feat(routines): enable tool access in lightweight routine execution (#257) (#730)

* Rebase onto staging

* fix(routines): prevent autonomy-escalation in lightweight routines

  - Add ROUTINE_TOOL_DENYLIST to block routine_create/update/delete/fire
    and restart from being callable by lightweight routines
  - Deduplicate sentinel logic by reusing handle_text_response() in the
    no-tools path
  - Filter tool definitions sent to LLM to only include callable tools,
    avoiding wasted tokens on tools that would be rejected
This commit is contained in:
Reid
2026-03-12 11:38:27 -07:00
committed by GitHub
parent 006c15e79c
commit 6bbf87ba3a
6 changed files with 260 additions and 27 deletions
+116 -3
View File
@@ -207,7 +207,7 @@ impl Trigger {
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(tag = "type", rename_all = "snake_case")]
pub enum RoutineAction {
/// Single LLM call, no tools. Cheap and fast.
/// Single LLM call (optionally with tools). Cheap and fast.
Lightweight {
/// The prompt sent to the LLM.
prompt: String,
@@ -217,6 +217,14 @@ pub enum RoutineAction {
/// Max output tokens (default: 4096).
#[serde(default = "default_max_tokens")]
max_tokens: u32,
/// Enable tool access (default: false for backward compatibility).
/// When true, the LLM can call tools during execution.
/// Tools requiring approval are automatically filtered out.
#[serde(default)]
use_tools: bool,
/// Max tool call rounds (default: 3). Only used when use_tools is true.
#[serde(default = "default_max_tool_rounds")]
max_tool_rounds: u32,
},
/// Full multi-turn worker job with tool access.
FullJob {
@@ -243,6 +251,19 @@ fn default_max_iterations() -> u32 {
10
}
fn default_max_tool_rounds() -> u32 {
3
}
/// Hard upper bound for max_tool_rounds to prevent runaway loops and cost explosion.
pub(crate) const MAX_TOOL_ROUNDS_LIMIT: u32 = 20;
/// Clamp max_tool_rounds to [1, MAX_TOOL_ROUNDS_LIMIT].
/// Accepts u64 to avoid truncation before clamping.
fn clamp_max_tool_rounds(value: u64) -> u32 {
value.clamp(1, MAX_TOOL_ROUNDS_LIMIT as u64) as u32
}
/// Parse a `tool_permissions` JSON array into a `Vec<String>`.
pub fn parse_tool_permissions(value: &serde_json::Value) -> Vec<String> {
value
@@ -290,10 +311,22 @@ impl RoutineAction {
.get("max_tokens")
.and_then(|v| v.as_u64())
.unwrap_or(default_max_tokens() as u64) as u32;
let use_tools = config
.get("use_tools")
.and_then(|v| v.as_bool())
.unwrap_or(false);
let max_tool_rounds = clamp_max_tool_rounds(
config
.get("max_tool_rounds")
.and_then(|v| v.as_u64())
.unwrap_or(default_max_tool_rounds() as u64),
);
Ok(RoutineAction::Lightweight {
prompt,
context_paths,
max_tokens,
use_tools,
max_tool_rounds,
})
}
"full_job" => {
@@ -339,10 +372,14 @@ impl RoutineAction {
prompt,
context_paths,
max_tokens,
use_tools,
max_tool_rounds,
} => serde_json::json!({
"prompt": prompt,
"context_paths": context_paths,
"max_tokens": max_tokens,
"use_tools": use_tools,
"max_tool_rounds": max_tool_rounds,
}),
RoutineAction::FullJob {
title,
@@ -504,7 +541,8 @@ pub fn next_cron_fire(
#[cfg(test)]
mod tests {
use crate::agent::routine::{
RoutineAction, RoutineGuardrails, RunStatus, Trigger, content_hash, next_cron_fire,
MAX_TOOL_ROUNDS_LIMIT, RoutineAction, RoutineGuardrails, RunStatus, Trigger, content_hash,
next_cron_fire,
};
#[test]
@@ -554,11 +592,13 @@ mod tests {
prompt: "Check PRs".to_string(),
context_paths: vec!["context/priorities.md".to_string()],
max_tokens: 2048,
use_tools: false,
max_tool_rounds: 3,
};
let json = action.to_config_json();
let parsed = RoutineAction::from_db("lightweight", json).expect("parse lightweight");
assert!(
matches!(parsed, RoutineAction::Lightweight { prompt, context_paths, max_tokens }
matches!(parsed, RoutineAction::Lightweight { prompt, context_paths, max_tokens, .. }
if prompt == "Check PRs" && context_paths.len() == 1 && max_tokens == 2048)
);
}
@@ -695,4 +735,77 @@ mod tests {
);
assert_eq!(Trigger::Manual.type_tag(), "manual");
}
#[test]
fn test_action_lightweight_backward_compat_no_use_tools() {
// Simulate old DB record without use_tools field
let json = serde_json::json!({
"prompt": "old routine",
"context_paths": [],
"max_tokens": 4096
});
let parsed = RoutineAction::from_db("lightweight", json).expect("parse lightweight");
assert!(
matches!(parsed, RoutineAction::Lightweight { use_tools, max_tool_rounds, .. }
if !use_tools && max_tool_rounds == 3),
"missing use_tools should default to false, max_tool_rounds to 3"
);
}
#[test]
fn test_max_tool_rounds_clamped_to_upper_bound() {
let json = serde_json::json!({
"prompt": "test",
"use_tools": true,
"max_tool_rounds": 9999
});
let parsed = RoutineAction::from_db("lightweight", json).expect("parse");
match parsed {
RoutineAction::Lightweight {
max_tool_rounds, ..
} => {
assert_eq!(
max_tool_rounds, MAX_TOOL_ROUNDS_LIMIT,
"should clamp to MAX_TOOL_ROUNDS_LIMIT"
);
}
_ => panic!("expected Lightweight"),
}
}
#[test]
fn test_max_tool_rounds_clamped_to_lower_bound() {
let json = serde_json::json!({
"prompt": "test",
"use_tools": true,
"max_tool_rounds": 0
});
let parsed = RoutineAction::from_db("lightweight", json).expect("parse");
match parsed {
RoutineAction::Lightweight {
max_tool_rounds, ..
} => {
assert_eq!(max_tool_rounds, 1, "should clamp 0 to 1");
}
_ => panic!("expected Lightweight"),
}
}
#[test]
fn test_max_tool_rounds_normal_value_passes_through() {
let json = serde_json::json!({
"prompt": "test",
"use_tools": true,
"max_tool_rounds": 10
});
let parsed = RoutineAction::from_db("lightweight", json).expect("parse");
match parsed {
RoutineAction::Lightweight {
max_tool_rounds, ..
} => {
assert_eq!(max_tool_rounds, 10, "normal value should pass through");
}
_ => panic!("expected Lightweight"),
}
}
}