eval(deltawire): integrate terminal v6 runtime boundary

.github/workflows/deltawire-runtime.yml

@@ -0,0 +1,37 @@
+name: Verify DeltaWire runtime integration
+
+on:
+  pull_request:
+    paths:
+      - "labs/20-deltawire/**"
+      - ".github/workflows/deltawire-runtime.yml"
+  push:
+    branches: [main]
+    paths:
+      - "labs/20-deltawire/**"
+      - ".github/workflows/deltawire-runtime.yml"
+
+permissions:
+  contents: read
+
+jobs:
+  validate-runtime:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-python@v6
+        with:
+          python-version: "3.12"
+      - uses: astral-sh/setup-uv@v7
+      - run: python3 labs/20-deltawire/eval/scripts/v6/verify_terminal_evidence_lock.py
+      - run: python3 labs/20-deltawire/eval/scripts/v6/test_terminal_evidence_lock.py
+      - run: python3 labs/20-deltawire/eval/scripts/v6/test_release.py
+      - run: python3 labs/20-deltawire/eval/scripts/v6/test_environment_receipt.py
+      - run: python3 labs/20-deltawire/eval/scripts/v6/test_artifact_manifest.py
+      - run: python3 labs/20-deltawire/eval/scripts/v6/test_setup_receipt.py
+      - run: python3 labs/20-deltawire/eval/scripts/v6/test_trial_status.py
+      - run: python3 labs/20-deltawire/eval/scripts/v6/test_shell_observation.py
+      - run: python3 labs/20-deltawire/eval/scripts/v6/negative_controls.py --self-test
+      - run: PYTHONPATH=labs/20-deltawire/eval uv run --with harbor==0.20.0 python labs/20-deltawire/eval/agents/test_preinstalled_gemini_cli.py

labs/20-deltawire/eval/agents/__init__.py

@@ -0,0 +1 @@
+"""Versioned Harbor custom agents for DeltaWire evaluation."""

labs/20-deltawire/eval/agents/preinstalled_gemini_cli.py

@@ -0,0 +1,73 @@
+"""Harbor Gemini CLI agent with a verification-only install boundary."""
+import json
+import shlex
+
+from harbor.agents.installed.gemini_cli import GeminiCli
+
+AGENT_NAME = "preinstalled-gemini-cli-v1"
+EXPECTED_NODE_PATH = "/usr/local/bin/node"
+EXPECTED_NODE_VERSION = "v22.23.1"
+EXPECTED_NPM_VERSION = "10.9.8"
+EXPECTED_GEMINI_PATH = "/usr/local/bin/gemini"
+EXPECTED_GEMINI_VERSION = "0.51.0"
+EXPECTED_PACKAGE_VERSION = "0.51.0"
+EXPECTED_IMAGE_ID = "sha256:589fce58d7ddf09910876afe0198bcee0086a840af6de7c08515bbce3e8367a7"
+RECEIPT_PATH = "/logs/artifacts/preinstalled-gemini/setup-receipt.json"
+
+
+class PreinstalledGeminiCli(GeminiCli):
+    @staticmethod
+    def name() -> str:
+        return AGENT_NAME
+
+    async def install(self, environment) -> None:
+        async def checked(command: str) -> str:
+            result = await self.exec_as_agent(environment, command=command)
+            if result.return_code != 0:
+                raise RuntimeError(f"preinstalled runtime verification failed: {command}")
+            return (result.stdout or "").strip()
+
+        node_path = await checked("command -v node")
+        node_version = await checked("node --version")
+        npm_version = await checked("npm --version")
+        gemini_path = await checked("command -v gemini")
+        gemini_version = await checked("gemini --version")
+        package_version = await checked(
+            "node -p \"require('/usr/local/lib/node_modules/@google/gemini-cli/package.json').version\""
+        )
+        checks = {
+            "agent_identity": self.name() == AGENT_NAME,
+            "gemini_executable": gemini_path == EXPECTED_GEMINI_PATH,
+            "gemini_version": gemini_version == EXPECTED_GEMINI_VERSION,
+            "image_id": EXPECTED_IMAGE_ID.startswith("sha256:") and len(EXPECTED_IMAGE_ID) == 71,
+            "node_executable": node_path == EXPECTED_NODE_PATH,
+            "node_version": node_version == EXPECTED_NODE_VERSION,
+            "npm_version": npm_version == EXPECTED_NPM_VERSION,
+            "package_identity": package_version == EXPECTED_PACKAGE_VERSION,
+        }
+        receipt = {
+            "agent": AGENT_NAME,
+            "checks": checks,
+            "gemini_path": gemini_path,
+            "gemini_version": gemini_version,
+            "image_id": EXPECTED_IMAGE_ID,
+            "node_path": node_path,
+            "node_version": node_version,
+            "npm_version": npm_version,
+            "package_version": package_version,
+            "schema_version": "preinstalled-gemini-setup-receipt.v1",
+            "status": "pass" if all(checks.values()) else "fail",
+        }
+        payload = json.dumps(receipt, sort_keys=True, separators=(",", ":")) + "\n"
+        parent = RECEIPT_PATH.rsplit("/", 1)[0]
+        command = (
+            f"mkdir -p {shlex.quote(parent)} && "
+            f"temporary={shlex.quote(RECEIPT_PATH + '.tmp')} && "
+            f"printf %s {shlex.quote(payload)} > \"$temporary\" && "
+            f"mv \"$temporary\" {shlex.quote(RECEIPT_PATH)}"
+        )
+        result = await self.exec_as_agent(environment, command=command)
+        if result.return_code != 0:
+            raise RuntimeError("preinstalled setup receipt write failed")
+        if receipt["status"] != "pass":
+            raise RuntimeError("preinstalled runtime identity mismatch")

