feat(cyber-guardian-agent): migrate recipe to contrib/python and align with repo standards by abhilash-adavi · Pull Request #2776 · google/adk-recipes

abhilash-adavi · 2026-10-07T05:39:42Z

Summary

Migrates the cyber-guardian-agent recipe from python/agents/cyber-guardian-agent to contrib/python/cyber-guardian-agent and updates it to meet all repository baseline standards.

Changes

Directory & Package Layout: Moved recipe to contrib/python/cyber-guardian-agent with standardized app/ package structure, Dockerfile, and .dockerignore.
Manifest Metadata: Added manifest.yaml with ownership (team, POC, and contributors), architecture, dependencies, and tags.
Model & Environment Variables: Upgraded model to gemini-3.5-flash configured via MODEL_NAME in .env.example with load_dotenv() in app/__init__.py (no hardcoded model literals in source).
No Import-Time Side Effects: Replaced module-import-time BigQuery table initialization and ADC lookups with lazy runtime initialization on first tool invocation.
Testing & Linting: Added tests/test_runnability.py and tests/conftest.py; formatted and linted with ruff.

Verification

uv run python tools/validate.py all contrib/python/cyber-guardian-agent — Passed
uv run ruff check && uv run ruff format --check — Passed
uv run --directory contrib/python/cyber-guardian-agent pytest -v — Passed
End-to-end multi-agent execution verified against Vertex AI (gemini-3.5-flash) and BigQuery.

…n with repo standards - Move cyber-guardian-agent recipe to contrib/python/cyber-guardian-agent with app/ package layout - Add manifest.yaml with ownership, architecture, dependencies, and tags metadata - Add Dockerfile, .dockerignore, and runnability test suite (tests/test_runnability.py, tests/conftest.py) - Upgrade model to gemini-3.5-flash read via MODEL_NAME from environment with load_dotenv() bootstrap - Remove import-time BigQuery table initialization and ADC lookup in favor of lazy runtime initialization

github-actions

Automated House Rules review — 1 finding(s).

Also, on lines this PR does not change:

contrib/python/cyber-guardian-agent/app/tools.py:1 — [MINOR] the .py files carry two different headers: 0 a shorter notice, 16 none, and 2 the full Apache block (contrib/python/cyber-guardian-agent/app/tools.py; and 5 more)

github-actions

Automated Hygiene review — 3 finding(s).

Round 1 · 0 of this PR's 25 automated comments used · this round is capped at 20 across all reviewers.

github-actions

Automated Correctness review — 0 finding(s).

Also, on lines this PR does not change:

contrib/python/cyber-guardian-agent/app/tools.py:111 — [MAJOR] If asset_rows is empty because the hostname was not found, the subsequent return statement accessing asset_rows[0] will raise an IndexError. Can we check if asset_rows is non-empty before accessing its elements?

Round 1 · 0 of this PR's 25 automated comments used · this round is capped at 20 across all reviewers.

github-actions

Automated Maintainability review — 2 finding(s).

Round 1 · 0 of this PR's 25 automated comments used · this round is capped at 20 across all reviewers.

github-actions

Automated Security review — 2 finding(s).

Round 1 · 0 of this PR's 25 automated comments used · this round is capped at 20 across all reviewers.

github-actions · 2026-10-07T05:52:47Z

This pull request isn't ready for review yet.

I removed the review request for tommywagz until the items below are fixed.

Failing checks (1)

Dependency CVE Scan / Scan lockfiles for known vulnerabilities

Fix the failing checks. Once everything is green I will re-request the review automatically, so there is no need to ping anyone. This is re-checked every 15 minutes.

_{Applied by the review gate (.github/review-gate-config.yml); the status/not-ready-for-review label is removed on release.}

…OSV CVEs

github-actions

Automated Correctness review — 1 finding(s).

Round 2 · 7 of this PR's 25 automated comments used · this round is capped at 4 across all reviewers · after this round only Security and Correctness run.

github-actions · 2026-10-07T15:38:06Z

+        client.query(context_query, job_config=job_config).result()
+    )
+    asset_context = (
+        dict(asset_rows[0]) if asset_rows else "No asset context found."


