Fix #1128: strip CLAUDE_CODE_EFFORT_LEVEL from the nested claude -p environment

changelog.d/1128.fixed.md

@@ -0,0 +1,16 @@
+- Fixed: the nested `claude -p` the summarizer spawns for a save, an NDC
+  compression, or a staging consolidation inherited `CLAUDE_CODE_EFFORT_LEVEL`
+  from the parent environment. Anyone who sets Claude's own effort to `max` in
+  `~/.claude/settings.json` (the only way to keep it at `max` across sessions)
+  had every nested Haiku call run at max effort instead of the plugin's own
+  choice: small saves still finished but cost about five times the output
+  tokens, and the larger NDC and consolidation calls stopped finishing inside
+  their 180s timeout at all. Memory kept working but got slower on every save,
+  then `now.md` stopped rotating once NDC itself always timed out, grew past
+  `session_start_max_bytes`, and SessionStart silently stopped injecting it --
+  with nothing in the logs explaining why. `CLAUDE_CODE_EFFORT_LEVEL` is now
+  removed from the nested call's environment the same way the plugin already
+  removes the parent session's own identity variables, and restored afterward.
+  `CLAUDE_EFFORT` is deliberately left alone: it is not a Claude Code session
+  variable, and unsetting only `CLAUDE_CODE_EFFORT_LEVEL` was enough to fix
+  the measured timeouts (#1128).

docs/configuration.md

@@ -110,10 +110,14 @@ does not pass as that session
 back afterwards: `CLAUDECODE`, `CLAUDE_JOB_DIR`, `CLAUDE_PROJECT_DIR`,
 `CLAUDE_CODE_SESSION_ID`, `CLAUDE_CODE_ENTRYPOINT`, `CLAUDE_CODE_CHILD_SESSION`,
 `CLAUDE_CODE_SESSION_ATTENDED`, `CLAUDE_CODE_EXECPATH`, `CLAUDE_CODE_MESSAGING_SOCKET`,
-`CLAUDE_CODE_SSE_PORT` -- the reason for each is a comment beside the code in
-`pipeline/summarizer.py`. `CLAUDE_CODE_MESSAGING_TOKEN`, the handshake paired with the messaging
-socket, is not on the list and reaches the nested call: with the socket removed, the nested
-call has no channel to present it on. Your login credential is never on that list
+`CLAUDE_CODE_SSE_PORT`, `CLAUDE_CODE_EFFORT_LEVEL` -- the reason for each is a comment beside
+the code in `pipeline/summarizer.py`. `CLAUDE_CODE_MESSAGING_TOKEN`, the handshake paired with
+the messaging socket, is not on the list and reaches the nested call: with the socket removed,
+the nested call has no channel to present it on. `CLAUDE_EFFORT` is not on the list either and
+also reaches the nested call -- only `CLAUDE_CODE_EFFORT_LEVEL` outranks `--effort` and every
+settings-file source, so only it needs stripping
+([#1128](https://github.com/Digital-Process-Tools/claude-remember/issues/1128)). Your login
+credential is never on that list
 ([#131](https://github.com/Digital-Process-Tools/claude-remember/issues/131)); provider
 settings such as Bedrock selection are not either, so they reach the nested call
 ([#316](https://github.com/Digital-Process-Tools/claude-remember/issues/316)). To keep any

pipeline/summarizer.py

@@ -382,6 +382,15 @@ def _without_session_env():
     saved_messaging_socket = os.environ.pop("CLAUDE_CODE_MESSAGING_SOCKET", None)
     # The parent's IDE connection; the child would attach to the user's IDE.
     saved_sse_port = os.environ.pop("CLAUDE_CODE_SSE_PORT", None)
+    # #1128: outranks --effort and every settings-file source, so a parent
+    # session that pins "max" (the only way to keep max effort across
+    # sessions, since effortLevel stops at xhigh) sends Haiku into max-effort
+    # thinking on a summarization task -- NDC and consolidation then blow
+    # their 180s call timeout. CLAUDE_EFFORT is deliberately NOT on this
+    # list: the reporter measured that unsetting only this one variable fixed
+    # it, and CLAUDE_EFFORT is not a Claude Code session variable the child
+    # would otherwise look like a resumable part of.
+    saved_effort_level = os.environ.pop("CLAUDE_CODE_EFFORT_LEVEL", None)
     try:
         yield
     finally:
@@ -405,6 +414,8 @@ def _without_session_env():
             os.environ["CLAUDE_CODE_MESSAGING_SOCKET"] = saved_messaging_socket
         if saved_sse_port is not None:
             os.environ["CLAUDE_CODE_SSE_PORT"] = saved_sse_port
+        if saved_effort_level is not None:
+            os.environ["CLAUDE_CODE_EFFORT_LEVEL"] = saved_effort_level
 
