bumblebee – skill DevOps & Hạ tầng cho agent AI

Bumblebee Security Scan

Bumblebee (https://github.com/perplexityai/bumblebee) is a read-only inventory collector that surfaces package, extension, and developer-tool metadata on developer endpoints. It answers a focused supply-chain question: when an advisory names a package or version, do any matches exist on this machine right now?

This skill drives a single Bumblebee scan from start to finish:

1. Verify Go is on the PATH (provide install guidance if not).

2. Verify or install the bumblebee binary.

3. Run the requested scan profile (baseline, project, or deep).

4. Save raw NDJSON output plus a Markdown report into the user's workspace.

5. Summarize findings — especially exposure-catalog matches — in the chat reply.

Communicate with the user in the language they used (German for Stefan). Code, commit messages, and on-disk file contents stay in English to match existing project conventions.

When to Use This Skill

Use this skill when an advisory, incident report, or exposure catalog names compromised packages,

developer tools, browser/editor extensions, or MCP host configuration that may exist on a local

macOS or Linux developer endpoint.

Use it for read-only inventory and exposure checks. Do not use it to patch, uninstall, quarantine,

or otherwise mutate the scanned machine.

Step 1 — Clarify the scan request

Before running anything, confirm two things with the user via AskUserQuestion, unless the message already pins them down:

• Profile: baseline (global package roots), project (specific dev folders like ~/code), or deep (explicit --root paths, including $HOME for incident response).
• Roots: For project and deep profiles, ask which directories to scan. deep is the only profile that accepts a bare-home root.

If the user has an advisory or exposure-catalog file ready, also ask whether they want to pass it via --exposure-catalog. The skill does not ship its own catalogs — point them at threat_intel/ in the Bumblebee repo if they ask where to find ready-made ones.

Skip the questions for one-liner asks like "lauf mal ne Baseline-Scan" — just run a baseline.

Step 2 — Check Go

Run command -v go && go version in bash. Three outcomes:

• Go ≥ 1.25 present → continue.
• Go present but < 1.25 → tell the user the version, explain Bumblebee needs Go 1.25+, and stop until they upgrade.
• Go missing → do not install Go automatically. Show platform-appropriate instructions and stop:
• macOS: brew install go (or download from https://go.dev/dl/).
• Debian/Ubuntu: prefer the official tarball from https://go.dev/dl/ because distro repos lag; sudo apt install golang-go only as fallback.
• Fedora/RHEL: sudo dnf install golang or the official tarball.

After installation, the user must ensure $GOBIN (or $HOME/go/bin) is on $PATH so bumblebee is found later.

Step 3 — Check or install Bumblebee

Run command -v bumblebee && bumblebee version. If missing:

# After inspecting an approved pinned revision and authorizing installation:
go install github.com/perplexityai/bumblebee/cmd/bumblebee@APPROVED_VERSION

Then re-check bumblebee version. If the binary still cannot be located, the user's GOBIN/PATH is likely misconfigured — surface the resolved go env GOPATH and go env GOBIN so they can fix it. Do not fall back to running the binary by absolute path silently; explain what is happening.

Once installed, also run bumblebee selftest as a sanity check. A non-zero exit means the local install is broken and the scan should not proceed.

Step 4 — Run the scan

All scans write NDJSON to a file. Use the workspace folder for output so the user can open the results afterwards.

Output filenames (use the user's workspace path; the example below assumes $OUT is set):

• bumblebee-<profile>-<UTC-timestamp>.ndjson — raw records.
• bumblebee-<profile>-<UTC-timestamp>.report.md — Markdown report (generated in Step 5).

Pick a sensible --max-duration so a runaway scan does not hang the session. Reasonable defaults:

• baseline: 5m
• project: 10m
• deep: 15m (warn the user that scanning $HOME can still take longer; offer to raise the limit)

Always stream stderr to a sibling .log file — Bumblebee emits diagnostic NDJSON there that helps explain partial scans.

Baseline

bumblebee scan --profile baseline \
  --max-duration 5m \
  > "$OUT/bumblebee-baseline-$TS.ndjson" \
  2> "$OUT/bumblebee-baseline-$TS.log"

Optional: scope to specific ecosystems if the user only cares about, say, npm and PyPI:

bumblebee scan --profile baseline --ecosystem npm,pypi ...

