Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
A component deployed to ~/.hermes/plugins/, ~/.hermes-cortex/scripts/,
or ~/.hermes/mcp-servers/ can be either a symlink to the repo source
or a copy. After git pull, repo source is updated — but copies are not.
An agent running with a stale copy silently runs old code.
Two Approaches — Context Matters
Which approach is correct depends on whether the component needs
immutability (chattr +i) for security hardening.
Symlink — simpler, always fresh, no immutability
A symlink shares an inode with the repo source. After git pull, the
deployed component is automatically fresh — no copy needed.
Use when: The component does NOT require chattr +i immutability.
Copy — required for immutability, needs lifecycle management
A chattr +i on a symlink target ALSO locks the repo's tracked file,
because they share an inode. git pull would fail when trying to update
the repo file. Therefore, to secure a component with chattr +i, it MUST
be deployed as a standalone copy.
Use when: The component needs chattr +i immutability (governance
enforcer plugin, critical enforcement hooks).
When a component is deployed as a copy, updates require an explicit lifecycle
to handle immutability. This is handled by cortex-update.sh's
deploy_governance_plugin() function:
1. Check if deployed file has chattr +i
2. If immutable → sudo hermes-plugin-lock unlock --cortex-update (sanctioned token; removes +i)
3. Copy updated files from repo
4. chmod 444 on __init__.py
5. If sudoers configured → sudo hermes-plugin-lock lock (re-applies +i)
Since 2026-07-31unlock/update are gated: they require the
--cortex-update token (cortex-update.sh) or an orchestrator Unix account
(--orchestrator). Plain sudo hermes-plugin-lock unlock is REFUSED + audit-logged.
Blanket Lock at End of Deploy
For the full enforcement chain, a blanket lock runs at the END of
cortex-update.sh main() after ALL files are deployed. This ensures
every enforcement file is locked atomically, not just the ones the
plugin deployer handled:
This pattern handles the bootstrap problem: old files get chattr +i via
the old helper, new files get chmod 444 until the helper is updated.
Deployment Pitfall: Self-Lock Loop
When the helper binary lists ITSELF as a target (for self-protection),
every deploy must unlock the helper before overwriting it. Two mechanisms:
check_each_mapped_file() does a blanket sudo hermes-plugin-lock unlock --cortex-update
at the start of the copy loop.
deploy_system_scripts() has special handling for hermes-plugin-lock:
it tries sudo -n "$dest" update --cortex-update first (self-update), then falls back
to sudo cp if the self-update isn't available.
Without this, the helper locks itself on deploy N, and deploy N+1 can't
overwrite it — cp fails with "Operation not permitted" even via sudo.
If the sudoers entry for hermes-plugin-lock is not set up, steps 2 and 5
fail gracefully with a warning telling the user what to do.
Restricted Helper Pattern
For any agent operation that requires sudo, DO NOT give the agent raw
sudo chattr access. Instead:
Write a restricted helper script to /usr/local/sbin/ that:
Hardcodes the exact path it can operate on
Only accepts a fixed set of arguments (e.g., lock, unlock, status)
Has no argument injection surface
Add a NOPASSWD sudoers entry for ONLY that script path
Canonical example:/usr/local/sbin/hermes-plugin-lock (or deployed to
~/.hermes-cortex/scripts/hermes-plugin-lock for user-writable paths):
A production lock helper typically manages MULTIPLE critical files — not
just the enforcer plugin. The real /usr/local/sbin/hermes-plugin-lock
on this system has 8 targets:
Each target gets chattr +i (Linux) or chflags uchg (macOS) atomically.
The helper also supports a --self-update command: when cortex-update.sh
has a new version, the helper reads from the repo source and overwrites
itself.
Deploying the Helper: Bootstrap Problem
The helper is the ONLY NOPASSWD sudo entry on the system (by design — see
sudoers below). This creates a bootstrap problem: how do you update the
helper binary if sudo cp requires a password?
Three approaches, in order of preference:
Register in MAP (best) — Deploy the helper to ~/.hermes-cortex/scripts/
via cortex-update.sh's register() call. This path is user-writable,
so no sudo needed for the copy. The blanket lock at end of deploy then
applies sudo hermes-plugin-lock lock which covers all targets.
