huskies: merge 652_story_pass_resume_session_id_on_agent_respawn_so_new_sessions_inherit_prior_reasoning

server/src/agents/runtime/claude_code.rs

@@ -7,6 +7,7 @@ use tokio::sync::broadcast;
 
 use crate::agent_log::AgentLogWriter;
 use crate::io::watcher::WatcherEvent;
+use crate::slog_warn;
 
 use super::{AgentEvent, AgentRuntime, RuntimeContext, RuntimeResult, RuntimeStatus};
 
@@ -49,18 +50,51 @@ impl AgentRuntime for ClaudeCodeRuntime {
             &ctx.cwd,
             &tx,
             &event_log,
-            log_writer,
+            log_writer.clone(),
             ctx.inactivity_timeout_secs,
             Arc::clone(&self.child_killers),
             self.watcher_tx.clone(),
             ctx.session_id_to_resume.as_deref(),
         )
-        .await?;
+        .await;
 
-        Ok(RuntimeResult {
-            session_id: pty_result.session_id,
-            token_usage: pty_result.token_usage,
-        })
+        match pty_result {
+            Ok(result) => Ok(RuntimeResult {
+                session_id: result.session_id,
+                token_usage: result.token_usage,
+            }),
+            Err(e) if ctx.session_id_to_resume.is_some() && ctx.fresh_prompt.is_some() => {
+                // Resume failed — fall back to a fresh session without --resume.
+                slog_warn!(
+                    "[agents] Resume failed for {}:{}, retrying without --resume: {}",
+                    ctx.story_id,
+                    ctx.agent_name,
+                    e
+                );
+                let fresh = ctx.fresh_prompt.unwrap();
+                let fallback_result = super::super::pty::run_agent_pty_streaming(
+                    &ctx.story_id,
+                    &ctx.agent_name,
+                    &ctx.command,
+                    &ctx.args,
+                    &fresh,
+                    &ctx.cwd,
+                    &tx,
+                    &event_log,
+                    log_writer,
+                    ctx.inactivity_timeout_secs,
+                    Arc::clone(&self.child_killers),
+                    self.watcher_tx.clone(),
+                    None, // no --resume on fallback
+                )
+                .await?;
+                Ok(RuntimeResult {
+                    session_id: fallback_result.session_id,
+                    token_usage: fallback_result.token_usage,
+                })
+            }
+            Err(e) => Err(e),
+        }
     }
 
     fn stop(&self) {

server/src/agents/runtime/gemini.rs

@@ -705,6 +705,7 @@ mod tests {
             inactivity_timeout_secs: 300,
             mcp_port: 3001,
             session_id_to_resume: None,
+            fresh_prompt: None,
         };
 
         let instruction = build_system_instruction(&ctx);
@@ -723,6 +724,7 @@ mod tests {
             inactivity_timeout_secs: 300,
             mcp_port: 3001,
             session_id_to_resume: None,
+            fresh_prompt: None,
         };
 
         let instruction = build_system_instruction(&ctx);
@@ -803,6 +805,7 @@ mod tests {
             inactivity_timeout_secs: 300,
             mcp_port: 3001,
             session_id_to_resume: None,
+            fresh_prompt: None,
         };
 
         // The model extraction logic is inside start(), but we test the

server/src/agents/runtime/mod.rs

@@ -32,6 +32,13 @@ pub struct RuntimeContext {
     /// than `claude -p <full_prompt>`.  The agent re-enters the previous
     /// conversation and receives the `prompt` (if non-empty) as a new message.
     pub session_id_to_resume: Option<String>,
+    /// Full rendered prompt for a fresh session, kept as fallback if resume fails.
+    ///
+    /// When `session_id_to_resume` is `Some`, `prompt` contains only the
+    /// resume context (e.g. gate failure output).  If the CLI rejects the
+    /// resume (session expired, file missing, version mismatch), the runtime
+    /// retries with this full prompt and no `--resume` flag.
+    pub fresh_prompt: Option<String>,
 }
 
 /// Result returned by a runtime after the agent session completes.
@@ -101,6 +108,7 @@ mod tests {
             inactivity_timeout_secs: 300,
             mcp_port: 3001,
             session_id_to_resume: None,
+            fresh_prompt: None,
         };
         assert_eq!(ctx.story_id, "42_story_foo");
         assert_eq!(ctx.agent_name, "coder-1");

server/src/agents/runtime/openai.rs

@@ -617,6 +617,7 @@ mod tests {
             inactivity_timeout_secs: 300,
             mcp_port: 3001,
             session_id_to_resume: None,
+            fresh_prompt: None,
         };
 
         assert_eq!(build_system_text(&ctx), "Custom system prompt");
@@ -634,6 +635,7 @@ mod tests {
             inactivity_timeout_secs: 300,
             mcp_port: 3001,
             session_id_to_resume: None,
+            fresh_prompt: None,
         };
 
         let text = build_system_text(&ctx);
@@ -684,6 +686,7 @@ mod tests {
             inactivity_timeout_secs: 300,
             mcp_port: 3001,
             session_id_to_resume: None,
+            fresh_prompt: None,
         };
         assert!(ctx.command.starts_with("gpt"));
     }
@@ -700,6 +703,7 @@ mod tests {
             inactivity_timeout_secs: 300,
             mcp_port: 3001,
             session_id_to_resume: None,
+            fresh_prompt: None,
         };
         assert!(ctx.command.starts_with("o"));
     }