v1.2.3: Opus 4.7 default, private Cloud Run, generalized deploy, chunking bytes fix

.dockerignore

@@ -0,0 +1,6 @@
+*
+!Dockerfile
+!package.json
+!tsconfig.json
+!src/
+!src/**

.gcloudignore

@@ -0,0 +1,9 @@
+# Allowlist: only upload what the Dockerfile consumes.
+# Everything else (test fixtures, binaries, terraform, docs, .git, etc.)
+# is excluded to keep the Cloud Build context small.
+*
+!Dockerfile
+!package.json
+!tsconfig.json
+!src/
+!src/**

.gitignore

@@ -23,9 +23,18 @@ npm-debug.log*
 
 # Environment
 .env
-.env.local
+.env.*
 .envrc
 
+# Cloud credentials (defensive; never commit service-account keys or similar)
+*credentials*.json
+*service-account*.json
+*-sa-key*.json
+gcp-key*.json
+
+# TypeScript incremental build cache
+*.tsbuildinfo
+
 # Test coverage
 coverage/
 
@@ -42,3 +51,6 @@ deploy/terraform.tfstate*
 skills-lock.json
 
 .private/
+
+# Playwright MCP artifact directory
+.playwright-mcp/

CHANGELOG.md

@@ -4,6 +4,25 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/).
 
+## [1.2.3] - 2026-04-18
+
+### Added
+- `anthropic` and `anthropic-vertex` providers now expose `claude-opus-4-7` as the default flagship model. `claude-opus-4-6` remains selectable via `PDF_ANALYZER_MODEL`.
+- Cloud Run deployment supports all five providers (`google-vertex`, `anthropic-vertex`, `google`, `anthropic`, `openai`) via a single `PDF_ANALYZER_PROVIDER` knob; direct-API providers read their key from Secret Manager at runtime.
+
+### Fixed
+- `analyzePdf` no longer crashes when a `gs://` source falls back to chunked processing. The chunking branch now handles the `bytes` source kind via a new exhaustive `resolveSourceBytes` helper.
+- MCP Streamable HTTP: the `/mcp` route handles any method (GET/POST/DELETE) instead of POST-only. GET was previously returning 404 and causing clients to misreport "SDK auth failed".
+
+### Changed
+- Cloud Run deploys are now private by default (`--no-allow-unauthenticated`). Connect via `gcloud run services proxy` locally.
+- Cloud Build context slimmed from ~20 MiB to ~120 KiB via allowlist `.gcloudignore` / `.dockerignore`.
+- Deploy scripts (`deploy/gcloud.sh`, `deploy/main.tf`) and templates generalized for any provider + auth mode; see `deploy/README.md` for the matrix.
+- HTTP transport tests now drive the real production request handler via a new exported `createRequestHandler`; the old tests ran against an inline copy that could diverge from production.
+
+### Security
+- `.gitignore` hardened with defensive patterns for common credential, env, and build-cache leaks (`*credentials*.json`, `*service-account*.json`, `.env.*`, `*.tsbuildinfo`, etc.).
+
 ## [1.2.2] - 2026-04-08
 
 ### Added

CLAUDE.md

@@ -23,7 +23,7 @@ git push -u origin <branch-name>  # Push and create PR
 Models per provider (do not change without discussion). Users choose during `--setup`:
 
 - **Google Gemini**: `gemini-3-flash-preview` (fast) / `gemini-3.1-pro-preview` (flagship)
-- **Anthropic Claude**: `claude-sonnet-4-6` (fast) / `claude-opus-4-6` (flagship)
+- **Anthropic Claude**: `claude-sonnet-4-6` (fast) / `claude-opus-4-7` (flagship) / `claude-opus-4-6` (previous flagship, still selectable)
 - **OpenAI**: `gpt-5.4-mini` (fast) / `gpt-5.4` (flagship)
 
 Thinking/reasoning is set to minimum for all models (document analysis doesn't benefit from extended thinking).
@@ -85,6 +85,26 @@ npm run type-check && npm run lint && npm test
 
 Always use `test/fixtures/1-pager.pdf` for MCP tool testing. It is small and cheap on LLM API calls. Never use `test/fixtures/oversized-doc.pdf` or other large PDFs unless the user gives explicit approval.
 
+## Deploying to Cloud Run
+
+The deploy scripts (`deploy/gcloud.sh` and `deploy/main.tf`) support every provider and both auth modes; which one runs is decided by `PDF_ANALYZER_PROVIDER` in `deploy/env` (gcloud) or `provider_id` in `terraform.tfvars`:
+
+- `google-vertex`, `anthropic-vertex` → ADC via attached service account, no API key required
+- `google`, `anthropic`, `openai` → API key pulled from a Secret Manager secret named in `API_KEY_SECRET_NAME` / `api_key_secret_name`
+
+See `deploy/README.md` for the full matrix, required IAM roles per provider, and the one-time `gcloud secrets create` command for the direct-API providers. The service is always deployed `--no-allow-unauthenticated` (private).
+
+### Running the remote MCP locally
+
+Because the service requires authenticated invocation, MCP clients connect through a local proxy that mints fresh identity tokens per request:
+
+```bash
+gcloud run services proxy <service-name> \
+  --project=<project-id> --region=<region> --port=8080
+```
+
+Point `.mcp.json`'s HTTP MCP entry at `http://localhost:8080/mcp`. When the proxy stops, the MCP disconnects until you start it again. No secrets live in `.mcp.json` — auth is handled per-request by the proxy against your ADC.
+
 ## Release Process
 
 Branch protection requires releases to go through a PR: