Add dbhq-uk/heliograph-skill to community agent skills

[grinidx]

Adds dbhq-uk/heliograph-skill under Community Discovery and Skills.

The operational use case. Debugging a host the agent cannot reach at all: air-gapped, client-owned or change-controlled estates where only the customer's own staff can log in. It is not another investigator agent - it is the transport and capture layer for when SSH is off the table. Steps go out over a private git repo, an on-site operator runs one command that never changes (or starts an unattended runner), and the whole run comes back as a committed log. Every line carries a UTC timestamp, so a hang shows up as a gap rather than being indistinguishable from slow progress, and the log is pushed with its exit code even when the step fails.

Classification, and why.

Field	Value	Reasoning
action_level	write-capable	Steps are arbitrary shell. Nothing constrains them to reads, so the stricter label is the honest one
human_approval	true	The documented default is the operator pulling and running each step. Disclosed in risk_notes: the optional unattended runner executes whatever is pushed to the branch it watches, so per-step review only holds in the manual mode
evidence_tracing	yes	The committed log is the deliverable, not a side effect. Every run is retained in git, timestamped and untruncated
maturity	prototype	Published 11 Aug 2026. Stricter signal, per the contributing guide

Against the operator safety checklist. No credentials of any kind: no cloud auth, no API keys, no tokens beyond the git remote. Nothing to give the agent, so it can be evaluated with no real cloud credentials at all - bootstrap.sh into a scratch repo and run ./run.sh env against any box you already have. Blast radius is whatever account the operator runs it as, which is stated in risk_notes. Secrets handling is the one real caveat and it is documented in the skill's own SECURITY.md as well as here: logs are committed, so anything a command prints lands in git history permanently. cap_redact masks the common shapes on the way out and the README says plainly that this is a safety net, not a guarantee.

It does not tunnel, proxy or hold a connection open. The only thing that crosses the gap is a git commit, run by someone who already has legitimate access.

Checklist

• Repo URL is public and reachable.
• Specific category: community-agent-skills, an existing catalog slug.
• type: skill, an existing validator-backed artifact kind.
• risk_notes explains what could go wrong.
• operator_note explains why an infrastructure operator should care.
• Labels match observed behaviour: prototype, approval, evidence, write. No prod/prototype combination; write paired with write-capable; evidence paired with evidence_tracing: yes.
• data/repos.yaml and README.md both updated, plus data/catalog.json and the README counts via scripts/sync_catalog_json.py and scripts/sync_readme_counts.py.
• python3 scripts/validate_repos_yaml.py passes (81 entries); pytest passes (149 tests).
• Open source, MIT. Nothing commercial-only, no paid tier, no gated functionality.

Disclosure: I am the author.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud