fix: staging CI review issues (batch 1) (#883)

* fix: address staging-ci-review issues (batch 1)

- #811: Fix unreachable error handling in worker — restructure .await?
  to explicit match on nested Result so token budget errors are properly
  logged and marked as failed
- #813: Combine metadata + token budget into single update_context()
  call to prevent concurrent worker observing partial state
- #814: Persist max_tokens and total_tokens_used to both PostgreSQL and
  libSQL backends — add V12 migration, update save_job/get_job
- #815: Cap user-supplied max_tokens at configured max_tokens_per_job
  to prevent budget bypass via metadata injection
- #869: Release locks before async I/O in webhook handler (http.rs) and
  SIGHUP handler (main.rs) to prevent blocking concurrent requests

Fixes: #811, #813, #814, #815, #869

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address PR #883 review feedback

- Fix min(user_val, 0) bug: guard for unlimited config (max_tokens_per_job == 0)
- Remove duplicate columns from libSQL base SCHEMA (v12 migration is sole source)
- Use get_i64() helper for consistency in libsql/jobs.rs
- Add regression tests for scheduler token budget capping

Co-Authored-By: Claude Opus 4.6 <[email protected]>

---------

Co-authored-by: Claude Opus 4.6 <[email protected]>

src/agent/scheduler.rs

@@ -160,24 +160,36 @@ impl Scheduler {
             .create_job_for_user(user_id, title, description)
             .await?;
 
-        // Apply token budget from config, allowing per-job metadata override.
-        let max_tokens = metadata
+        // Apply metadata and token budget in a single atomic update.
+        // This prevents concurrent workers from observing partial state.
+        // Cap user-supplied max_tokens at the configured limit (Issue #815).
+        let user_max_tokens = metadata
             .as_ref()
             .and_then(|m| m.get("max_tokens"))
-            .and_then(|v| v.as_u64())
+            .and_then(|v| v.as_u64());
+
+        let max_tokens = user_max_tokens
+            .map(|user_val| {
+                if self.config.max_tokens_per_job == 0 {
+                    // Config is "unlimited": use the user-supplied value directly.
+                    user_val
+                } else {
+                    std::cmp::min(user_val, self.config.max_tokens_per_job)
+                }
+            })
             .unwrap_or(self.config.max_tokens_per_job);
 
