docs: harden examples for end-to-end use

examples/README.md

@@ -29,7 +29,7 @@ application-facing Python SDK contract:
 - resolving relative workspace and skill paths with `base_dir`;
 - running maintained Hermes Agent, Codex, Claude, and Deep Agents adapters through the Python SDK.
 
-Start with:
+Complete the [code-review setup](code_review_agent/README.md#set-up), then run:
 
 ```bash
 just build-all

examples/code_review_agent/README.md

@@ -21,12 +21,19 @@ just build-all
 ```
 
 The default variant uses Hermes Agent with an NVIDIA-hosted model. Follow the
-[Hermes Agent quick start](../../README.md#quick-start-hermes-agent) to install
-Hermes Agent, then set `NVIDIA_API_KEY` and `ADAPTER_PYTHON` as described there.
+[Hermes Agent adapter setup](../../adapters/hermes/README.md) with Python
+3.11-3.13, then set `NVIDIA_API_KEY`. If Hermes Agent is installed in a
+separate environment, select its interpreter only for the Hermes command:
 
-The config also demonstrates a harness-native GitHub MCP server. Set
-`GITHUB_MCP_URL` when you want to use that server; the review prompt below does
-not require it.
+```bash
+ADAPTER_PYTHON="$PWD/.tmp/hermes-venv/bin/python" \
+  .venv/bin/python -m examples.code_review_agent \
+  --input "Reply with exactly: fabric works"
+```
+
+Do not export a Hermes-only `ADAPTER_PYTHON` globally before running the Codex,
+Claude, or Deep Agents variants. Those variants use the project interpreter
+unless you select another interpreter explicitly.
 
 ## Inspect the plan
 
@@ -67,7 +74,11 @@ Add `--relay` to any variant to enable the Relay ATOF and ATIF configuration:
 
 Relay requirements depend on the selected adapter. The Codex and Claude
 adapters require an external `nemo-relay` CLI in the supported `0.6.x` range;
-the Python package named `nemo-relay` does not install that command.
+the Python package named `nemo-relay` does not install that command. Hermes
+Agent and Deep Agents require the Relay Python package in their selected
+adapter environment. Refer to the
+[installation guide](../../docs/getting-started/install.mdx#install-nemo-relay)
+for the current compatibility requirements.
 
 ```bash
 .venv/bin/python -m examples.code_review_agent \
@@ -89,15 +100,17 @@ application-owned composition:
 from examples.code_review_agent import (
     BASE_DIR,
     hermes_config,
+    with_fabric_managed_github_mcp,
     with_opensandbox,
     with_relay,
 )
 
 config = hermes_config()
 relay_config = with_relay(config)
 sandbox_config = with_opensandbox(config)
+github_config = with_fabric_managed_github_mcp(config)
 ```
 
-Each function returns a deep copy. `config`, `relay_config`, and
-`sandbox_config` can therefore be planned or run independently with
-`base_dir=BASE_DIR`.
+Each function returns a deep copy. The four configs can therefore be planned or
+run independently with `base_dir=BASE_DIR`. Set `GITHUB_MCP_URL` before running
+`github_config`; the default smoke does not configure or contact that server.

examples/code_review_agent/config.py

@@ -58,12 +58,6 @@ def base_config() -> FabricConfig:
         telemetry=TelemetryConfig(),
     )
     config.add_skill_path(SKILL_PATH)
-    config.add_mcp_server(
-        "github",
-        transport="streamable-http",
-        url="${GITHUB_MCP_URL}",
-        exposure="harness_native",
-    )
     return config
 
 
@@ -122,7 +116,6 @@ def codex_config() -> FabricConfig:
         workspace=WORKSPACE,
         artifacts="./artifacts/codex",
     )
-    config.remove_mcp_server("github")
     config.remove_skill_path(SKILL_PATH)
     return config
 
@@ -149,7 +142,6 @@ def deepagents_config() -> FabricConfig:
         workspace=WORKSPACE,
         artifacts="./artifacts/deepagents",
     )
-    config.remove_mcp_server("github")
     config.remove_skill_path(SKILL_PATH)
     return config
 
@@ -188,7 +180,6 @@ def claude_config() -> FabricConfig:
         workspace=WORKSPACE,
         artifacts="./artifacts/claude",
     )
-    config.remove_mcp_server("github")
     config.remove_skill_path(SKILL_PATH)
     return config
 

examples/code_review_agent/skills/code-review/SKILL.md

@@ -0,0 +1,14 @@
+---
+name: code-review
+description: Review Python changes for correctness risks and missing test coverage.
+---
+
+# Code Review
+
+1. Read the changed implementation and its nearest tests.
+2. Identify correctness, error-handling, and compatibility risks before style issues.
+3. Cite the smallest relevant file and line range for each finding.
+4. Suggest the narrowest correction and a focused regression test.
+
+If no correctness issue is present, say so directly and identify any remaining
+test-coverage uncertainty.

