feat(evaluators): add extensions/metadata for additional fields that we want to pass as context to execution of evaluator

[namrataghadi-galileo]

Summary

Pass the authenticated execution context through Agent Control to Galileo evaluators so scorer invocation can receive the Orbit-authorized organization, caller, project, and target information.

• Preserve upstream authorization response extras as opaque Principal.extensions metadata.
• Add an optional opaque extensions claim to newly minted runtime JWTs. Existing claims, expiration behavior, and tokens without extensions remain compatible.
• Pass verified principal extensions through a provider-neutral evaluator hook. Existing evaluators continue using their current evaluation path.
• In the Galileo evaluator only, translate the trusted extension envelope into Galileo's execution_context. For the current Orbit contract, caller_id is the authenticated user ID. A log_stream target's authenticated target ID becomes run_id.
• Return a clear evaluator error when required authenticated fields are missing.

Scope

• API/runtime contract: additive opaque extensions claim; no named Galileo fields are added to generic principal or JWT models.
• Internal: carry verified extensions from auth exchange through runtime-token verification and engine evaluation.
• Galileo: serialize the local execution context with scorer invocation requests.
• Out of scope: scorer grants, scorer authorization/execution, API-key forwarding to Runners, record factory changes.

Trust and Compatibility

Execution metadata comes only from the verified runtime principal. Evaluator step payloads are not used to construct it, and user-supplied tenant identity is not trusted. No scorer grant is acquired or cached, and the application GALILEO_API_KEY is not sent to Runners. Legacy runtime tokens without extensions and evaluators that do not override the extension hook remain supported.

Testing and Coverage

• Engine tests: 120 passed, including opaque extension pass-through and legacy evaluator fallback.
• Built-in evaluator tests: 311 passed, including default extension-hook delegation.
• Galileo tests: 156 passed, including context serialization, missing caller handling, and API-key non-forwarding.
• Focused server auth tests: 6 passed, including malformed extension handling and signed-token round-trip.
• make check passed in PR CI; Codecov patch check passed.
• make lint and make typecheck passed locally.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[namrataghadi-galileo]

Tested in devstack:

(.venv) namratag@NAMRATAG-M-6QFJ ace-demo % TOKEN="$(jq -r '.token' /tmp/agent-control-auth.json)"      

python - "$TOKEN" <<'PY'
import base64
import json
import sys

payload = sys.argv[1].split(".")[1]
payload += "=" * (-len(payload) % 4)

claims = json.loads(base64.urlsafe_b64decode(payload))
print(json.dumps(claims, indent=2, sort_keys=True))
print("\nproject_id:", claims.get("extensions", {}).get("project_id"))
PY
{
  "actor_id": "4b0890ea-bc1d-47a0-9808-556f76a83835",
  "domain": "runtime",
  "exp": 1790819890,
  "extensions": {
    "project_id": "96da6d6d-a63f-40db-94c9-ccfcfa99c95c"
  },
  "iat": 1790819590,
  "iss": "agent-control/server",
  "jti": "77Bs7-hge_oVxRualkZytw",
  "namespace_key": "f3befdd7-6f82-4136-8a7d-529366231ce3",
  "scopes": [
    "runtime.use"
  ],
  "target_id": "eee67868-b583-4bda-b7ad-8bfd8ddb8348",
  "target_type": "log_stream"
}

project_id: 96da6d6d-a63f-40db-94c9-ccfcfa99c95c