labs/20-deltawire/eval/agents/test_preinstalled_gemini_cli.py

@@ -0,0 +1,60 @@
+#!/usr/bin/env python3
+import ast
+import asyncio
+import importlib
+import inspect
+from pathlib import Path
+from types import SimpleNamespace
+
+MODULE = importlib.import_module("agents.preinstalled_gemini_cli")
+CLASS = MODULE.PreinstalledGeminiCli
+
+
+def structural_test() -> None:
+    source = Path(inspect.getsourcefile(CLASS)).read_text()
+    tree = ast.parse(source)
+    node = next(item for item in tree.body if isinstance(item, ast.ClassDef) and item.name == CLASS.__name__)
+    methods = {item.name for item in node.body if isinstance(item, (ast.FunctionDef, ast.AsyncFunctionDef))}
+    inherited = {name for name in methods if hasattr(CLASS.__mro__[1], name)}
+    assert inherited == {"name", "install"}
+    lowered = inspect.getsource(CLASS.install).lower()
+    for forbidden in ("apt-get", "apt ", "curl", "wget", "nvm", "npm install", "npm update"):
+        assert forbidden not in lowered
+    assert CLASS.run is CLASS.__mro__[1].run
+
+
+def functional_test() -> None:
+    outputs = iter(
+        [
+            "/usr/local/bin/node\n",
+            "v22.23.1\n",
+            "10.9.8\n",
+            "/usr/local/bin/gemini\n",
+            "0.51.0\n",
+            "0.51.0\n",
+            "",
+        ]
+    )
+    commands = []
+
+    async def fake_exec(_environment, command):
+        commands.append(command)
+        return SimpleNamespace(return_code=0, stdout=next(outputs), stderr="")
+
+    instance = object.__new__(CLASS)
+    instance.exec_as_agent = fake_exec
+    original = MODULE.EXPECTED_IMAGE_ID
+    MODULE.EXPECTED_IMAGE_ID = "sha256:" + "1" * 64
+    try:
+        asyncio.run(instance.install(object()))
+    finally:
+        MODULE.EXPECTED_IMAGE_ID = original
+    assert len(commands) == 7
+    assert "setup-receipt.json.tmp" in commands[-1]
+    assert '"status":"pass"' in commands[-1]
+
+
+if __name__ == "__main__":
+    structural_test()
+    functional_test()
+    print("Preinstalled Gemini agent inheritance and install tests passed.")

labs/20-deltawire/eval/conformance/preinstalled-gemini-setup-v1/environment/.gitkeep

@@ -0,0 +1,2 @@
+# Harbor task discovery requires the environment directory even when task.toml
+# selects a prebuilt docker_image and no Dockerfile is used.

