Skip to content

0010-plugin-execution-context: plugin-execution-context

Context

All trusted native plugins share the RLS-bound platform_app role, so tenant isolation is enforced but plugin isolation is not. A bug in one plugin, or a confused kernel code path, could:

  • read another plugin's private tables;
  • write inbox markers under another plugin's consumer identity;
  • publish events as another plugin;
  • settle another plugin's job schedule.

Per-plugin database roles would close this, but they multiply provisioning, connection pools, grants, and credential rotation for every plugin.

Decision

  • A transaction-local plugin execution context. The host sets app.plugin_id with platform_db::begin_plugin_tx(pool, tenant_id, plugin_id) before running plugin code, next to app.tenant_id. The identifier is validated before any statement runs.
  • The Contacts service opens every transaction this way.
  • The worker opens every consumer delivery and job run in the owning plugin's context.
  • A consumer id must sit inside that plugin's namespace.
  • Restrictive RLS policies enforce it. Each policy is ANDed with the tenant policy, so plugin work needs both the right workspace and the right plugin:
  • Plugin tables: every table in a p_* schema requires its owning plugin's context for reads and writes. The tenant harness fails if any plugin table lacks such a policy.
  • Job schedules: reads, writes, and lease settlement are limited to the current plugin.
  • Inbox: a marker can only be written under the current plugin's consumer namespace.
  • Outbox: an event can only be written with the current plugin as source_plugin, and in its event-type namespace.
  • Reads of the inbox and outbox stay tenant-scoped, for diagnostics.
  • Kernel code without a plugin context sees no plugin rows.

Consequences

  • A plugin's data, events, and background state can only be used inside that plugin's context, in the right tenant. Cross-plugin access must go through public APIs, events, or governed read models.
  • New plugins must add a restrictive app.plugin_id policy to each table and open transactions with begin_plugin_tx.
  • The boundary guards against bugs and confused deputies, not hostile native code: in-process code can set the variable itself. Isolation from hostile code remains the job of the Wasm and external-service runtimes (ADR 0002).
  • There is still one runtime role, one pool, and no extra credentials.

Alternatives considered

  • Per-plugin database roles. Strongest, but operationally heavy at this stage.
  • Schema USAGE grants per role. Still requires per-plugin roles.
  • Rust-only checks. Bypassed by any raw query.

Future revisit triggers

Revisit when third-party native plugins are allowed, when audit needs per-plugin database identity, or when a plugin needs a governed read of another plugin's data.