메뉴
Use when changes touch auth, permissions, secrets, env files, deploy config, external messages, data export/import, production DB writes, or customer-visible behavior; surfaces sensitive-path findings + approval-required list. Do NOT use as general code review.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | security-permissions-review |
| description | Use when changes touch auth, permissions, secrets, env files, deploy config, external messages, data export/import, production DB writes, or customer-visible behavior; surfaces sensitive-path findings + approval-required list. Do NOT use as general code review. |
| source_inspiration | conceptual-only |
| license_status | clean-room-original |
| fusebase_flow_version | 2.1 |
| risk_level | high |
| invocation | manual |
| expected_outputs | ["security review summary (chat or docs/verification/<slug>-security.md)","sensitive-path findings list","approval-required list (operations needing approval artifact)","mitigation steps"] |
| related_workflows | ["verification-gate.md"] |
| hook_dependencies | ["permission_request","pre_tool_use"] |
Targeted review for changes that touch security-sensitive surfaces. Distinct from general code-review because the failure modes (credential leak, auth bypass, customer data exposure) have different response thresholds — even non-blocking findings here may require approval gates.
.env, secrets handling, credential storage, encryption code| Input | Where it lives | If missing |
|---|---|---|
| Diff | git diff <baseline>..HEAD | Stop; ask which diff to review |
| Sensitive paths config | policies/protected-paths.yml (auth + secrets sections) | Use built-in defaults from policy template; flag policy as needing customization |
| Secret patterns | policies/secret-patterns.yml | Use built-in defaults; flag policy as needing customization |
| Approval policy | policies/approval-policy.yml (+ optional .local.yml override) | Stop; FR-12 cannot enforce without policy |
| CLI edition map, for Fusebase Apps work | docs/fusebase-cli-edition.md | Continue with generic security review, but mark app-domain auth/secret assumptions unknown |
| App access + permission grants (Fusebase Apps diffs) | deploy handoff / fusebase app list / registration step in ticket | Ask operator for current grants; never assume the app is org-only — an inherited visitor principal changes every finding's severity |
policies/protected-paths.yml.app-secrets, handling-authentication-errors, fusebase-gate, fusebase-dashboards) as runtime-specific security context.AGENTS.md. Flag missing gates as blocker.fusebase app create/update --access: visitor = any unauthenticated user; --access="" = ALL org roles (NOT lockdown). --access REPLACES the entire principal list — principals are type or type:id — visitor, or orgRole:<id> with ids guest, client, member, manager, owner — verify the change neither silently widens (adds visitor, or adds more org roles than the feature needs; the widest grants are all-org-roles or a visitor principal) nor drops a restriction the operator relies on. Any widening → approval-required list; verify every surface the app exposes (dashboard views, backend endpoints, gate stores) is safe for the widest principal granted.--permissions="dashboardView.<dashboardId>:<viewId>.<privileges>": least-privilege — write only on views the runtime actually mutates (batchPutDashboardData / addRelationRows); read-only SDK usage (getDashboardViewData) gets read. Flag write-everywhere grants and grants with no matching SDK call in runtime code.@fusebase/fusebase-gate-sdk: fusebase app update --sync-gate-permissions must run after Gate-SDK call changes — Gate soft mode silently DEGRADES tokens to the allowed subset, so unsynced grants surface as 403s / missing data in smoke, not as deploy failures. Flag any user-facing flow that falls back to service-token auth on Gate 401/403 (blocker per fusebase-gate security rule — service tokens live only in explicit system/admin endpoints with audit logging).isolated_store.* permissions from fusebase env create): verify writes are org/tenant-scoped and requested store permissions match the operations the runtime performs — no broader store grant than the feature needs.getDashboardViewData fan-out, CSV/JSON dumps) or bulk imports: (a) export scope = minimum views/columns needed; (b) enumerate the PII fields crossing the boundary (names, emails, phones, client identifiers) in the review summary; (c) tenant isolation — exported/imported rows respect the org/workspace boundary, and import paths validate ownership before write. approval-policy.yml defines no data-export key — flag bulk export/import in the review summary as requiring explicit operator go-ahead before deploy; map to real approval keys only where the same diff genuinely triggers one (auth_or_permission_change, database_migration, secret_file_write).external_customer_visible_message per approval-policy.yml.require_approval in approval-policy.yml and lacks a corresponding artifact in state/approvals/.| Artifact | Path / location | Mode |
|---|---|---|
| Security summary | chat output | Mode A |
| Optional persistent record | docs/verification/<slug>-security.md | Mode B (full) |
| Approval-required list | embedded in summary | Mode B (full, table) |
| Failure mode | Detection | Response |
|---|---|---|
