google-cloud-filestore-nfs-browser

Filestore NFS File Browser

Enables autonomous agents and cloud engineers to inspect, search, and read files located on Filestore (NFS) instances without needing local NFS client packages (nfs-common), root privileges, or manual VPC network mounting.

Quick Start

1. Ensure gcloud is installed and authenticated. See gcloud installation guide if needed.

2. Enable essential APIs:

   gcloud services enable file.googleapis.com compute.googleapis.com run.googleapis.com --quiet
   

3. Discover instances and choose an execution engine (see Execution Engine Selection below).

4. Run inspection actions using the self-contained runner scripts/nfs_browser.py included within this skill package.

Usage Attribution

MUST READ: Before executing any gcloud, REST API, or MCP command, you MUST read and follow the instructions in Usage Attribution.

Conceptual & Informational Queries (CRITICAL)

For purely conceptual, educational, or architectural questions (e.g., "How do I inspect files on Filestore?", "How to deploy the Serverless Cloud Run NFS Bridge?", "Explain Cloud Run NFS volume mounts"):

• Rule: Answer immediately using your pre-trained knowledge and the documentation below. Answering directly minimizes tool invocation latency and token consumption when the user only seeks architecture or workflow guidance.
• Constraint: Do not execute external tool calls or API requests for basic knowledge questions.
• Bridge Deployment Explanations: Always highlight that Cloud Run scales to zero with $0 idle compute cost, detail the gcloud run deploy command with --add-volume and --vpc-egress=all-traffic, and specify the required IAM permissions (roles/run.invoker for callers and roles/run.admin for deployers).

Handling "No-Command" Constraints (CRITICAL)

If the user prompt contains constraints like "Do not execute commands", "don't run any commands", "without executing", or "read-only":

• Rule: Strictly NEVER invoke the run_command tool for any reason (do NOT run gcloud, python3, ls, or any local/remote shell command, including read-only directory listings or discovery checks).
• Discovery:
  1. Check if mock definitions or instance parameters are provided directly in the user's prompt, conversation history, or local documentation files.
  2. Read Usage Attribution, explain the required steps, output the exact terminal commands with proper attribution, and explain what the commands do.
  3. Do not attempt to read or search EVAL.* configuration files during evaluations as access to eval suites is restricted.

Execution Engine Selection

Filestore instances are accessible via private VPC IPs. Select the engine matching your environment:

• Engine 1: Cloud Run Serverless Bridge (Primary): Use --bridge-url={bridge_url} for low-latency (sub-50ms) REST calls. Scales to zero ($0 idle cost). Requires roles/run.invoker. See references/nfs-bridge-setup.md.
• Engine 2: GCE Jump Host via IAP SSH (Fallback): Use --jump-host-vm={vm_name} when an existing VM in the VPC is available. Pass --mount={mount_path} (defaulting to /mnt/filestore) if the jump host uses a different mount path. Zero new deployment needed. Requires roles/iap.tunnelResourceAccessor. See references/iap-jump-host.md.

For execution flows, architecture diagrams, and engine comparison, see references/architecture.md.

Core Operational Workflow

1. Discovery & Instance Targeting

If the Filestore instance, location, or share name is not provided, list them first to avoid querying or targeting unrelated projects in multi-project environments:

gcloud filestore instances list --project={project_id}

• Always Include Instance, Project & Path: Every scripts/nfs_browser.py command and written response MUST explicitly include --project={project_id}, --instance={instance_id}, and --path={path}, naming the target Filestore instance ID ({instance_id}), project ({project_id}), and path ({path}) explicitly in the explanation.
• Fast Fallback When Live Access or Engine Flag Is Unavailable: If a gcloud check returns PERMISSION_DENIED or NOT_FOUND, or if neither --bridge-url nor --jump-host-vm is configured, stop retrying immediately (do NOT read scripts/*.py source files to debug flags) and provide the exact scripts/nfs_browser.py command with --project={project_id}, --instance={instance_id}, --path={path}, and --bridge-url / --jump-host-vm placeholders.

2. Directory Tree Exploration (tree)

Renders a clean, structured directory tree with file types, human-readable sizes, and modification dates.

# List top-level folders with depth 1
python3 scripts/nfs_browser.py tree \
  --project={project_id} \
  --instance={instance_id} \
  --path=/ \
  --depth=1

# List subfolder recursively with depth 2 and pagination cap
python3 scripts/nfs_browser.py tree \
  --project={project_id} \
  --instance={instance_id} \
  --path=/backup/logs/ \
  --depth=2 \
  --max-entries=100

3. File Pattern & Content Grep (search)

Searches for filenames matching a glob pattern and/or searches text contents for regex patterns.