-        // Apply metadata if provided
+        // Apply both metadata and token budget in one closure (Issue #813: atomic update)
         if let Some(meta) = metadata {
             self.context_manager
                 .update_context(job_id, |ctx| {
                     ctx.metadata = meta;
+                    if max_tokens > 0 {
+                        ctx.max_tokens = max_tokens;
+                    }
                 })
                 .await?;
-        }
-
-        // Set token budget (separate update to avoid overwriting metadata)
-        if max_tokens > 0 {
+        } else if max_tokens > 0 {
             self.context_manager
                 .update_context(job_id, |ctx| {
                     ctx.max_tokens = max_tokens;
@@ -685,8 +697,140 @@ impl Scheduler {
 mod tests {
     use super::*;
     use crate::config::SafetyConfig;
+    use crate::llm::{
+        CompletionRequest, CompletionResponse, LlmError, LlmProvider, ToolCompletionRequest,
+        ToolCompletionResponse,
+    };
     use crate::safety::SafetyLayer;
     use crate::tools::{ApprovalRequirement, Tool, ToolError, ToolOutput};
+    use rust_decimal_macros::dec;
+
+    /// Minimal LLM provider stub for scheduler tests that don't exercise LLM calls.
+    struct StubLlm;
+
+    #[async_trait::async_trait]
+    impl LlmProvider for StubLlm {
+        fn model_name(&self) -> &str {
+            "stub"
+        }
+        fn cost_per_token(&self) -> (rust_decimal::Decimal, rust_decimal::Decimal) {
+            (dec!(0), dec!(0))
+        }
+        async fn complete(&self, _req: CompletionRequest) -> Result<CompletionResponse, LlmError> {
+            Err(LlmError::RequestFailed {
+                provider: "stub".into(),
+                reason: "not implemented".into(),
+            })
+        }
+        async fn complete_with_tools(
+            &self,
+            _req: ToolCompletionRequest,
+        ) -> Result<ToolCompletionResponse, LlmError> {
+            Err(LlmError::RequestFailed {
+                provider: "stub".into(),
+                reason: "not implemented".into(),
+            })
+        }
+    }
+
+    /// Create a Scheduler for token-budget tests. The LLM stub will fail if a
+    /// worker actually tries to call it, but `dispatch_job` sets the token
+    /// budget *before* spawning the worker so we can inspect the context
+    /// immediately after dispatch.
+    fn make_test_scheduler(max_tokens_per_job: u64) -> Scheduler {
+        let config = AgentConfig {
+            name: "test".to_string(),
+            max_parallel_jobs: 5,
+            job_timeout: std::time::Duration::from_secs(30),
+            stuck_threshold: std::time::Duration::from_secs(300),
+            repair_check_interval: std::time::Duration::from_secs(3600),
+            max_repair_attempts: 0,
+            use_planning: false,
+            session_idle_timeout: std::time::Duration::from_secs(3600),
+            allow_local_tools: true,
+            max_cost_per_day_cents: None,
+            max_actions_per_hour: None,
+            max_tool_iterations: 10,
+            auto_approve_tools: true,
+            default_timezone: "UTC".to_string(),
+            max_tokens_per_job,
+        };
+        let cm = Arc::new(ContextManager::new(5));
+        let llm: Arc<dyn LlmProvider> = Arc::new(StubLlm);
+        let safety = Arc::new(SafetyLayer::new(&SafetyConfig {
+            max_output_length: 100_000,
+            injection_check_enabled: false,
+        }));
+        let tools = Arc::new(ToolRegistry::new());
+        let hooks = Arc::new(HookRegistry::default());
+
+        Scheduler::new(config, cm, llm, safety, tools, None, hooks)
+    }
+
+    #[tokio::test]
+    async fn test_dispatch_job_caps_user_max_tokens() {
+        let sched = make_test_scheduler(1000);
+        let meta = serde_json::json!({ "max_tokens": 5000 });
+        let job_id = sched
+            .dispatch_job("user1", "test", "desc", Some(meta))
+            .await
+            .unwrap();
+
+        let ctx = sched.context_manager.get_context(job_id).await.unwrap();
+        assert_eq!(ctx.max_tokens, 1000, "should cap at configured limit");
+    }
+
+    #[tokio::test]
+    async fn test_dispatch_job_unlimited_config_preserves_user_tokens() {
+        let sched = make_test_scheduler(0); // 0 = unlimited
+        let meta = serde_json::json!({ "max_tokens": 5000 });
+        let job_id = sched
+            .dispatch_job("user1", "test", "desc", Some(meta))
+            .await
+            .unwrap();
+
+        let ctx = sched.context_manager.get_context(job_id).await.unwrap();
+        assert_eq!(
+            ctx.max_tokens, 5000,
+            "unlimited config should preserve user value"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_dispatch_job_no_user_tokens_uses_config() {
+        let sched = make_test_scheduler(2000);
+        let job_id = sched
+            .dispatch_job("user1", "test", "desc", None)
+            .await
+            .unwrap();
+
+        let ctx = sched.context_manager.get_context(job_id).await.unwrap();
+        assert_eq!(
+            ctx.max_tokens, 2000,
+            "should use config default when no user value"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_dispatch_job_atomic_metadata_and_tokens() {
+        let sched = make_test_scheduler(10_000);
+        let meta = serde_json::json!({
+            "max_tokens": 3000,
+            "custom_key": "custom_value"
+        });
+        let job_id = sched
+            .dispatch_job("user1", "test", "desc", Some(meta))
+            .await
+            .unwrap();
+
+        let ctx = sched.context_manager.get_context(job_id).await.unwrap();
+        assert_eq!(ctx.max_tokens, 3000, "should use user value within limit");
+        assert_eq!(
+            ctx.metadata.get("custom_key").and_then(|v| v.as_str()),
+            Some("custom_value"),
+            "metadata should be set atomically with token budget"
+        );
+    }
 
     #[test]
     fn test_scheduler_creation() {