jared/cc-os

Watch

Fork 0

lsp-probe keeps one warm ruby-lsp daemon per worktree root, with warm and stop commands #600

New issue

Closed

opened 2026-09-16 19:52:56 +00:00 by jared · 2 comments

jared commented

2026-09-16 19:52:56 +00:00

Owner

Context

Branch lsp-probe-cli (ADR-0197) added bin/lsp-probe, a Thor CLI that spawns ruby-lsp over stdio per call (lib/os_sdlc/lsp_probe/query_runner.rb:21, Open3.popen3). Measured 2026-09-16: one lsp-probe call 4-9 s; bin/evidence collect --root plugins/os-sdlc/lib Model::Db 209 s because lib/os_sdlc/evidence/fingerprint.rb:83-91 runs one LSP query, and so one spawn, per method with no ast-grep sender. Daemon lifecycle management reduces per-call spawn overhead.

Owner sketch

class Daemon
  IDLE_TIMEOUT = 30 * 60
  def self.connect(root:) = new(root: root).connect
  def connect = alive? ? attach : spawn_and_attach
  def stop = ...
  private def alive? = pid_file.exist? && Process.kill(0, pid)
end

Cases

Daemon.connect on a root with no .cc-os/lsp/ spawns once and writes sock and pid files (spawner double).
Second Daemon.connect on the same root spawns nothing and attaches to the existing socket.
Stale pid file whose process is dead: connect removes the files and spawns fresh.
lsp-probe stop --root X kills the pid and removes the two files; running it twice is a no-op with exit 0.
lsp-probe warm --root X returns within 1 s and a following lsp-probe defs call completes without a spawn.
Idle timer: a daemon with no request for IDLE_TIMEOUT exits (clock injected).

Origin

Trigger: evidence performance audit 2026-09-16
Improvised this session: none
Chain: per-call spawn overhead ← QueryRunner always popen3 ← no daemon persistence
Root candidate: n/a
Where: lib/os_sdlc/lsp_probe/daemon.rb
Session: 01153d94-6459-43dc-bd20-e68b757af56f
Transcript: https://claude.com/claude-code

## Context Branch lsp-probe-cli (ADR-0197) added bin/lsp-probe, a Thor CLI that spawns ruby-lsp over stdio per call (lib/os_sdlc/lsp_probe/query_runner.rb:21, Open3.popen3). Measured 2026-09-16: one lsp-probe call 4-9 s; `bin/evidence collect --root plugins/os-sdlc/lib Model::Db` 209 s because lib/os_sdlc/evidence/fingerprint.rb:83-91 runs one LSP query, and so one spawn, per method with no ast-grep sender. Daemon lifecycle management reduces per-call spawn overhead. ## Owner sketch ```ruby class Daemon IDLE_TIMEOUT = 30 * 60 def self.connect(root:) = new(root: root).connect def connect = alive? ? attach : spawn_and_attach def stop = ... private def alive? = pid_file.exist? && Process.kill(0, pid) end ``` ## Cases - `Daemon.connect` on a root with no `.cc-os/lsp/` spawns once and writes sock and pid files (spawner double). - Second `Daemon.connect` on the same root spawns nothing and attaches to the existing socket. - Stale pid file whose process is dead: `connect` removes the files and spawns fresh. - `lsp-probe stop --root X` kills the pid and removes the two files; running it twice is a no-op with exit 0. - `lsp-probe warm --root X` returns within 1 s and a following `lsp-probe defs` call completes without a spawn. - Idle timer: a daemon with no request for IDLE_TIMEOUT exits (clock injected). ## Origin - Trigger: evidence performance audit 2026-09-16 - Improvised this session: none - Chain: per-call spawn overhead ← QueryRunner always popen3 ← no daemon persistence - Root candidate: n/a - Where: lib/os_sdlc/lsp_probe/daemon.rb - Session: 01153d94-6459-43dc-bd20-e68b757af56f - Transcript: https://claude.com/claude-code

jared added the

plugin/os-sdlc

filed-by/agent

labels

2026-09-16 19:52:56 +00:00

jared referenced this issue

2026-09-16 19:53:02 +00:00

worktree start warms lsp-probe daemon, worktree finish stops it #601

jared referenced this issue

2026-09-16 20:46:24 +00:00

evidence run shares one ruby-lsp session instead of one spawn per gapped method #599

jared commented

2026-09-16 20:47:50 +00:00

Author

Owner

Work started via os-sdlc pipeline on branch ticket-600 (worktree off lsp-probe-cli, after #599 cab8b5ea). ADR-0198 recorded. Session 01153d94.

