Prevent stale process state after failures

CHANGELOG.md

@@ -4,6 +4,18 @@ All notable changes to `devloop` will be recorded in this file.
 
 ## [Unreleased]
 
+## [0.10.2] - 2026-08-26
+
+### Fixed
+
+- Invalidated process-output state before every start attempt and when the
+  process stops or exits, preventing stale values such as tunnel URLs from
+  surviving a missing executable or dead process.
+- Made `wait_for_process` reject stopped and failed processes even when their
+  readiness state persists from an earlier run.
+- Logged complete workflow failure chains while explicitly continuing the
+  runtime in degraded mode.
+
 ## [0.10.1] - 2026-08-26
 
 ### Changed

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "devloop"
-version = "0.10.1"
+version = "0.10.2"
 edition = "2024"
 
 [dependencies]

docs/behavior.md

@@ -116,8 +116,13 @@ Managed processes are long-running child commands.
   escapes portable Unix process-group containment. Such commands must
   provide their own shutdown integration instead of daemonizing beneath
   `devloop`.
-- `wait_for_process` waits on the configured readiness probe, not just
-  on successful spawning.
+- `wait_for_process` requires the named process to be running while it
+  waits on the configured readiness probe. Persisted readiness state from
+  an earlier process instance cannot make a stopped or failed process ready.
+- State keys populated by a process's output rules belong to that process
+  instance. `devloop` clears them before every start attempt and when the
+  process stops or exits, so a failed dependency cannot leave a stale URL or
+  other process-derived value available to later workflows.
 - `restart = "always"` restarts a child after any exit unless
   `devloop` is shutting down.
 - `restart = "on_failure"` restarts only after unsuccessful exit.
@@ -139,6 +144,8 @@ Managed processes are long-running child commands.
   probe is checked.
 - Missing or malformed environment references fail loudly with the
   field name so the configuration error is visible.
+- Workflow failures include their complete causal error chain and leave the
+  runtime watching in degraded mode when it is safe to continue.
 
 Liveness probes are checked on the configured interval while the process
 is running. If a liveness probe fails and the restart policy allows it,

docs/configuration.md

@@ -197,7 +197,10 @@ Use `plain` when subprocess color or exact body rendering matters. Use
 ### Output rules
 
 Each rule extracts a value from process output and writes it into the
-session state.
+session state. The value belongs to the current process instance: `devloop`
+clears it before each start attempt and when the process stops or exits.
+Consequently, a missing executable or other startup failure cannot leave a
+value captured by an earlier run available to later workflows.
 
 ```toml
 output = { rules = [{ state_key = "tunnel_url", extract = "url_token" }] }

src/engine.rs

@@ -217,7 +217,7 @@ impl WorkflowEffectAdapter for LiveWorkflowAdapter<'_, '_> {
     }
 
     async fn stop_process(&mut self, process: &str) -> Result<()> {
-        self.processes.stop_named(process).await
+        self.processes.stop_named(process, self.state).await
     }
 
     async fn restart_process(&mut self, process: &str) -> Result<()> {
@@ -421,8 +421,8 @@ async fn execute_runtime_effects<A: RuntimeEffectAdapter>(
                 if let Err(error) = adapter.run_workflow(&workflow_name, &changed_files).await {
                     error!(
                         workflow = %workflow_name,
-                        error = %error,
-                        "workflow failed; continuing runtime"
+                        error = %workflow_failure_chain(&error),
+                        "workflow failed; continuing runtime in degraded mode"
                     );
                 }
             }
@@ -446,6 +446,10 @@ async fn execute_runtime_effects<A: RuntimeEffectAdapter>(
     Ok(false)
 }
 
+fn workflow_failure_chain(error: &anyhow::Error) -> String {
+    format!("{error:#}")
+}
+
 fn forward_watcher_event(
     tx: &tokio::sync::mpsc::UnboundedSender<notify::Result<Event>>,
     shutting_down: &AtomicBool,
@@ -745,6 +749,17 @@ mod tests {
         std::env::temp_dir().join(format!("devloop-engine-state-{unique}.json"))
     }
 
+    #[test]
+    fn workflow_failure_chain_includes_root_cause() {
+        let error = anyhow!("executable 'cloudflared' was not found")
+            .context("guardian failed to start process 'tunnel'");
+
+        let message = workflow_failure_chain(&error);
+
+        assert!(message.contains("guardian failed to start process 'tunnel'"));
+        assert!(message.contains("executable 'cloudflared' was not found"));
+    }
+
     #[test]
     fn classify_changes_by_workflow() {
         let root = PathBuf::from("/tmp/example");