Fix: distinguish timeouts from usage limits, enable seamless resumption

src/clayde/claude.py

@@ -7,6 +7,7 @@
 import shutil
 import subprocess
 import time
+import uuid
 from abc import ABC, abstractmethod
 from pathlib import Path
 
@@ -61,6 +62,18 @@ def __init__(self, message: str, cost_eur: float = 0.0):
         self.cost_eur = cost_eur
 
 
+class InvocationTimeoutError(Exception):
+    """Raised when a Claude invocation exceeds the configured timeout.
+
+    Distinct from UsageLimitError — timeouts are not rate limits.
+    WIP is committed and conversation state is saved for seamless resumption.
+    """
+
+    def __init__(self, message: str, cost_eur: float = 0.0):
+        super().__init__(message)
+        self.cost_eur = cost_eur
+
+
 def format_cost_line(cost_eur: float) -> str:
     """Format a cost line for inclusion in GitHub comments.
 
@@ -338,6 +351,17 @@ def invoke(
                     repo_path=repo_path, span=span, timeout_s=tool_loop_timeout_s,
                     token_counter=token_counter,
                 )
+            except TimeoutError as e:
+                log.error("Claude API tool loop timed out after %ds", tool_loop_timeout_s)
+                if branch_name:
+                    commit_wip(repo_path, branch_name)
+                if conversation_path:
+                    self._save_conversation(conversation_path, messages)
+                partial_cost_eur = _calculate_cost_usd(model, token_counter["input"], token_counter["output"]) * _EUR_PER_USD
+                span.set_attribute("claude.timeout", True)
+                timeout_exc = InvocationTimeoutError(str(e), cost_eur=partial_cost_eur)
+                span.record_exception(timeout_exc)
+                raise timeout_exc from e
             except anthropic.APIConnectionError as e:
                 log.error("Claude API connection error: %s", e)
                 raise self._build_usage_limit_error(
@@ -477,23 +501,37 @@ def invoke(
             span.set_attribute("claude.backend", "cli")
             span.set_attribute("claude.cli_bin", cli_bin)
 
+            # Determine session ID: load existing or generate new
+            session_id = None
+            resumed = False
+            if conversation_path:
+                session_id = self._load_session_id(conversation_path)
+                if session_id:
+                    resumed = True
+
+            if not session_id:
+                session_id = str(uuid.uuid4())
+
             cmd = [
                 cli_bin, "-p", prompt,
                 "--append-system-prompt", identity,
                 "--output-format", "json",
                 "--dangerously-skip-permissions",
             ]
 
-            # Resume from a previous session if available
+            if resumed:
+                cmd.extend(["--resume", session_id])
+                span.set_attribute("claude.resumed", True)
+                span.set_attribute("claude.resumed_session_id", session_id)
+                log.info("Resuming CLI session %s", session_id)
+            else:
+                cmd.extend(["--session-id", session_id])
+                span.set_attribute("claude.resumed", False)
+                span.set_attribute("claude.session_id", session_id)
+
+            # Save session ID immediately so it survives timeouts
             if conversation_path:
-                session_id = self._load_session_id(conversation_path)
-                if session_id:
-                    cmd.extend(["--resume", session_id])
-                    span.set_attribute("claude.resumed", True)
-                    span.set_attribute("claude.resumed_session_id", session_id)
-                    log.info("Resuming CLI session %s", session_id)
-                else:
-                    span.set_attribute("claude.resumed", False)
+                self._save_session_id(conversation_path, session_id)
 
             try:
                 result = subprocess.run(
@@ -505,7 +543,7 @@ def invoke(
                 if branch_name:
                     commit_wip(repo_path, branch_name)
                 span.set_attribute("claude.timeout", True)
-                exc = UsageLimitError(f"Claude CLI timed out after {timeout_s}s")
+                exc = InvocationTimeoutError(f"Claude CLI timed out after {timeout_s}s")
                 span.record_exception(exc)
                 raise exc
 
@@ -515,14 +553,15 @@ def invoke(
                 and "no conversation found with session id" in (result.stderr or "").lower()
                 and conversation_path
             ):
-                log.warning("CLI session not found — deleting stale conversation file and retrying fresh")
-                conversation_path.unlink(missing_ok=True)
-                # Rebuild command without --resume
+                log.warning("CLI session not found — retrying with new session")
+                session_id = str(uuid.uuid4())
+                self._save_session_id(conversation_path, session_id)
                 cmd = [
                     cli_bin, "-p", prompt,
                     "--append-system-prompt", identity,
                     "--output-format", "json",
                     "--dangerously-skip-permissions",
+                    "--session-id", session_id,
                 ]
                 span.set_attribute("claude.stale_session_retry", True)
                 try:
@@ -535,7 +574,7 @@ def invoke(
                     if branch_name:
                         commit_wip(repo_path, branch_name)
                     span.set_attribute("claude.timeout", True)
-                    exc = UsageLimitError(f"Claude CLI timed out after {timeout_s}s")
+                    exc = InvocationTimeoutError(f"Claude CLI timed out after {timeout_s}s")
                     span.record_exception(exc)
                     raise exc
 

