.

This skill checks the health of a project's memory-bank governance plane: whether T0 laws/current state exist, whether T1 supporting context is present, whether T2 execution mirrors exist when the project has been initialized, and whether T3 archive indexes exist once evidence is produced. It is lightweight — not a full 4-gate CDD audit. For a complete project scan, use /project-stage-detect.

---

Step 0: Parse Audit Scope

Accept exactly one mode: no argument (all-law summary), full (all-law verbose), or one positive displayed principle number N or #N. Reject 0, negatives, decimals, arbitrary tokens and combinations such as full 2; show this usage and record invalid selection without silently switching to all laws.

Parsing sets requested scope but does not skip global checks. Continue read-only T0-T3, required-input, tech-context and adapter observations even for invalid, missing or ambiguous selectors. If the constitution is absent, Step 1 still reports that original missing, recommends /constitute and stops. No selector itself authorizes a write or offers an adapter-state record.

Step 1: Verify Constitution Exists

Check for the minimum memory-bank artifacts:

Artifact	Required?	How to check
memory_bank/t0_core/basic_law_index.md	Yes	Read if exists
memory_bank/t0_core/active_context.md	Yes	Read if exists
memory_bank/t0_core/current_state.md	Yes	Read if exists
memory_bank/t1_axioms/tech_context.md	Yes	Read if exists
memory_bank/t1_axioms/system_patterns.md	Yes	Read if exists
memory_bank/t1_axioms/behavior_context.md	Yes	Read if exists
memory_bank/t1_axioms/architecture_context.md	Recommended	Read if exists
memory_bank/t1_axioms/ux_accessibility_context.md	Recommended	Read if exists
memory_bank/t1_axioms/qa_context.md	Recommended	Read if exists
memory_bank/t1_axioms/knowledge_graph.md	Recommended	Read if exists
memory_bank/t1_axioms/module_support_map.yaml	Recommended	Read if exists
memory_bank/t2_execution/workflow_contract.md	Recommended	Read if exists
memory_bank/t2_execution/current_roadmap.md	Recommended	Read if exists
memory_bank/t2_execution/framework_contract.md	Recommended	Read if exists
memory_bank/t2_execution/adapter_state.yaml	Recommended	Read if exists
memory_bank/t3_archive/README.md	Recommended	Read if exists
memory_bank/t3_archive/gate_runs/README.md	Recommended	Read if exists
memory_bank/t3_archive/release_evidence/README.md	Recommended	Read if exists
memory_bank/t3_archive/reviews/README.md	Recommended	Read if exists
memory_bank/t3_archive/reviews/review-index.md	Recommended	Read if exists
memory_bank/t3_archive/sprint_snapshots/README.md	Recommended	Read if exists
memory_bank/t3_archive/sprint_snapshots/story-closure-index.md	Recommended	Read if exists
memory_bank/t3_archive/amendments/README.md	Recommended	Read if exists
memory_bank/README.md	Recommended	Read if exists

If basic_law_index.md is missing: "No constitution detected. Run /constitute to establish one." Stop here — nothing else to check.

If basic_law_index.md exists but older projects only have T0/T1 files, report NEEDS ATTENTION with a migration recommendation. Do not mark old T0/T1-only projects CRITICAL unless a required T0 file or tech_context.md is missing.

If memory_bank/t0_core/knowledge_graph.md exists, report it as a deprecated compatibility path. The canonical path is now memory_bank/t1_axioms/knowledge_graph.md.

---

Step 2: Read the Constitution

Read memory_bank/t0_core/basic_law_index.md and extract:

• Core thesis (Support ID: BL-01)
• All laws/ principles (Support IDs: BL-02 through BL-06)
• Each law's current-state requirement

Build the selector map from substantive displayed Law N/Principle N headings and each entry's actual Support ID. Core Thesis BL-01, unnumbered boundary text, BL suffixes and list positions do not supply principle numbers. In the normal fixture, Law 2 maps to BL-03; if only Laws 1 and 3 exist, 3 still maps to Law 3 / BL-04 rather than the second entry becoming Law 2.

