Deploy VAKRA to IBM Code Engine, data-backed by COS (additive, admin-…

[anupamamurthi]

What & why

Adds an optional path to run the VAKRA benchmark environment on IBM Code Engine,
with the ~35 GB dataset in IBM COS instead of on every machine. Purely additive — the
local make/docker-compose flow, the MCP servers, and existing CE/COS deployments are
untouched. All resources are prefixed vakra-*; every mutating script is gated behind
VAKRA_CE_ADMIN=1.

Two seams bridged (CE-only)

• Transport: locally MCP is reached via docker exec (stdio); CE has none, so a new
  docker/mcp_http_bridge.py serves MCP over HTTP at /mcp/<domain> when SERVE_MODE=http.
  It spawns a fresh, domain-pinned mcp_dispatch.py subprocess per request — same
  MCP_DOMAIN scoping as docker exec -e, so list_tools() stays scoped and
  verify_checksum still passes. Default (local) path unchanged.
• Data: DATA_SOURCE=mount (default) mounts the COS bucket via a CE persistent data
  store at the exact paths the servers expect; DATA_SOURCE=sync is a download-to-disk
  fallback (docker/cos_sync.py).

Contents

• deploy/ce/ — config.sh, 0_push_data_to_cos.py, 1_create_bucket_and_pds.sh,
  2_build_push_image.sh, 3_deploy_apps.sh, 4_deploy_explorer.sh, 5_test_ce.py,
  teardown.sh, Dockerfile.explorer, and docs (README/RUNBOOK/USING/CHEATSHEET).
• Additive edits: mcp_client.py (+http transport), benchmark_runner.py (log server URL),
  tools_explorer/app.py (MCP_CONFIG env), docker/{Dockerfile.unified,entrypoint-unified.sh}
  (opt-in bridge + mcp==1.27.0 pin), .dockerignore (anchor top-level /data/),
  benchmark/mcp_connection_config.ce.yaml.

Benchmarking against CE

Unchanged runner — only --mcp-config benchmark/mcp_connection_config.ce.yaml differs;
per-domain scoping is automatic.

Testing

• 5_test_ce.py: 4/4 capabilities green on CE (cap1=9, cap2=139, cap3/bpo=32, cap4=144 tools).
• Runner --list-tools against CE verified. Full LLM run is provider-key-gated (openai/watsonx).
• Bridge concurrency fix validated locally against a dummy MCP server before deploy.

Usage

Deploy it (admins)

From the repo root, one-time: ibmcloud login --sso, then create a venv
(python3.12 -m venv .venv && pip install -e ".[agents,mcp,init]" boto3 openai). Then:

1. python deploy/ce/0_push_data_to_cos.py --stream # data → COS (~35 GB, resumable)
2. VAKRA_CE_ADMIN=1 YES=1 deploy/ce/1_create_bucket_and_pds.sh
3. VAKRA_CE_ADMIN=1 YES=1 deploy/ce/2_build_push_image.sh # cloud build → ICR (~15 min)
4. VAKRA_CE_ADMIN=1 YES=1 deploy/ce/3_deploy_apps.sh # writes deploy/ce/.ce_urls.env
   Verify: source deploy/ce/.ce_urls.env && python deploy/ce/5_test_ce.py → expect 4/4.
   Full detail + gotchas: deploy/ce/RUNBOOK.md. One-pager: deploy/ce/CHEATSHEET.md.

Use the deployed env (everyone else)

Get deploy/ce/.ce_urls.env (the 4 URLs) from an admin, then pick one:
• Any MCP client → connect to <base>/mcp/<domain> (streamable HTTP; health at <base>/healthz).
• Benchmark runner → source .ce_urls.env then
python benchmark_runner.py --capability_id 2 --domain california_schools \ --mcp-config benchmark/mcp_connection_config.ce.yaml (add a provider + key for a real run).
• Hosted explorer → open the vakra-explorer URL and browse/invoke tools in the browser.
Endpoints are public (no auth) — treat URLs as internal. Guide: deploy/ce/USING.md.

SUMMARY: https://ibm-netlify-drop.1gxwxi8kos9y.us-east.codeengine.appdomain.cloud/s/6229b0be/

[anupamamurthi]

Tool Explorer (+ all the 4 services are up and running in CE)