hermes-official-plugins
☤ The official plugins for Hermes Agent, built and maintained by Nous Research: Home Assistant, Spotify, Claude Subscription DirectSDK and more, one install away from the Hermes plugin catalog.
Checkpoint, compact, rehydrate on your terms. Every number measured.
Translation pending
PluginsGitHub @ryanonline1234Availability not checked yetOwner unconfirmed
Have you tried it? One tap is enough. No sign-in needed.
Repository information has not been collected yet.
No updates detected yet. GitHub releases will appear here.
Claude Code plugin: checkpoint the session, compact automatically, rehydrate on your terms with a measured estimate.
Loading the full README.
Read the full README ↗Enforcement for Claude Code's compaction: what you decided and told it to keep survives, checked rather than hoped for.
Site: smart-compact.vercel.app
When Claude Code compacts, one summary decides what survives. Measured over the author's sessions, what got lost was standing rules ("never deploy without asking") and pending decision lists, the numbered questions a later "go with recs" answers. smart-compact keeps both, and it checks that they were kept:
/smart-compact:save writes a checkpoint before compacting. The guard refuses a new checkpoint that silently drops a rule or open decision the previous one held. At a compaction you didn't start, a fork of the session writes the checkpoint alongside the summary, and the plugin checks and corrects it like a save./smart-compact:rehydrate would put back and what it costs, minus the files Claude Code already re-attached./smart-compact:save checkpoint to disk; compaction queued for the end of the turn
(automatic) the plugin runs /compact with a pointer to the checkpoint
/smart-compact:rehydrate re-read the checkpoint and its manifest, up to a budget
After the compaction, the plugin's notice reads like this:
smart-compact: compacted (91k → 28k tokens); the system prompt and tools come on top. Checkpoint: ~/.claude/checkpoints/…/20261103T233304Z-bug-hunt.ckpt.md
Rehydrate would add ≈14k tokens, measured at save, budget 150k:
files: 2 ≈14k measured — docs/STATE.md 7.6k (why it is listed); src/auth.ts 6.7k (why it is listed)
checkpoint itself: already back
already back after compaction, so skipped: src/decoder.py
Send /smart-compact:rehydrate to rebuild context from it, or keep working from the summary.
Requires Claude Code 2.1.309 or later with mods (function hooks) enabled, the default from 2.1.287. Tested end to end on 2.1.310 in stream-json mode, the mode the desktop app runs sessions in.
claude plugin marketplace add ryanonline1234/smart-compact
claude plugin install smart-compact@smart-compact
Inside a session: /plugin marketplace add ryanonline1234/smart-compact, then /plugin install smart-compact@smart-compact, then /reload-plugins.
| Command | What it does |
|---|---|
/smart-compact:save [budget] | Writes a checkpoint, then compacts when the turn ends. budget caps what a later rehydrate re-reads; the default is 150k. |
/smart-compact:rehydrate [path] [budget] | Reads the checkpoint and its manifest, skipping anything already in context, and reports what it re-read, what it skipped, and why. You start it, or the plugin does after a compaction when you turn on autoRehydrate. |
Checkpoints are plain markdown in ~/.claude/checkpoints/<project-slug>/, keyed by the project's git root rather than the shell's directory. The newest ten are kept; older ones move to archive/.
Built for sessions that run for days: many rounds of feedback, several compactions, decisions answered with a short "go with recs". Over 229 of the author's interactive sessions, agents reading the transcripts judged what was lost after compactions: standing rules and pending asks, in 7 of 35 compactions, 5 of them automatic. A scripted re-measurement (docs/05-session-study.md) reproduces the population and finds that 78% of "go with recs" replies answer a numbered list. The loss count does not hold up: two automated re-judges found 1 and 2 losses in 47 compactions, but missed at least one clear loss the agents had found. The rate in these sessions is uncertain, somewhere between about 4% and 20% (details in docs/05). It is one person's sessions, not a general result.
This is the beginning of the README. Read the rest at the original. Read the full README ↗