Self-update command — If the helper is already at /usr/local/sbin/,
its update command reads from the repo source ($REAL_HOME/hermes-cortex/...)
and overwrites itself. Requires the update command to NOT use local
keyword (which breaks with set -e in bash case statements outside functions).
Direct deploy — If the user has interactive sudo access, sudo cp repo/src /usr/local/sbin/hermes-plugin-lock works for one-time bootstrap.
chattr +i Requires Root (CAP_LINUX_IMMUTABLE)
On Linux, chattr +i on ext4 requires the CAP_LINUX_IMMUTABLE
capability, which is only available to root — even when the file
is owned by the calling user:
# ❌ FAILS even on user-owned file if called as non-root
chattr +i ~/myfile.txt
# → "chattr: Operation not permitted while setting flags on ..."# ✅ WORKS when called from a root-owned process (e.g., via sudo helper)sudo hermes-plugin-lock lock
This means the restricted helper MUST be called via sudo (with NOPASSWD)
for chattr +i to work. Running the same script directly (without sudo)
will fail on the chattr +i calls, even though the file is user-owned
and the script has permission to read/write it normally.
Fallback for files that cannot get chattr +i:chmod 444 (read-only
for owner). This is weaker than chattr +i (root can still override), but
stronger than the default 755. The doctor checks for chattr +i at FAIL
level, so missing immutability is always visible.
The agent can then call sudo hermes-plugin-lock lock — it can ONLY lock
that one file, on that one path, with no argument injection possible.
Decision Table: Symlink vs Copy
Factor
Symlink
Copy
Freshness after git pull
Auto-fresh
Stale (needs a deploy)
Freshness after deploy
Auto-fresh
Fresh (re-copied)
Supports chattr +i?
❌ Locks repo file
✅ Locks only deployed copy
Deployment complexity
Simple (ln -sf)
Lifecycle required (unlock→copy→relock)
Doctor checks
PASS → FAIL if immutable flag missing (since Jul 2026)
PASS + SHA256 + perms + immutability
Sudo needed
No
Yes (for chattr, via restricted helper)
Doctor Behavior (current as of Jul 2026)
When the doctor finds a symlink, it now emits a WARN suggesting conversion
to copy (for chattr +i hardening). When it finds a copy, it verifies:
Content matches repo (SHA256)
Perms are 0o444
chattr +i is set
When to Choose What
Governance enforcer plugin — COPY. Needs chattr +i to prevent agents
from modifying their own enforcement rules. The deployment lifecycle is
handled by cortex-update.sh.
User-installed plugins — SYMLINK. No immutability needed, and
symlinks keep them fresh without explicit re-deploy.
Hooks (pre-commit, pre-push, post-commit) — COPY. These are
enforcement entry points and should be immutable.
Running Code ≠ Deployed File — the Gateway Memory Trap
The doctor's staleness checks compare files on DISK (repo vs deployed copy).
They do NOT prove the RUNNING process executes the new code. The Hermes
gateway loads plugins (including the governance enforcer) into memory at
STARTUP. A gateway that started before a deploy keeps running the OLD
enforcer logic even when the deployed file hash matches the repo.
Confirmed 2026-07-31: after deploying the P1-A enforcer fix
(sticky-lock + atomic lock writes), write tools were still blocked with
"8/8 skills loaded ✅" — the gateway (started 11:33) was running the
pre-fix enforcer loaded at boot, while the on-disk file (deployed 14:04)
had the fix. The doctor reported "enforcer deployed==repo hash MATCH"
because it checks disk, not memory. The gateway cannot self-restart
(pattern-blocked from inside); schedule it detached:
systemd-run --user --on-active=3s --unit=hermes-gw-restart hermes gateway restart.
Rule: after ANY enforcer/hook/plugin change that affects governance,
restart the gateway before testing the new behavior. "Hash match" proves
deployment, not activation.
DOGFOOD Compare: use git show HEAD:, never the dirty working tree
The pre-commit DOGFOOD gate must compare COMMITTED HEAD vs deployed, not
the working tree vs deployed. A dirty working tree:
Triggers false blocks that cortex-update.sh can't resolve (it deploys
from origin, not local edits) — HEAD==deployed but tree dirty.
