[fdaviddpt]

Closes #1128.

The summarizer's nested claude -p inherited CLAUDE_CODE_EFFORT_LEVEL from the parent session's environment. At max effort the small save calls still finished but cost roughly 5x the output tokens, and the two larger calls (NDC compression, staging consolidation) never finished inside their 180s timeouts -- so now.md stopped rotating, grew past session_start_max_bytes, and SessionStart stopped injecting it.

_without_session_env() in pipeline/summarizer.py now pops/restores CLAUDE_CODE_EFFORT_LEVEL the same way it already pops/restores the other ten Claude Code session variables stripped before the single nested claude -p spawn site (call_haiku/_run, which save, NDC compression and staging consolidation all route through). CLAUDE_EFFORT is deliberately NOT stripped -- the issue reporter measured that unsetting only CLAUDE_CODE_EFFORT_LEVEL fixed the timeouts, and CLAUDE_EFFORT is not a Claude Code session variable the child would otherwise look like a resumable part of.

tests/test_summarizer_env_898.py gained CLAUDE_CODE_EFFORT_LEVEL in its EXPECTED_SESSION_NAMES tuple, which several existing AST-based and behavioral tests already iterate over -- this alone turned 5 existing tests red before the fix and green after -- plus one new explicit test with a positive control (asserts CLAUDE_EFFORT survives in the child's env while CLAUDE_CODE_EFFORT_LEVEL does not, and that the parent's own environment is restored afterward).

A second commit fixes two stale-prose findings a spawned reviewer caught on the diff: docs/configuration.md's enumeration of the ten stripped variables (now eleven, with a sentence on why CLAUDE_EFFORT is not also listed), and the test file's module comment claiming the tuple matches round 16's shipped list verbatim (now notes the #1128 addition and its different provenance).

Test evidence

RED (before the pipeline/summarizer.py fix, only the test-tuple edit applied): python3 -m pytest tests/test_summarizer_env_898.py -q -> 5 failed, 45 passed.

GREEN (after the fix): same command -> 50 passed, then 51 passed once the dedicated positive-control test was added. The one non-zero exit code seen on these runs is the repo's own coverage-threshold gate on a single-file targeted subset (fail-under=80) -- expected and unrelated to this change.

Also ran clean: tests/test_haiku.py, tests/test_literal_env_reads_898.py (the repo-wide AST scanner for non-literal env-var names), tests/test_arith_base_lint_332.py, tests/test_pep604_floor_guard.py.

Full pytest (the repository's whole test_command) was deliberately not run locally -- targeted only; CI is the full-suite gate across 3 OSes and 4 interpreters.

Review

Two spawns reviewed the fix commit: oss:auditor (classes A/B/C plus platform, all clean) and Explore (2 findings, both fixed in the follow-up commit, described above). No correctness issues found in the pop/restore logic itself; the Codex child-env route was confirmed never to have carried this variable, so the fix's single-spawn-site scope matches the bug.

Docs

README.md (this repo's configured docs target) never documented the env-stripping mechanism, so nothing in it needed updating. docs/configuration.md, which does document it in detail, was updated in the second commit even though it isn't in the configured docs_targets, because it was stale on the diff itself.

Not implemented here, flagged as an optional follow-up the issue itself marks as a "could also": having the repo's doctor command warn when CLAUDE_CODE_EFFORT_LEVEL is set in the environment the summarizer will inherit.

[AI-generated]