Record structured security evidence for sandbox executions

[eshulman2]

Problem

Forge records workflow and model telemetry, but it does not produce a consolidated security record for sandbox execution. Operators cannot readily answer which policy was enforced, which image ran, what capabilities the driver supported, whether egress was denied, or which output validations passed.

Proposed solution

Produce a structured, redacted security evidence artifact for every sandbox run containing:

• ticket, repository, workflow stage, execution/trace ID, and driver
• immutable image reference/digest when available
• requested and effective sandbox policy fingerprints
• driver capability/preflight result
• timestamps, limits, exit/timeout/cancellation state, and cleanup result
• summarized network/policy denials
• safe-output, secret-scan, and deterministic-validation results
• credential delivery mode (never credential values)

Expose this through logs/traces and Prometheus counters; retain full evidence according to configurable policy.

Acceptance criteria

• Every sandbox execution emits a versioned security evidence model.
• Evidence contains no secret values or unbounded agent output.
• Requested/effective policy and image identity can be correlated with the workflow trace.
• Failed policy application, validation, or cleanup is visible and queryable.
• Prometheus metrics summarize denials and security-gate outcomes without high-cardinality labels.
• Unit tests cover redaction and schema compatibility.

Dependencies

• Define sandbox driver capabilities and execution requirements #265
• Add a fail-closed pre-push validation gate in GitOperations #263

Estimate

AI-assisted: 3–5 engineering days.