Allows an uncommitted edit to be pushed into the deployed enforcer by
the auto-deploy path (dirty-tree injection).
Orchestrator-authoring exception (host detected via _detect_orch AND staged
enforcer change) auto-deploys the working tree — the orchestrator is the
sanctioned author. Everyone else with HEAD≠deployed gets the hard block.
(AGENT_ID is obsolete; the hook only uses it to log claimed-identity
impersonation attempts.)
Update-from-working-tree is root code execution
hermes-plugin-lock update reads its source from the repo working tree
(user-writable). Without a committed-source guard, ANY agent can edit the
helper copy, then sudo hermes-plugin-lock update --cortex-update (the
token is forgeable) to install their modified code as root. Guard:
git diff --quiet HEAD -- ops/install/deploy/nginx/hermes-plugin-lock —
refuse with exit 1 + audit entry when uncommitted changes exist.
Components Requiring Staleness Checks
Most components deployed by cortex-update.shregister() are freshly copied
on every deploy and need no staleness check. The exceptions are
manually symlinked components:
Component
Deploy Path
Repo Source
Governance enforcer plugin
~/.hermes/plugins/governance-enforcer/
plugins/governance-enforcer/
Any manually linked Hermes plugin
~/.hermes/plugins/*/
plugins/*/
MCP server scripts (if manually copied)
~/.hermes/mcp-servers/
mcp-servers/
Integration: Doctor Check Pattern
The canonical implementation is in cortex-doctor.pycheck_governance():
Always name the sudoers file hermes (under /etc/sudoers.d/hermes), NOT
hermes-plugin-lock. The file can contain multiple rules for different
helpers. Using a single hermes file avoids:
Cluttering /etc/sudoers.d/ with one-file-per-helper
Confusion when cortex-update.sh's deploy warning references a different name
Having to remember which filename each agent used
On this system, the file at /etc/sudoers.d/hermes contains all Hermes-related
sudoers entries (plugin lock helper, nginx management, apt cleanup, etc.).
NEVER mutate live sudoers from an installer — verify read-only, fail closed
Luke directive 2026-08-05: "the one that exists SHOULD NOT BE WRITTEN
OVER." /etc/sudoers.d/hermes is the single live policy file; its owner is
deploy-sudoers.sh (backup → visudo validate → restore-on-failure). A deploy
script for a DIFFERENT artifact must not rm/sed/append/chmod it.
Anti-patterns seen in the old deploy-fix-blocked-ips.sh:
rm -f /etc/sudoers.d/hermes-security — dead code under the single-file
policy (2026-07-31); the split file does not exist
sed -i sweep of EVERY /etc/sudoers.d/* + /etc/sudoers deleting lines
matching an old path — blind rewrite of live sudoers
append to /etc/sudoers.d/hermes + chmod 0440 — wrote to the file
owned by deploy-sudoers.sh
Correct shape (committed b33a9fad): install the binary (unlock old
immutable copy → install -o root -g root -m 755 → chattr +i), then
verify read-only that the NOPASSWD entry exists:
grep -qF "/usr/local/sbin/fix-blocked-ips.py" /etc/sudoers.d/hermes; if
missing, exit 1 and direct the operator to deploy-sudoers.sh.
sudo env_reset breaks $HOME-based path resolution in root-owned copies — pass paths via argv
Symptom (2026-08-05):nginx-threat-pipeline committed+pushed new
blocked IPs every run but the deploy step failed silently for 5 days: repo
blocked_ips.add had 18,353 lines, live /etc/nginx/blocked_ips.conf had
17,287 deny rules (~1,066 IPs never applied). Log: ✗ Source not found: /usr/local/sbin/ops/install/deploy/nginx/blocked_ips.add.
Root cause chain:
P1-A hardening moved the NOPASSWD target to a root-owned immutable copy
at /usr/local/sbin/fix-blocked-ips.py (install-only, chattr +i).
The script's repo_dir() walks up from script location for .git —
none above /usr/local/sbin → falls back to $HOME/hermes-cortex.
sudo's env_reset default sets HOME to the TARGET user's home
(/root when target is root) — /root/hermes-cortex doesn't exist →
returns script_dir → source path is wrong → exit 1.
sudo blocks env overrides entirely: sudo -n HOME=/home/moses cmd
and sudo -n CORTEX_REPO=... cmd both fail with sorry, you are not allowed to set the following environment variables.
