Governance Flags

Control agent behavior with a single header or SDK parameter.

agent = Agent("https://tappass.example.com", "tp_...", flags={
    "mode": "observe",
    "pii": "mask",
    "email": "internal:company.com",
})

Or via HTTP header from any language:

X-TapPass-Flags: mode=observe, pii=mask, email=internal:company.com

Flag reference

`mode`

Overall governance posture.

Mode	Behavior
`observe`	Full pipeline runs, everything logged, nothing blocked.
`warn`	Detections logged and flagged, traffic flows through.
`enforce`	Detections trigger configured actions (block, redact, notify). Default.
`lockdown`	Any detection triggers a block.

`email`

Email tool call restrictions.

Mode	Behavior
`mirror:<address>`	All emails redirect to `<address>` with `[AI TEST]` prefix.
`internal:<domain>`	Only `@<domain>` recipients allowed.
`block`	All email tool calls blocked.
`allow`	No restrictions. Default.

`pii`

PII handling in responses.

Mode	Behavior
`mask`	PII replaced with `[MASKED_EMAIL]`, `[MASKED_SSN]`, etc.
`block`	Response blocked if PII detected.
`flag`	PII logged but passed through.
`off`	No PII scanning. Default.

`budget`

Cost and token caps.

Mode	Behavior
`dev`	$1/session, $5/day.
`standard`	$10/session, $50/day.
`custom:<per_call>:<per_session>`	Your own limits.
`unlimited`	No caps. Default.

`tools`

Tool call restrictions.

Mode	Behavior
`allowlist:<tool1>:<tool2>`	Only listed tools can execute.
`denylist:<tool1>:<tool2>`	Listed tools blocked.
`log`	All tool calls logged with full arguments. Default.
`block`	No tool calls. Chat only.

`files`

File operation restrictions.

Mode	Behavior
`read_only`	Reads allowed, all writes blocked.
`project`	Writes only within the workspace directory.
`block`	No file operations.
`allow`	Unrestricted. Default.

`db`

Database operation restrictions.

Mode	Behavior
`read_only`	SELECT only.
`safe_write`	SELECT, INSERT, UPDATE. No DELETE/DROP/TRUNCATE. Default.
`block`	No database operations.

`secrets`

Secret handling in responses.

Mode	Behavior
`redact`	API keys, tokens, passwords replaced with `[REDACTED]`. Default.
`block`	Response blocked if secrets detected.
`flag`	Secrets logged but passed through.

Setting flags

Python SDK (all calls)

agent = Agent("https://tappass.example.com", "tp_...", flags={
    "pii": "mask",
    "email": "internal:company.com",
})

Per-call override

response = agent.chat(
    "Send the salary report",
    flags={"email": "block", "mode": "lockdown"},
)

OpenAI SDK

from openai import OpenAI

client = OpenAI(base_url="https://tappass.example.com/v1", api_key="tp_...")
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello"}],
    extra_headers={"X-TapPass-Flags": "pii=mask, email=block"},
)

Environment variable

export TAPPASS_FLAGS="mode=observe, pii=mask"

Examples

Safe testing:

agent = Agent(url, key, flags={
    "mode": "observe",
    "email": "mirror:me@company.com",
    "budget": "dev",
})

Production:

agent = Agent(url, key, flags={
    "email": "internal:company.com",
    "pii": "mask",
    "secrets": "redact",
})

Incident response:

agent = Agent(url, key, flags={
    "mode": "lockdown",
    "email": "block",
    "tools": "block",
    "files": "read_only",
})

Customer-facing chatbot:

agent = Agent(url, key, flags={
    "tools": "block",
    "pii": "mask",
    "secrets": "redact",
    "budget": "custom:0.10:5",
})

Header format

X-TapPass-Flags: flag1=mode1, flag2=mode2:param1

Comma-separated key=value pairs. Values can include colon-separated parameters. Unknown flags and invalid modes are silently ignored.