 
 @contextlib.contextmanager

tests/test_summarizer_env_898.py

@@ -55,7 +55,9 @@
 # #204 already covered, and the IDE port (reasoned: set when an IDE is
 # attached, not observed in that dump). Same names, same order as round 16's
 # shipped `haiku.strip_session_env`, minus the messaging handshake (round
-# 18). A future Claude Code session variable that is not here reaches the
+# 18), plus CLAUDE_CODE_EFFORT_LEVEL (#1128 -- not from round 16's dump or
+# `strip_session_env`; it outranks --effort and every settings-file source).
+# A future Claude Code session variable that is not here reaches the
 # summarizer until a release adds it to `_without_session_env` AND here.
 EXPECTED_SESSION_NAMES = (
     "CLAUDECODE",
@@ -68,7 +70,15 @@
     "CLAUDE_CODE_EXECPATH",
     "CLAUDE_CODE_MESSAGING_SOCKET",
     "CLAUDE_CODE_SSE_PORT",
+    # #1128: outranks --effort and every settings-file source; a parent that
+    # pins "max" for its own sessions sends Haiku into max-effort thinking on
+    # a summarization task, which blows the 180s call timeout.
+    "CLAUDE_CODE_EFFORT_LEVEL",
 )
+# #1128: the *other* effort variable. The reporter measured that unsetting only
+# CLAUDE_CODE_EFFORT_LEVEL fixed the timeouts and left this one alone -- it is
+# not a Claude Code session variable, so it is inherited like any other.
+OTHER_EFFORT_VAR = "CLAUDE_EFFORT"
 # Round 18: inherited, no longer on the list.
 MESSAGING_HANDSHAKE = "CLAUDE_CODE_MESSAGING_TOKEN"
 
@@ -265,6 +275,25 @@ def test_child_inherits_an_arbitrary_variable(mock_run, monkeypatch, isolated_co
     assert seen[-1]["env"].get("REMEMBER_TEST_ARBITRARY_898") == "kept"
 
 
+@patch("pipeline.summarizer.subprocess.run")
+def test_effort_level_is_stripped_other_effort_var_is_inherited(
+        mock_run, monkeypatch, isolated_config):
+    """#1128: CLAUDE_CODE_EFFORT_LEVEL outranks --effort and every settings
+    file, so a parent that pins "max" sends Haiku into max-effort thinking on
+    a summarization task and blows the call's own timeout. CLAUDE_EFFORT is
+    the positive control: the reporter measured that unsetting only the
+    _LEVEL variable fixed it, and it is not a Claude Code session variable,
+    so it must reach the child like any other."""
+    monkeypatch.setenv("CLAUDE_CODE_EFFORT_LEVEL", "max")
+    monkeypatch.setenv(OTHER_EFFORT_VAR, "max")
+    seen = _capture(mock_run)
+    call_haiku("p")
+    env = seen[-1]["env"]
+    assert "CLAUDE_CODE_EFFORT_LEVEL" not in env
+    assert env.get(OTHER_EFFORT_VAR) == "max", "positive control: not everything is stripped"
+    assert os.environ.get("CLAUDE_CODE_EFFORT_LEVEL") == "max", "restored after the spawn"
+
+
 @pytest.mark.parametrize("name", EXPECTED_SESSION_NAMES)
 @patch("pipeline.summarizer.subprocess.run")
 def test_each_listed_session_variable_is_removed(mock_run, name, monkeypatch, isolated_config):