src/clayde/tasks/implement.py

@@ -8,7 +8,7 @@
 import subprocess
 from pathlib import Path
 
-from clayde.claude import UsageLimitError, format_cost_line, invoke_claude
+from clayde.claude import InvocationTimeoutError, UsageLimitError, format_cost_line, invoke_claude
 from clayde.config import DATA_DIR, get_github_client, get_settings
 from clayde.git import ensure_repo
 from clayde.prompts import collect_comments_after, render_template
@@ -85,6 +85,15 @@ def run(issue_url: str) -> None:
                 "interrupted_phase": IssueStatus.IMPLEMENTING,
             })
             return
+        except InvocationTimeoutError as e:
+            log.warning("[%s: %s] Timed out during implementation — will resume next cycle", issue_ref(owner, repo, number), issue.title)
+            accumulate_cost(issue_url, e.cost_eur)
+            span.set_attribute("implement.status", "timeout")
+            update_issue_state(issue_url, {
+                "status": IssueStatus.INTERRUPTED,
+                "interrupted_phase": IssueStatus.IMPLEMENTING,
+            })
+            return
 
         total_cost = pop_accumulated_cost(issue_url) + result.cost_eur
 

src/clayde/tasks/plan.py

@@ -11,7 +11,7 @@
 from github import Github
 from github.Issue import Issue
 
-from clayde.claude import UsageLimitError, format_cost_line, invoke_claude
+from clayde.claude import InvocationTimeoutError, UsageLimitError, format_cost_line, invoke_claude
 from clayde.config import get_github_client
 from clayde.git import ensure_repo
 from clayde.github import (
@@ -65,10 +65,11 @@ def run_preliminary(issue_url: str) -> None:
         log.info("[%s: %s] Invoking Claude for preliminary plan", issue_ref(owner, repo, number), issue.title)
         try:
             result = invoke_claude(prompt, repo_path)
-        except UsageLimitError as e:
-            log.warning("Usage limit hit during preliminary planning #%d", number)
+        except (UsageLimitError, InvocationTimeoutError) as e:
+            label = "Timed out" if isinstance(e, InvocationTimeoutError) else "Usage limit hit"
+            log.warning("%s during preliminary planning #%d", label, number)
             accumulate_cost(issue_url, e.cost_eur)
-            span.set_attribute("plan.status", "limit")
+            span.set_attribute("plan.status", "timeout" if isinstance(e, InvocationTimeoutError) else "limit")
             update_issue_state(issue_url, {
                 "status": IssueStatus.INTERRUPTED,
                 "interrupted_phase": IssueStatus.PRELIMINARY_PLANNING,
@@ -161,10 +162,11 @@ def run_thorough(issue_url: str) -> None:
         log.info("[%s: %s] Invoking Claude for thorough plan", issue_ref(owner, repo, number), issue.title)
         try:
             result = invoke_claude(prompt, repo_path)
-        except UsageLimitError as e:
-            log.warning("Usage limit hit during thorough planning #%d", number)
+        except (UsageLimitError, InvocationTimeoutError) as e:
+            label = "Timed out" if isinstance(e, InvocationTimeoutError) else "Usage limit hit"
+            log.warning("%s during thorough planning #%d", label, number)
             accumulate_cost(issue_url, e.cost_eur)
-            span.set_attribute("plan.status", "limit")
+            span.set_attribute("plan.status", "timeout" if isinstance(e, InvocationTimeoutError) else "limit")
             update_issue_state(issue_url, {
                 "status": IssueStatus.INTERRUPTED,
                 "interrupted_phase": IssueStatus.PLANNING,
@@ -278,10 +280,11 @@ def run_update(issue_url: str, phase: str) -> None:
         log.info("[%s: %s] Invoking Claude for plan update (%s phase)", issue_ref(owner, repo, number), issue.title, phase)
         try:
             result = invoke_claude(prompt, repo_path)
-        except UsageLimitError as e:
-            log.warning("Usage limit hit during plan update #%d", number)
+        except (UsageLimitError, InvocationTimeoutError) as e:
+            label = "Timed out" if isinstance(e, InvocationTimeoutError) else "Usage limit hit"
+            log.warning("%s during plan update #%d", label, number)
             accumulate_cost(issue_url, e.cost_eur)
-            span.set_attribute("plan.update_status", "limit")
+            span.set_attribute("plan.update_status", "timeout" if isinstance(e, InvocationTimeoutError) else "limit")
             update_issue_state(issue_url, {
                 "status": IssueStatus.INTERRUPTED,
                 "interrupted_phase": IssueStatus.PRELIMINARY_PLANNING if phase == "preliminary" else IssueStatus.PLANNING,

