Sandbox

Understand how the sandbox workspace gives OnCall AI a secure code execution environment for analyzing data during investigations.

  4 minute read  

Overview

The sandbox is a secure, isolated code execution environment that OnCall AI uses during investigations. When OnCall AI encounters large or complex datasets, it writes and runs Python or Bash scripts in the sandbox to analyze the data programmatically — computing statistics, correlating patterns across logs, and validating hypotheses with code.

The sandbox is enabled for every organization automatically. There is no configuration required and no new UI — OnCall AI activates it transparently when needed. Investigations that use the sandbox produce richer, more data-driven analysis.

When the sandbox activates

The sandbox uses lazy activation. It spins up only when OnCall AI encounters data that benefits from programmatic analysis — specifically, when tool results exceed approximately 10,000 characters or 200 lines. At that point, the data is stored in the sandbox workspace and OnCall AI writes scripts to process it.

The workspace persists across sub-steps within the same investigation, so OnCall AI can build on previous analysis results without re-fetching or reprocessing data.

What the sandbox provides

CapabilityDescription
Script executionWrite and run Python or Bash scripts for data analysis
File system accessRead and write files in an isolated environment
Local data storageStore large tool outputs and API responses for processing
Bash commandsExecute shell commands to inspect, parse, or transform data
Repository cloningClone Git repositories to analyze full codebases

How OnCall AI uses the sandbox

During a typical investigation:

  1. OnCall AI queries data sources. A monitor alert, PagerDuty incident, or event connector triggers an investigation thread, and OnCall AI pulls logs, metrics, or traces.
  2. Large results activate the sandbox. When the returned data exceeds the context window threshold, it is stored in the sandbox workspace.
  3. OnCall AI writes analysis scripts. Rather than trying to reason over raw text, OnCall AI writes Python or Bash scripts that parse, filter, aggregate, or correlate the data.
  4. Results feed back into the investigation. Script outputs — statistics, filtered results, correlations — are returned to OnCall AI for synthesis into findings and recommendations.

This approach means OnCall AI can process volumes of data that would otherwise overwhelm the model context window.

Token efficiency

Without the sandbox, analyzing large datasets requires loading the full content into the model context window, where token consumption scales with data size.

The sandbox offloads data to a local file system. OnCall AI reads and processes files with scripts, sending only relevant findings back to the model. This substantially reduces token consumption for investigations involving large log volumes or verbose API responses.

The investigation depth setting controls how aggressively OnCall AI uses the sandbox for deep analysis versus summarized results. Deeper investigations produce more thorough findings but consume more tokens.

Token usage per thread is visible in channel thread details and tracked in Performance.

Error handling

The sandbox enables OnCall AI to handle errors programmatically rather than failing the investigation.

  • Large responses. When an API response is too large for the context window, OnCall AI downloads it to disk and processes it with a script.
  • Unreliable calls. If an external API is intermittent or rate-limited, OnCall AI writes retry logic instead of abandoning the query.
  • Complex analysis. OnCall AI writes Python scripts to parse, filter, or summarize datasets before returning structured results to the model.

Memory in sandbox investigations

During sandbox investigations, OnCall AI surfaces stored knowledge from prior work.

  • Organization-wide memories come from previous analysis findings. If OnCall AI has investigated a similar pattern before, it retrieves those findings to accelerate the current investigation.
  • Personal memories reflect individual user preferences, such as preferred analysis approaches or prior decisions.

You can toggle organization-wide and personal memories independently and configure retention policies in Settings.

Security and isolation

Each investigation gets its own isolated code interpreter instance. Sandboxes do not share state between investigations. When the investigation completes, the sandbox is deprovisioned.

All actions within the sandbox follow the same permission controls as other teammate operations. For the full permission model, see Security Best Practices.

For details on data handling and AI model provider policies, see the Edge Delta Trust Center.

See also