Secret Handling Agent Skill

secret-handling

The secret-source gate. Routed to when changed code reads, writes, generates, stores, or passes a secret. If a value's secret status is uncertain, treat it as a secret.

Pre-flight

Read these, or STOP and surface the gap — never guess the policy:

• ${CLAUDE_PROJECT_DIR}/.codearbiter/security-controls.md — the approved secret store, the access method (IAM role, workload identity, service-account token — never long-lived static keys), and any required reference format. If this file is unreadable, BLOCK; do not infer the store.

Phase 1 — Identify · gate: BLOCK

Scan the changed code for secret-bearing names: password, secret, token, key, credential, api_key, apikey, private, cert, passphrase. For each match, record its source (where the value originates) and every sink (where it flows). No candidate may remain unclassified.

Gate: every candidate secret is listed with its source and sinks.

Phase 2 — Source · gate: BLOCK

Each secret MUST originate from the approved store in security-controls.md, accessed via the approved method. The following sources BLOCK unconditionally:

• Hardcoded string literal.
• process.env or .env for a secret value — these are for non-sensitive config only (ports, log levels). .env files MUST be gitignored and secrets-scanned on every PR.
• A database column holding the raw value (a stored *reference* is allowed — see Phase 3).
• Any store endpoint not named in security-controls.md.

Stakes: a hardcoded or unapproved-source secret ships to every clone, every CI log, and the

history; rotation is the only fix once it lands. State that when you block, not just "unapproved

source." A genuine catch the user then rotates earns exactly one warm sentence at the close (per

the orchestrator register), never on a clean pass.

Gate: every secret is sourced from the approved store via the approved access method.

Phase 3 — Sinks and persistence · gate: BLOCK

Trace each secret to all sinks. These are prohibited regardless of project, with no log-redaction excuse:

• Any logger call, at any level.
• Client-facing error messages or HTTP response bodies.
• Telemetry, metrics, tracing, span attributes.
• Any LLM prompt or agent context — treat the provider as out-of-boundary; verify shape with length, prefix, or hash, never the value.
• Serialized state, session storage, or JWT payload.
• A database column, except the store *reference* (path, ID, ARN) in the format security-controls.md requires. A migration adding a reference column without a format check constraint BLOCKs.

Secrets MUST NOT outlive the request that uses them — no module-level variable, no instance field, no cross-request cache holding a secret value.

Dispatch the auth-crypto-reviewer agent (${CLAUDE_PLUGIN_ROOT}/agents/auth-crypto-reviewer.md) to confirm these findings against security-controls.md.

Gate: no secret reaches a prohibited sink and no secret persists beyond its request.

On pass — record the gate: follow ${CLAUDE_PLUGIN_ROOT}/includes/security-gate-record.md (the shared record mechanism). For this gate the relevant commit hook is H-10b (secrets). On any BLOCK, do NOT record the pass.

Out-of-scope finding: do not act on it and do not author an ADR (ADRs are user-attributed, via /adr only). Mark it inline with [NEEDS-TRIAGE]; never silently drop it.

Hard rules

• MUST read security-controls.md for the approved store before Phase 2 — BLOCK if it cannot be read.
• MUST NOT source a secret from a hardcoded literal, process.env, a .env file, or any unapproved store.
• MUST NOT let a secret reach a logger, error response, telemetry, LLM prompt, serialized state, or session/JWT payload.
• MUST NOT store a secret value in the database — store the approved-store reference only, with a format check constraint.
• MUST NOT let a secret persist beyond the request boundary.
• MUST record the security-gate-passed marker (via hooks/security-pass.py) ONLY when the gate genuinely passes — the marker is what unblocks the commit (hook H-10b), so a premature or unconditional recording defeats the gate.