forked from cobusgreyling/loop-engineering
-
Notifications
You must be signed in to change notification settings - Fork 0
Production Patterns
Marius Egerhei Torjusen edited this page Oct 3, 2026
·
1 revision
Loop engineering provides seven battle-tested patterns designed for real-world development workflows. Each pattern provides defined scheduling, risk profiles, verification gates, and human hand-off protocols.
| Pattern | Recommended Cadence | Risk Level | Primary Objective |
|---|---|---|---|
| Daily Triage | 2h β 24h | Low | Scans open PRs, stale issues, and updates STATE.md with team priorities. |
| CI Sweeper | 5m β 15m | Medium | Detects failing CI runs, isolates root cause, generates minimal reproducible fix. |
| PR Babysitter | 5m β 15m | Medium | Keeps PRs rebased against main, resolves merge conflicts, ensures CI checks pass. |
| Dependency Sweeper | 6h β 24h | Medium | Audits npm/pip advisories, bumps patch versions, verifies lockfiles against tests. |
| Issue Triage | 2h β 24h | Low | Categorizes issues, detects duplicates, requests missing reproduction steps. |
| Changelog Drafter | 24h | Low | Parses merged PRs and commit logs to draft release notes grouped by feature. |
| Post-Merge Cleanup | 6h β 24h | Low | Deletes merged remote branches, closes resolved tickets, updates documentation. |
- Problem Solved: Standup overhead and uncoordinated agent work.
-
Mechanism: Queries GitHub/Linear for assigned items, inspects blockers, reconciles
STATE.md, and prepares the task queue for downstream specialist agents. - Safety Gate: Read-only analysis. Never modifies production code or closes tickets without explicit label rules.
- Problem Solved: Flaky tests, broken trunk builds, and slow developer feedback loops.
- Mechanism: Triggered on CI failure webhook or polled schedule. Clones the failing commit into an isolated worktree, reproduces the failure locally, applies a minimal fix, and verifies via local test suite.
- Verification: The fix must pass 100% of existing tests plus add a targeted regression test.
- Problem Solved: PR bit-rot, merge conflicts, and stale review comments.
-
Mechanism: Monitors active PRs. When
mainadvances, fetches the PR branch into a worktree, rebases cleanly, resolves trivial conflicts, runs verification, and force-pushes with lease. - Human Gate: If non-trivial semantic conflicts occur, tags the author and exits without destructive action.
- Problem Solved: Security vulnerabilities (Dependabot spam) and stale package versions.
-
Mechanism: Runs
npm audit/cargo audit/pip audit. Identifies non-breaking semver patches. Updates dependencies one at a time and runs the full test suite. - Safety Rule: Major version bumps and lockfile anomalies are strictly escalated to human engineers.
- Problem Solved: Unstructured bug reports and feature requests cluttering the backlog.
- Mechanism: Reads incoming issues. Checks against an issue template schema. Prompts the author for missing reproduction steps, logs environment details, and assigns component labels.
- Problem Solved: Late, inaccurate, or missing release notes.
-
Mechanism: Analyzes the delta between the latest git tag and
HEAD. Extracts PR titles and conventional commit prefixes (feat:,fix:,perf:). Drafts release markdown indocs/releases/.
- Problem Solved: Branch sprawl and lingering tracking issues.
-
Mechanism: Scans remote branches whose commits have been merged into
main. Deletes the remote branch if delete-on-merge was not triggered, and marks corresponding tickets as resolved.
WEEK 1: L1 REPORT-ONLY MODE
βββ Loop executes triggers and runs checks
βββ Generates reports and diffs to STATE.md
βββ Prohibits direct pushes and PR merges
WEEK 2+: GRADUATED ACTUATION
βββ Only after 7 days of 100% report accuracy
βββ Enable automated branch creation and PR opening
βββ Maintain human gate on auto-merge
STOP PROMPTING / DESIGN THE LOOP / GET A SCORE
Loop Engineering Β· Autonomous Code Agent Tooling, Strict Eval Frameworks & Local-First CLI Utilities
Sololys Systems & Engineering Β· sololys/loop-engineering