영어 원문

이 설계 기록은 역사적 맥락과 원문의 의미를 보존하기 위해 영어로 제공합니다.

ADR-019: Bedrock Guardrails data-plane application + per-team override (D6)

Date: 2026-07-08 Status: Accepted (implemented). Related: ADR-016 (teams as keystore records — this ADR extends the teams table and reuses its fresh-per-request lookup posture); ADR-009 (the PII-masking filter seam — the only other content-policy mechanism in the repo, and the thing this ADR is explicitly NOT reusing, since Guardrails are provider-side, not a request-transform).

Context

The spec names Bedrock Guardrails bypass as the substantive problem: a client configures a Guardrail in the AWS console, but a gateway sitting between the client and Bedrock can silently not apply it — because applying a guardrail is a parameter on the InvokeModel/Converse API call itself, not something a proxy gets "for free" by forwarding bytes. Before this change, inferplane's Bedrock provider (providers/bedrock/) had no guardrail parameter anywhere — every invocation went out guardrail-free regardless of what was configured in AWS. The spec is explicit that the console UI is not the fix; the data plane (the actual SDK call) is.

User-confirmed scope (2026-07-08): both provider-level default AND per-team override — full spec compliance, not just the anti-bypass minimum.

Decision

1. Provider-level default (the actual anti-bypass fix) + per-team override, no opt-out

providers/bedrock/bedrock.go's factory reads guardrail_id/guardrail_version from the provider's config Settings (populated by internal/live.BuildState from config.ProviderConfig.GuardrailID/GuardrailVersion) into a defaultGuardrail field on provider. Every one of the four call paths (Invoke, InvokeStream, Converse, ConverseStream) resolves an effective guardrail via provider.guardrailFor(req): a per-team override (threaded through providers.ProxyRequest.GuardrailID/GuardrailVersion) wins over the default; with neither, no guardrail fields are set on the SDK call at all (today's pre-D6 behavior, unchanged when nothing is configured).

Deliberately no per-team opt-out. A team record can select a different guardrail (a stricter one, or one scoped to that team's use case) but cannot clear the provider's default to get an unguarded call — guardrailFor has no "none" branch. This matches the anti-bypass framing: the mechanism this ADR exists to fix is a client (or, here, an intermediate layer) making the guardrail not apply; letting the team itself disable it would reopen the identical hole one layer up.

2. providers.ProxyRequest.GuardrailID/GuardrailVersion — a narrow, explicit exception to provider isolation

CLAUDE.md §8 says a provider PR touches only providers/<name>/. A per-team override, by construction, has to travel from the team record (keystore) → the ingress handler (which knows the calling team) → the specific provider call. Two new string fields on the shared ProxyRequest transport struct are the minimal core-diff needed to carry that value across the ingress/provider boundary; every other provider ignores them (they're plain data, not a behavior change to any code path outside providers/bedrock). This is called out explicitly as an intentional, narrow exception — not a precedent for routing arbitrary per-provider config through core structs.

3. Governor-lookup pattern reused for a non-governance purpose: SetTeamPolicy

ADR-016 established Governor.SetTeamLookup: a fresh-per-request keystore read, no caching, so a console edit takes effect on the very next request. That pattern is exactly what a guardrail override needs too — but a guardrail override is not a governance decision, so it doesn't belong on Governor. Both anthropicapi.MessagesHandler and openaiapi.ChatHandler gain a separate setter, SetTeamPolicy(func(team string) (keystore.TeamRecord, bool)), resolved once per request (before the fallback loop, reused across every fallback attempt) and threaded into each ProxyRequest. D7/ADR-020's region-lock reuses this exact same setter — one lookup, two independent per-team overrides read off the same TeamRecord — rather than adding a second near-identical setter.

cmd/inferplane/gateway.go wires teamPolicy as its own closure over store.GetTeam, sitting right next to (but independent from) the governance SetTeamLookup closure. This means a request now does two keystore point reads instead of one when a team record exists — accepted as negligible next to the SQLite read KeyAuth already does to resolve the principal itself; a shared per-request cache was considered and rejected as unwarranted complexity for a sub-millisecond local read.

4. teams table migration: the first ALTER-TABLE path it has ever needed

ADR-016 shipped teams as a brand-new table (CREATE TABLE IF NOT EXISTS only — no pre-existing table to migrate). D6 adds guardrail_id/ guardrail_version columns to that now-real, possibly-populated table, so internal/keystore/sqlite.go's ensureSchema gains an ALTER-if-missing block for teams, mirroring the one keys already had for its D2 governance columns. The PRAGMA-scan-then-ALTER logic was factored into two small shared helpers (existingColumns, applyMigrations) so keys and teams share the same migration mechanics instead of duplicating the scan loop a second time.

5. Version defaulting: empty → "DRAFT"

The Bedrock SDK rejects a GuardrailIdentifier with no GuardrailVersion set (a ValidationException). Guardrail.versionOrDraft() defaults an empty version to "DRAFT" — Bedrock's own default working version — so an operator who only sets guardrail_id (the common case while iterating on a guardrail) doesn't have to also always specify "DRAFT" explicitly.

6. Capability flag: guardrails keeps its current (PII-masking) meaning

Capabilities.Guardrails (spec §4.4) is currently wired as masking != nil — it answers "is the PII-masking filter on," not "is a Bedrock Guardrail configured." Reshaping it into the richer object the spec sketches (guardrails: {...}) is deferred; per-team guardrail state is instead fully discoverable through the existing /admin/teams surface (gated on teams_records, which is unconditionally true). Documented here rather than silently left inconsistent with the spec's eventual shape.

Deferred

• ~~providerstore (UI-registered providers) guardrail columns / provider-form fields~~ — implemented (2026-07-11): providerstore.ProviderRow carries GuardrailID/GuardrailVersion (TEXT columns, ALTER-TABLE migration mirroring auth_header); overlay.go's providerConfigFromRow/rowFromProviderConfig thread both fields (this also fixed a real bug — rowFromProviderConfig previously dropped them, so a config-file-declared bedrock provider's guardrail was silently lost the moment SeedIfEmpty imported it into the DB); configapi/write.go's ParseProviderWrite validates guardrail shape/format/type-is-bedrock (same six checks the team path and file-config path each enforce, duplicated per this codebase's established per-surface pattern); configapi/config.go's ProviderView/ViewFrom echo both fields back (non-secret, console prefill); the admin console's provider form gained pf-guardrail-id/ pf-guardrail-version. A DB-registered bedrock provider can now carry its own default guardrail exactly like a file-config-declared one; a team override still takes precedence when present.
• Trace / StreamProcessingMode guardrail knobs (SDK fields left at zero value) — add as Settings/team-record fields when a real need for guardrail trace output appears.
• Config-file-declared per-team guardrail (config.TeamConfig has no guardrail field) — the console (/admin/teams) is the write path; a config-declared team has no override slot, matching D3's own precedent (config teams have no DB-only fields either).
• Spec §4.4's guardrails: {...} capability object shape.