src/clayde/tasks/review.py

@@ -2,7 +2,7 @@
 
 import logging
 
-from clayde.claude import UsageLimitError, format_cost_line, invoke_claude
+from clayde.claude import InvocationTimeoutError, UsageLimitError, format_cost_line, invoke_claude
 from clayde.config import DATA_DIR, get_github_client, get_settings
 from clayde.git import ensure_repo
 from clayde.prompts import render_template
@@ -112,10 +112,11 @@ def run(issue_url: str) -> None:
                 branch_name=branch_name,
                 conversation_path=conversation_path,
             )
-        except UsageLimitError as e:
-            log.warning("[%s] Usage limit hit during review handling", issue_label)
+        except (UsageLimitError, InvocationTimeoutError) as e:
+            label_msg = "Timed out" if isinstance(e, InvocationTimeoutError) else "Usage limit hit"
+            log.warning("[%s] %s during review handling", issue_label, label_msg)
             accumulate_cost(issue_url, e.cost_eur)
-            span.set_attribute("review.status", "limit")
+            span.set_attribute("review.status", "timeout" if isinstance(e, InvocationTimeoutError) else "limit")
             update_issue_state(issue_url, {
                 "status": IssueStatus.INTERRUPTED,
                 "interrupted_phase": IssueStatus.ADDRESSING_REVIEW,

tests/test_claude.py

@@ -11,6 +11,7 @@
     ApiBackend,
     CliBackend,
     InvocationResult,
+    InvocationTimeoutError,
     UsageLimitError,
     _calculate_cost_usd,
     _get_backend,
@@ -333,9 +334,39 @@ def fake_monotonic():
              patch.object(backend, "_get_client", return_value=mock_client), \
              patch.object(ApiBackend, "_execute_tool", return_value="output"), \
              patch("clayde.claude.time.monotonic", side_effect=fake_monotonic):
-            with pytest.raises(TimeoutError):
+            with pytest.raises(InvocationTimeoutError):
                 backend.invoke("implement", str(tmp_path))
 
+    def test_tool_loop_timeout_saves_conversation(self, tmp_path):
+        (tmp_path / "CLAUDE.md").write_text("identity")
+        conv_path = tmp_path / "conv.json"
+        tool_block = _make_tool_use_block("bash", "tool-1", {"command": "echo loop"})
+        tool_response = _make_tool_response([tool_block])
+        mock_client = MagicMock()
+        mock_client.beta.messages.create.return_value = tool_response
+        backend = ApiBackend()
+
+        call_count = [0]
+        def fake_monotonic():
+            call_count[0] += 1
+            if call_count[0] <= 1:
+                return 0.0
+            return 2000.0
+
+        with patch("clayde.claude.APP_DIR", tmp_path), \
+             patch("clayde.claude.get_settings", return_value=_mock_settings()), \
+             patch.object(backend, "_get_client", return_value=mock_client), \
+             patch.object(ApiBackend, "_execute_tool", return_value="output"), \
+             patch("clayde.claude.time.monotonic", side_effect=fake_monotonic), \
+             patch("clayde.claude.commit_wip") as mock_wip:
+            with pytest.raises(InvocationTimeoutError) as exc_info:
+                backend.invoke("implement", str(tmp_path),
+                              branch_name="branch", conversation_path=conv_path)
+            mock_wip.assert_called_once_with(str(tmp_path), "branch")
+
+        assert conv_path.exists()
+        assert exc_info.value.cost_eur >= 0.0
+
     def test_token_usage_accumulated_across_turns(self, tmp_path):
         (tmp_path / "CLAUDE.md").write_text("identity")
         tool_block = _make_tool_use_block("bash", "t1", {"command": "echo x"})
@@ -652,14 +683,15 @@ def test_saves_session_id(self, tmp_path):
 
         assert conv_path.exists()
         data = json.loads(conv_path.read_text())
+        # Session ID from response overwrites the pre-generated one
         assert data["session_id"] == "my-session-id"
 
     def test_resumes_from_session_id(self, tmp_path):
         (tmp_path / "CLAUDE.md").write_text("identity")
         conv_path = tmp_path / "conv.json"
         conv_path.write_text(json.dumps({"session_id": "prev-session"}))
         mock_result = MagicMock()
-        mock_result.stdout = self._cli_json_output("resumed")
+        mock_result.stdout = self._cli_json_output("resumed", "prev-session")
         mock_result.stderr = ""
         mock_result.returncode = 0
         backend = CliBackend()
@@ -731,7 +763,7 @@ def test_limit_saves_session_before_raising(self, tmp_path):
         data = json.loads(conv_path.read_text())
         assert data["session_id"] == "limit-session"
 
