Sand Table — Protocol Meta Skill

Design new sand tables, scaffold project-local skills, extract agent-ops traces, and validate event streams. This skill knows the protocol and all existing implementations.

When to Use

• /sand-table design <use-case> — Design a sand table for a new domain
• /sand-table scaffold — Generate a project-local skill + domain invariant
• /sand-table extract <project-path> — Extract agent-ops from Claude session logs
• /sand-table validate <json-path> — Check event stream against protocol
• "How should I build a sand table for X?"
• "What sand table implementations exist?"

What This Skill Knows

1. The Protocol — Read references/protocol-spec.md for the event envelope, temporal models, normalization contract, execution tiers (Tier 1-3), multi-run comparison, and replay injection patterns.

2. Existing Implementations — Read references/implementations.md for the registry of Substack readership, AIEnablement training, and Agent-Ops implementations with their paths, event types, and temporal models.

3. Domain Design — Read references/domain-invariant-template.md for the scaffold template. Read references/examples.md for real annotated events from all three domains.

4. Reusable Patterns — Read references/patterns.md for the 6 battle-tested design recipes: program invariant structure, agent design, event schema enforcement (including the known LLM drift catalog), output structure, new domain creation, and multi-run comparison.

5. Reliability — Read references/reliability.md for impossible narrative detection, timeout/abort rules, module batching, and cross-session context resolution patterns.

6. Multi-Session Continuity — Read references/multi-session.md for the exit context schema, context loading chain, cohort matching, and context accumulation model.

Commands

design <use-case>

1. Read references/protocol-spec.md, references/domain-invariant-template.md, and references/patterns.md
2. Recommend: temporal model, persona count, event types, scoring dimensions, execution tier (Tier 1-3)
3. Identify closest existing implementation (from references/implementations.md) as a reference pattern
4. Output a filled domain invariant for the proposed domain
5. Flag domain-specific drift risks using the known drift catalog from references/patterns.md
6. Recommend reliability configuration: timeout thresholds, abort conditions, narrative integrity checks (from references/reliability.md)
7. If multi-session: recommend exit context schema and context loading strategy (from references/multi-session.md)

scaffold

1. Ask which domain invariant to use (or design one first)
2. Generate into the current project:
   • A project-local skill (.claude/skills/sand-table.md or similar)
   • A drift-mappings.json for the domain
   • A replay generator stub
3. Register in references/implementations.md
4. Optionally generate a manifest.json for discovery

extract <project-path>

Run the shared extractor:

python ~/Projects/leegonzales/AISkills/SandTable/sand-table/scripts/extract_agent_ops.py \
    --project <project-path> --since <date> -o <output.json>

Then validate the output:

python ~/Projects/leegonzales/AISkills/SandTable/sand-table/scripts/validate_stream.py <output.json>

validate <json-path>

python ~/Projects/leegonzales/AISkills/SandTable/sand-table/scripts/validate_stream.py <json-path>

Validation includes:

• Schema compliance (required fields, valid types, enum values)
• Drift correction using domain drift-mappings.json (see known drift catalog in references/patterns.md)
• Impossible narrative detection for multi-agent simulations (see references/reliability.md)
• Score range clamping and derived field computation

For legacy-format files (pre-protocol), normalize first:

python ~/Projects/leegonzales/AISkills/SandTable/sand-table/scripts/normalize.py \
    --wrap-legacy <json-path> -o <output.json>

reliability <json-path>

Run the full reliability analysis on a simulation output:

1. Read references/reliability.md for the detection patterns
2. Scan all agent events for impossible narrative signals (5 checks)
3. Check for timeout/NR events and assess simulation completeness
4. If multi-session: validate exit context files against schema (from references/multi-session.md)
5. Output a reliability report:
   • Narrative integrity status (CLEAN / N warnings / INTEGRITY CONCERN)
   • Data completeness (NR count, missing events)
   • Context chain validity (multi-session only)
   • Recommendations (re-run, accept, investigate specific agents)

Key Principle

This meta skill is a guide, not a gatekeeper. Project-local sand tables work standalone. The meta skill adds wisdom when consulted — protocol awareness, cross-domain patterns, and normalization infrastructure.