docs(guardrails): describe streaming and post-stage blocks as the gateway runs them

[nik13]

Summary

The guardrail pages described streaming and blocked responses differently from how the gateway behaves:

1. Streaming. The pages said post-stage guardrails check a streamed response before it reaches the caller, and stop the stream when they block. In the gateway:
   • pre-stage checks run before the stream opens, so a block there returns 403;
   • only rules in the gateway's own config file with guardrails.streaming.enabled check a stream while it is generated, and only on /v1/chat/completions;
   • every other post-stage check, including those set up in the dashboard, runs after the last chunk and before data: [DONE], and can't stop or recall what was already sent.
2. Block status codes. The pages said a Block always returns 403. Only a pre-stage Block does:
   • a post-stage Block from a dashboard guardrail returns 200 with the x-agentcc-guardrail-triggered header;
   • a synchronous config.yaml post rule makes /v1/chat/completions return 500.
3. Stage. A guardrail's Stage setting isn't applied by the gateway today. Each check runs at its built-in stage: post for Data Leakage Prevention and Hallucination Detection, pre for the rest.

Every statement was checked against the gateway source on main, so these pages can merge now.

Related:

• feat(gateway): run external guardrails on the model response (per-check stage) future-agi#3233 makes the Stage setting apply and lets post-stage checks run on dashboard guardrails. The Stage note, the stage table and the post-stage status rows need revisiting when it ships.
• fix(frontend): stop offering a mask action the gateway runs as log future-agi#3237 removes the Mask action from the dashboard. The matching docs change (actions are Block, Warn and Log; masking PII is the PII check's Remediation) will be a separate PR, to merge once #3237 is deployed.

What changed

• Command Center → Guardrails (command-center/features/guardrails.mdx):
  • the Streaming Guardrails section says what each stage does to a stream, and documents the self-hosted guardrails.streaming mode (check_interval, stop or disclaimer), which uses only config.yaml rules;
  • the Enforcement Modes table's Enforce row now applies to the pre stage only;
  • the Built-in Guardrail Types table shows each check's built-in stage, and says a config.yaml rule uses its own stage.
• Command Center → Streaming (command-center/features/streaming.mdx): the guardrails paragraph and its Next Steps card, which now links to the streaming section.
• Command Center → API reference (command-center/concepts/api-reference.mdx): the streaming note and the post-stage statuses.
• Protect (protect/concepts/understanding-protect.mdx, protect/reference/guardrail-checks.mdx): the Stage note, and when the gateway returns 403, 200 with the triggered header, or 500.

Checks

• node scripts/audit-links.mjs: 0 broken nav links, 0 broken content links. The orphan-page warnings were already on main.
• node scripts/check-deleted-pages.mjs main: no pages deleted.
• npx astro build: every page compiled.
• git diff --check: clean.

🤖 Generated with Claude Code