Default and full assess all applicable laws. Numeric focus assesses only its unique matching law. No match reports NOT FOUND; duplicate displayed numbers report AMBIGUOUS with actual candidates and no arbitrary selection. Invalid, no-match or ambiguous requests assess no principle alignment. Label unselected laws NotAssessed, preserving existing evidence/history; do not infer ALIGNED or all-law HEALTHY from focused/unresolved scope. Global findings and severity remain visible in every mode.

Read memory_bank/t0_core/active_context.md and extract:

• Current working state (State A-E)
• Module status summaries
• Active risks and open decisions

Read memory_bank/t0_core/current_state.md when present and extract:

• Current phase
• Current blocker
• Next command
• Stage source

Read memory_bank/t1_axioms/tech_context.md and extract:

• Language, framework, platform
• Performance budgets
• External dependencies

Read T2 files when present:

• memory_bank/t2_execution/workflow_contract.md
• memory_bank/t2_execution/current_roadmap.md
• memory_bank/t2_execution/framework_contract.md
• memory_bank/t2_execution/adapter_state.yaml

Read T3 indexes when present:

• memory_bank/t3_archive/qa_evidence_index.md
• memory_bank/t3_archive/gate_runs/
• memory_bank/t3_archive/release_evidence/
• memory_bank/t3_archive/reviews/
• memory_bank/t3_archive/reviews/review-index.md
• memory_bank/t3_archive/sprint_snapshots/
• memory_bank/t3_archive/sprint_snapshots/story-closure-index.md
• memory_bank/t3_archive/amendments/

---

Step 3: Adapter Freshness Check

For the submission boundary, inspect actual diffs, the Git index and intended delivery/publication contents under the governing Source/Target rules. Establish CDD framework provenance, ownership and delivery role before classifying complete CDD bodies and required support. Project-owned development skills, independent roles, tools, documents, configuration, business code/tests/QA retain project rules; a CDD reference does not change their ownership. Split mixed CDD/project content per its actual parts and support. Required local inputs and Target Git candidates are separate: confirmed CDD templates, generic standards/workflow originals, skill-testing definitions and maintenance tools stay outside Target Git but may remain locally and be discovered, loaded, take effect and be used under existing authorization. Locally readable ignored inputs are not missing Git deliverables and gain no Git eligibility merely because a workflow requires them. Optional authorized maintenance is allowed; ordinary installed-input use needs no toolkit. Ignore rules are not runtime loading controls and do not remove tracked/staged content. Existing tracked CDD reconciliation needs its governance gate, separate exact-path tracking-only migration authority and preservation of local originals, required closure, customizations, recoverable before bytes and Git history. Location/presence alone is not a finding. This audit does not authorize moving/ deleting files, untracking or changing log locations.

First read memory_bank/t2_execution/framework_contract.md when it exists. Its project/Source binding section records the actual project role, sole Source, installed revision and runtime, exact installed paths, required neutral inputs and project-owned customizations. Unknown values remain unconfigured.

When this audit includes import/update or layout reconciliation, read the existing scoped plan/report and apply standards/directory-structure.md, section "Consumer layout resolution for imports and updates". Bind the latest applicable Source inputs and compare actual installed paths, active readers/references, discovery/configuration and generation/comparison chain with the resolved tree and mapping. Internal agreement between legacy inputs and their own adapters does not prove current alignment. Do not treat a planned destination as installed or automatically retain an old path as a customization exemption. Distinguish historical identifiers with actual recoverable originals/witnesses from current dependencies: do not fail mere historical path text or silently waive a stale live reader. Local copy presence alone is not a layout violation. Unresolved layout/ownership/conflicts leave affected alignment pending; only valid applicable comparison evidence can establish fresh/stale. Unselected layout alignment is NotAssessed; ordinary offline installed-input observations still work and do not imply a Source update or new global gate.

