Preserve resume dispatch failures

Author: cailmdaley

AGENTS.md

@@ -187,8 +187,17 @@ shuttle/
 mix test                   # full Elixir suite (110 tests, ~7s)
 mix test --only focus      # tagged subset
 go test ./pkg/schema/...   # Go schema tests
+
+# Opt-in real harness smoke. Opens real Claude/Codex/Pi CLIs in tmux,
+# sends no prompt, captures the idle pane, then kills the smoke sessions.
+SHUTTLE_REAL_HARNESS_SMOKE=1 mix test --only integration test/shuttle/real_harness_smoke_test.exs
 ```
 
+The real harness smoke is deliberately outside ordinary `mix test`. It uses
+tmux session names like `shuttle-harness-smoke-<harness>-<unique>`, records
+captures under `_build/test/shuttle_harness_smoke/`, and skips harnesses that
+are not available in `bash -l`.
+
 ## Contributing
 
 See `CONTRIBUTING.md`.

README.md

@@ -192,6 +192,10 @@ make clean    # rm _build and stray .beam files
 
 mix test                    # Elixir suite (110 tests)
 go test ./pkg/schema/...    # Go schema tests
+
+# Opt-in real harness smoke: opens Claude/Codex/Pi in tmux, sends no prompt,
+# captures the idle pane, then kills the smoke-owned sessions.
+SHUTTLE_REAL_HARNESS_SMOKE=1 mix test --only integration test/shuttle/real_harness_smoke_test.exs
 ```
 
 ## License

lib/shuttle/poller.ex

@@ -421,13 +421,8 @@ defmodule Shuttle.Poller do
                 {:reply, {:ok, "human"}, state}
 
               dispatch_eligible?(fiber, state, opts) ->
-                new_state = do_dispatch_fiber(state, fiber, opts)
-
-                if Map.has_key?(new_state.running, fiber_id) do
-                  {:reply, {:ok, session}, new_state}
-                else
-                  {:reply, {:error, :dispatch_failed}, new_state}
-                end
+                {new_state, result} = do_dispatch_fiber(state, fiber, opts)
+                {:reply, result || {:ok, session}, new_state}
 
               true ->
                 {:reply, {:error, :not_eligible}, state}
@@ -628,7 +623,8 @@ defmodule Shuttle.Poller do
 
       Enum.reduce(dispatchable, state, fn fiber, state_acc ->
         if available_slots(state_acc) > 0 do
-          do_dispatch_fiber(state_acc, fiber)
+          {new_state, _result} = do_dispatch_fiber(state_acc, fiber)
+          new_state
         else
           state_acc
         end
@@ -1284,7 +1280,7 @@ defmodule Shuttle.Poller do
         # the kanban shows the card in inFlight (status:active, enabled:true)
         # without any tmux session to watch.
         Logger.info("Human-worker fiber #{fiber_id} accepted; no watcher started")
-        state
+        {state, {:ok, "human"}}
 
       {:ok, session} ->
         agent_name = fetch_shuttle_agent_name(fiber_id, state)
@@ -1325,23 +1321,26 @@ defmodule Shuttle.Poller do
             }
 
             broadcast_snapshot(state)
-            state
+            {state, {:ok, session}}
 
           {:error, reason} ->
             Logger.error("Failed to start watcher for #{fiber_id}: #{inspect(reason)}")
 
-            schedule_retry(state, fiber_id, 1, %{
-              error: "watcher start failed: #{inspect(reason)}"
-            })
+            state =
+              schedule_retry(state, fiber_id, 1, %{
+                error: "watcher start failed: #{inspect(reason)}"
+              })
+
+            {state, {:error, :watcher_start_failed}}
         end
 
       {:error, :already_running} ->
         # Session exists but we don't have a watcher — adopt it
-        adopt_session(state, fiber_id)
+        {adopt_session(state, fiber_id), {:error, :already_running}}
 
       {:error, reason} ->
         Logger.warning("Dispatch failed for #{fiber_id}: #{inspect(reason)}")
-        state
+        {state, {:error, reason}}
     end
   end
 
@@ -1538,7 +1537,8 @@ defmodule Shuttle.Poller do
       {:ok, fiber} ->
         state =
           if eligible?(fiber, state) do
-            do_dispatch_fiber(state, fiber)
+            {new_state, _result} = do_dispatch_fiber(state, fiber)
+            new_state
           else
             Logger.debug("Retry no longer eligible: #{fiber_id}")
             release_claim(state, fiber_id)

test/shuttle/dispatch_integration_test.exs

@@ -1,7 +1,7 @@
 defmodule Shuttle.DispatchIntegrationTest do
   use ExUnit.Case, async: false
 
-  alias Shuttle.Dispatcher
+  alias Shuttle.{Dispatcher, Poller}
 
   # ── Integration Runner ─────────────────────────────────────────────────────
   # Passes `felt` commands to the real felt CLI (with -C felt_store).
@@ -459,6 +459,42 @@ defmodule Shuttle.DispatchIntegrationTest do
              )
   end
 
+  test "poller dispatch API preserves missing session UUID errors", %{host: host} do
+    write_fiber(host, "tests/poller-resume-no-uuid", """
+    ---
+    name: Poller resume no UUID
+    status: active
+    tags:
+      - constitution
+    shuttle:
+      enabled: true
+      kind: oneshot
+      agent: claude-sonnet
+    ---
+    No session UUID in shuttle block.
+    """)
+
+    append_review_comment(host, "tests/poller-resume-no-uuid",
+      summary: "Please resume",
+      resume_mode: "previous"
+    )
+
+    {:ok, poller} =
+      Poller.start_link(
+        name: :test_poller_resume_no_uuid,
+        runner: IntegrationRunner,
+        poll_interval_ms: 600_000,
+        felt_stores: [host]
+      )
+
+    assert {:error, :missing_session_id} =
+             Poller.dispatch_fiber(poller, "tests/poller-resume-no-uuid", [])
+
+    refute Enum.any?(IntegrationRunner.commands(), fn {cmd, args} ->
+             cmd == "tmux" and Enum.at(args, 0) == "new-session"
+           end)
+  end
+
   # review-comment with resume_mode=fresh explicitly requests a new session
   # even when a prior session UUID is stored.
   test "resume_mode=fresh takes fresh path even with stored session.id", %{host: host} do