AEP-006: Artifact Management for Reports & Logs¶

Gap Analysis¶

Current Implementation¶

Agent tools return data as plain text or dicts in the LLM response. There is no mechanism to: - Save a generated incident report as a downloadable file - Store pod log dumps for later reference - Attach Prometheus query results as a CSV - Share health check snapshots between sessions

What ADK Provides¶

ADK has an Artifacts system (since v0.1.0): - Named, versioned binary data associated with a session or user - Represented as google.genai.types.Part objects with MIME types - tool_context.save_artifact(filename, part) saves artifacts - tool_context.load_artifact(filename) retrieves them - Versioning: each save with the same filename creates a new version - Scoping: session-level (default) or user-level (persistent across sessions)

Gap¶

DevOps agents generate valuable outputs that should be preserved as artifacts: - Incident triage reports (PDF/Markdown) - Kafka consumer lag snapshots (JSON/CSV) - Pod log exports (plain text) - Health check dashboards (HTML) - Prometheus query results (JSON)

Currently these are ephemeral text in the chat — they cannot be downloaded, shared, or referenced later.

Proposed Solution¶

Step 1: Add Artifact Saving to Key Tools¶

from google.genai import types


async def get_kafka_cluster_health(tool_context: ToolContext) -> dict:
    health_data = await _run_sync(_fetch_cluster_health)

    # Save as artifact for later reference
    artifact = types.Part(
        inline_data=types.Blob(
            mime_type="application/json",
            data=json.dumps(health_data, indent=2).encode(),
        )
    )
    tool_context.save_artifact(
        filename=f"kafka_health_{datetime.now().isoformat()}.json",
        artifact=artifact,
    )

    return health_data

Step 2: Add Report Generation to Triage Summarizer¶

The triage summarizer currently writes a text summary. Enhance it to also save a structured report:

async def generate_triage_report(tool_context: ToolContext) -> dict:
    report_md = _build_markdown_report(tool_context.state)
    artifact = types.Part.from_bytes(
        data=report_md.encode(),
        mime_type="text/markdown",
    )
    tool_context.save_artifact("incident_report.md", artifact)
    return {"status": "Report saved", "filename": "incident_report.md"}

Step 3: Add Artifact Retrieval Tool¶

async def get_report(filename: str, tool_context: ToolContext) -> dict:
    """Retrieve a previously saved report or data snapshot."""
    artifact = tool_context.load_artifact(filename)
    if artifact and artifact.inline_data:
        return {"data": artifact.inline_data.data.decode()}
    return {"error": f"Artifact '{filename}' not found"}

Affected Files¶

Acceptance Criteria¶

• [ ] Health check tools save snapshots as JSON artifacts
• [ ] Triage summarizer generates downloadable Markdown report
• [ ] Pod log retrieval saves logs as text artifacts
• [ ] Artifacts visible and downloadable in ADK web UI
• [ ] Artifact filenames include timestamps for versioning
• [ ] At least 3 tool types save artifacts (health, logs, queries)

Notes¶

• Artifacts are stored by the ArtifactService. In-memory by default, but GcsArtifactService is available for production persistence.
• Consider artifact retention policies — DevOps data can accumulate quickly.
• Artifacts can be used with the Memory service: save an artifact, then reference it in memory for cross-session access.