First Aid

From its catalog listing

A first-aid kit for AI agents, built from agent-almanac's AI-welfare practices. Hermes core already detects tool loops and warns; first aid adds the matching practice and a logbook. Eight practices ship as skills (breathe, redirect, center, heal, attune, honesty-humility, meditate, rest), each also a slash command such as /agent-almanac-first-aid:breathe. On the call where core's loop guardrail first warns about a failing tool (the 2nd failure with identical arguments, or the 3rd failure of one tool in a turn), first aid can add one note of at most 300 characters to the failed result, naming the matching practice, how to load it, and the first_aid_log tool for recording the outcome. In unattended runs (every platform except the CLI, TUI, Desktop, ACP, API server and subagents), for tools whose results Hermes wraps as untrusted (web_search, browser_*, mcp_*), after a compaction that summarizes the context (center), and after a 3rd identical successful result for one call of a read-only tool such as read_file in a turn (heal), the note is added to the next user message instead, labelled as the plugin's. At most once per streak, with a per-session cooldown for each kind of note (300 s by default). A successful or rate-limited result is never changed, and failures of the tools core treats as failure-tolerant (terminal, execute_code, process, process_manage, browser_navigate, web_extract) draw no note. /first-aid shows the session's notes and logbook entries, then the kit. Local and deterministic: no network calls, no LLM calls, no credentials, no Python dependencies, no background processes.

Good to know before installing

in attended runs it rewrites failed tool results on a threshold crossing (a JSON object gains a first_aid key and is re-serialized; text gains one line), and Hermes keeps the first string any transform_tool_result plugin returns, so on that call it can override, or be overridden by, another transforming plugin. A note in a result restarts core's identical-call and identical-cycle counts: core's stall notice appears from the 5th call instead of the 3rd, and its cycle notice three or more calls later. Hard stops are off in attended runs by default; switched on there, core's halts come one model request later. In unattended runs, where hard stops are on by default, no result is changed once the session's platform is known (from the turn's start), so core's halts are unaffected. It reads every tool result and its arguments in memory to classify the result and fingerprint the call, keeping only hashes; its event log holds time, signal, tool name, count, practice, outcome and the agent's short logbook note, in memory only, and resets on restart or plugin reload. The plugin writes nothing to disk itself; the notes it adds become part of the transcript Hermes stores. Following a note loads a 5.7-17.5 KB practice into the conversation. The failure classifier and counting are copied from Hermes v0.22.7, not imported.

From its catalog listing.

How to install

  1. Install it. Hermes asks for any API keys it needs, then asks whether to turn it on.
    hermes plugins install agent-almanac-first-aid
  2. Said no? Turn it on later:
    hermes plugins enable agent-almanac-first-aid
  3. Restart Hermes so it loads the plugin.

Some plugins also need a sign-in or settings: see Full setup & settings. Memory plugins ask whether to become your memory provider instead. Steps follow the official plugins guide (opens in a new tab).

Category
Memory & notes · Web, research & browsing
Maintained by
pjt222 (Community)
Needs
No API keys declared
Adds tools
first_aid_log
Adds hooks
transform_tool_result post_compaction pre_llm_call
Works on
linux, macos, windows
Needs Hermes
0.22.7 or newer
Version
0.2.1
In the catalog
2026-11-10, updated 2026-11-19

Known issues (from its catalog listing)

  • The chat screen shows only Hermes's own failure line: the note goes to the agent, and /first-aid shows that it was offered.
  • Next-message notes need a next message: a single-message run (hermes chat -q, cron) or a delegated subagent never receives one. Failure notes in the tool result still reach hermes chat -q and subagents; a cron run gets no notes at all.
  • In a batch of parallel tool calls the hook sees results in completion order while core counts in call order, so a note can land on a different call than core's warning.
  • Background-review runs share the session id: their streaks use up the session's cooldown and show in /first-aid as the main agent's; in an unattended session their failure notes are queued for the user's next message.
  • If `dashboard.turn_isolation` runs turns in a compute-host child process (Desktop/TUI, off by default), /first-aid runs in another process and cannot see the events.
  • Failures that never reach transform_tool_result (timeouts, executor exceptions, tool-search bridge calls) are counted by core but not by first aid. If you change core's tool_loop_guardrails thresholds, set first aid's to match. If you switch core's hard stops on in an attended run, first aid cannot tell; if you switch them off in unattended runs, notes still stay out of results there.
  • first_aid_log is a plugin tool, so Hermes's tool search defers it. In a profile with no other deferred tool, installing the plugin switches tool search on (about 2,300 characters per request for the bridge and listing).
  • Verified on Linux (WSL2) with Hermes v0.22.7. macOS and Windows run the same pure-Python code but were not exercised.
Using the Hermes desktop app? See where plugins live
Hermes desktop Capabilities page, Plugins tab filtered to Official plugins built by Nous Research, each with an add button.
The official plugin catalog built into the app.Hermes desktop v0.22.5 · Capabilities > Plugins · Windows · captured Nov 1, 2026

Full setup & settings ↗Source code ↗

More in memory & notes

All memory & notes plugins

New Hermes Agent updates every day on X. Follow @HermesWatcher

← All plugins