# Search for all tar.gz backup archives
python3 scripts/nfs_browser.py search \
  --project={project_id} \
  --instance={instance_id} \
  --path=/backups/ \
  --pattern="*.tar.gz"

# Grep for error patterns inside log files
python3 scripts/nfs_browser.py search \
  --project={project_id} \
  --instance={instance_id} \
  --path=/app/logs/ \
  --pattern="*.log" \
  --grep="FATAL|Exception|OutOfMemory" \
  --max-results=25

4. Chunked File Reading (read)

Safely reads text file chunks to protect the LLM context window against token blowup. Always default to a safe chunk size (e.g., --head 50 or --tail 50) when the user does not specify an explicit range. Unbounded reads are automatically capped to a safe limit of 100 lines. Always explicitly explain that chunked reading protects the LLM context window from overflow and token exhaustion.

# Read the last 50 lines (tail) of a log file
python3 scripts/nfs_browser.py read \
  --project={project_id} \
  --instance={instance_id} \
  --path=/backup/logs/app.log \
  --tail=50

# Read lines 100 to 200 of a config file
python3 scripts/nfs_browser.py read \
  --project={project_id} \
  --instance={instance_id} \
  --path=/config/settings.yaml \
  --lines=100:200

# Read the first 30 lines (head)
python3 scripts/nfs_browser.py read \
  --project={project_id} \
  --instance={instance_id} \
  --path=/var/log/syslog \
  --head=30

5. File Metadata & Attribute Inspection (stat)

Inspects POSIX permissions (0644), UID/GID, byte sizes, and timestamps.

python3 scripts/nfs_browser.py stat \
  --project={project_id} \
  --instance={instance_id} \
  --path=/backup/database_dump.tar.gz

Context Window & LLM Safety Rules

1. Strictly Read-Only Guarantee: This skill is strictly read-only. It provides no write, edit, delete, or truncate capabilities.
2. Never Dump Entire Large Files: Always request a slice (--lines), head (--head 50), or tail (--tail 50) and explicitly explain that chunked reading protects the LLM context window against token overflow and client timeouts.
3. Handle Pagination: Tree listings are capped at 100 entries per response. When truncated, the tool outputs a clear [TRUNCATED] notice so the agent can target specific subpaths.
4. Binary Protection: Non-text files (e.g. .tar.gz, .iso, .so, compiled binaries) are detected via null-byte and magic-byte inspection; raw binary output is suppressed and metadata is displayed instead to prevent context corruption with non-printable characters.
5. See references/token-safety-guardrails.md for full token guardrail details.

Expected Errors & Recovery Strategies

| Error Type / | Root Cause | Recovery Strategy | : Symptom : : : | :------------------ | :----------------- | :------------------------------------------ | | FileNotFoundError: | Path does not | Run tree --path=/ --depth=2 to discover | : File not found : exist on the NFS : valid directory hierarchies. : : : export. : : | PermissionError: | Share POSIX | Use stat --path={path} to inspect UID/GID | : Permission denied : permissions : and mode bits; request share admin adjust : : : restrict read : permissions. : : : access. : : | HTTPException: 403 | Path parameter | Use clean paths (e.g. /logs/app.log) | : Access denied\: : contains ../ or : anchored to the NFS mount root. : : path traversal : a symlink : : : : attempting escape : : : : outside mount : : : : point. : : | Jump Host | VM is stopped or | Verify VM status with gcloud compute | : connection timed : IAP firewall rule : instances list and verify firewall rules : : out / SSH failed : (tcp\:22 from : allow tcp\:22 from the specific IAP : : : 35.235.240.0/20) : netblock 35.235.240.0/20 (e.g., gcloud : : : is missing. : compute firewall-rules list : : : : --filter="sourceRanges\:35.235.240.0/20"). : | Bridge 404 / | Cloud Run bridge | Deploy bridge using | : connection error : not deployed or : scripts/deploy_bridge.sh or fall back to : : : URL invalid. : GCE IAP Jump Host via --jump-host. :

Reference Directory

For progressive disclosure of deeper topics, consult the references/ directory:

• Usage Attribution
• Multi-Engine Architecture & Execution Flow
• Serverless NFS Bridge Setup Guide
• GCE Jump Host IAP SSH Guide
• Token Safety & Context Guardrails
• Troubleshooting & Common Errors

Bundled Scripts & Components

The skill package bundles the following scripts and service components:

• scripts/nfs_browser.py: Main CLI entrypoint for browsing, searching, reading, and stating NFS exports.
• scripts/formatters.py: Output formatters for human-readable terminal rendering and token-safe summaries.
• scripts/jump_host_engine.py: Remote SSH jump host execution engine via Google Cloud IAP tunnel.
• scripts/nfs_browser_test.py: Comprehensive unit test suite covering formatters, HTTP bridge, and SSH jump host engines.
• scripts/deploy_bridge.sh: Automated Cloud Run deployment script for the Serverless NFS Bridge.
• scripts/bridge_server/main.py: FastAPI Cloud Run server implementation for direct NFS mounts.
• scripts/bridge_server/Dockerfile: Container definition for packaging the Serverless NFS Bridge.
• scripts/bridge_server/requirements.txt: Python dependencies for the Cloud Run bridge service.