Fix — argv is the ONLY channel sudo allows:
sudo -n "$FIX_SCRIPT" --repo "${CORTEX_REPO}"# wrapper
def repo_dir(repo_arg=None): # scriptfor c in (repo_arg, os.environ.get("CORTEX_REPO")):
if c and os.path.isdir(os.path.join(c, ".git")):
return c
... # then walk-up, then $HOME fallback
Verify the LIVE artifact, not the pipeline log — a pipeline reporting
"Committed + Pushed" while the deploy step fails leaves repo ahead of the
running service, and the doctor's Deploy sync check does NOT cover
/usr/local/sbin copies:
wc -l ops/install/deploy/nginx/blocked_ips.add # repo source
grep -c "^deny" /etc/nginx/blocked_ips.conf # live — mismatch = silent gap
Why only the orchestrator host hit it: only hosts that installed the
P1-A root-owned copy run the script from /usr/local/sbin; hosts using the
repo copy work because the walk-up from inside the repo finds .git.
Why copy, not symlink: a symlink /usr/local/sbin/foo.py → ~/hermes-cortex/ops/... re-opens the root-ACE hole P1-A closed — the repo
copy is user-writable, so an agent could edit it and sudo would execute
arbitrary code as root. The immutable root-owned copy is the point.
Deploy ≠ refresh:cortex-update.sh does NOT own /usr/local/sbin.
After pushing changes to a script with a root-owned immutable copy, the
operator must re-run the dedicated installer (sudo bash .../deploy-fix-blocked-ips.sh) — password-gated by design, agents cannot.
Doctor Check Pattern for Verifying a Restricted Helper
When adding a doctor check for a deployed restricted helper, follow this pattern:
def_check_helper(res, path, label, sudo_test_cmd):
"""
Generic helper-verification check.
Args:
res: Results collector (from results.py)
path: Path to the helper binary (Path object)
label: Display label for the check
sudo_test_cmd: List of args for sudo -n functional test
(e.g. ["sudo", "-n", "/usr/local/sbin/hermes-plugin-lock", "status"])
"""# 1. Binary existsifnot path.exists():
res.add(label, "FAIL",
f"{path.name} not found at {path.parent}",
f"Run: sudo cortex-update.sh or deploy manually")
return# 2. Binary is executableifnot os.access(str(path), os.X_OK):
res.add(label, "FAIL",
f"{path.name} exists but is not executable",
f"Fix: sudo chmod 755 {path}")
return# 3. Functional test via sudo -n (proves sudoers entry + helper work)import subprocess
try:
result = subprocess.run(
sudo_test_cmd,
capture_output=True, text=True, timeout=5,
)
if result.returncode == 0:
res.add(label, "PASS",
f"helper + sudoers OK ({result.stdout.strip()[:60]})")
else:
res.add(label, "FAIL",
f"sudo -n {path.name} status exited {result.returncode}: "f"{result.stderr.strip()[:60]}",
"Check sudoers entry for the helper path")
except (subprocess.TimeoutExpired, OSError) as e:
res.add(label, "FAIL",
f"sudo -n {path.name} status failed: {e}",
"Check sudoers entry and helper binary")
Key design decisions:
NO separate sudoers file scan — /etc/sudoers.d/ files are r--r-----
owned by root and NOT world-readable. Reading them as the moses user
always fails with PermissionError. The sudo -n functional test IS the
proof that the sudoers entry exists.
FAIL, not WARN — A missing helper is a hard failure, not advisory.
If the helper is missing, the immutability lifecycle is broken and
cortex-update.sh's lock/unlock fails silently.
One check, three states — binary missing (FAIL), binary present but
sudo-n fails (FAIL), everything works (PASS). No partial states.