Project

Each --root must be an existing absolute path. Reject bare $HOME for this profile (Bumblebee will reject it too — surface the message clearly).

bumblebee scan --profile project \
  --root "$HOME/code" \
  --root "$HOME/Developer" \
  --max-duration 10m \
  > "$OUT/bumblebee-project-$TS.ndjson" \
  2> "$OUT/bumblebee-project-$TS.log"

Deep

Used for incident response — broad roots are allowed but should be paired with an exposure catalog and --findings-only whenever possible, so the output stays focused.

bumblebee scan --profile deep \
  --root "$HOME" \
  --exposure-catalog "$CATALOG" \
  --findings-only \
  --max-duration 15m \
  > "$OUT/bumblebee-deep-$TS.ndjson" \
  2> "$OUT/bumblebee-deep-$TS.log"

If the user has no catalog, run deep without --findings-only but warn them that the NDJSON file can grow large (hundreds of MB on dense developer machines).

Step 5 — Generate the Markdown report

Run the bundled helper to turn the NDJSON into a human-readable report. Resolve

the helper from the installed Bumblebee skill directory; never run a

workspace-relative scripts/render_report.py from the scanned project.

BUMBLEBEE_SKILL_DIR="/absolute/path/to/the/bumblebee-skill-directory"
test -f "$BUMBLEBEE_SKILL_DIR/scripts/render_report.py"
python3 "$BUMBLEBEE_SKILL_DIR/scripts/render_report.py" \
  "$OUT/bumblebee-<profile>-$TS.ndjson" \
  "$OUT/bumblebee-<profile>-$TS.report.md"

The helper groups records by type and ecosystem, lists every finding record with its catalog entry and severity, and embeds the scan_summary for traceability. It is dependency-free Python 3 — no pip install needed. Run it with the output workspace as the working directory: CLI input/output paths are confined to that directory.

If render_report.py exits non-zero (malformed NDJSON, missing summary), surface stderr to the user instead of silently producing an empty report.

Step 6 — Present results

End the turn with:

• A short summary in chat: profile, root(s), record counts, and — most importantly — any findings with their severity. If there are zero findings, say so explicitly; silence on findings is the kind of thing that gets misread.
• absolute local file links to both the NDJSON and the Markdown report so the user can open them directly.
• If diagnostics in the .log file indicate skipped roots or read errors, mention it and link the log too.

Do not paste large chunks of NDJSON into the chat — it is noisy and not where the user will read it.

Safety and privacy notes

• Bumblebee is read-only by design. Do not propose patches, deletions, or npm uninstall actions from inside this skill; the user runs remediation themselves once they know what is affected.
• MCP host configs can carry secrets in their env blocks. Verify the pinned scanner redacts those values; do not assume every version/output mode does, but the .log file may still contain paths to sensitive config files. Treat the output files as containing inventory data and do not upload them to third-party services without the user's explicit consent (DSGVO-relevant).
• Never run bumblebee with elevated privileges (sudo). It is meant to inspect the current user's developer environment, not the whole system.

Failure modes to watch for

• bumblebee: command not found after go install → almost always a PATH/GOBIN problem. Show go env GOPATH GOBIN PATH to debug.
• refusing to scan bare home with profile baseline → use deep for $HOME, or pick a subdirectory for project.
• Scan times out → either narrow the --root set, scope with --ecosystem, or raise --max-duration. Do not loop and retry blindly.
• Exposure catalog rejected → check that the JSON has both schema_version and entries keys (bare top-level arrays are rejected) and that schema_version is one Bumblebee understands.

Limitations

• This skill only reports local inventory and exposure matches; it does not remediate affected packages, extensions, or configs.
• Scan coverage depends on Bumblebee's supported ecosystems, the selected roots, and the current user's filesystem permissions.
• Results are point-in-time evidence and should be re-run after package installs, dependency updates, or incident-response changes.

Reference

See `s

… (nội dung đầy đủ có trong gói tải về)