examples/harbor/calculator/.gitignore

@@ -0,0 +1,6 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+runs/
+task/environment/vendor/
+task/environment/.vendor.*/

examples/harbor/calculator/README.md

@@ -85,7 +85,7 @@ uv run --extra runtime --extra harbor harbor run \
   --ak fabric_adapter_id=nvidia.fabric.hermes \
   --ak fabric_config_base_dir=/opt/fabric-calculator \
   --ak fabric_workspace=/app \
-  --ak 'fabric_harness_settings={"cwd":"/app","base_url":"https://integrate.api.nvidia.com/v1","max_iterations":20}' \
+  --ak 'fabric_harness_settings={"base_url":"https://integrate.api.nvidia.com/v1","max_iterations":20}' \
   --ae "NVIDIA_API_KEY=$NVIDIA_API_KEY" \
   --job-name fabric-hermes \
   --jobs-dir "$RUNS_DIR" \
@@ -110,7 +110,7 @@ uv run --extra runtime --extra harbor harbor run \
   --ak fabric_config_base_dir=/opt/fabric-calculator \
   --ak fabric_workspace=/app \
   --ak fabric_telemetry=relay \
-  --ak 'fabric_harness_settings={"cwd":"/app","base_url":"https://integrate.api.nvidia.com/v1","max_iterations":4,"terminal_timeout":120}' \
+  --ak 'fabric_harness_settings={"base_url":"https://integrate.api.nvidia.com/v1","max_iterations":4,"terminal_timeout":120}' \
   --ae "NVIDIA_API_KEY=$NVIDIA_API_KEY" \
   --job-name fabric-hermes-relay \
   --jobs-dir "$RUNS_DIR" \
@@ -135,8 +135,9 @@ find "$RUNS_DIR/fabric-hermes-relay" \
 ## 4. Claude
 
 The Claude Agent SDK supplies its compatible Claude Code executable. Harbor
-passes the API key into the task environment; NeMo Fabric forwards only the model's
-configured credential variable to Claude.
+passes the API key into the task environment. NeMo Fabric forwards the supported
+Claude authentication variables selected by the adapter; this command uses
+`ANTHROPIC_API_KEY`.
 
 ```bash
 : "${ANTHROPIC_API_KEY:?Export ANTHROPIC_API_KEY before running Claude}"
@@ -173,8 +174,9 @@ uv run --extra runtime --extra harbor harbor view "$RUNS_DIR"
 ```
 
 Check NeMo Fabric status, harness and adapter identity, runtime and invocation IDs,
-artifacts, telemetry, Harbor exceptions, and reward. A successful smoke run has
-NeMo Fabric status `succeeded` and Harbor mean reward `1.0`.
+artifacts, telemetry, Harbor exceptions, and reward. A successful calculator run
+has one completed trial, zero errored trials, NeMo Fabric status `succeeded`, and
+Harbor mean reward `1.0`.
 
 After the runs, remove the generated build-context copy:
 

examples/harbor/prepare_swebench.sh

@@ -19,9 +19,10 @@ repo_root="$(git rev-parse --show-toplevel)"
 bundle_dir="$repo_root/examples/harbor/swebench"
 wheelhouse="$bundle_dir/.wheelhouse"
 relay_root="$bundle_dir/.relay"
+relay_cli_version="0.6.0"
 
-if [[ "$(uname -m)" != "x86_64" ]]; then
-    echo "This SWE-Bench example currently targets x86_64 task images." >&2
+if [[ "$(uname -s)" != "Linux" || "$(uname -m)" != "x86_64" ]]; then
+    echo "This SWE-Bench example currently requires an x86_64 Linux host." >&2
     exit 1
 fi
 
@@ -46,14 +47,14 @@ docker run --rm \
 
 # TEMP: Keep only this block if PyPI wheels land before Relay CLI distribution.
 relay_version="$({ "$relay_root/bin/nemo-relay" --version 2>/dev/null || true; })"
-if [[ "$relay_version" != "nemo-relay 0.5.0" ]]; then
+if [[ "$relay_version" != "nemo-relay $relay_cli_version" ]]; then
     docker run --rm \
         --user "$(id -u):$(id -g)" \
         -e CARGO_HOME=/tmp/cargo \
         -e CARGO_TARGET_DIR=/tmp/target \
         -v "$relay_root:/out" \
         rust:1.94-bullseye \
-        cargo install nemo-relay-cli --version 0.5.0 --locked --force --root /out
+        cargo install nemo-relay-cli --version "$relay_cli_version" --locked --force --root /out
 fi
 
 # TEMP: Record the source-built meta-wheel path for installation in the task.

examples/harbor/swebench/README.md

@@ -42,7 +42,7 @@ export RUNS_DIR="$PWD/.tmp/harbor/fabric-swebench"
 ```
 
 The preparation script builds the native runtime in a manylinux2014 container,
-builds the standalone Relay 0.5.0 CLI in Debian Bullseye, and writes the exact
+builds the standalone Relay 0.6.0 CLI in Debian Bullseye, and writes the exact
 task-local package requirement to `.fabric-package`. The generated files are
 ignored by Git. The uploaded bundle contains adapter descriptors, the example
 MCP server, generated wheels, and Relay; it does not contain a persisted

examples/harbor/swebench/sample-artifacts/README.md

@@ -20,3 +20,9 @@ Each harness directory contains:
 
 Use these samples to compare Harbor's canonical trajectory with Relay's direct
 ATOF and ATIF output before running the example yourself.
+
+The checked-in trajectories record Relay 0.5.0 and the earlier
+`nvidia/nemotron-3-nano` model identifier because they are historical evidence
+from the original runs. New runs use the compatible Relay CLI version pinned by
+[`prepare_swebench.sh`](../../prepare_swebench.sh) and the model selected in the
+current walkthrough; do not treat sample metadata as installation guidance.