{"session": "28585272-a6c8-4e89-9ca9-459772bc64e0", "env": "main-hedy-lockwell", "command": "journal helper done 4153 \"branch parallel-answers 7d5649870\" 2>&1|tail -1; journal work end 129 --how parallel-answers 2>&1|tail -1; journal helper report \"Alfred: 4153 is committed on branch parallel-answers (7d5649870, on origin/main 2.268.4); I ran no tests. What I chose: the free-threaded build first, as the report says, and it is enough, so I did not shard. Measured with a scratch in-process server (not on 8420-8429) holding 8 environments of 300 rows each, a separate client process asking dashboard, message, to-do and agent lists, each client on its own environment, 8 seconds a run. Today's setup (CPython 3.14.7 with the 1 ms switch interval): 1 client 177 requests/s, p95 23 ms; 8 clients 296 requests/s, p50 14 ms, p95 79 ms, so eight environments get 1.7 times one. Free-threaded 3.14.8 (python3.14t from uv, default switch interval): 1 client 286 requests/s, p95 8 ms; 8 clients 725 requests/s, p50 5 ms, p95 44 ms, which is 2.4 times today's throughput with eight clients, half the p95, and no errors in the server's output. What changed in the code: the server can start on a free-threaded interpreter named in JOURNAL_SERVER_PYTHON (engine/package.py server_entry, used by the launch in engine/viewer.py), and serve.py keeps the 1 ms switch interval only while the GIL is on. It is opt-in on purpose: the installed shim's Python is chosen at install, no free-threaded Python is on a user's machine by default, and I only measured reads; a soak with writes and hooks under load on python3.14t should come before it is the default. The module caches you named (SUMMARIES, STANDING, RECENT, session files) held under eight concurrent readers; I did not make them lock-safe for concurrent writers, so that audit is still owed before defaulting. Sharding stays the fallback if that audit finds a hard problem. Tests I wrote and did not run: test_the_server_starts_on_the_free_threaded_interpreter_when_one_is_named in tests/test_it_boots.py. python3.14t is installed by uv at ~/.local/share/uv/python/ and can be removed with uv python uninstall 3.14t.\" 2>&1|tail -1", "parts": ["journal helper done 4153 \"branch parallel-answers 7d5649870\" 2>&1|tail -1", "journal work end 129 --how parallel-answers 2>&1|tail -1", "journal helper report \"Alfred: 4153 is committed on branch parallel-answers (7d5649870, on origin/main 2.268.4); I ran no tests. What I chose: the free-threaded build first, as the report says, and it is enough, so I did not shard. Measured with a scratch in-process server (not on 8420-8429) holding 8 environments of 300 rows each, a separate client process asking dashboard, message, to-do and agent lists, each client on its own environment, 8 seconds a run. Today's setup (CPython 3.14.7 with the 1 ms switch interval): 1 client 177 requests/s, p95 23 ms; 8 clients 296 requests/s, p50 14 ms, p95 79 ms, so eight environments get 1.7 times one. Free-threaded 3.14.8 (python3.14t from uv, default switch interval): 1 client 286 requests/s, p95 8 ms; 8 clients 725 requests/s, p50 5 ms, p95 44 ms, which is 2.4 times today's throughput with eight clients, half the p95, and no errors in the server's output. What changed in the code: the server can start on a free-threaded interpreter named in JOURNAL_SERVER_PYTHON (engine/package.py server_entry, used by the launch in engine/viewer.py), and serve.py keeps the 1 ms switch interval only while the GIL is on. It is opt-in on purpose: the installed shim's Python is chosen at install, no free-threaded Python is on a user's machine by default, and I only measured reads; a soak with writes and hooks under load on python3.14t should come before it is the default. The module caches you named (SUMMARIES, STANDING, RECENT, session files) held under eight concurrent readers; I did not make them lock-safe for concurrent writers, so that audit is still owed before defaulting. Sharding stays the fallback if that audit finds a hard problem. Tests I wrote and did not run: test_the_server_starts_on_the_free_threaded_interpreter_when_one_is_named in tests/test_it_boots.py. python3.14t is installed by uv at ~/.local/share/uv/python/ and can be removed with uv python uninstall 3.14t.\" 2>&1|tail -1"]}