Fledge

From its catalog listing

Skills have to earn their wings. Before Hermes saves or rewrites a skill, Fledge checks what actually happened in the run it learned from (tools still failing when the turn ended, a failed or interrupted turn, the user correcting the agent in the next message, /undo or a /rollback restore right after, denied approvals), using only facts Hermes reports through hooks: no model grades the model. A save learned from a bad run is held in Hermes' own skills pending queue with the reasons (review from any chat with /fledge pending|diff|approve|reject, or /skills pending in the CLI); a foreground save after errors goes through Hermes' approval prompt, and unattended runs queue instead of waiting. New and rewritten skills start on probation: clean real runs graduate them, repeated bad runs ground them, and a save from a run the user corrected right after is grounded at once; a grounded skill Hermes created on its own is archived (hermes curator restore brings it back), skills the user asked for are only reported. Skills the user edited by hand are fingerprinted, and autonomous rewrites or deletes of them become proposals in the same queue (optionally Hermes' content lock). Modes: enforce, watch (records only), off.

Good to know before installing

nine hooks (pre_llm_call, pre_tool_call, post_tool_call, on_session_end, on_session_finalize, on_session_reset, on_skill_lifecycle, pre_command, post_approval_response); pre_tool_call blocks or asks for approval on skill_manage only and lets the write through on its own errors (a write already queued for review stays held); reads hook payloads (tool statuses and error text, the first 280 characters of the next user message) and skill folders to fingerprint them; writes its own SQLite under each profile's plugin-data/fledge (skill names, fingerprints, verdict reasons, session/turn ids; reasons quote the matched correction phrase and up to 80 characters of a tool error, hold summaries the skill's one-line description; no other message text, no skill bodies), held writes into pending/skills/, and archives grounded agent-created skills into skills/.archive/; calls tools.skill_provenance, tools.write_approval and tools.skill_usage (the functions behind /skills pending and hermes curator archive/lock) because no public seam exists yet, each guarded; no network, no model calls, no credentials, no shell commands, never approves anything. Needs plugins.isolation in_process to tell background-review writes from foreground ones. Measured on 2026-11-10: 153 tests passed on Python 3.14 and 3.11; `hermes plugins validate` passes on Hermes 0.23.1, 0.23.0 and 0.22.7 with only the private-import warnings above; live sessions with a real model in isolated HERMES_HOMEs (bad-run hold, approve onto probation, grounding after a correction, hand-edit proposal and approval prompt, watch mode, 13 clean writes with 0 holds, fail-open under a locked database) in VALIDATION.md, raw output in tests/smoke-2026-11-05.txt and tests/profiles-2026-11-10.txt.

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 fledge
  2. Said no? Turn it on later:
    hermes plugins enable fledge
  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
Security & privacy
Maintained by
12burb (Community)
Needs
No API keys declared
Adds hooks
on_session_end on_session_finalize on_session_reset on_skill_lifecycle post_approval_response post_tool_call pre_command pre_llm_call pre_tool_call
Needs Hermes
0.22.7 or newer
Version
0.1.0
In the catalog
2026-11-10
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 security & privacy

All security & privacy plugins

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

← All plugins