Work started via os-sdlc pipeline on branch ticket-600 (worktree off lsp-probe-cli, after #599 cab8b5ea). ADR-0198 recorded. Session 01153d94.

jared referenced this issue

2026-09-16 21:27:38 +00:00

pipeline passes tickets whose cases are all satisfiable by doubles; production wiring never lands #604

jared commented

2026-09-17 06:16:41 +00:00

Author

Owner

Resolution

Done: LspProbe::Daemon with Paths/Spawner/Server/Query/Client/Connection/Idle/Stopper; Server routes defs/refs/symbols/outline through Daemon.connect; CLI gains warm, stop, hidden serve; Fingerprint's SharedRunner uses the daemon; QueryRunner deleted. Commit 4a348c3f on ticket-600, merges to lsp-probe-cli. ADR-0198 amended (hashed path, spawn not fork, abandoned-socket exit, TERM/KILL stop).

Evidence: Suite 1295 green, rubocop clean. Measured: symbols 4.0 s cold -> 0.26 s warm; evidence collect Model::Db 10.5 s -> 5.2 s warm; after stop no ruby-lsp process and hashed dir empty. Pipeline impl 24: 15 dispatches, ended unexpected_pass escalation; pipeline output was a placeholder daemon (fork+sleep, empty sock file, connect returned nil, QueryRunner untouched); real daemon hand-finished. See #604.

Follow-ups: Kill the daemon process group (Spawner already sets pgroup: true; Stopper should signal -pid) so a hung ruby-lsp child cannot outlive lsp-probe serve: filed as a line here, ticket after user review. Daemon exit when the root tree disappears, not only the socket: dropped, abandoned-socket check covers the observed case. lsp-probe stop --all reaper for crashed daemons: dropped until seen. SessionEnd hook to stop daemons: covered by #601 worktree finish for pipeline roots; main-checkout daemons rely on idle exit.

## Resolution **Done:** LspProbe::Daemon with Paths/Spawner/Server/Query/Client/Connection/Idle/Stopper; Server routes defs/refs/symbols/outline through Daemon.connect; CLI gains warm, stop, hidden serve; Fingerprint's SharedRunner uses the daemon; QueryRunner deleted. Commit 4a348c3f on ticket-600, merges to lsp-probe-cli. ADR-0198 amended (hashed path, spawn not fork, abandoned-socket exit, TERM/KILL stop). **Evidence:** Suite 1295 green, rubocop clean. Measured: symbols 4.0 s cold -> 0.26 s warm; evidence collect Model::Db 10.5 s -> 5.2 s warm; after stop no ruby-lsp process and hashed dir empty. Pipeline impl 24: 15 dispatches, ended unexpected_pass escalation; pipeline output was a placeholder daemon (fork+sleep, empty sock file, connect returned nil, QueryRunner untouched); real daemon hand-finished. See #604. **Follow-ups:** Kill the daemon process group (Spawner already sets pgroup: true; Stopper should signal -pid) so a hung ruby-lsp child cannot outlive lsp-probe serve: filed as a line here, ticket after user review. Daemon exit when the root tree disappears, not only the socket: dropped, abandoned-socket check covers the observed case. lsp-probe stop --all reaper for crashed daemons: dropped until seen. SessionEnd hook to stop daemons: covered by #601 worktree finish for pipeline roots; main-checkout daemons rely on idle exit.

jared closed this issue

2026-09-17 06:16:41 +00:00

~~jared referenced this issue 2026-09-17 14:39:57 +00:00~~

os-sdlc owns the lsp-probe daemon lifecycle per implementation; worktree plugin stops managing it #606

jared referenced this issue

2026-09-17 14:46:51 +00:00

os-sdlc owns the lsp-probe daemon lifecycle per implementation; worktree plugin stops managing it #606

jared referenced this issue from a commit

2026-09-18 16:23:53 +00:00

feat(os-sdlc): lsp-probe warm ruby-lsp daemon per root over a UNIX socket; warm/stop/serve; evidence runs reuse it (#600, ADR-0198)

jared referenced this issue from a commit

2026-09-18 16:23:53 +00:00