remote-hosts

Imported from AlaisterLeung/hermes-remote-tools (skills/remote-hosts/SKILL.md). Install upstream with npx skills add AlaisterLeung/hermes-remote-tools --skill remote-hosts. Copyright stays with the author.

Remote hosts via hermes-remote-tools

When to use

Use this plugin's tools for anything that must run or read on a host other than the one Hermes runs on. The builtin terminal, read_file, write_file, patch and execute_code are host-local: a remote path given to them reads the wrong machine's filesystem or dies with a not-found error. The local machine is served by the builtins; every other configured host goes through the six remote tools.

The six tools:

• terminal_remote - shell on a configured remote host, with the same approval stack as the builtin terminal. background=true registers the process for notifications and standard process management.
• read_remote_file / search_remote_files - reads and regex/glob searches.
• write_remote_file / patch_remote_file - mutations, snapshotted before the edit (undoable) and verified by a read-back hash.
• undo_remote_file - restores the previous content of a file, one step at a time.

Call shape

The tools are local deferred tools behind the tool_search bridge, so every parameter sits INSIDE the inner call's arguments object:

{"calls": [{"name": "terminal_remote", "arguments": {"target": "devbox", "command": "uname -a"}}]}

Both failure shapes are verified: omitting target from arguments, and placing target as a sibling of arguments. One entry per tool_call: batching two invocations is rejected outright, so a batched call returns nothing and wastes a round-trip. On a schema refusal, tool_describe the tool once, verify the shape, retry once, then report the blocker instead of looping.

Configuration

Targets live in config.yaml under the plugin's settings:

plugins:
  entries:
    hermes-remote-tools:
      settings:
        targets:
          devbox:
            host: 203.0.113.10
            user: alaister
            port: 22
            key: ~/.ssh/devbox
            cwd: "~"

A machine reachable on more than one network takes a list of addresses in host, tried in order, the first that connects wins; the pooled session is probed before every reuse, so a machine that moved networks re-dials automatically. Connection-relevant fields are fingerprinted, so a config edit rotates the pooled session on next use. Plugin code loads once per process: after a config edit restart the gateway (and the dashboard when desktop sessions use the plugin), then prove reachability with ONE live call. A green hermes plugins doctor hermes-remote-tools proves registration, not reachability.

Execution semantics

• Remote commands run under bash -c even when the remote login shell is fish, csh or tcsh: the plugin probes $SHELL once per session and, on a non-POSIX login shell, base64-wraps the script through the command line so the login shell never parses it. Write POSIX/bash chains inside the script.
• cd inline (cd ~/path && ...), not through the workdir parameter: a failing cd short-circuits the chain, so nothing runs in the wrong directory.
• Results carry failed: true only on a nonzero exit; success omits the key.
• Keep timeout well under the background-promotion cap: an over-cap timeout promotes the command to a tracked background process.
• Never put secrets in a command string: the approval layer redacts credential-shaped substrings. Write any credential-touching script with write_remote_file and execute the written file.

Verification discipline

• After every remote write, prove the bytes on the target: ls -la <path> plus sha256sum through terminal_remote. A write can report success and a matching read-back hash while landing on the wrong host or nowhere at all.
• terminal_remote is the ground truth for remote file state; cross-check any read_remote_file result against it before building on the read.
• Never use scp. Ship bytes through write_remote_file (new or whole-file rewrite, absolute path, never ~) or the base64 heredoc through terminal_remote, which no shell on any hop parses. For a whole-file rewrite of an existing file, write <path>.new, sha256sum and diff -u it against the target, then mv only after the hash matches. diff exits 1 on differences, so failed: true on a successful diff is expected.
• Build patch payloads as a local file first with anchor assertions, then ship them: hand-inlining multi-line replacement text into a tool_call body trips the control-character schema check.

Pitfall: never leave shell tracing on inside a remote command

set -x (or bash -x) in the command body poisons the pooled session's working directory. The env's per-command wrapper emits its CWD marker with printf '<marker>%s<marker>' "$(pwd -P)", xtrace echoes that printf line, and the marker parser picks the traced FORMAT STRING (the %s sitting between two marker occurrences), so self.cwd becomes the literal %s. Every later call on that target then dies before the command runs: bash: line 3: cd: %s: No such file or directory with exit 126 from the wrapper's own cd ... || exit 126, and the file tools answer working directory '%s' does not exist on the active terminal backend. The real marker line survives in the output unstripped, which is the tell. A long set -x command also returns visibly mangled output for the same reason (the strip span eats the traced line and glues its neighbours).

Recovery needs no restart and no config edit: pass an explicit absolute workdir on the next call. It overrides the stored cwd (effective_cwd = cwd or self.cwd), so the command runs and the wrapper's trailing marker re-parses to the real path. Verified 2026-10-18 on ljair (macOS, git 2.54): poisoning reproduced on demand, one-call recovery after it.

Trace a nested shell instead (bash -x ./script.sh), and print debug lines plainly rather than turning on set -x in the remote command body.

Classifying failures

Two layers, do not conflate: a wrapper refusal (missing required argument(s): target, or a batched call) is call-shape and fires before any shell sees the command; mangled output or a nonzero exit is execution (shell semantics or quoting). Classify before retrying anything.