CIDER-style nREPL: jack-in, per-project connections, completion, eldoc, doc, M-.

[nooga]

CIDER-style Clojure/let-go tooling for let-go-mode: an nREPL client, one-command jack-in, completion, eldoc, doc lookup and goto-definition. Each buffer routes all of it to either the editor's own VM or its project's nREPL server.

What you get

Key	What
C-c M-j	Jack in: find the project root, start its nREPL server on a free port, connect. Stopped again by C-c C-q or editor exit
C-c M-c / C-c C-q	connect to a running server (pre-filled from .nrepl-port) / disconnect
C-x C-e C-j C-c C-v C-c C-e	unchanged keys; they evaluate on the server when the buffer's project is connected (C-c C-e = load-file)
TAB / M-TAB	completion. TAB completes only when indenting would change nothing and point ends a symbol
(automatic)	eldoc: ns/name: [args] in the echo area while typing a call
C-c C-d d	arglists, doc and source location in *Help*
M-. / M-,	goto definition (files, open buffers, or source inside a Clojure jar, read-only) and back

Jack-in launch configs are data (register-launch-config!):

Project	Marker	Server
lgx	lgx.edn	lgx nrepl --port N
Clojure CLI	deps.edn	clojure -Sdeps '{nrepl + cider-nrepl}' -M -m nrepl.cmdline …
Babashka	bb.edn	bb nrepl-server N
let-go	deps.edn, or none	lg -n -p N

When the markers are ambiguous (deps.edn for both CLI and lg, bb.edn next to deps.edn), the buffer's extension and a priority decide, and a real tie prompts. M-x nrepl-jack-in-command edits the command first, and M-x nrepl-server-log shows the server output.

Connections are per project root. File buffers under a connected root use that server. Fileless buffers (*scratch*, *repl*) always stay in-process, so editor scripting keeps working next to two projects on two servers. The lighter shows nrepl:clj / :bb / :lg / :lgx.

How it's built

• legmacs.modes.letgo: every eval command builds one backend request. eval-backend-for picks the backend per buffer through registered resolvers. With :in-process, C-x C-e/C-j/C-c C-e stay synchronous, exactly as before; any other backend runs them through spawn-task. Remote results come back as printed-result. C-j and C-c C-v check that the form is still in the same buffer before writing. New after-command-hooks let eldoc run after paren-matching.
• legmacs.nrepl: bencode over net, one blocking round trip per request, serialized by a channel-token lock inside the job's future. There's no long-lived reader goroutine for loop-wait to keep polling. Connections are keyed by project root; ops are discovered with describe, preferring cider-nrepl's complete/info/eldoc when present. Server errors are cleaned to one line (message plus caused by: chain), and an eval in a namespace the server hasn't loaded gets a note, because let-go silently falls back to user.
• legmacs.jack-in: let-go interop can't reach cmd.Process, so servers run under sh -c. The script writes a pidfile, logs to a file (an unread pipe would block the server) and execs the command. The editor holds its stdin open (lg -n exits on EOF), and stopping it is SIGTERM plus closing stdin. Liveness uses ps -o stat=, because kill -0 reports a zombie as alive. The editor picks the port, since bb doesn't write .nrepl-port for port 0. New bufs/add-exit-hook! stops everything after the terminal is restored.
• legmacs.modes.letgo-complete / letgo-lookup: completers and lookups registries per backend. In-process answers come from ns-map/ns-publics/aliases (including the ns form's :requires) and var metadata. eldoc scans only back to the top-level form and caches remote answers for 15s. A new :visit buffer-command and a shared :xref-stack implement M-./M-,.

Also in here

• test/run.lg ran 0 tests. let-go's run-tests now defaults to *ns*, so the suite had been reporting success without running anything, in CI too. It now uses run-all-tests and runs 401 tests. CLAUDE.md's single-namespace command is fixed the same way.
• The let-go pin moves to main (36d57e3), because the old pin predates the net/bencode namespaces. build/go.mod/go.sum are tidied against it. CI's test step puts lg on PATH, since the jack-in tests start a real lg -n server.

What was tested

• The suite: 401 tests, 0 failures, on let-go 36d57e3 (the new pin). The new tests use a fake bencode server on a loopback socket, a real lg -n jack-in, and a server that dies at startup.
• Headless end-to-end against real servers:
  • lg, Babashka and Clojure CLI (+cider-nrepl): jack-in, load, eval, completion, disconnect (server gone afterwards).
  • Clojure only: eldoc, doc, and M-. into clojure/core.clj inside the Clojure jar.
• The real TUI under a pty, from source and as the native binary: jack-in, eval, M-TAB, M-./M-,, doc, and no server left running after C-x C-c.
• build.sh plus the plan9/amd64 and plan9/arm cross-builds against the new pin. The generated Go has no closure-hoisting pattern (gogen: a capturing closure inside a conditional is emitted as the function's return value, discarding the branch let-go#766).
• lgx is untested: it isn't installed here.

Known gaps, from let-go

Filed together as nooga/let-go#961:

• AOT builds drop def metadata (plus an option to elide it on purpose), so the native binary's in-process eldoc/doc/M-. know names only.
• Core vars have no arglists or docs.
• let-go's nREPL info returns name/ns only, and it has no eldoc, streamed output, headless mode, plain-text errors, or ns in completions.
• No subprocess spawn with real pipes, pid and kill.
• No symbol/end positions or structured error positions from the reader and compiler.

Not done yet from CIDER: test runner, stacktrace view, interrupt (nooga/let-go#589), macroexpand, and a REPL buffer tied to the connection.

🤖 Generated with Claude Code