[MAJOR] It seems the fallback path here returns a string while the success path returns a dictionary. Can we return an empty dictionary instead to keep the JSON schema consistent?

Fixed in b5e8918 — updated the fallback in triageQueryTool to return {} when asset_rows is empty so asset_context is consistently a dictionary.

…ssing

happyhuman

Just reviewed the PR and left a few comments. PTAL

happyhuman · 2026-10-07T17:04:10Z

+*   **`network_connection_log`**:
+    *   **Schema:** `log_timestamp, source_host, source_ip, destination_ip, destination_port, protocol, process_pid_ref`
+    *   **Purpose:** Correlate network egress for EDR, confirm `SourceHost` to `DestinationIP` for IOC_MATCH.
+*   *(Implicit `mail_logs` for PHISHING_EMAIL - assumed schema for sender, recipient, URL, SMTP relay)*


Where does mail_logs come from? I only see six CSVs in sample_data/ and none of them is mail logs.

Fixed in 036cc4d — removed the non-existent mail_logs reference and updated the PHISHING_EMAIL step to analyze extracted email entities and correlate with network_connection_log.

happyhuman · 2026-10-07T17:04:10Z

+4.  **Construct Output and Transfer:**
+    *   Create the JSON object containing the `recommended_actions` list.
+    *   Return the JSON object.
+    *   *(Note: `responseExecutionTool` is used by the Orchestrator *after* HITL approval for actual (simulated) execution, not by the Response Agent directly in this prompting context.)*


This note says responseExecutionTool isn't used by the Response Agent, but it's listed under Tools above. Which one is right?

Fixed in 036cc4d — removed the contradictory note and updated Step 3 to clarify that response_agent calls responseExecutionTool for auto-approved actions (requires_approval: false) while deferring requires_approval: true actions for HITL approval.

happyhuman · 2026-10-07T17:04:10Z

+    If classification is: IOC-heavy [IOC_MATCH, PHISHING_EMAIL]
+        Do NOT output intermediate conversational messages or wait for user confirmation before transferring. Perform all agent transfers immediately and autonomously.
+        Always follow this path "CALL threatintel_agent FIRST AND THEN CALL investigation_agent" with the below instructions
+        Delegate to Threat Intel (threat_intel_agent): Call threatintel_agent(indicators=list_of_extracted_iocs).


uses both threat_intel_agent and threatintel_agent here

Fixed in 036cc4d — standardized all references in app/prompt.py to threat_intel_agent.

happyhuman · 2026-10-07T17:04:10Z

+
+**Input:**
+*   `indicators`: A list of strings, where each string is an IOC value (e.g., IP address, hash, URL) (list[str]).
+    Note: Always send the input to the tool as a list even if there is only 1 IOC value or Empty string if there are no IOC values.


Should this be an empty list rather than an empty string? indicators is typed list[str] just above.

Fixed in 036cc4d — updated the note to specify an empty list ([]) instead of an empty string.

happyhuman · 2026-10-07T17:04:10Z

+        rows = list(client.query(query, job_config=job_config).result())
+        return bq_rows_to_json(rows)
+
+    return "Query did not match a known investigation type. Please provide more specific parameters."


Is there a reason this branch returns a plain string? The other two return JSON.

Fixed in 036cc4d — wrapped the fallback error message in json.dumps({"error": ...}) so all branches return JSON.

happyhuman · 2026-10-07T17:04:10Z

+
+# --- Configure Logging ---
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)


logger isn't used in this file

Fixed in 036cc4d — removed the unused logging import and logger variable.

happyhuman · 2026-10-07T17:04:10Z

+Step 2: Delegate to Triage (Always First)
+    Action: Call the triage_agent.
+    Agent Call: triage_agent(hostname=extracted_hostname, alert_type=classification)
+    Critical Check: If the triage_agent returns is_duplicate: true, you MUST stop all further processing. Log this result and terminate the workflow.,


stray comma after workflow.

