secondme-skill（第二我技能）

secondme-skill

secondme-skill is a complete pipeline for building your AI Second Me — distill your identity from owned data, grow a private knowledge base, train a local model that speaks like you, and govern what gets shared. Local-first, privacy-first, fully yours.

This is an orchestration skill package. It does not replace lower-level capabilities.

Source of truth

• Persona source declaration: persona.json
• Pipeline runtime state: state/pipeline-state.json
• Product and governance spec: references/product-report.md
• Generated runtime pack: generated/persona-secondme-skill/
• Regeneration script: scripts/regenerate-pack.sh

Dependency chain

• Foundation: openpersona -> persona pack creation and lifecycle baseline
• Orchestration: secondme-skill -> workflow gates, state, and report contracts
• Capability chain:

1. anyone-skill -> identity extraction and evidence grading 2. persona-knowledge -> data ingestion, deduplication, wiki/KG, versioned export 3. persona-model-trainer -> local training, evaluation, export, integration

Required execution policy

• Use non-interactive generation and scripts where possible.
• Keep local-first and least-privilege defaults.
• Keep stage outputs auditable with version/hash references.
• Treat persona-secondme-skill/ as generated output (read-only baseline).
• When persona.json changes, regenerate the runtime pack before release.
• Before publishing outside this repository, run scripts/publish-check.sh.
• Preferred release check path: scripts/run-gates.sh (regenerate + sync + model gate + publish gate).

Stage contract

init

• Validate toolchain and directories.
• Initialize or load state/pipeline-state.json.

ingest

• Ingest user-owned data with PII scanning.
• Require explicit source authorization from user.

distill

• Build structured persona extraction artifacts.
• Ensure minimum persona input for OpenPersona is complete.

train

• Route by hardware tier:

- Apple Silicon: mlx - NVIDIA: unsloth - No local GPU: colab

eval

• Check thresholds:

- voice_score >= 3.5 - probe_score >= 0.8 - perplexity degradation <= 20% vs last viable version

integrate

• Integrate model artifacts only when eval gate passes.
• Require runtime pack persona model integration before marking stage pass.

report

• Emit three reports under reports/data, reports/model, reports/deploy.
• Keep report and deployment recommendation in blocked state if persona model gate fails.

Failure routing

• Data gate fail -> return to ingest and request source expansion.
• Train fail -> change backend or reduce model size.
• Eval fail -> augment data or retune hyperparameters, then retrain.

Always update error_code, last_error, and retry_count in pipeline state before retry.

Human approval gates

Require explicit human approval for:

• financial/legal commitments
• account-changing write actions
• external publishing/sharing of identity artifacts

Persona model gate

secondme requires trained persona model integration, not only host default model fallback.

Pass criteria:

• generated/persona-secondme-skill/persona.json contains body.runtime.models.
• body.runtime.models has at least one model entry.
• scripts/check-model-integration.sh returns success.

If this gate fails, report must not be marked pass and deployment recommendation remains blocked.

Sync discipline

1. Edit root persona.json and orchestration docs first.
2. Run scripts/regenerate-pack.sh.
3. Run scripts/check-sync.sh to validate root and generated pack alignment.
4. Verify runtime pack path generated/persona-secondme-skill/ exists and updated.
5. Only then produce release reports under reports/.