Portable Agent Plugins v1 package from glasser-ai/hermes-plugin-glasser:
a stdio bridge through mcp-remote to Glasser's hosted MCP (browser
sign-in, no key in the package) plus a bundled setup and usage skill.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
TROVE now self-heals the multi-process WAL index-drift fault at open
(REINDEX in place, stale -shm cleanup) instead of silently corrupting.
See misterdas/hermes-trove@f0aa40a.
Desktop-only plugin that renders PriceWin hotel and flight results as
inline transcript cards. Registers two transcript directives and nothing
else: no tools, hooks, middleware or env. Pinned at v0.1.1.
Third commit on this branch. 8aea272 is a technical-writing review of
README and all three docs checked line by line against the source: the
Hermes action table listed actions vault_tool rejects and omitted one it
supports, five detectors were graded against the wrong severity, the cron
snippet pointed at a script the package does not ship, scan_text findings
were described as logged when the log line carries only counts and
detector names, and the sub-key tree missed two v0.7.0 sub-keys.
Two over-claims in code text are fixed in the same commit, both from the
#118456 review: the tool description said read actions never expose
plaintext while listing consume_token as a read action, and the
fingerprint docstring called an unsalted SHA-256 "safe to log".
387 passed (text-only change), ruff clean, hermes plugins validate 13/13
from a fresh clone at 8aea272, structural validator OK, catalog lane
green, all five sha-pinned URLs 200. v0.8.3 tagged.
Follow-up to the audit note on #118456: inject_env was advertised as
allow_write-gated by five documents plus the plugin tool schema, but
hermes_tool.py kept it in READ_ACTIONS, so the agent tool path ran it
ungated. It now sits in WRITE_ACTIONS, with a test that is red on 0.8.1.
README safety rule 1 also claimed unconditionally that Hermes never sees
plaintext; it now names the two operator-controlled routes that do
(raw consume_token under an authorising policy, inject_env).
387 passed, ruff clean, hermes plugins validate green from a fresh clone
at 20a0059, structural validator OK, catalog lane 35 passed, all five
sha-pinned URLs 200. Tag v0.8.2 pushed; v0.8.1 left where it was.
JoaoMarcos44's review on #118456 was right: the entry advertised outbound
secret-value DLP that the pinned hook did not perform, because
on_transform_llm_output() only ran the 12 solicitation-phrase patterns.
VaultKnox 0.8.1 (29af5cf) runs the 28-pattern value detector over
assistant replies before the solicitation rewrite, with a regression test
that is red on 899fd8a. Description now also carries the consume_token
policy caveat the review asked for.
