THI-455: output below an API error clears the flag

periscope/panes.py

@@ -368,18 +368,13 @@ def recency_stamps_for(target: str) -> dict:
 # active-op / prose by the no-space-before-`(` pattern.
 TOOL_CALL_RE = re.compile(r"^[^\x00-\x7f]\s+\w+\(")
 
-# Tool-result lines start with `⎿`. Claude Code renders rate-limit and
-# transport failures as a normal tool result whose body begins with
-# "API Error:" and then STOPS — it does not auto-retry; the turn just
-# ends with the error sitting there until the user types something
-# (`keep going`, etc.) to restart. When the most recent `⎿` line is an
-# API Error the pane is silently blocked waiting on a human nudge. A
-# later non-error `⎿` means a subsequent turn succeeded and the flag
-# should clear. A stream cut mid-response ("The response stopped
-# arriving") renders as an assistant `⏺` line instead (THI-436), so both
-# glyphs count as the error.
+# A tool result (`⎿ …`). An API error is a tool result (rate limit) or an
+# assistant `⏺` line (stream cut, THI-436). A collapsed tool group
+# (`  Called fdy 6 times (ctrl+o to expand)`) prints no `⎿` (THI-455); its
+# capital excludes a result's `… +3 lines` row and `› Message from @…`.
 TOOL_RESULT_RE = re.compile(r"^\s*⎿\s+")
 API_ERROR_RE = re.compile(r"^\s*[⎿⏺]\s+API Error:")
+COLLAPSED_TOOL_GROUP_RE = re.compile(r"^\s+[A-Z]\w* .*\(ctrl\+o to expand\)\s*$")
 
 
 def all_pane_ids() -> set[str]:
@@ -1046,23 +1041,21 @@ def _detect_asked_question(lines: list[str]) -> bool:
 
 
 def _detect_api_error(lines: list[str]) -> bool:
-    """True when the most recent `⎿` tool-result line is an API Error,
-    or a `⏺ API Error:` line sits below it.
+    """True when Claude has written nothing since the newest API Error
+    line: no `⎿` tool result, `⏺` line or collapsed tool group below it.
 
-    Claude Code renders rate-limit / transport failures as a normal tool
-    result whose body begins "API Error:" (or, for a stream cut, as an
-    assistant `⏺ API Error:` line) and then STOPS — no auto-retry;
-    the turn just ends until the user nudges it (`keep going`, etc.). The
-    flag tracks "pane is silently blocked". A later non-error `⎿` means a
-    subsequent turn succeeded and the flag clears.
+    A rate-limit error ends the turn until the user nudges it (`keep
+    going`); after a stream cut Claude Code may carry on by itself. Either
+    way, output below the error means the API answered and the pane isn't
+    blocked.
 
     Caller invokes this only for Claude panes — prose mentioning "API
     Error" in a shell pane (logs, grep output) must not trip it.
     """
     for line in reversed(lines):
         if API_ERROR_RE.match(line):
             return True
-        if TOOL_RESULT_RE.match(line):
+        if TOOL_RESULT_RE.match(line) or COLLAPSED_TOOL_GROUP_RE.match(line) or line.lstrip().startswith("⏺"):
             return False
     return False
 

static/src/split/Detail.jsx

@@ -264,7 +264,7 @@ export function PaneHeader({ w }) {
     parts.push(<><span class="hsep">·</span><span class={`header-mem ${mh.cls}`} title={mh.title}>↻ {mh.label}</span></>);
   }
   if (w.api_error) {
-    parts.push(<><span class="hsep">·</span><span class="header-api-error" title="last tool result was an API error">⚠ API error</span></>);
+    parts.push(<><span class="hsep">·</span><span class="header-api-error" title="Claude hit an API error and has written nothing since">⚠ API error</span></>);
   }
   // After ↻ and API-error: its fixed width (styles.css .header-res) can't
   // shift either chip.

tests/test_panes.py

@@ -738,15 +738,16 @@
 # Claude Code renders rate-limit / transport failures as a normal `⎿` tool
 # result whose body starts with "API Error:". We flag panes whose MOST
 # RECENT tool result is an error so the dashboard can show a stuck-on-
-# retry pane at a glance. A subsequent successful tool result clears it.
+# retry pane at a glance. A subsequent successful tool result clears it,
+# and so does a `⏺` reply to a prompt sent after the error.
 
