kodeblue-self-integrity — Claude Code skill: install and overview
kodeblue-self-integrity is a Claude Code skill that investigates changes to kodeblue binaries, engine, and configuration. It analyzes whether modifications or watcher downtime stem from tampering.
Apache-2.0 ★ 71 updated 2026-11-15
npx skills add Problem99-io/kodeblue --skill kodeblue-self-integrity -a claude-codeWhat kodeblue-self-integrity does
This skill inspects integrity findings when /usr/local/bin/kodeblue, /usr/local/lib/kodeblue/engine, or configs in /etc/kodeblue change or stop running. Upgrades driven by install.sh open a journaled 10-minute maintenance window that absorbs binary swaps without triggering alerts. Any change outside that window indicates manual intervention or tampering.
The skill executes verification steps via host_exec, checking root ownership, 0755 file modes, and SHA-256 hashes against recorded baselines. It inspects running processes (kodeblue run, kodeblue mcp), checks login history via last, and tracks sudo activity around the change timestamp.
The agent never classifies these changes as benign on its own. It resolves the event as inconclusive and prompts the operator to verify the action with kodeblue ack <finding id>.
What it helps with
- investigate modifications to kodeblue binaries and engine files
- verify file permissions and SHA-256 hashes of the watcher
- differentiate planned installer upgrades from manual binary swaps
- resolve tampering alerts with an operator confirmation command
Example prompts for the agent
- Investigate the finding about changes to /usr/local/bin/kodeblue
- Check why kodeblue stopped during the incident
- Verify whether the recent kodeblue binary modification was an upgrade
How to install kodeblue-self-integrity
1Skills CLI (Vercel)
npx skills add Problem99-io/kodeblue --skill kodeblue-self-integrity -a claude-code2Manually: put the files in place
~/.claude/skills/kodeblue-self-integrity/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/kodeblue-self-integrity/
1Skills CLI (Vercel)
npx skills add Problem99-io/kodeblue --skill kodeblue-self-integrity -a codex2Manually: put the files in place
.agents/skills/kodeblue-self-integrity/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add Problem99-io/kodeblue --skill kodeblue-self-integrity -a cursor2Manually: put the files in place
.cursor/skills/kodeblue-self-integrity/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install Problem99-io/kodeblue blue/skills/kodeblue-self-integrity2Manually: put the files in place
.github/skills/kodeblue-self-integrity/SKILL.md1Skills CLI (Vercel)
npx skills add Problem99-io/kodeblue --skill kodeblue-self-integrity -a gemini-cli2Platform CLI
gemini skills install https://github.com/Problem99-io/kodeblue --path blue/skills/kodeblue-self-integrityManually: put the files in place
.devin/skills/kodeblue-self-integrity/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about kodeblue-self-integrity
How to install kodeblue-self-integrity in Claude Code?
Run npx skills add Problem99-io/kodeblue --skill kodeblue-self-integrity -a claude-code.
What does kodeblue-self-integrity do?
It inspects changes to kodeblue binaries and configs, verifies permissions and SHA-256 sums, and ensures tampering alerts require operator acknowledgement.
Why does the skill refuse to mark integrity findings as benign?
Any change to the watcher might be an attempt to disable detection. The agent sets the status to inconclusive until an operator runs kodeblue ack <finding id>.
Source text
SKILL.md · Apache-2.0
---
name: kodeblue-self-integrity
description: Findings about kodeblue itself — its binaries, engine, config or state changing, or kodeblue having stopped — and why they are never benign without an operator.
metadata:
kodeblue-bundled: true
---
# kodeblue's own integrity
An attacker who controls the machine wants to blind or tamper with its
watcher. A change to kodeblue itself is therefore treated as tampering until an
operator confirms it.
## What a legitimate change looks like
- **Upgrades** through `install.sh` (a hub-driven upgrade or an operator
re-running the installer) open a journaled 10-minute maintenance window first,
so swapping kodeblue's binaries inside it is absorbed and does not alert.
`stats` shows the recent journal. A finding on kodeblue's binaries usually means
the change happened **outside** such a window: a manual copy, a build pushed by
hand, or someone else.
- **Config** in `/etc/kodeblue` is root-owned; a change there is an operator's
action or a root compromise.
## Check
1. `host_exec ["stat","/usr/local/bin/kodeblue"]` (and
`/usr/local/lib/kodeblue/engine`): owner, mode, change time. Owner and mode
must match the install (root, 0755).
2. `host_exec ["sha256sum","/usr/local/bin/kodeblue"]`: the integrity collector
records the same SHA-256 (findings may show it shortened), so you can match the
finding's old and new values.
3. `ps -eo pid,user,lstart,cmd`: the agent restarted shortly after the change
and runs cleanly (`kodeblue run`, `kodeblue mcp`, the engine). Old instances may
linger as zombies for a moment.
4. Who could have written it: logins and sudo around the change time
(`last`, the auth log), and other files changed then
(`find /etc /root /home /tmp /var/tmp /usr/local -newermt <time> -type f`).
5. Attack context: was an attacker active (request log, open incidents) near the
change time? A change during an active attack keeps high severity until an
operator confirms it.
Notes:
- The integrity baseline keeps the old hash until a finding is resolved, so a
second change the same day reports against the original value. Line the values
up before concluding there was only one change.
- A finding on kodeblue's binaries that is still open, waiting for an operator or
resolved `inconclusive` keeps those paths out of maintenance absorption: until
the operator acks it, every later upgrade alerts again. Say so when you ask for
the ack, and check for such an older finding when an upgrade alerted.
- `kodeblue upgrade` itself runs as root from the binary its installer replaces
(in the `kodeblue-upgrade.service` unit), which kodeblue recognises and
journals as `absorb.upgrade` instead of reporting a deleted executable.
- A gap finding ("kodeblue was not running for …") is the same question: was it
a planned restart or upgrade, or did something stop it?
## Verdict
Never `benign_change` on your own judgement, even when the pattern looks exactly
like an upgrade. Resolve `inconclusive`, send an alert asking whether someone deployed a
build at that time, and let the operator acknowledge it with
`kodeblue ack <finding id>`.
Without that answer, treat a change during attacker activity as a true_positive.
© Problem99-io · Apache-2.0 · Problem99-io/kodeblue