Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Model-mapping greedy-prefix misrouting (the claude-opus-4-8 → retired Opus 4.0 incident): the model cache matched requested IDs with starts_with ordered by prefix length, so a request for claude-opus-4-8 could shadow-match a row keyed claude-opus-4 and silently route to a retired inference profile, causing Bedrock 400s. Hardened across the read and write paths so the bug class cannot recur.

Exact-match model cache (Slice 1): ModelCache now stores mappings in a HashMap and looks them up by exact equality (O(1)), not prefix matching. The hardcoded fallback was likewise converted from greedy starts_with catch-alls to exact-match arms — a future claude-sonnet-4-8 now passes through to discovery instead of misrouting to retired Sonnet 4.0.
Read-time date-suffix fallback (Slice 2): on an exact miss, the cache retries once against the date-stripped form, but only when the stripped form is minor-version-bearing (e.g. claude-opus-4-8), never a bare major (claude-opus-4) — so dated aliases resolve without reopening the greedy shadow.
discover_model persists the exact requested model ID (Slice 3) as anthropic_prefix, not the lossy date-stripped form (the root cause of the original incident). Profile selection now runs three ordered passes — exact stem, versioned stem, fuzzy contains — to disambiguate variants like claude-opus-4-8-thinking.

Cold-start model seed (Slice 4): a curated model_seed.json is embedded at compile time and inserted at startup with ON CONFLICT DO NOTHING, so a request for a known model succeeds on a fresh deploy before any discovery tick. It is a floor, not a ceiling — discovery still handles anything unseeded. Explicitly not a runtime/call-home fetch.

Websearch admin control: three configurable modes for web search behavior:
- Enabled (default): per-user provider configuration, each user chooses their own search provider
- Disabled: web search tool silently stripped from requests server-side; setup script pushes permissions.deny: ["WebSearch"] to Claude Code clients
- Global: admin configures a single search provider (DuckDuckGo, Tavily, Serper, or Custom) used for all users; per-user provider config is overridden
Admin API: GET/PUT /admin/websearch-mode for reading and setting the mode, with provider config for global mode
Websearch mode exposed in GET /admin/settings so the portal and clients can read it
Portal Settings page: three-button mode selector with global provider config form (type, API key, URL, max results)
Portal nav: Web Search menu item hidden when mode is disabled
API key masking: global provider API key never returned in GET responses (replaced with has_api_key boolean)
SearchProvider::from_global_config(): constructs a search provider from admin-configured JSON with validation (provider_type required, api_key required for Tavily/Serper, api_url required for Custom, max_results clamped to 1-20)
extract_web_search_tool_with_mode(): mode-aware tool extraction that strips web_search tools when mode is disabled
Setup script WebSearch deny injection covers both SSO and virtual key setup paths
33 new tests (14 integration, 19 unit) covering all websearch admin control behavior

translate() now accepts a websearch_mode parameter; handler reads mode from DB settings on each request
Deploy script now uses GHCR (GitHub Container Registry) instead of ECR for image storage

Per-IDP user_claim configuration: controls which JWT claim is used as the user identifier. Configurable via portal UI, admin API, or OIDC_USER_CLAIM env var. Supports email, preferred_username, upn, oid, name, sub, or auto (default fallback chain).
Extract preferred_username, upn, oid, and name claims from OIDC JWTs (previously only email and sub were extracted).
OIDC login now requests email and profile scopes in addition to openid, for broader claim availability.

OIDC user identification now prefers email > preferred_username > upn > name > sub by default (was sub only). This fixes Entra ID (Azure AD) compatibility where sub is an opaque pairwise hash.
EventBridge notification detail no longer includes source, event_type, or timestamp fields — these are redundant with the EventBridge envelope (source, detail-type, time). Webhook and SNS payloads are unchanged. Breaking for consumers parsing these fields from EventBridge detail.

OIDC users from Entra ID no longer show up as opaque pairwise hashes. The gateway now uses the email or preferred_username claim for user identification.
CLI login (ccag login) redirect URL no longer breaks when the IDP audience is a UUID client_id (common with Entra ID). The audience is only used as a redirect hostname when it looks like a domain name.
IAM policy: split Bedrock invoke and ListInferenceProfiles into separate statements so the wildcard resource only applies to List, not Invoke.
Documentation: clarified OIDC_ISSUER is a one-time bootstrap seed (not a persistent override). Changing it after first startup has no effect unless the IDP is deleted from the portal first.
Documentation: expanded ADMIN_USERS description with OIDC subject identifier details and startup seeding behavior.
Documentation: renamed confusing “Notifications” env var section to “Infrastructure Alarms” and corrected inaccurate alarmWebhookUrl reference. Separated from app-level notifications.
Documentation: added separate EventBridge payload example showing envelope structure without redundant fields.

API gateway translating Anthropic Messages API to Amazon Bedrock
Streaming support (SSE translation from Bedrock binary events)
Multi-user authentication: virtual API keys, OIDC SSO (any provider), session tokens
Admin portal SPA with dashboard, key management, team management, and analytics
Budget enforcement with per-team and per-user spend limits
Rate limiting (per-key sliding window)
Web search interception (translates Anthropic’s web_search tool via DuckDuckGo)
Multi-endpoint support (route teams to different AWS accounts/regions)
Model ID mapping (automatic region-based Bedrock inference profile resolution)
Prometheus metrics and optional OTLP export
AWS CDK infrastructure (ECS Fargate + RDS Postgres + ALB)
Database migrations (automatic on startup)
CloudWatch alarms for operational monitoring
Autoscaling based on CPU and memory utilization