Use Profiles

ThoughtGate supports two configuration profiles that control how governance decisions are enforced. Profiles let you safely roll out governance rules by starting in development mode (observe-only) before switching to production mode (enforcing).

When to Use Each Profile

Scenario	Profile	Why
First time wrapping an agent	development	See what would be blocked without disrupting work
Testing new governance rules	development	Validate rules before enforcing
Production deployment	production	Full enforcement with Slack approvals
CI/CD pipelines	production	Ensure governance is enforced

Behavior Comparison

Behavior	Production	Development
Governance enforcement	Blocking — denied requests return errors	Log-only — denied requests are forwarded with a warning
Approval workflows	Required — Slack approval must be granted	Auto-approved — logged but not blocked
Decision log prefix	BLOCKED	WOULD_BLOCK
NDJSON smuggling detection	Reject — smuggled messages are dropped	WOULD_REJECT — logged but forwarded
Governance errors	Fail-closed — errors result in denial	Fail-open — errors result in forwarding

Starting with Development Mode

Use --profile development to observe governance decisions without blocking:

thoughtgate wrap --profile development -- claude-code

In your logs, you'll see entries like:

WOULD_BLOCK tool=delete_user rule=delete_* action=deny
WOULD_BLOCK tool=admin_console rule=admin_* action=deny

These tell you which requests would be blocked in production, without actually blocking them.

Reading WOULD_BLOCK Logs

Each WOULD_BLOCK entry includes:

Field	Description
tool	The tool name that triggered the rule
rule	The governance rule pattern that matched
action	What would happen in production (deny, approve)

Use these logs to:

1. Verify your rules match the expected tools
2. Check for false positives (rules that are too broad)
3. Ensure critical tools are governed

Switching to Production

Once you're confident in your rules, switch to production:

thoughtgate wrap --profile production -- claude-code

Or simply omit --profile (production is the default):

thoughtgate wrap -- claude-code

Profile Effects

Governance Decisions

In production, a deny rule returns an error to the agent:

{
  "jsonrpc": "2.0",
  "error": {
    "code": -32003,
    "message": "Denied by ThoughtGate policy"
  },
  "id": 1
}

In development, the same request is forwarded to the upstream server with a log warning.

Approval Adapter

In production, approval rules create Slack messages and wait for human reactions (👍/👎).

In development, approvals are auto-approved immediately with an audit log entry. No Slack messages are posted. This means you don't need a Slack token in development mode.

NDJSON Smuggling Detection

In production, detected smuggling attempts cause the message to be rejected.

In development, smuggling is logged as WOULD_REJECT but the message is still forwarded.

Governance Errors

If ThoughtGate encounters an error evaluating a governance rule:

• Production: Fails closed — the request is denied (fail-safe)
• Development: Fails open — the request is forwarded with an error log

See Also

• CLI Reference — --profile flag documentation
• Wrap MCP Agents — Agent-specific wrapping guides
• Security Model — Profile security implications