Add GitHub PR review agent example for macOS

[munnamihir]

What this adds:

A new community example: examples/github-pr-review-agent/

An agent that automatically reviews open GitHub pull requests using a local Ollama model and posts structured review comments back to GitHub.

Why it's useful:

• No API keys or credits needed — runs fully local via Ollama
• First Mac-native example in the community repo (tested on Apple Silicon, Sonoma 14.x)
• Inside OpenShell sandbox, inference routes to Nemotron via inference.local automatically
• policy.yaml locks egress to api.github.com and inference.local only — nothing else leaves the sandbox

Files added:

• agent.py — main agent logic
• policy.yaml — OpenShell network policy
• .env.example — environment variable template
• README.md — setup and usage guide

Tested:

• Fetches open PRs from a GitHub repo
• Downloads unified diff
• Reviews with local LLM
• Posts comment back to GitHub
• OpenShell policy enforced

[munnamihir]

Looks like the DCO workflow is failing due to gsactions/dco-check@v1.1.1 not being in the allowed actions list for this org — not related to this PR's changes. Happy to re-sign or adjust anything needed once the workflow issue is resolved on the repo side.

[nv-nedelman-1]

@munnamihir some concerns about this PR:

1. these examples should make use of a nemoclaw-supported agent harness (e.g. hermes, openclaw etc), as opposed to a simple agent script
2. it seems to me like the agent script doesn't run in an "always-on" fashion - it requires the user to explicitly run it. a better example would make use of an always-on agent that periodically/automatically reviews PRs

[nv-nedelman-1]

Closing this PR for now - please re-open when above items are addressed and we'll re-review. thanks!

[munnamihir]

Thanks for the detailed feedback this is a really helpful direction.
I'll rework this example to:

Use Hermes as the agent harness that is replace the standalone script with a Hermes skill that handles PR review, following the pattern in personal-community-sentiment-triage
Make it always-on: add a polling loop inside the Hermes skill that watches for new PRs on a configurable interval (e.g. every 15 minutes via a scheduled trigger), so it runs continuously without manual invocation
Add a SOUL.md: define the agent's behavior and PR review persona the same way hermes-launchable structures it

Give me a day or two to rebuild it properly. Happy to get this merged the right way.

[munnamihir]

Thanks for the detailed feedback, really appreciated.

I've reworked the example to address both points:

Hermes agent harness replaced the standalone script with a proper Hermes agent structure (agents/hermes/) including SOUL.md, a github-pr-review skill with SKILL.md and a helper script, following the pattern from personal-community-sentiment-triage.

Always-on — the skill now runs in watch mode, polling for new PRs every 15 minutes continuously. Reviewed PRs are tracked in /tmp/reviewed_prs.json to avoid double-reviewing. A scripts/bring-up.sh starts the full NemoClaw sandbox
Inference — switched from Ollama to integrate.api.nvidia.com with Nemotron via the NVIDIA Build API, routed through NemoClaw

Let me know if the structure needs further adjustment happy to iterate.