If that control-plane file is absent or lacks a binding, use only existing project-owned documentation/configuration selected by the project's governing instructions and defining those same fields and their owners. Do not invent a configuration path/schema, restore a maintenance profile/lock, or create the Memory Bank for this audit. Without a documented binding, report its affected scope as Unknown/NOT AVAILABLE and continue independent installed-input checks; no freshness state write is available. A local manifest or copied script alone does not establish that this is the CDD Source library.

For a consumer, directly read the selected installed entries and their actual required references. Report missing/unread originals with their affected scope. These observations work offline and need no local maintenance toolkit. They do not establish live Source freshness. A consumer freshness comparison must bind its exact installed allowlist against the actual sole Source revision/expected outputs and approved project differences; a Source-only check is insufficient.

If the Source is unavailable, the binding is unresolved, or no valid comparison covers the consumer's installed files, report Live adapter freshness: NOT AVAILABLE with the exact limitation. Preserve any recorded state and continue independent constitution/installed-input checks. Do not add general source mirrors/checkers to consumer Git delivery or label an uninitialized record fresh.

Use only an actually available, authorized read-only comparison that covers the applicable owner and exact installed paths. Without such a comparison, keep NOT AVAILABLE and continue at Step 4 without offering a state write. A missing optional checker does not hide a failed comparison that actually ran.

For an applicable comparison, bind the actual owner, Source revision, full input identities, complete compared paths, approved project differences, result and verification time. A consumer comparison must include its installed allowlist; a Source-library result alone is insufficient. Record the real comparison method without inventing an invocation. Missing, malformed or mismatched evidence is INVALID or NOT AVAILABLE for its affected scope; preserve recorded state and offer no state write. Valid evidence may establish fresh/stale only for its declared scope. Evidence used with the existing adapter-state schema must include schema_version: 1, a status of fresh or stale, full 64-character lowercase hexadecimal manifest_digest and source_digest, ISO-8601 UTC checked_at, the actual checked_commit, counts and check_command. Validate all fields and complete input scope; record the actual comparison method in check_command, never a fabricated invocation. Missing, malformed or scope-mismatched fields are INVALID and offer no state write. Manual observations that do not supply this valid record remain separate from CURRENT/OUTDATED and the optional YAML-record branch below.

Compare an applicable valid live result with memory_bank/t2_execution/adapter_state.yaml:

Recorded state	Meaning
MISSING	The memory bank exists but the adapter-state template was not initialized
UNINITIALIZED	The deterministic template exists and has no recorded evidence
CURRENT	Applicable checked owner/scope, recorded status and both digests match the live result
OUTDATED	A prior record exists but status or either digest differs
UNVERIFIED	A prior record exists but no applicable live comparison is available

Report both values separately: Live adapter freshness: FRESH/STALE/INVALID/NOT AVAILABLE and Recorded adapter state: CURRENT/OUTDATED/UNINITIALIZED/MISSING/UNVERIFIED. Include all non-zero counts and drift paths. A clean checked_commit is contextual; the two digests are the authoritative identity of the checked inputs.

Optional approved state record

Only for default/full or a resolved numeric scope, and a valid live result covering the actual checked owner/inputs, prepare the complete YAML draft with this field order and actual live values. Use its actual portable check description; do not store host-specific paths in a target project. No unavailable or Source-only consumer check is recordable:

schema_version: 1
status: fresh
manifest_version: 2
manifest_digest: "[live manifest_digest]"
source_digest: "[live source_digest]"
checked_commit: "[live checked_commit]"
checked_at: "[live checked_at]"
check_command: "[actual live portable check command/description]"

Use stale instead of fresh when the live check reports drift. Show the entire draft. Reuse existing matching authority for this exact record; otherwise ask exactly:

May I write this adapter freshness record to memory_bank/t2_execution/adapter_state.yaml?

If approved, write only that file. If declined, leave the repository unchanged and state that the report was read-only. Never create memory_bank/ from this skill; when the control plane or framework_contract.md is missing, recommend /constitute.

