hauntty: Stateful, observable shell sessions for LLM agents

Read full text
Authors: Stefano Schotten
Cited by: 0

Abstract

Agent-ergonomics release, driven by field notes from a full day of real infra ops through hauntty (docs/field-notes-2026-08-06.md). Added Session names — every session gets a memorable, daemon-generated adjective-noun name (spectral-willow), or an explicit one via connect --name infra / spawn --name build. Stored daemon-side (no client bookkeeping to go stale), unique per daemon, shown by list, printed as HAUNTTY_NAME=. Address sessions by name or host — -s accepts a SID, a name, or a host (-s ai02 → that daemon's primary session). Tiered resolution SID > name > host against live daemons; ambiguous refs error with candidates; empty -s falls back to $HAUNTTY_SID. exec --json — one machine-readable result object (sid, seq, state, rc, stdout_lines, stderr_lines, cwd, background_pids, elapsed_s, prompts, stdout inlined when ≤100 lines). exec --script — run a local script's contents in the session with zero shell-quoting surface. Changed Captured output is clean: the status ticker goes to stderr only when stderr is a TTY; the ^M^[[K erase-sequence corruption is gone; multi-line commands echo as first line + … (+N more lines). Faster completion for long commands: the rc-file gate polls every 100 ms decoupled from 1 s /proc sampling — done is reported in ~100 ms after the fast phase, any exit code. exec default deadline is now 10m (was 5m). Fixed uninstall no longer risks killing its own SSH session (bracketed remote pkill pattern). Full changelog: https://github.com/sch0tten/hauntty/blob/main/CHANGELOG.md Backward compatible: v0.5.0 daemons keep working with the new client (blank name column, SID addressing unchanged). sha256: 86cdf903214bc0a1019ce9cdc5c3c559acdfe4d92622ac9121544374e5e46040 hauntty-linux-amd64 02a9cf9a9c94a188237038349b3ee5643b775c5230b9c71186e800e3d495d5fb hauntty-linux-arm64

Fields of Study & Subjects

Software System Performance and Reliability Cloud Computing and Remote Desktop Technologies Scientific Computing and Data Management

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
11 March 2026
Language
EN