| name | discord-ralph-recap |
| description | Post a beautifully formatted recap of a Ralph tick-loop's progress to a Discord channel every time a pull request is merged, using the DISCORD_BOT_TOKEN and RALPH_CHANNEL_ID environment variables. Use when the user says "recap my Ralph loop", "post merge stats to Discord", "summarize the tick loop", "set up a Ralph progress bot", or wants merge-rate, review- iteration, ETA, and a ten-word "what this unlocked" headline pushed to Discord on each merge. Covers the GitHub Actions trigger, the stats engine, and the Discord embed. Do NOT use for iterating on a single PR's review feedback (use address-feedback skill), waiting on a verdict (use await-claude-review skill), debugging CI failures (use ci-debugging skill), or general GitHub backlog cleanup (use backlog-grooming skill). |
| metadata | {"author":"Geoff","version":"1.0.0"} |
Discord Ralph Recap
Whenever a PR merges, summarize how the Ralph tick loop is doing and post it to a Discord channel as a clean embed: PRs merged, merge rate, average review iterations before LGTM, estimated time remaining, and a ten-word headline for what the latest merge unlocked.
A "Ralph tick loop" is an autonomous agent grinding a backlog one issue at a time — each tick opens a PR, iterates against the Claude reviewer's verdict until LGTM, merges, and moves on. This skill turns that merge history into a recap.
What You Need
DISCORD_BOT_TOKEN — a Discord bot token. The bot must be in the server and have View Channel + Send Messages on the target channel.RALPH_CHANNEL_ID — the target channel's numeric ID (Developer Mode → right-click channel → Copy Channel ID).- A GitHub token. In GitHub Actions the built-in
GITHUB_TOKEN is enough (it needs contents: read, pull-requests: read, issues: read).
- Optional
ANTHROPIC_API_KEY — enables the Claude-written ten-word headline. Without it, the headline falls back to a cleaned PR title, and nothing else changes.
Files
discord-ralph-recap/
├── SKILL.md
├── scripts/
│ ├── recap.py # I/O shell: fetch GitHub data, generate headline, post embed
│ └── stats.py # pure, unit-tested statistics math
├── assets/
│ └── ralph-recap.yml # GitHub Actions workflow: fires on every merged PR
└── references/
└── metrics.md # exactly how each number is computed, and how to tune it
Instructions
Step 1: Confirm the trigger model
The recap should "fire every time a PR is merged". The robust way to do that is a GitHub Actions workflow keyed on pull_request with types: [closed], gated to merged PRs:
on:
pull_request:
types: [closed]
jobs:
recap:
if: github.event.pull_request.merged == true
closed fires on both merge and plain close; the merged == true guard keeps it to actual merges. Do not use push: branches: [main] — squash-merges land as one commit but a non-merge push (a hotfix, a revert) would also trip it.
Step 2: Install the workflow
Copy the bundled workflow into the repo and confirm the secrets are wired:
mkdir -p .github/workflows
cp .claude/skills/discord-ralph-recap/assets/ralph-recap.yml .github/workflows/ralph-recap.yml
Then add the repository secret DISCORD_BOT_TOKEN and the repository variable (or secret) RALPH_CHANNEL_ID. Add ANTHROPIC_API_KEY as a secret if you want the Claude headline. The workflow reads them as env vars of the same name — no code change needed.
Step 3: Dry-run locally before trusting the trigger
Verify the embed renders before it ever posts. --dry-run prints the payload instead of sending it:
DISCORD_BOT_TOKEN=unused RALPH_CHANNEL_ID=unused \
GITHUB_TOKEN="$(gh auth token)" \
python .claude/skills/discord-ralph-recap/scripts/recap.py \
--repo your-org/your-repo --dry-run
Inspect the JSON. If the numbers look right, drop --dry-run (and set the real DISCORD_BOT_TOKEN / RALPH_CHANNEL_ID) to post once manually.
Step 4: Tune the metrics if needed
The defaults work out of the box, but the backlog ETA and the iteration count make assumptions worth checking — see references/metrics.md. The two most common adjustments:
- Backlog size: ETA divides open GitHub issues (PRs excluded) by the merge rate. If the loop's backlog lives somewhere else (a
git-issues/ directory, a project board), point the count at that instead.
- Verdict detection: iteration counting keys on Claude review comments containing a
Verdict line, matching the iteration-trigger.yml convention. If the reviewer phrases verdicts differently, adjust normalize_verdict in stats.py.
Step 5: Let it run
After the first real merge, the recap posts automatically. Each post is self-contained — no state is stored between runs; every recap is recomputed from the live merge history.
Examples
Example 1: Stand it up on an existing Ralph repo
cp .claude/skills/discord-ralph-recap/assets/ralph-recap.yml .github/workflows/ralph-recap.yml
gh secret set DISCORD_BOT_TOKEN
gh variable set RALPH_CHANNEL_ID --body "123456789012345678"
gh secret set ANTHROPIC_API_KEY
git add .github/workflows/ralph-recap.yml
git commit -m "Add Discord Ralph recap on merge"
The next merged PR triggers the first recap.
Example 2: Post a one-off recap from the terminal
DISCORD_BOT_TOKEN="$BOT_TOKEN" RALPH_CHANNEL_ID="$CHANNEL_ID" \
GITHUB_TOKEN="$(gh auth token)" ANTHROPIC_API_KEY="$ANTHROPIC_API_KEY" \
python .claude/skills/discord-ralph-recap/scripts/recap.py --repo your-org/your-repo
Useful for backfilling a recap after enabling the bot mid-campaign.
Troubleshooting
Error: Discord API request failed: 401 Unauthorized
The bot token is wrong or revoked. Confirm DISCORD_BOT_TOKEN is the bot token (from the Bot tab), not the application/client secret, and that it is not prefixed with Bot — the script adds that prefix itself.
Error: Discord API request failed: 403 Forbidden
The bot is not in the server, or lacks View Channel / Send Messages on RALPH_CHANNEL_ID. Invite the bot with the bot scope and grant the two permissions on that channel.
Error: Discord API request failed: 404 Not Found
RALPH_CHANNEL_ID is wrong. Enable Developer Mode in Discord, right-click the channel, Copy Channel ID — it is an all-digits snowflake, not the channel name.
The headline is just the PR title
ANTHROPIC_API_KEY is unset or the SDK isn't installed, so the heuristic fallback is used. Add the key and pip install anthropic (the bundled workflow already installs it) to get the Claude-written headline.
Iteration count says "no LGTM verdicts found yet"
No merged PR has a Claude review comment containing a Verdict line. Either the reviewer isn't running, or it phrases verdicts differently — see Step 4 and references/metrics.md.
Backlog ETA is "unknown (loop is stalled)"
The merge rate computed to zero — there are no merges in the window, so no rate to extrapolate. This resolves itself once merges resume.
See Also
references/metrics.md — precise definition and tuning for every number in the recap.claude/skills/address-feedback/ — the per-PR feedback loop this recap aggregates over.claude/skills/await-claude-review/ — how the Verdict comments the iteration count reads get produced
