consensus-permission-escalation-guard（共识权限升级守卫）

consensus-permission-escalation-guard

consensus-permission-escalation-guard is the final safety gate before privilege elevation is applied.

What this skill does

• validates escalation requests against a strict input schema (reject unknown fields)
• evaluates hard-block and rewrite policy flags for IAM risk patterns
• runs persona-weighted voting (or aggregates external votes)
• returns one of: ALLOW | BLOCK | REQUIRE_REWRITE
• writes decision artifacts for replay/audit

Decision policy shape

Hard-block examples:

• wildcard permissions (*, : *, broad owner/admin jumps)
• missing ticket reference when required
• break-glass escalation without incident reference
• separation-of-duties conflicts (e.g., create + approve authority)

Rewrite examples:

• weak or non-actionable justification
• temporary duration exceeds policy limit
• production escalation requires explicit human confirmation gate

Runtime and safety model

• runtime binaries: node, tsx
• network behavior: none in deterministic guard logic
• environment config read by this package: CONSENSUS_STATE_FILE, CONSENSUS_STATE_ROOT
• filesystem writes: consensus board/state artifacts under configured state path

Invoke contract

• invoke(input, opts?) -> Promise<OutputJson | ErrorJson>

Modes:

• mode="persona" (default): uses local deterministic persona defaults for internal voting
• mode="external_agent": consume external_votes[], then aggregate and enforce policy deterministically

Install

npm i consensus-permission-escalation-guard

Quick start

node --import tsx run.js --input ./examples/input.json

Tests

npm test

Test coverage includes schema rejection, hard-block paths, rewrite paths, allow paths, idempotent retries, and external-agent aggregation behavior.

Note: this skill depends on consensus-guard-core for aggregation/state helpers; review that package alongside this one for full runtime auditability.

See also: SECURITY-ASSURANCE.md for threat model, runtime boundaries, and deployment hardening guidance.