Setup Automation

Set up the automatic analysis pipeline: WandB Automation → GitHub Actions → Claude Code. This enables automatic result analysis and PR posting when WandB runs finish.

Prerequisites

• /init-project

  must have been run first (WandB and GitHub are configured)
• User must have a GitHub PAT with

  repo

  scope ready
• User must have an Anthropic API key ready

Workflow

Phase 1: Gather Credentials

1. Ask for credentials (these will be stored as GitHub Secrets, NOT in the repo):
   • Anthropic API key (

     ANTHROPIC_API_KEY

     )
   • GitHub PAT with

     repo

     scope (for WandB webhook → GitHub dispatch)
   • Confirm WandB API key is available (should already be in

     .claude/settings.local.json

     )

Phase 2: GitHub Repository Configuration

2. Get repo info

   gh repo view --json owner,name,url
   

3. Set GitHub Secrets

   gh secret set ANTHROPIC_API_KEY
   gh secret set WANDB_API_KEY
   

   Claude cannot set these programmatically (they require interactive input or piped values). Guide the user to set them:

   • Option A:

     echo "<value>" | gh secret set ANTHROPIC_API_KEY

   • Option B: Set via GitHub UI → Settings → Secrets and Variables → Actions

4. Set GitHub Variables (non-secret, read from

   .claude/settings.json

   ):

   gh variable set WANDB_ENTITY --body "<entity>"
   gh variable set WANDB_PROJECT --body "<project>"
   

5. Create GitHub labels for experiment tracking:

   # Workflow state labels
   gh label create "experiment:in-progress" --color "FFA500" --description "Experiment branch is being developed" --force
   gh label create "experiment:done" --color "6E6E6E" --description "Experiment cycle completed" --force
   
   # Verdict labels
   gh label create "experiment:winner" --color "00AA00" --description "Experiment outperformed baseline" --force
   gh label create "experiment:loser" --color "DD0000" --description "Experiment did not outperform baseline" --force
   
   # Run state labels (set by check-results post-pr)
   gh label create "experiment:finished" --color "0E8A16" --description "WandB run finished successfully" --force
   gh label create "experiment:failed" --color "B60205" --description "WandB run failed" --force
   gh label create "experiment:crashed" --color "B60205" --description "WandB run crashed" --force
   gh label create "experiment:running" --color "1D76DB" --description "WandB run is in progress" --force
   gh label create "experiment:killed" --color "FBCA04" --description "WandB run was killed" --force
   
   # Baseline labels (used by rollback-baseline)
   gh label create "regression" --color "D93F0B" --description "Baseline regression detected" --force
   gh label create "baseline" --color "5319E7" --description "Related to baseline management" --force
   

Phase 3: Verify GitHub Actions Workflow

6. Check workflow file exists

   cat .github/workflows/wandb-run-finished.yml
   

   • If missing, warn the user and suggest re-initializing from the template

7. Verify workflow is enabled

   gh workflow list
   

   • If the workflow is disabled, enable it:

     gh workflow enable wandb-run-finished.yml

Phase 4: WandB Automation Setup

8. Guide user through WandB Automation creation

   WandB Automations must be configured in the WandB UI. Print the following instructions:

   === WandB Automation Setup ===
   
   1. Go to: https://wandb.ai/<entity>/<project>/automations
   2. Click "Create automation"
   3. Configure:
      - Name: "Auto-analyze on run finish"
      - Event type: "Run state changes"
      - Trigger on: finished, failed, crashed
      - Action: "Webhook"
      - URL: https://api.github.com/repos/<owner>/<repo>/dispatches
      - Method: POST
      - Headers:
        Authorization: Bearer <YOUR_GITHUB_PAT>
        Accept: application/vnd.github+json
      - Body: (copy from misc/wandb_webhook_payload.json)
   4. Save the automation
   

   Ask user to confirm when done.

Phase 5: Test the Pipeline

9. Test GitHub dispatch (dry run):

   gh api repos/<owner>/<repo>/dispatches \
     -f event_type=wandb_run_finished \
     -f 'client_payload[branch]=test' \
     -f 'client_payload[run_id]=test-run' \
     -f 'client_payload[run_state]=finished'
   

10. Verify the workflow ran:

    gh run list --workflow=wandb-run-finished.yml --limit 1
    

    • If it shows a run triggered by

      repository_dispatch

      , the pipeline works
    • The test run will fail (no real branch/run) — that's expected, it just proves the trigger works

Phase 6: Summary

11. Print verification checklist:

    • ANTHROPIC_API_KEY

      set in GitHub Secrets

    • WANDB_API_KEY

      set in GitHub Secrets

    • WANDB_ENTITY

      set in GitHub Variables

    • WANDB_PROJECT

      set in GitHub Variables
    • All GitHub labels created (workflow, verdict, run state, baseline)

    • .github/workflows/wandb-run-finished.yml

      exists and is enabled
    • WandB Automation created with webhook to GitHub dispatch
    • Test dispatch triggered the workflow

12. Print how it works:

    Pipeline: WandB run finishes → WandB Automation fires webhook →
    GitHub Actions receives dispatch → Claude Code analyzes results →
    Results posted to PR automatically
    

Rules

• Never store secrets in the repo — only in GitHub Secrets or

  .claude/settings.local.json

• The GitHub PAT for the webhook belongs to the user — Claude should never ask for it in plaintext. Guide the user to paste it directly into WandB UI.
• If any step fails, explain what went wrong and how to fix it manually
• This skill is idempotent — safe to re-run if setup was partial