P3 backend: files, skills, routines, claw chat with LIVE taint plumbing

- tc-files: BlobStore trait + LocalBlobStore (traversal-proof keys); wired
  through Runtime (config storage.data_dir in deployments)
- File tools: files.write/files.list (workspace-internal) + files.delete
  (gated FileDeletion — tested: file survives pending, gone after approve);
  GET /api/openclaw/files + /api/shared-drive/files (drive/agent scoped)
- Skills: catalog/library + idempotent install with counter, uninstall;
  GET /api/skills[?clawId=], POST install/uninstall
- tc-scheduler: croner cron math (clock-controlled tests), SKIP LOCKED
  claim-and-advance firing REAL runs into dedicated ' name' sessions
  (reused, exactly-once), paused routines skipped; routines API + agent
  tool routine.schedule; loop spawned in server
- Claw chat: 1:1 threads, chat.send enforcing the target's Other-Claws
  policy, chat.inbox whose output carries inter_agent taint; the run loop
  now ACCUMULATES taint from tool outputs into LoopState, classifies with
  it, and stamps steps + approvals — a poisoned inbox followed by
  email.send produces an approval whose taint_sources says inter_agent
- ScriptedProvider scenario selection now keys on the most recent marker
  (session history kept earlier markers alive)

132 Rust tests green.

Co-Authored-By: Claude Fable 5 <[email protected]>
This commit is contained in:
Omar Sobh
2026-06-10 05:28:50 -05:00
co-authored by Claude Fable 5
parent ea5162ac65
commit 67f918439c
69 changed files with 3651 additions and 183 deletions
+74
View File
@@ -0,0 +1,74 @@
//! Routines (spec §7.6): agent-created scheduled tasks. Due routines are
//! claimed atomically (one firing even with replicas) and each firing
//! drives a REAL run through the runtime in the routine's dedicated
//! session — gated tools inside a routine still hit the approval queue.
use sqlx::PgPool;
use tc_db::repo::{agents, routines, sessions};
use tc_runtime::Runtime;
use time::OffsetDateTime;
pub use tc_runtime::scheduling::next_occurrence;
#[derive(Debug, thiserror::Error)]
pub enum ScheduleError {
#[error(transparent)]
Pattern(#[from] tc_runtime::scheduling::ScheduleError),
#[error(transparent)]
Db(#[from] tc_db::DbError),
}
pub struct Scheduler {
pool: PgPool,
runtime: Runtime,
}
impl Scheduler {
pub fn new(pool: PgPool, runtime: Runtime) -> Scheduler {
Scheduler { pool, runtime }
}
/// Fires every due routine once and reschedules it. Returns how many
/// fired. Time is a parameter so tests control the clock.
pub async fn tick(&self, now: OffsetDateTime) -> Result<usize, ScheduleError> {
let due = routines::claim_due(&self.pool, now).await?;
for routine in &due {
// Reschedule first: a firing failure must not stall the clock.
let next = next_occurrence(&routine.schedule_cron, now).ok();
routines::set_next_run(&self.pool, routine.id, next).await?;
let message = routine.action["message"].as_str().unwrap_or_default();
if message.is_empty() {
continue;
}
let agent_id = tc_domain::AgentId::from(routine.agent_id);
let Ok(agent) = agents::get(&self.pool, agent_id).await else {
continue; // deleted agent: routine is orphaned
};
// Each routine runs in one dedicated, recognizable session.
let title = format!("{}", routine.name);
let session = match sessions::list_by_agent(&self.pool, agent_id)
.await?
.into_iter()
.find(|s| s.title == title)
{
Some(existing) => existing,
None => sessions::create(&self.pool, agent_id, agent.workspace_id, &title).await?,
};
let _ = self.runtime.send_message(session.id, message).await;
}
Ok(due.len())
}
/// The production loop: ticks on an interval with the real clock.
pub fn spawn(self, interval: std::time::Duration) {
tokio::spawn(async move {
let mut tick = tokio::time::interval(interval);
loop {
tick.tick().await;
let _ = self.tick(OffsetDateTime::now_utc()).await;
}
});
}
}