Step 4: Alignment Check

Check alignment only for principles selected at Step 2. Unselected principles remain NotAssessed; invalid/no-match/ambiguous scope runs no alignment scan. Continue independent global health and tech-context checks in every mode:

3a: Principle → Code Alignment

For each selected law in basic_law_index.md, perform a quick alignment scan:

1. Read the law's current-state requirement
2. Check whether the codebase respects it
3. Flag violations or uncertainties

Example checks:

• Law requires "no function over 20 lines" → grep for functions, check length
• Law requires "API stability: no breaking changes" → check for deprecation markers
• Law requires "all public APIs documented" → check docstring coverage in src/
• Law requires "tests before merge" → check test file count vs source file count

For each law, produce one of:

• ALIGNED — evidence found that the law is being followed
• CONCERN — potential violation or insufficient evidence
• UNABLE TO CHECK — law is abstract or can't be auto-verified (e.g., "prioritize user experience")

3b: Tech Context → Reality Check

Compare tech_context.md against actual project state:

• Is the declared language actually used? (Glob src/** for language files)
• Are declared dependencies actually in the project? (Check config files)
• Are performance budgets plausible given the codebase size?

---

Step 5: Gap Detection

Identify what's missing or stale:

Gap	Detection
Missing T0 current state	memory_bank/t0_core/current_state.md does not exist
Missing T1 required docs	tech_context.md, system_patterns.md, or behavior_context.md missing
Missing T1 recommended docs	architecture_context.md, ux_accessibility_context.md, qa_context.md, knowledge_graph.md, or module_support_map.yaml missing
Deprecated knowledge graph path	memory_bank/t0_core/knowledge_graph.md exists
Missing T2 workflow contract	memory_bank/t2_execution/workflow_contract.md does not exist
Missing T2 current roadmap	Project has roadmap evidence but memory_bank/t2_execution/current_roadmap.md is missing
Missing framework contract	memory_bank/t2_execution/framework_contract.md does not exist
Missing adapter state	memory_bank/t2_execution/adapter_state.yaml does not exist
Uninitialized adapter state	State exists with status: uninitialized; only valid applicable live evidence can initialize the record
Unavailable live comparison	Exact Source/binding/check evidence unavailable; preserve recorded state and continue independent checks
Stale adapters	Valid live comparison evidence reports status: stale or non-zero stale/missing/extra/invalid counts
Outdated adapter record	Recorded status or digests differ from the live result
Missing T3 archive README	memory_bank/t3_archive/README.md does not exist
Missing T3 gate archive	Gate evidence exists but memory_bank/t3_archive/gate_runs/README.md is missing
Missing QA evidence index	QA evidence exists but memory_bank/t3_archive/qa_evidence_index.md is missing
Missing release evidence archive	Release evidence exists but memory_bank/t3_archive/release_evidence/README.md is missing
Missing review index	Review artifacts exist but memory_bank/t3_archive/reviews/review-index.md is missing
Missing story closure index	Completed stories exist but memory_bank/t3_archive/sprint_snapshots/story-closure-index.md is missing
Missing README	memory_bank/README.md does not exist
Stale tech context	Tech context mentions old versions or unused dependencies
Drifted principles	Law says X, code does Y
No review mode set	production/review-mode.txt does not exist

---

Step 6: Present Report

Keep it concise. Use this format:

## Memory Bank Health: [Project Name]

**Core thesis:** [from BL-01, one line]
**Constitution established:** [date from active_context.md or file timestamp]
**Current phase:** [from current_state.md or production/stage.txt]
**Principle scope:** [all / Law N and actual Support ID / invalid / not found / ambiguous]
**Alignment coverage:** [assessed / total; unselected NotAssessed]

### T0 Current Truth
- Laws: [present/missing/stale]
- Active context: [present/missing/stale]
- Current state: [present/missing/stale]

### T1 Supporting Context
- Tech context: [present/missing/stale]
- System patterns: [present/missing/stale]
- Behavior context: [present/missing/stale]
- Recommended context: [summary]

