Autonomous Workflow Patterns

duyetbot can run autonomously within a session (via /loop) or as a one-shot task (via claude -p). These patterns reduce human-in-the-loop friction for routine maintenance.

Context compaction

Auto-compaction is explicitly welcomed when the conversation grows long. The user confirmed on 2026-05-25: "remember -> can compact to save tokens when needed."

Before compaction happens, save anything future turns will need into a memory file or commit body. Never repeat goal text verbatim in responses — the goal hook re-surfaces it on every stop. Commit messages are the most durable form of session memory.

Background deploys

Always run cf:deploy:prod in the background after pushing. Continue with the next task immediately. The harness notifies when the deploy completes; verify production after the notification.

Do NOT run two background deploys in parallel — cf-deploy-prod.sh renames .env.local → .env.local.deploy-bak during execution. Parallel invocations collide on that rename and lose CLOUDFLARE_API_TOKEN propagation.

Recurring task pattern (/loop)

/loop 1h "run PLAN.md maintenance: benchmark → assess → act → verify → ship"

Each iteration:

1. Measure current state
2. Fix the highest-priority issue
3. Verify (lint + test + build pass)
4. Commit and push
5. Deploy in background
6. Log to memory/maintenance-log.md

Autonomous decision priority

When running unattended, address issues in this order:

1. Build fails → fix immediately
2. Tests fail → fix immediately
3. Lint errors → auto-fix
4. Deploy needed → verify then ship
5. Code quality → improve if no higher-priority issues open
6. New features → only from memory/roadmap.md

Improvement rotation

Nine improvement cycles completed as of March 2026 (~155 total fixes). A cycle typically runs as a scoped /loop targeting one category (security, dead code, design, tests). Cycle 9 added 104 new tests and unified the design system across llm-timeline, cv, and homelab.