Claude Code Cloud Routines: When outcomes Pushes to a Feature Branch Instead of main

TL;DR

Cloud routine outcomes config makes the cloud session work on a feature branch. Even if the skill says git push origin main, the agent is already on the branch and pushes there. Fix: remove outcomes + add git checkout main in every skill. Along the way, hit a RemoteTrigger list API pagination bug where the cursor never advances — had to get trigger IDs manually from the Web UI.

Context

The quidproquo blog runs 15 Claude Code cloud routines to auto-generate daily articles (arxiv digest, GitHub trending, security alerts, model cards, etc.). Each routine reads its corresponding skill file and pushes to main at the end.

On 2026-09-01, the arxiv digest appeared on the site but the other 7 daily articles were missing.

Problem

Comparing the successful arxiv session's run log against the failed GitHub digest session using RemoteTrigger's get_run_log revealed the key difference:

Arxiv session (success): worked on main, ran git push -u origin main directly.

GitHub digest session (failure): worked on claude/trusting-cannon-npgqrz, pushed to the feature branch. Articles never reached main.

Both routines had identical skill instructions saying git push origin main. Why the different behavior?

Investigation

Step 1: Compare routine configs

Pulled both routine configs with RemoteTrigger's get action. Both had an outcomes field:

"outcomes": [{
  "git_repository": {
    "git_info": {
      "branches": ["claude/trusting-cannon"],
      "repo": "vincentxuu/quidproquo"
    }
  }
}]

outcomes makes the cloud environment create a feature branch with a random suffix (e.g., claude/trusting-cannon-npgqrz) and checks it out. When the agent runs git pull origin main, it only merges main into the feature branch — it does not switch to main.

Step 2: Assess blast radius

Scanned all remote branches with git ls-tree and found 8 branches with 09-01 daily articles that never made it to main:

for b in $(git branch -r); do
  count=$(git ls-tree --name-only "$b" -- src/content/posts/daily/ \
    | grep "2026-09-01" | wc -l)
  [ "$count" -gt 0 ] && echo "$b: $count files"
done

Step 3: List API pagination bug

Tried to list all daily routines via RemoteTrigger's list action to batch-update their configs. The API's next_cursor always returned the same value MTc4NzUwMDcwNTE0NDI2MzAwMHw2YjhmZmRiOC0wN2JkLTQ1ZWUtOTI3NS04Mjc4YTZkNTM5NjA= (base64 for 1787500705144263000|6b8ffdb8-07bd-45ee-9275-8278a6d53960). Pagination was stuck.

The newest 20 results were all one-shot PR re-check routines (created Aug 23-27), completely hiding the daily-digest routines (created Aug 16). Tried fabricating earlier cursors — same page returned. The server ignores the cursor parameter entirely.

Workaround: user manually copied 14 trigger IDs from https://claude.ai/code/routines.

Fix

Three-layer repair, applied simultaneously:

1. Rescue stuck articles: cherry-pick

Each branch had only 1-2 unique commits. Cherry-picked all 8 back to main with zero conflicts:

git cherry-pick 5b2622f5  # github digest
git cherry-pick f5b419b5  # product builder interview
git cherry-pick 9ee6b4be  # security alert
# ... 8 commits total
pnpm verify  # all green
git push origin main

2. Add git checkout main to skills (belt)

Added git checkout main before git pull origin main in all 15 daily-digest skills:

# Step 1: Preparation
git checkout main        # ← added
git pull origin main

Batch-applied with sed:

for f in .agents/skills/daily-digest-*/SKILL.md; do
  sed -i '' '/^git pull origin main$/i\
git checkout main
' "$f"
done
pnpm skills:sync

3. Remove outcomes from routine configs (suspenders)

Updated all 15 routines via RemoteTrigger's update action, rewriting session_context without outcomes:

{
  "job_config": {
    "ccr": {
      "session_context": {
        "allowed_tools": ["Bash", "Read", "Write", "Edit", "Glob", "Grep"],
        "model": "claude-sonnet-5",
        "sources": [{"git_repository": {"url": "https://github.com/vincentxuu/quidproquo"}}]
      }
    }
  }
}

All 13 remaining updates sent in parallel, all returned HTTP 200.

Root Cause

outcomes is a cloud routine config designed to direct session output to a specific feature branch — useful when you want PR-based review. For a personal blog that pushes directly to main, it creates an implicit trap:

1. The cloud environment checks out the outcomes branch after cloning
2. git pull origin main on a feature branch is a merge, not a branch switch
3. The agent may follow the branch it's on rather than the skill's git push origin main instruction
4. The arxiv routine "accidentally succeeded" because its agent happened to stay on main

The outcomes field's branch-checkout behavior is not documented — it was deduced by comparing run logs.

Lessons Learned

• Cloud routine outcomes changes the session's starting branch. If your workflow pushes directly to main, don't configure outcomes.
• git pull origin main ≠ git checkout main. On a non-main branch, pull-from-main is just a merge.
• get_run_log is the most useful debugging tool for cloud routines — it shows every command the agent ran, which branch it was on, and where it pushed.
• The RemoteTrigger list API has a pagination bug (cursor never advances). Work around it with get and known trigger IDs.

References

• Claude Code Routines Management UI
• Claude Code Official Documentation
• RemoteTrigger API — built-in reference from /schedule skill — API reference loaded when the /schedule skill is invoked