Output Controls

Output controls filter and limit what your agent returns to callers. Two mechanisms are available: regex-based redaction and character truncation.

Redaction (redact_patterns)

Configuration

policies:
  safety:
    redact_patterns:
      - '\b\d{3}-\d{2}-\d{4}\b'                                    # SSN
      - '\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b'    # email
      - '\b\d{4}[\s-]?\d{4}[\s-]?\d{4}[\s-]?\d{4}\b'             # credit card

Field	Type	Default	Description
redact_patterns	list[str]	[]	Python regex patterns. Matches are replaced with [REDACTED].

How It Works

After the adapter returns the agent’s output, apply_output_policies() in the RuntimePolicyEngine:

1. Serializes the output dict to a JSON string
2. For each pattern in redact_patterns, runs re.sub(pattern, "[REDACTED]", text)
3. Deserializes back to a dict (if the output was a dict)

This means redaction applies to all string values in the output, including nested fields.

Example

Agent output before redaction:

{
  "answer": "Contact john@company.com or call 123-45-6789"
}

After redaction (with the patterns above):

{
  "answer": "Contact [REDACTED] or call [REDACTED]"
}

Truncation (max_output_chars)

Configuration

policies:
  safety:
    max_output_chars: 10000

Field	Type	Default	Description
max_output_chars	int	null	Maximum characters in the serialized output. Must be > 0 if set.

How It Works

After redaction, if the serialized output exceeds max_output_chars, it is truncated to that length. This is a hard limit — the output is cut without regard to JSON structure.

Truncation runs after redaction, so redaction patterns are applied to the full output before any truncation.

Combined Example

policies:
  safety:
    redact_patterns:
      - '\b\d{3}-\d{2}-\d{4}\b'
    max_output_chars: 5000
    halt_on_violation: false

Processing order:

1. Agent produces output
2. Redaction: SSN patterns replaced with [REDACTED]
3. Truncation: output capped at 5000 characters
4. Result returned to caller

halt_on_violation

When halt_on_violation is true, policy violations raise a PolicyViolationError instead of silently filtering. This affects:

• Redaction: if patterns match, an error is raised instead of redacting
• The default (false) silently applies filters, which is usually what you want in production

Source: SafetyPolicy in packages/schema/dockrion_schema/dockfile_v1.py; redact() in packages/policy-engine/dockrion_policy/redactor.py; RuntimePolicyEngine.apply_output_policies() in packages/runtime/dockrion_runtime/policies.py

---

Previous: Prompt Injection | Next: Tool Gating →