Ask when it helps
For substantial, ambiguous work, Joy asks what should remain yours. Clear or trivial tasks keep moving.
An Agent Skill for thoughtful delegation
Keep the joy.
Let go of the toil.
Joy helps you decide what to delegate without giving away the parts of programming you value. It asks a focused question before an AI agent takes over.
- Claude Code
- GitHub Copilot CLI
- Agent Skills
AI can remove friction.
Joy keeps it from removing your part before you choose.
A movable boundary
Not less AI.
More intention.
Meaningful challenge and incidental toil are not the same thing. Joy lets you keep the design, discovery, or craft you want—and hand off the rest without guilt.
No moral hierarchy
Learning, craftsmanship, exploration, and fast delivery are equally valid goals. Delegating everything is valid too.
A boundary that listens
Your latest instruction wins. Move work between Keep, Pair, and Delegate without ceremony or pressure.
Choose the session you need
Choose a mode. Your call.
Be explicit, or let Joy infer the lightest useful boundary from your request.
pairdefaultWork through it together
Focused, understandable increments with consequential choices discussed before implementation.
learnProtect the discovery
Questions, explanations, and graduated hints while the central solution stays yours.
craftKeep hands-on authorship
You own architecture and interesting implementation; the agent handles support work.
shipOptimize for completion
Safe autonomous execution, a small complete change, validation, and a concise handoff.
tidyDecide what stays
Evidence-based Keep, Refine, and Let go recommendations before destructive cleanup.
offStep back immediately
Return to normal assistant behavior without carrying a previous Joy boundary forward.
See the boundary move
Joy in the terminal.
Watch real CLI sessions. Pause, seek, change the pace, or select terminal text while Joy protects a different kind of ownership.
An end-to-end case study
A crash window.
An intentional boundary.
Follow a real Claude Code review of an executable webhook worker. The delivery semantics remain developer-owned while Joy pairs on the failure invariant and delegates verification.
Crash-safe webhook delivery
What happens when delivery succeeds, then the worker crashes?
Executable fixture
Jump to
Protected implementation
Open worker.py
The consequential lines stay visible.
examples/webhook_worker/worker.pydeveloper-owned core
claim = self._repository.claim_due(now=now, lease_seconds=self._policy.lease_seconds)
+ self._transport.send(
+ claim.endpoint,
+ claim.payload,
+ idempotency_key=claim.idempotency_key,
+ )
+ if crash_after_send:
+ raise CrashAfterSend("delivery succeeded before its local acknowledgement")
+ if not self._repository.mark_delivered(claim, now=now):
+ raise LostLease(claim.delivery_id)Why it matters: the remote effect and local acknowledgement cannot be atomic. The chosen contract permits redelivery with the same idempotency key rather than pretending exactly-once delivery exists.
Deterministic fault injection
Open tests
Prove the uncomfortable boundary.
✓
Crash after sendRedelivers after lease expiry with the same key.
✓
Transport failureWaits for exponential backoff before retrying.
✓
Stale acknowledgementCannot overwrite a newer claim generation.
condensed unittest proof
test_crash_after_send_redelivers_with_one_receiver_effect ... ok
test_stale_worker_cannot_acknowledge_a_newer_claim ... ok
test_transport_failure_waits_for_backoff_before_retry ... ok
Ran 3 tests
OKDecision log
Read the rationale
The trade-offs are part of the artifact.
At-least-once delivery
A successful remote effect may be repeated after a crash. The stable key is part of the contract.
Attempts are generations
A stale worker cannot acknowledge or reschedule work after a newer lease has been claimed.
Retry forever—for now
The example has no dead-letter policy. That omission remains an explicit product decision, not a hidden default.
Verification plumbing
SQLite persistence, a deduplicating fake receiver, fault injection, and the test harness support the chosen semantics.
A small skill, no runtime
Bring Joy to your next session.
The canonical skill is plain Markdown, follows the Agent Skills standard, and has no runtime dependencies.
Install globally with skills.sh
npx skills add JGalego/Joy --skill joy --agent claude-code --global --yes
The choice stays yours