When to Add a Doctor Helper Check
Add a check like _check_plugin_lock_helper() when:
A restricted helper script is deployed to /usr/local/sbin/
The helper requires a sudoers NOPASSWD entry to function
A component's lifecycle depends on the helper working (lock/unlock cycles)
Multiple agents in the fleet need to independently verify their setup
Copy Deployment Lifecycle — Platform Detail
On Linux:
Check chattr +i via lsattr — if set → sudo hermes-plugin-lock unlock --cortex-update
Copy updated files from repo
chmod 444 on __init__.py
sudo hermes-plugin-lock lock
On macOS:
Check chflags uchg via ls -lO — if set → sudo hermes-plugin-lock unlock --cortex-update
Copy updated files from repo
chmod 444 on __init__.py
sudo hermes-plugin-lock lock
The helper script must handle both platforms transparently.
Related
cortex-update.shdeploy_governance_plugin() — the canonical deploy lifecycle
/usr/local/sbin/hermes-plugin-lock — the restricted sudo helper
hermes-cortex-deployment — full deployment lifecycle for Cortex agents
sudoers-audit — sudoers rule audit and debugging
SOURCE Header Format — Deploy and Doctor Strip Must Agree (2026-08-02)
cortex-update.sh's copy_file() adds a SOURCE header to every deployed .sh/.py.
Since 2026-08-02 the header sits BELOW the shebang (a header above it breaks direct
./script.py execution — the kernel only honors a shebang at line 1). Deployed layout:
#!/usr/bin/env python3
# SOURCE: <repo-path>
# Do NOT edit this file — edit the source above and run: bash cortex-update.sh
(blank)
The doctor strips this header in THREE places in ops/scripts/manage/cortex_doctor/checks.py
(~lines 656, 842, and the _content_md5() helper). All three must handle BOTH layouts:
New (shebang-first): keep lines[0] (the shebang), drop lines[1:4] (SOURCE + Do NOT edit + blank)
Legacy (header-first): drop lines[0:3]
Coordinated-change rule: when changing the header format in cortex-update.sh, update all
three strip sites in the SAME commit and verify with a one-liner asserting
strip(new_layout) == strip(legacy_layout) == repo_source_bytes. Changing only one side makes
every deployed .sh/.py falsely flag as checksum-mismatched in the doctor. Raw sha256sum
comparison against repo will ALWAYS differ for deployed copies (they carry the header) — always
strip via _content_md5() first.
Stale Deploy Detection
Beyond individual component freshness, the deploy directory as a whole
can accumulate orphaned files when register() entries are removed from
cortex-update.sh without cleaning up the deployed copy.
The Pattern
cortex-update.sh has a MAP array built by register() calls. On deploy,
it iterates MAP and copies each source to its destination. If a register()
line is removed, the deployed file stays on disk — an orphan.
Detection: check_stale_deploys() in the doctor
The doctor now includes a check_stale_deploys() function that:
Reads cortex-update.sh and parses all active register() calls
Scans ~/.hermes-cortex/scripts/ for .py and .sh files
Flags any file NOT in the expected destinations as stale
Also flags files that are symlinks (should be copies) and source files
that don't exist
Implementation sketch:
import re
from pathlib import Path
content = Path("cortex-update.sh").read_text()
deploy_home = Path.home() / ".hermes-cortex"
destinations = set()
for line in content.splitlines():
line = line.strip()
ifnot line.startswith("register ") or line.startswith("#"):
continue
m = re.match(r'register\s+"([^"]+)"\s+"([^"]+)"', line)
if m:
dest_str = m.group(2)
dest_str = dest_str.replace("${CORTEX_DEPLOY_HOME}", str(deploy_home))
dest_str = dest_str.replace("${HOME}", str(Path.home()))
destinations.add(Path(dest_str))
for f in (deploy_home / "scripts").rglob("*"):
if f.suffix in (".py", ".sh") and f notin destinations:
print(f"STALE: {f}")
Auto-Cleanup: cortex-update.sh --clean-stale
After every deploy, cortex-update.sh now auto-runs
clean_stale_deploys():
Builds destinations from MAP entries (same as doctor check)
The scripts/ directory under CORTEX_DEPLOY_HOME is the primary target
for stale detection because:
It contains only files deployed via register()
It has no mixed-content (other files from other sources)
It's the largest deploy surface (~200 files)
Other deploy paths (dashboard/, bus/, offline/) have mixed content,
so stale detection there is not implemented — but the same pattern would
apply if needed.