labs/20-deltawire/eval/conformance/preinstalled-gemini-setup-v1/instruction.md

@@ -0,0 +1 @@
+Verify the preinstalled Gemini CLI agent runtime. Do not execute an agent or model.

labs/20-deltawire/eval/conformance/preinstalled-gemini-setup-v1/task.toml

@@ -0,0 +1,30 @@
+schema_version = "1.3"
+artifacts = []
+
+[task]
+name = "operatorstack/preinstalled-gemini-setup-v1"
+description = "Credential-free install-only conformance for the v6 agent runtime"
+authors = []
+keywords = []
+
+[metadata]
+benchmark_result = false
+conformance_version = "v1"
+
+[verifier]
+timeout_sec = 60.0
+
+[[verifier.collect]]
+service = "main"
+command = "python3 /usr/local/bin/environment_receipt.py --expectations /deltawire-environment-expectations.json --receipt /logs/artifacts/deltawire/environment-receipt.json"
+timeout_sec = 30.0
+
+[agent]
+timeout_sec = 60.0
+
+[environment]
+docker_image = "deltawire-preflight-v6:runtime-v1"
+network_mode = "no-network"
+build_timeout_sec = 60.0
+os = "linux"
+mcp_servers = []

labs/20-deltawire/eval/conformance/preinstalled-gemini-setup-v1/tests/test.sh

@@ -0,0 +1,3 @@
+#!/usr/bin/env bash
+set -euo pipefail
+exit 0

labs/20-deltawire/eval/docs/10-preinstalled-gemini-boundary.md

@@ -0,0 +1,29 @@
+# Preinstalled Gemini CLI boundary
+
+DeltaWire preflight v6 separates agent bootstrap from paid treatment behavior. The task image contains the runtime; Harbor's custom agent performs local identity checks only.
+
+## Harbor 0.20.0 grounding
+
+- Custom agents use the `module.path:ClassName` import form. V6 uses `agents.preinstalled_gemini_cli:PreinstalledGeminiCli` with the eval directory on `PYTHONPATH`.
+- `harbor run --install-only` performs environment and agent setup without agent execution or model verification. `--model` is optional on that path.
+- Harbor's official `GeminiCli.install` begins at line 109 of `harbor/agents/installed/gemini_cli.py` and performs dynamic `apt-get`, NVM, and npm installation. Its SHA-256 in Harbor 0.20.0 is `dc8ce81e8f41b78f00abb54cd91fd7840d147f442960ca32e68c7a760e060a2a`.
+- The inherited official `GeminiCli.run` begins at line 775 in that same frozen source. The custom class overrides only `name` and `install`; execution, authentication, settings, skills, prompt, trajectory, ATIF, token, cost, and resume behavior remain Harbor's implementation.
+
+Grounding commands:
+
+```text
+uvx --from harbor==0.20.0 harbor run --help
+uv run --with harbor==0.20.0 python <inspect-source-script>
+```
+
+## Deterministic setup proof
+
+The image pins the Linux/amd64 base manifest, Node, npm, Gemini CLI package and DeltaWire binary. The custom install boundary runs only local path/version/package checks and writes a canonical setup receipt under `/logs/artifacts`. In Harbor 0.20.0, `--install-only` implies disabled verification, so verifier collect hooks do not run and `artifacts/manifest.json` is not emitted; Harbor still performs its automatic `/logs/artifacts` collection after agent setup. The paid path retains the main-service collect hook and artifact manifest as separate environment gates. Three credential-free install-only Harbor trials must agree on setup-receipt hash and image identity, finish setup within 60 seconds, and show no model, credential, skill, agent execution, verifier execution, provider use, or dynamic installation.
+
+Harbor process exit and trial exception are normalized independently. A zero Harbor process exit cannot turn a trial exception into a pass.
+
+## Authorization boundary
+
+The v6 runner exposes only `probe-v6` and `pair`. The paid probe is unavailable until the exact phrase `approve the paid v6 probe` is separately recorded after manifest-head review. The pair requires the later exact phrase `approve the frozen range-large/r1 pair`. Neither authorization is implied by building or validating v6.
+
+All terminal states keep `READY_FOR_72=false`, `pair_launched=false`, `six_run_canary_started=false`, and `full_run_started=false` until evidence says otherwise; v6 never exposes a full-run mode.