-    def test_timeout_raises_usage_limit_error(self, tmp_path):
+    def test_timeout_raises_invocation_timeout_error(self, tmp_path):
         (tmp_path / "CLAUDE.md").write_text("identity")
         backend = CliBackend()
 
@@ -740,10 +772,48 @@ def test_timeout_raises_usage_limit_error(self, tmp_path):
              patch("clayde.claude._resolve_cli_bin", return_value="/usr/bin/claude"), \
              patch("clayde.claude.subprocess.run", side_effect=__import__("subprocess").TimeoutExpired("claude", 1800)), \
              patch("clayde.claude.commit_wip") as mock_wip:
-            with pytest.raises(UsageLimitError):
+            with pytest.raises(InvocationTimeoutError):
                 backend.invoke("prompt", "/repo", branch_name="branch")
             mock_wip.assert_called_once_with("/repo", "branch")
 
+    def test_timeout_saves_session_id_for_resumption(self, tmp_path):
+        """When a fresh CLI session times out, the pre-generated session ID is saved for resumption."""
+        (tmp_path / "CLAUDE.md").write_text("identity")
+        conv_path = tmp_path / "conv.json"
+        backend = CliBackend()
+
+        with patch("clayde.claude.APP_DIR", tmp_path), \
+             patch("clayde.claude.get_settings", return_value=_mock_settings(backend="cli")), \
+             patch("clayde.claude._resolve_cli_bin", return_value="/usr/bin/claude"), \
+             patch("clayde.claude.subprocess.run", side_effect=__import__("subprocess").TimeoutExpired("claude", 1800)), \
+             patch("clayde.claude.commit_wip"):
+            with pytest.raises(InvocationTimeoutError):
+                backend.invoke("prompt", "/repo", branch_name="branch", conversation_path=conv_path)
+
+        # Session ID should be saved even though the process timed out
+        assert conv_path.exists()
+        data = json.loads(conv_path.read_text())
+        assert data["session_id"]  # a UUID was generated and saved
+
+    def test_timeout_preserves_session_id_for_resumed(self, tmp_path):
+        """When a resumed CLI session times out, the session ID is preserved for next resumption."""
+        (tmp_path / "CLAUDE.md").write_text("identity")
+        conv_path = tmp_path / "conv.json"
+        conv_path.write_text(json.dumps({"session_id": "my-session"}))
+        backend = CliBackend()
+
+        with patch("clayde.claude.APP_DIR", tmp_path), \
+             patch("clayde.claude.get_settings", return_value=_mock_settings(backend="cli")), \
+             patch("clayde.claude._resolve_cli_bin", return_value="/usr/bin/claude"), \
+             patch("clayde.claude.subprocess.run", side_effect=__import__("subprocess").TimeoutExpired("claude", 1800)), \
+             patch("clayde.claude.commit_wip"):
+            with pytest.raises(InvocationTimeoutError):
+                backend.invoke("prompt", "/repo", branch_name="branch", conversation_path=conv_path)
+
+        # Session ID should still be in the conversation file
+        data = json.loads(conv_path.read_text())
+        assert data["session_id"] == "my-session"
+
     def test_fallback_on_non_json_stdout(self, tmp_path):
         (tmp_path / "CLAUDE.md").write_text("identity")
         mock_result = MagicMock()
@@ -760,8 +830,8 @@ def test_fallback_on_non_json_stdout(self, tmp_path):
 
         assert result.output == "plain text output"
 
-    def test_stale_session_retries_fresh(self, tmp_path):
-        """When CLI reports 'No conversation found', delete conv file and retry without --resume."""
+    def test_stale_session_retries_with_new_session_id(self, tmp_path):
+        """When CLI reports 'No conversation found', retry with a new session ID."""
         (tmp_path / "CLAUDE.md").write_text("identity")
         conv_path = tmp_path / "conv.json"
         conv_path.write_text(json.dumps({"session_id": "stale-session"}))
@@ -788,12 +858,13 @@ def test_stale_session_retries_fresh(self, tmp_path):
             result = backend.invoke("prompt", str(tmp_path), conversation_path=conv_path)
 
         assert result.output == "fresh output"
-        # First call should have --resume, second should not
+        # First call should have --resume, second should have --session-id (new UUID)
         first_cmd = mock_run.call_args_list[0][0][0]
         second_cmd = mock_run.call_args_list[1][0][0]
         assert "--resume" in first_cmd
         assert "--resume" not in second_cmd
-        # Conv file should now have the new session ID
+        assert "--session-id" in second_cmd
+        # Conv file should now have the new session ID from the response
         data = json.loads(conv_path.read_text())
         assert data["session_id"] == "new-session"