### T2 Execution Control
- Workflow contract: [present/missing]
- Current roadmap: [present/missing/not started]
- Framework contract: [present/missing]
- Live adapter freshness: [fresh/stale/invalid/not available]
- Installed inputs: [observed complete / affected missing or unread paths]
- Recorded adapter state: [current/outdated/uninitialized/missing/unverified]

### T3 Archive Indexes
- Archive README: [present/missing]
- QA evidence index: [present/missing/not applicable yet]
- Release evidence: [present/missing/not applicable yet]

### Principles ([aligned]/[assessed] aligned; [unselected] NotAssessed)
- [Support ID]: ALIGNED — [brief evidence]
- [Support ID]: ALIGNED — [brief evidence]
- [Support ID]: CONCERN — [what's wrong, suggestion]

### ✓ Tech Stack
- Language: [X] ✓
- Framework: [Y] ✓
- Platform: [Z] ✓
- [Any concerns]

### → Gaps
- [Gap 1 — with concrete fix, e.g. "Run /constitute-check for a read-only audit; updates need separate authority"]
- [Gap 2]

### → Recommendations
1. [Highest priority action]
2. [Secondary action]

---
**Global T0-T3/input health:** [HEALTHY / NEEDS ATTENTION / CRITICAL]
**Overall:** [HEALTHY / NEEDS ATTENTION / CRITICAL / PARTIAL — actual alignment scope]

Severity levels:

• HEALTHY: Required T0/T1 artifacts present, T2/T3 indexes appropriate for stage, all checkable laws ALIGNED
• NEEDS ATTENTION: Old T0/T1-only memory bank, stale/uninitialized/outdated adapter state, 1-2 CONCERNs, or missing recommended T1/T2/T3 indexes
• CRITICAL: Missing basic_law_index.md, missing required T0/T1 files, 3+ CONCERNs, or stale constitution

Numeric focus shows the displayed number/actual ID, selected evidence and unselected NotAssessed entries. Global required-input gaps keep their severity, including CRITICAL, regardless of selected alignment. If global observations have no concerns but laws remain unassessed, report Overall PARTIAL for alignment scope; never certify all-law HEALTHY. Invalid/no-match/ambiguous selection retains its unresolved finding and cannot offer a state write. full adds detail for all assessed laws; default summarizes all. Neither mode waives a global gap.

---

Edge Cases

• No constitution: "No constitution detected. Run /constitute to establish your project's governing principles." Stop.
• Old T0/T1-only memory bank: "Your constitution exists, but the T2/T3 governance control plane is not initialized. Run /constitute to refresh the memory-bank skeleton." Continue with NEEDS ATTENTION.
• Missing adapter templates in an older memory bank: Recommend /constitute to initialize both framework files; do not create the control plane from this audit.
• Unavailable Source/comparison: Report NOT AVAILABLE and any historical recorded state separately; preserve it. Complete installed-input and constitution observations offline without requiring Source trees or tools.
• Stale live check: Report every drift class for the actual compared owner/paths and offer to record status: stale; regeneration remains a separate, explicitly requested action.
• Invalid comparison evidence: Report INVALID, preserve any prior state file, and do not write partial or fabricated values.
• Constitution exists but is very old: Note the date — "Your constitution was established [N months] ago. Much may have changed. Consider running /constitute to refresh it."
• Path D project (existing code, new constitution): Flag likely drift — "Your constitution was recently established. Some existing code may not yet align. This is normal for brownfield adoption — prioritize alignment iteratively."
• All laws are abstract (can't auto-check): "Your principles are abstract — they can't be automatically verified. Consider adding falsifiable current-state requirements to each law for future audits."

---

Collaborative Protocol

• Read-only by default. The sole exception is an exact adapter-state YAML record after explicit approval.
• No hidden writes. Running the comparison, declining the draft, or encountering invalid comparison evidence must leave the repository unchanged.
• Be honest. Don't sugarcoat drift or gaps.
• Be specific. Every gap must come with a concrete fix suggestion.
• One primary recommendation. The user should leave knowing exactly one thing to do next.