Re Frame Pair — Development skill for Claude Code

re-frame-pair

Lets an AI (Claude, Codex?) debug and develop against your **running re-frame application**.

Claude can inspect `app-db`, dispatch events, trace the 6 dominoes, hot-swap handlers, time-travel, and map visible UI back to source code — all live against the runtime, with no source edits required for probes.

A coding agent working with just static code is working with a limited perspective. It can read handlers, subscriptions, and views, but it has to guess what happened in the browser at runtime. With re-frame-pair, Claude can ask the running app: which event fired, what changed in `app-db`, which effects fired, which subscriptions re-ran or cache-hit, which views rendered, and where the relevant source lives.

The intended loop is empirical:

1. Observe the current runtime state.
2. Inspect the relevant epoch.
3. Form a hypothesis.
4. Undo state, probe with a new dispatch, read app-db, hot-swap handlers, side-effect stub, or REPL eval.
5. Compare the new epoch with the baseline.
6. Only then edit source.

REPL changes are **ephemeral** (try-and-discard); source edits **stick**.

Post-mortems are a core use case. Instead of asking the developer to describe every click and state transition that led to a bug, Claude can inspect the recent epoch trail and identify the event that introduced the bad value, including its parent dispatch, effects, `app-db` diff, subscriptions, renders, and source call sites.

Examples

**You**: What changes did the last event make to `app-db` under `:shopping-cart`?

**Claude**: 2 items, subtotal 47.50, no coupon. `{:items [{:sku "BK-42" :qty 2} {:sku "TS-9" :qty 1}] :total 47.50 :coupon nil}`

---

**You**: I dispatched `[:profile/save ...]` but the name in the header isn't changing.

**Claude**: The header reads `[:user/display-name]`, derived from `[:user/profile]`. Your handler wrote to `[:user :pending-profile]`, not `[:user :profile]`. Layer 2 returned `=`, the Layer 3 short-circuited, the header didn'