| Detected secret in diff | regex match on policies/secret-patterns.yml | BLOCK deploy. Redact value in output. Recommend git reset --soft HEAD~1 + rotation. Per FR-12, log incident in docs/problem-catalog/<slug>/. |
Missing approval artifact for require_approval op | scan state/approvals/ against approval-policy.yml: require_approval | BLOCK deploy. Surface which artifact is missing and how to author it. |
| Auth gate missing on new endpoint | endpoint added without matching auth middleware import/decorator | BLOCK deploy. Surface specific endpoint + project auth pattern to apply. |
| Customer-visible message added without approval | outbound message API call in diff + no external_customer_visible_message approval artifact | BLOCK deploy. Per FR-12 + approval-policy. |
| Production DB write outside repository pattern | direct connection / raw SQL outside *Repository.ts / *_repository.py | BLOCK deploy. Surface migration path through repository layer. |
| App access widened without approval | --access gains visitor (public), or --access="" (empty = EVERY org role incl. guest/client; NOT visitors, NOT lockdown — broad internal exposure) in diff/deploy notes, no approval artifact | BLOCK deploy. visitor = any unauthenticated user → confirm every reachable surface is public-safe; empty access → confirm org-wide exposure is intended. Require explicit approval either way. |
| Dashboard write scope broader than SDK usage | --permissions grants write on views the runtime only reads (no mutation SDK calls) | Finding (non-blocking): narrow to read per least-privilege; blocker if the grant spans client data views. |
| Gate permissions out of sync with runtime Gate-SDK calls | Gate-SDK calls changed in diff, no --sync-gate-permissions in deploy steps | BLOCK deploy until sync runs — soft mode masks the gap by silently degrading tokens. |
| Implicit service-token fallback | user-facing flow catches Gate 401/403 and retries with service-token auth | BLOCK. Explicit system/admin endpoints + audit logging only (fusebase-gate security rule). |
| Export crosses tenant boundary / leaks PII | export path reads without org/workspace filter, or dumps PII columns not needed downstream | BLOCK deploy. Require scoped query + PII field list in summary + explicit operator go-ahead (approval-policy.yml has no data-export key; auth_or_permission_change applies only if the diff also changes auth/permission code). |
docs/problem-catalog/incident-<date>/problem.md, alert operator immediatelypermission_request hook — if it's wired, it's part of the contractOriginal Fusebase Flow content. Designed after reviewing public AI coding workflow patterns; no third-party code, prompts, skill files, or hook scripts are copied. See docs/source-map.md.
Use ONLY when the operator runs "/find-wasted-code" (or explicitly asks to "scan the repo for dead-end tool calls / broken links / missing helpers / footgun configs / silent push-through"). Statically scans THIS repo for friction footguns and writes a tracked report to docs/wasted-code/report.md. Manual-trigger ONLY — do NOT auto-invoke; the skill carries disable-model-invocation and no hook wires it. Read-only except the one report write. Findings are review candidates, never auto-fixes. Do NOT use for token/transcript economy (/token-waste-audit), for process ceremony (/find-wasted-effort), or for CLI/Flow drift health (/fusebase-health).
Use at Specify to classify a ticket Full vs Lightweight, and whenever a change looks small / reversible / low-risk ("small fix", "tweak", "hotfix", "drop pretty-printing", "bump a constant"). Operationalizes FR-21 — ceremony proportional to change size. Defines the eligibility gate, the change-note artifact, the one-pass build→verify→deploy procedure, and mid-flight promotion. Do NOT use for risky/uncertain work, schema/data migrations, security/permission changes, new public contracts needing a decision, unknown-root-cause diagnostics, or large features — those take the Full lane.
ALWAYS load at session start. After self-attesting, Read references/<role>.md for your role (Product Owner / AI Developer / Architect / Deploy phase) — role-specific don't-list, exact refusal phrasing for FR violations, and recovery pointers; shared cross-role protocols live in this file. Mandatory; not on-demand.
Use when the operator types /product-owner, asks to activate Product Owner, start a Product Owner or PO session, or searches /product in Codex for Fusebase Flow project management. Do NOT use for implementing code; route implementation to the AI Developer role.
Use when planning tasks that touch big source files, when an edit would push a file past the module-size ceiling, when the pre-commit module-size ratchet warns or blocks, or when generating policies/module-size-baseline.txt. Operationalizes FR-25 — module-size ratchet (ceiling default 800; over-ceiling files may shrink, never grow; extraction along a responsibility seam is in-scope, not scope creep). Do NOT use for docs/markdown size (FR-23 owns docs), to force decomposition of an existing monolith (that is its own explicit ticket), or to judge split quality by regex (seam-vs-mechanical is semantic — code-review owns it).
Use ONLY when verification passed AND operator explicitly says "prepare deploy" / "draft deploy" / "ship it"; drafts deploy handoff, captures deploy hash + probes + smoke, advises spec DRAFT→DONE flip. Do NOT auto-invoke; operator triggers explicitly.