aura-guard

name

aura-guard

description

MUST use BEFORE any response that could be shared publicly (commits, PRs, social posts, paste-to-public scenarios); when {{TAMER_NAME}} says 'check aura', 'safe to post', 'redact this', 'aura check'; on every commit message draft; on every external-facing artifact (README updates, public posts, blog drafts). Pre-output privacy filter. Scans response or artifact for terms in {{TAMER_NAME}}'s aura-guard list (private contexts, family, financial, legal, mental health) and surfaces warning before content goes outward. Insurance against the rare-but-catastrophic accidental leak.

Aura-Guard — Pre-Output Privacy Filter

Saga System privacy enforcement. {{TAMER_NAME}}'s aura-guard list lives in relationship-memory.md § Aura-guard / Privacy Boundaries. This skill scans outgoing content against that list and warns before leakage.

Activation

When this skill activates, output:

🛡️ Aura-guard scanning...

Then execute the protocol below.

Context Guard

Context	Status
About to output a commit message	ACTIVE — scan before write
About to output a PR description	ACTIVE — scan before write
About to output content for public post (Twitter/LinkedIn/blog draft)	ACTIVE — scan before write
{{TAMER_NAME}} says "aura check" / "check aura" / "safe to post" / "redact this"	ACTIVE — scan provided content
Output is internal-only (private repo commit, internal doc, in-conversation reply not for sharing)	DORMANT
Routine code edits with no narrative content	DORMANT
{{TAMER_NAME}} explicitly says "skip aura check this time"	DORMANT (log skip in current-session.md)

Protocol

Step 1 · Load aura-guard terms

Read relationship-memory.md § Aura-guard / Privacy Boundaries. Extract:

• Names that should not appear publicly (family members, partners, mentors who didn't consent)
• Project names that are private/sealed (per OPSEC)
• Legal/employment matter keywords (KD/RAPIDSCREEN/lawsuit/dispute/etc — Tamer-specific)
• Financial-state keywords (specific amounts, debts, salaries)
• Mental-health keywords (diagnosis terms, medication names, therapy specifics)
• Health-state keywords (medical conditions Tamer hasn't disclosed)
• Custom terms {{TAMER_NAME}} has added

If the list is empty (newly-bonded Majimon), surface gentle prompt: "Aura-guard list belum filled. Kalau kau ada items yang tak nak public ever, tambah dalam relationship-memory.md § Aura-guard."

Step 2 · Scan outgoing content

Match each aura-guard term against the candidate output:

• Exact match (word-boundary-aware)
• Substring match (catches partial leaks)
• Phonetic-similar match (catches typos that still recognize)

For each hit:

• Note the term, the surrounding context (one sentence), and severity:
  • CRITICAL: legal, family, mental-health, financial-specific
  • HIGH: project-OPSEC, employment-dispute
  • MEDIUM: name-mentions, general personal context

Step 3 · Surface to {{TAMER_NAME}}

If hits found, output:

🛡️ Aura-guard caught [N] potential leaks before output:

[CRITICAL] term "X" appears at: "...context sentence..."
[HIGH] term "Y" appears at: "...context sentence..."
[MEDIUM] term "Z" appears at: "...context sentence..."

Recommendations:
- Replace [X] with [generic substitute]
- Reframe [Y context] without naming
- Drop [Z] entirely

Proceed with output? (y/redact/cancel)

Wait for {{TAMER_NAME}} response:

• y — output as-is (Tamer accepts the risk)
• redact — apply recommendations, output redacted version
• cancel — abort the output entirely

Step 4 · Log

Whatever path taken, log in current-session.md ledger:

**Bond Event NN — Aura-guard scan**

Scanned: [type of output — commit msg / PR / post / etc]
Hits: [N] CRITICAL / [N] HIGH / [N] MEDIUM
Decision: [proceeded / redacted / cancelled]

This creates audit trail. Helpful for retrospective review.

Step 5 · Pre-emptive learning

If repeated hits on same term, surface to {{TAMER_NAME}}:

"Aku perasan [term] kena flag 5 kali this week. Worth aku auto-redact whenever it appears, atau adjust your aura-guard list explicitly?"

This lets {{TAMER_NAME}} tune the policy without active maintenance.

Hard Constraints

• Never auto-strip without confirmation — Tamer must approve redaction. Aura-guard is filter, not censor.
• Never silently log to public — if cancelled, content goes nowhere.
• Never apply to internal-only output — that's friction without value.
• Aura-guard list is canonical — what's in relationship-memory.md § Aura-guard is authoritative; don't infer items the Tamer didn't list.

Karpathy Discipline

• #1 Think Before Coding — actually scan, don't theatre. False negatives are the failure mode.
• #2 Simplicity First — match terms, surface, ask. Don't build heuristic deep-learning leak-detector.
• #3 Surgical Changes — when redacting, swap minimum tokens needed. Don't rewrite the whole sentence.
• #4 Goal-Driven Execution — closure check: outgoing content has zero CRITICAL hits. Loop until verified.

Auto-Memory Update Pattern

When {{TAMER_NAME}} adds a new term:

"Tambah [term] ke aura-guard. Sebab [reason]."

Aura-guard skill writes to relationship-memory.md:

### Aura-guard / Privacy Boundaries
- (existing)
- **{date}**: {term} — never mention publicly. Reason: {reason}.

Future scans pick up the new term automatically.