Fixed in 036cc4d — removed the stray comma after workflow..

happyhuman · 2026-10-07T17:04:10Z

+version = "0.1.0"
+description = "AI-driven agent for cybersecurity incident response"
+authors = [
+    { name = "Cyber Guardian Team" },


Can you please put real names here? Cyber Guardian Team and Google POC look like placeholders.

Fixed in 036cc4d — updated authors to Abhilash Adavi and Anas Aslam.

github-actions

Automated Security review — 1 finding(s).

Round 3 · 8 of this PR's 25 automated comments used · this round is capped at 1 across all reviewers · only Security and Correctness still run.

github-actions · 2026-10-14T04:57:07Z

+            *   `action_command`: Value from `ActionCommand`.
+            *   `target`: Value from `TargetDescription`.
+            *   `requires_approval`: Value from `RequiresApproval`.
+        *   If `requires_approval` is `false`, call `responseExecutionTool(action=action_command, target=target)` to simulate immediate execution. If `requires_approval` is `true`, do not execute it until Human-In-The-Loop (HITL) approval is granted.


[MAJOR] Is there a reason we are allowing the response agent to autonomously execute actions without Human-In-The-Loop approval when requires_approval is false? This could lead to unauthorized tool execution if the triggering inputs are manipulated.

abhilash-adavi added 2 commits October 7, 2026 10:54

chore(cyber-guardian-agent): add abhilash-adavi to manifest contributors

38ab3f0

abhilash-adavi requested a review from happyhuman as a code owner October 7, 2026 05:39

github-actions Bot reviewed Oct 7, 2026

View reviewed changes

Comment thread contrib/python/cyber-guardian-agent/app/tools.py Resolved

github-actions Bot reviewed Oct 7, 2026

View reviewed changes

Comment thread contrib/python/cyber-guardian-agent/.env.example Resolved

Comment thread contrib/python/cyber-guardian-agent/manifest.yaml Resolved

Comment thread contrib/python/cyber-guardian-agent/.env.example Resolved

github-actions Bot reviewed Oct 7, 2026

View reviewed changes

Comment thread contrib/python/cyber-guardian-agent/app/__init__.py Resolved

Comment thread contrib/python/cyber-guardian-agent/app/__init__.py Resolved

github-actions Bot reviewed Oct 7, 2026

View reviewed changes

Comment thread contrib/python/cyber-guardian-agent/app/tools.py Resolved

Comment thread contrib/python/cyber-guardian-agent/app/tools.py Resolved

github-actions Bot added the status/not-ready-for-review Review requests held until checks pass and bot comments are resolved label Oct 7, 2026

abhilash-adavi and others added 3 commits October 7, 2026 18:20

fix(cyber-guardian-agent): address PR review comments and house rules

a58b40c

chore(cyber-guardian-agent): upgrade uv.lock dependencies to resolve …

bfd7ec1

…OSV CVEs

Merge branch 'main' into feat/migrate-cyber-guardian-agent

0c07086

github-actions Bot reviewed Oct 7, 2026

View reviewed changes

Merge branch 'main' into feat/migrate-cyber-guardian-agent

be4b205

github-actions Bot removed the status/not-ready-for-review Review requests held until checks pass and bot comments are resolved label Oct 7, 2026

fix(cyber-guardian-agent): return empty dict when asset context is mi…

b5e8918

…ssing

happyhuman reviewed Oct 7, 2026

View reviewed changes

fix(cyber-guardian-agent): address maintainer review feedback

036cc4d

abhilash-adavi requested a review from tommywagz as a code owner October 14, 2026 04:55

github-actions Bot reviewed Oct 14, 2026

View reviewed changes

Merge branch 'main' into feat/migrate-cyber-guardian-agent

23eb1ea

github-actions Bot added the status/not-ready-for-review Review requests held until checks pass and bot comments are resolved label Nov 2, 2026

Conversation

abhilash-adavi commented Oct 7, 2026

Summary

Changes

Verification

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

github-actions Bot commented Oct 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions Bot Oct 7, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

happyhuman left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions Bot Oct 14, 2026

Choose a reason for hiding this comment

github-actions Bot commented Oct 7, 2026 •

edited

Loading