Memory Index Lint

Memory Index Lint

Purpose / When to Activate

Validate that index.md and archive/superseded-decisions.archive.md satisfy the discoverability invariants required for AI agents to reliably navigate the memory corpus without missing context.

Activate when:

• Pre-commit (via git hook)
• Post-merge (via CI)
• Before invoking memory-archive-superseded (precondition check)
• After invoking memory-archive-superseded (post-condition check)
• After manual edits to index.md or archive files
• As part of memory-index-sync post-step

Two execution paths

L4 — Capability-dependent script (preferred when shell available)

When the agent has Bash/shell access (Claude Code, Cursor with terminal, daemon claude -p) :

python3 scripts/validate-memory-index.py
# Exit 0 = CLEAN
# Exit 1 = VIOLATIONS (printed to stderr/stdout)
# Exit 2 = USAGE / IO error

# Strict mode: treat soft warnings as failures (use post-migration)
python3 scripts/validate-memory-index.py --strict

# Quiet mode: print only final summary line
python3 scripts/validate-memory-index.py --quiet

Pre-commit hook invokes the script automatically — no manual invocation needed in most flows.

L3 — Tool-agnostic in-session prompt (fallback when no shell)

When the agent is on a surface without shell access, or when invoking the script is not appropriate, perform these checks manually by reading the files :

Invariant checks (mandatory)

1. Table column consistency. For every Markdown table in index.md and the archive, every row has the same number of |-delimited cells. If a row has different cell count → format corruption, FAIL. (This catches the kind of manual-edit corruption where two table rows get concatenated into one — empirically observed in earlier index revisions.)

2. Active ↔ archive symmetry. Every DEC ID in the "Superseded (archived)" pointer table of index.md MUST have exactly one matching row (column 1) in archive/superseded-decisions.archive.md § "Superseded DEC entries", and vice versa. Dangling pointer (active claims archived but no archive row) = FAIL. Orphan archive row (archive has DEC but active doesn't reference it) = FAIL.

3. DEC frontmatter source-of-truth. For every DEC referenced in the archive index, read decisions/DEC-XX.md frontmatter and verify :

   • status: superseded (not active, not candidate, not draft)
   • superseded_by: is set (DEC-X or [DEC-X, ...] — non-empty, non-null)
   • archived_to: archive/superseded-decisions.archive.md
   • archived_at: is a date string

4. Path stability. Every DEC referenced in active index OR archive index has its file at decisions/DEC-XX.md on disk. Exception: a DEC ID may be reserved-no-file (declared in the index as such — e.g. a contingency pivot reservation that intentionally holds a slot without a body file).

5. Supersedeur backlink (V2 — soft check V1). For every supersession edge DEC-X → DEC-Y, the supersedeur DEC-Y's frontmatter should declare supersedes: [DEC-X, ...]. Missing or asymmetric backlink = WARN (not FAIL in V1 ; promotes to FAIL in V2).

Affordance signal check (soft)

6. "How agents read" section present. If index.md references "Superseded" or has a "Superseded (archived)" section, it MUST contain an affordance block telling agents :

   • Files are still readable at original path
   • Read both endpoints for "why" questions
   • Follow chains end-to-end
   • Use the archive skill to archive new supersedeses

   Missing this block = soft warning (non-blocking) in V1. Promotes to hard fail post full-migration.

Reporting

After running all checks, produce an inline report :

# Memory Index Lint — {date}

## Hard violations
- [<category>] <description> ... (or "none")

## Soft warnings
- [<category>] <description> ... (or "none")

## Result: CLEAN | VIOLATIONS_FOUND | SOFT_ONLY

If hard violations found, the operation that triggered the lint should be reverted or escalated. Soft warnings are informational and do not block.

Worktree + parallel safety

• Linter is stateless — reads filesystem snapshot at invocation time, produces report, exits
• Idempotent — running twice in succession gives identical result (no side effects)
• Safe in parallel — multiple worktrees can run the linter concurrently ; each sees its own working tree state
• Cross-worktree backstop — pre-commit catches per-worktree commits ; CI catches merged-state anomalies (when WT-A and WT-B independently commit valid changes that produce invalid merge)

Quality Checks

• Linter is read-only — never modifies any file
• Zero external dependencies (Python stdlib only — no PyYAML, no markdown parser)
• Fast — full validation completes in <2s on a corpus of ~100 DECs
• Returns structured exit codes for hook integration (0 = clean, 1 = violations, 2 = usage error)
• Both modes (--strict and default) coexist — pre-commit uses default tolerant mode during in-flight migration ; post-migration switch to --strict for hard enforcement of affordance signals

Non-Goals

This skill must NOT :

• Modify any file
• Auto-fix violations (would mask integrity issues)
• Validate semantic correctness of decisions or rationales
• Replace memory-index-sync (which heals drift) or memory-archive-superseded (which migrates rows)
• Run on decisions/DEC-XX.md content beyond frontmatter
• Validate _log.md or other auxiliary files

memory-index-lint validates structural invariants. It does not enforce policy.

Related skills

• memory-archive-superseded — invokes lint as pre + post condition
• memory-index-sync — heals drift between disk and registry ; runs lint as final step
• memory-compact — compaction operations should run lint before commit