GitHub Release Analyzer#

Use this skill as a fixed pipeline:

text

prepare -> summarize -> render -> deliver -> commit

Only summarize is free-form. prepare, render, and commit are deterministic script steps.

Inputs#

• Repo input: https://github.com/<owner>/<repo> or <owner>/<repo>
• Default stateKey: {owner}__{repo}
• Default template: default
• Default behavior:
  • manual: latest formal release only, no state read/write
  • cron: formal releases only, first run selects latest only, later runs select unprocessed only

Workflow#

Before running any scripts/... command, set the command working directory to the directory containing the loaded SKILL.md (the Skill root). Do not run these commands from the OpenClaw workspace root. This does not require a shell SKILL_DIR environment variable: OpenClaw uses the loaded Skill's installed directory, and Hermes uses the skill_dir returned by skill_view.

1. Read references/execution-modes.md and choose manual or cron.
2. From the Skill root, run:

bash

python3 scripts/run.py prepare --repo <repo> --mode <manual|cron>

Useful flags:

• --state-key <key>
• --limit <n>
• --include-prerelease
• --output-template default
• --initial-cron-behavior latest-only

1. If status=has_updates, read references/summary-contract.md and produce one summary string per selected release. Keep the summary language aligned with the primary language of the invocation instruction.

2. Render the final message:

bash

python3 scripts/run.py render --repo <repo> --mode <manual|cron> < payload.json

When building the render payload, preserve the release fields required by render, especially tag_name, published_at, and html_url.

1. In cron mode, after successful delivery, commit state:

bash

python3 scripts/run.py commit \
  --repo <repo> \
  --mode cron \
  --processed-tags <comma-separated-tags> \
  --latest-release-id <id> \
  --latest-published-at <iso8601>

Rules#

• Treat prepare output as the source of truth for selection, ordering, first-run behavior, repo normalization, and state path.
• The output language must follow the primary language of the invocation instruction.
• The final outgoing message must come from render.
• In cron mode:
  • if status=no_update or status=no_release, reply with NO_REPLY
  • if status=has_updates, emit only the final rendered message
• Update processed state only after successful delivery.

Hard prohibitions#

• Do not skip prepare.
• Do not bypass render.
• Do not reorder or silently drop releases selected by prepare.
• Do not call commit in manual mode.
• Do not call commit before delivery succeeds in cron mode.
• Do not replace the summary contract with ad hoc prose outside the required summaries[] string-array shape.
• In cron mode, when status=no_update or status=no_release, reply exactly NO_REPLY.
• In cron mode, do not emit progress chatter, intermediate JSON, step labels, or workflow narration.

Files to read#

• references/execution-modes.md
• references/summary-contract.md
• references/templates.md
• references/state-format.md