discourse/plugins/discourse-ai/lib/agents/tools/lock_post.rb
Rafael dos Santos Silva fe5e4a27e9
FEATURE: Add human-in-the-loop approval queue for AI agent tool actions (#38446)
## Summary

AI agents have 13 moderation tools (close_topic, delete_topic,
edit_tags, edit_post, etc.) that currently execute immediately without
human oversight. This adds an optional approval queue that routes these
tool actions through Discourse's review queue for moderator approval
before execution.

- **New `require_approval` toggle** on AI agents — when enabled,
moderation tool calls are intercepted and sent to the review queue
instead of executing immediately
- **Review queue integration** — moderators see the agent name, tool
name, parameters, and a rendered snippet of the triggering post, then
approve or reject
- **Loop prevention** — approved tool execution is wrapped in
`DiscourseAutomation.set_active_automation` to prevent automation
re-trigger loops (e.g., `edit_tags` → `topic_tags_changed` → automation
fires again)

### New files
- `AiToolAction` model — stores tool name, parameters (JSONB), agent/bot
user refs, and triggering post ID
- `ReviewableAiToolAction` — Reviewable subclass with approve (executes
tool) and reject (discards) actions
- `ReviewableAiToolActionSerializer` — serializes target tool data and
payload context
- Review queue frontend component — displays tool action details and
post snippet
- Two migrations: `ai_tool_actions` table and `require_approval` column
on `ai_agents`

### Modified files
- `Tool` base class gains `requires_approval?` (default `false`),
overridden to `true` on all 13 moderation tools
- `Bot#invoke_tool` — intercepts tools when both tool and agent opt in
to approval
- Agent admin editor — new "Require approval" checkbox
- Agent REST model — `require_approval` added to attribute whitelists
for save payloads
- Serializer, controller, plugin.rb — wired up for the new field and
reviewable type
2026-03-13 12:46:59 -03:00

71 lines
1.9 KiB
Ruby
Vendored

# frozen_string_literal: true
module DiscourseAi
module Agents
module Tools
class LockPost < Tool
def self.signature
{
name: name,
description: "Locks or unlocks a post based on the locked parameter.",
parameters: [
{
name: "post_id",
description: "The ID of the post",
type: "integer",
required: true,
},
{
name: "locked",
description: "true to lock the post, false to unlock it",
type: "boolean",
required: true,
},
{
name: "reason",
description: "Short explanation of why the post is being locked or unlocked",
type: "string",
required: true,
},
],
}
end
def self.name
"lock_post"
end
def self.requires_approval?
true
end
def invoke
post = Post.find_by(id: parameters[:post_id])
return error_response(I18n.t("discourse_ai.ai_bot.lock_post.errors.not_found")) if !post
if !guardian.can_lock_post?(post)
return error_response(I18n.t("discourse_ai.ai_bot.lock_post.errors.not_allowed"))
end
if reason.blank?
return error_response(I18n.t("discourse_ai.ai_bot.lock_post.errors.no_reason"))
end
locker = PostLocker.new(post, acting_user)
if !!parameters[:locked]
locker.lock
else
locker.unlock
end
{ status: "success", message: I18n.t("discourse_ai.ai_bot.lock_post.success") }
end
def description_args
{ post_id: parameters[:post_id], locked: parameters[:locked] }
end
end
end
end
end