Nightly Routine: draft PRs for low-risk silent bugs

Why this matters

The silent-bugs recipe is the on-demand version: you sit at your desk, ask “what’s silently breaking right now,” the agent finds them. Great but bounded — you have to be there.

This is the lights-out version. The Routine wakes at 2am, picks the top silent bug, investigates it, and either:

• Opens a draft PR (if the fix is small, contained, and well-understood)
• Files a one-paragraph triage note (if the fix needs human judgment)

You wake up to a Slack notification: “I drafted PR #420 for the silent NullPointer in BillingForm — 14 LOC, no schema, no config change. Review and merge if it looks right.” Or: “I triaged the runaway image-decoder bug but the fix needs a new env var — see the note.” Either way, the work moved forward overnight.

Most teams never get to nightly bug-fixes because the labor cost is real (an engineer attached to each one). With a Routine, the engineer-attached-cost is “5 minutes reviewing a PR while drinking coffee.”

What you get

• A scheduled cloud agent that runs nightly
• Per run: picks the top silent bug, investigates, drafts a PR or a note
• Strict guardrails (PR only if <20 LOC + no schema/config touched + tests still pass)
• A morning notification with the result

Walk through it

/schedule "0 2 * * *" "
1. agentry_list_cases { status: 'open', has_agent_summary: false, min_users_affected: 5,
                        order_by: 'users_affected desc', limit: 1 } → pick TARGET_CASE
   If empty: log 'no silent bugs to triage', exit silently.

2. agentry_get_case { case_id: TARGET_CASE.id } → get stack, message, env, fingerprint

3. Reproduce locally: cd into the project, search for the file:line in the stack,
   read 50 lines around the call site. Read related files.

4. Hypothesize the fix. Validate against the audit:
   - LOC touched < 20?
   - Any schema migrations needed?     → if yes, abort PR, file triage note
   - Any config / env / secret changes?  → if yes, abort PR, file triage note
   - Tests for the affected file exist?  → if no, abort PR
   - Does the change pass typecheck and existing tests?  → if no, abort

5. If all guards pass:
   a) Create a branch: fix/<fingerprint>
   b) Apply the fix
   c) Add a regression test
   d) gh pr create --draft --title 'fix(<area>): null-guard in <file>' --body '<full
      summary including: case link, root cause, fix description, test added>'
   e) agentry_remember case_id=TARGET_CASE.id memo='Auto-PR #<num> opened nightly; review tomorrow'

6. Otherwise:
   a) agentry_remember case_id=TARGET_CASE.id memo='Auto-triage: <one-paragraph
      hypothesis + why I couldn't auto-fix it (e.g. 'needs new env var X')>'

7. Post to wh_morning Slack: a one-line summary of step 5 or step 6.
"

{ "routine_id": "rt_nightly", "next_run_at": "2026-05-17T02:00:00Z" }

🌙 Nightly triage — May 17

  ✅ Drafted PR #420
     Case: c_npe_1 (BillingForm.tsx:118 — null user.address)
     Fix: 14 LOC, added null-guard + regression test
     Tests pass: ✓     Typecheck: ✓
     Review: https://github.com/you/repo/pull/420

Or:

🌙 Nightly triage — May 17

  📝 Triage note (no PR — needs human)
     Case: c_oom_2 (ImageDecoder OOM at 64MB)
     Hypothesis: image dimensions exceeded ImageDecoder default budget
     Why no auto-PR: fix needs a new env var IMAGE_MAX_BUDGET_MB and
       a config change to deploy/staging.yaml — out of guardrail scope.
     Read the note: https://agentry.sh/case/c_oom_2#memory

The output

auto-fix Routine (rt_nightly)

  ✓ Schedule:        nightly at 02:00 UTC
  ✓ Target:          top silent bug (≥5 users, no prior summary)
  ✓ PR guardrails:
      max 20 LOC, no schema, no config, tests must pass
  ✓ Slack:           wh_morning   → #eng-mornings
  ✓ First run:       2026-05-17T02:00:00Z

Expected morning notification (2 forms):

  ✅ Drafted PR #N (full fix)
       Case + diff link
       Test added, all checks green

  📝 Triage note only (humans needed)
       Hypothesis + why auto-fix was blocked
       Read note: agentry.sh/case/...

Setting it up

1. The Routine itself. The /schedule prompt above is the full Routine. The agent running it has access to the MCP toolbox + a shell with git and gh configured.

2. GitHub CLI auth. The Routine needs gh authenticated to a bot account or scoped PAT. Set up the credentials in the Routine’s execution context — see your Routines docs for secret injection.

3. Tune the guardrails. Defaults are conservative for a reason. After a week of green PRs, you might relax LOC to 50 or allow simple package upgrades. Don’t relax “no schema / no config” — that’s where Routines get genuinely dangerous.

4. Pair with weekly-agent-digest. Friday’s digest will include “drafted 4 PRs, 2 merged, 2 closed” — a quick read on whether the Routine is actually shipping value or just generating PRs you ignore.

5. Add a kill switch. Set up a feature flag (disable_nightly_auto_fix) that the Routine checks before acting. If you ever need to pause it, flip the flag instead of editing the schedule:

// At the top of the Routine prompt:
//   if agentry_evaluate_feature_flag('disable_nightly_auto_fix', 'system') is true,
//   log and exit silently.

Variations

• “Same Routine but scoped to my mobile app only — production-Android cases ≥ 100 users.”
• “Hourly variant: don’t wait for 2am, just run continuously. Each hour pick the top untriaged case.”
• “For cases the Routine drafts a PR on, auto-assign to the engineer who last modified the file (CODEOWNERS lookup).”
• “Add a ‘maximum PRs per day’ cap (3?) so you don’t wake up to a flood when your error rate spikes.”
• “On Slack message, include a ‘snooze 24h’ button that adds the case fingerprint to a no-touch list.”