sub-claude

sub-claude

A session-oriented pool of persistent Claude TUI slots backed by tmux. Solves a fundamental problem: launching a new claude process while another Claude session has a Bash tool call in flight silently destroys that call's output. sub-claude pre-starts sessions and reuses them — start, followup, and wait never launch new claude processes.

Not claude-spawn: pool sessions are persistent TUIs (not claude -p), support multi-turn conversation via followup, and never launch new claude processes.

Output parsing: Use -v / --verbosity to filter output. Levels: response (final model answer only), conversation (all prompts + responses), full (everything minus TUI chrome), raw (default, unfiltered). Example: sub-claude -v response wait "$id"

Quick Start

# Fire-and-forget
id=$(sub-claude start "refactor auth module")

# Blocking — wait for result (parsed to just the response)
result=$(sub-claude -v response start "summarize this file" --block)

# Multi-turn
id=$(sub-claude start "analyze the codebase")
sub-claude wait "$id"
sub-claude followup "$id" "now suggest improvements" --block

# See what a running session is doing
sub-claude capture "$id"

# Parallel work
id1=$(sub-claude start "task one")
id2=$(sub-claude start "task two")
sub-claude wait "$id1"
sub-claude wait "$id2"

# Cross-project: manage a pool in another directory
sub-claude -C ~/obsidian pool init --size 3
sub-claude -C ~/obsidian start "organize notes"
sub-claude -C ~/obsidian pool status

CLI

Run sub-claude --help for full CLI reference.

Output Contract

Command	stdout	stderr	blocks?
start	job ID	—	no
start --block	terminal output	job ID + warnings	yes (unless queue pressure)
followup	job ID	—	no
followup --block	terminal output	job ID + warnings	yes (unless queue pressure)
capture	live terminal or snapshot	offload note (if applicable)	no
result	terminal output	—	no (errors if running)
wait <id>	terminal output	warnings	yes
wait (no ID)	result of first child to finish	--- <id> --- header	yes

Session Lifecycle

queued → processing → finished(idle) → finished(offloaded)
             │
             └→ error (slot crashed)

• queued — waiting for a slot (all slots busy)
• processing — Claude is working on it in a slot
• finished(idle) — done, session still loaded in a slot, ready for followup
• finished(offloaded) — slot was reclaimed; terminal snapshot stored, session can be resumed

Multi-Turn Conversations

followup resumes a finished session — conversation history is preserved.

• Only works on finished sessions. Errors if still processing (use wait first).
• If the session was offloaded, a slot is allocated and resumed automatically.
• Reuses the existing job ID (same session, same conversation).

Blocking vs Non-Blocking

Non-blocking (default): start prints the ID and returns immediately. Use wait or capture to observe progress.

--block: waits inline, prints terminal output to stdout. ID goes to stderr. Under queue pressure, --block degrades to non-blocking — see pool-management.md.

Best Practices

Prefer fire-and-forget + explicit wait over --block for parallel work.

Use followup for multi-turn, not input. input is for raw terminal interaction (menus, interactive prompts).

Clean completed sessions to free slots: sub-claude clean --completed

Auto-init: if no pool exists when you call start, it initializes one automatically (5 slots). Prefer explicit pool init --size N with a larger size before starting work.

Always overprovision. Sub-agents, other Claude instances, and hooks all consume slots. A pool that looks big enough at launch saturates quickly once nested work kicks in.

Sub-Skills

Situation	File
Interactive key sequences, pinning, multi-instance isolation	interactive-sessions.md
Pool setup, resizing, queue pressure, environment variables	pool-management.md
Reusable custom agents via command scripts	custom-agents.md
sub-claude not found, install/update/uninstall	installation.md

Feedback & Issues

File bugs and feature requests at github.com/EliasSchlie/sub-claude/issues.