 API_ERROR_CASES: list[tuple[str, str, bool]] = [
     # The screenshot Tom hit: three consecutive rate-limit hits, the user
     # repeatedly typed `keep going`, and the bottom shows Claude resuming.
-    # The latest tool result above the resumed turn is still the API
-    # error → flag should be on so the card surfaces the stuck retry.
+    # The `⏺` under the last `keep going` is a reply → the errors belong to
+    # earlier turns, flag off (THI-455).
     (
-        "rate-limited-stuck",
+        "rate-limited-resumed",
         textwrap.dedent(f"""\
         Reshaping the query — payload columns are huge.
 
@@ -773,6 +774,57 @@
 
         ⏺ Skipping reminder. One last lookup — there's a closest_store_by_address variant I haven't traced yet.
         {STATUS_LINE}
+    """),
+        False,
+    ),
+    # The same pane before anything answers the nudge: `keep going` alone
+    # doesn't clear it.
+    (
+        "rate-limited-nudged",
+        textwrap.dedent(f"""\
+        ❯ keep going
+          ⎿  API Error: Server is temporarily limiting requests (not your usage limit) · Rate limited
+
+        ✻ Brewed for 0s
+
+        ❯ keep going
+
+        ✻ Crunching…
+        {STATUS_LINE}
+    """),
+        True,
+    ),
+    # The nudged turn has so far printed only a collapsed tool group: the
+    # API answered → off.
+    (
+        "rate-limited-resuming-calls",
+        textwrap.dedent(f"""\
+        ❯ keep going
+          ⎿  API Error: Server is temporarily limiting requests (not your usage limit) · Rate limited
+
+        ✻ Brewed for 0s
+
+        ❯ keep going
+
+          Called fdy 6 times (ctrl+o to expand)
+
+        ✻ Crunching…
+        {STATUS_LINE}
+    """),
+        False,
+    ),
+    # A long error's `… +N lines` row is part of the error, not output after it.
+    (
+        "error-truncated-result",
+        textwrap.dedent(f"""\
+          Called fdy (ctrl+o to expand)
+          ⎿  API Error: 500 {{"type":"error","error":{{"type":"api_error","message":"Internal server error"}}}}
+             … +3 lines (ctrl+o to expand)
+
+        ✻ Brewed for 2s
+
+        ❯
+        {STATUS_LINE}
     """),
         True,
     ),
@@ -818,8 +870,7 @@
     """),
         False,
     ),
-    # API error mid-buffer with no later `⎿` tool results — Claude is
-    # presumably retrying / waiting. Flag on.
+    # API error, then a `⏺` reply in the same turn: the API answered → off.
     (
         "error-then-only-text",
         textwrap.dedent(f"""\
@@ -832,7 +883,7 @@
         ❯
         {STATUS_LINE}
     """),
-        True,
+        False,
     ),
     # THI-436: a stream cut mid-response renders as an assistant `⏺` line,
     # not a `⎿` tool result — and a successful tool result sits above it,
@@ -889,6 +940,61 @@
     """),
         False,
     ),
+    # THI-455, from a live pane: collapsed tool calls leave no `⎿` to clear
+    # the flag, and a later prompt got a full reply → off.
+    (
+        "collapsed-calls-later-reply",
+        textwrap.dedent(f"""\
+        ❯ compare the clinic rankings from both sheets
+
+          Called faraday_internal 2 times (ctrl+o to expand)
+
+        ⏺ API Error: The response stopped arriving. The response above may be incomplete.
+
+          Called faraday_internal 2 times (ctrl+o to expand)
+
+        ⏺ Rankings by headroom count are nearly identical and track catchment size.
+
+        ❯ add the clinic names to these codes
+
+          Called faraday_internal (ctrl+o to expand)
+
+        ⏺ Here are the clinic codes from the comparison with their names.
+
+        ✻ Churned for 34s · done 8:33 AM
+
+        {RULE}
+        ❯ add this comparison to the handoff
+        {RULE}
+        {CUSTOM_STATUS_LINE}
+          ⏵⏵ auto mode on (shift+tab to cycle) · ← 2 agents
+    """),
+        False,
+    ),
+    # The same pane when that turn ended: Claude Code carried on after the
+    # stream cut with no prompt in between → off.
+    (
+        "stream-cut-recovered-in-turn",
+        textwrap.dedent(f"""\
+        ❯ compare the clinic rankings from both sheets
+
+          Called faraday_internal 2 times (ctrl+o to expand)
+
+        ⏺ API Error: The response stopped arriving. The response above may be incomplete.
+
+          Called faraday_internal 2 times (ctrl+o to expand)
+
+        ⏺ Rankings by headroom count are nearly identical and track catchment size.
+
+        ✻ Brewed for 2m 10s · done 8:07 AM
+
+        {RULE}
+        ❯
+        {RULE}
+        {CUSTOM_STATUS_LINE}
+    """),
+        False,
+    ),
 ]