Bash tool: output-end and out-of-memory notes for Claude, test outcomes and program lists in analytics

What

The Bash tool is the tool Claude uses to run shell commands. Several connected changes affect what it reports after a command runs:

• When a command fails, the end of its output can be re-attached to what Claude sees, and notes can be added to the command's error output (stderr).
• A command's result can now carry memoryLimitNote, which says the system killed a process because it ran out of memory.
• A result can carry triggeredNote, a piece of advice set by a rule from a server-side feature flag.
• A result can carry stdoutHoldsOutputEnd, which marks output that was shortened by leaving out the middle.
• These notes apply only to commands Claude runs in the main session. They do not apply to subagents (helper agents Claude starts), remote sessions or commands you type yourself, and some of these paths sit behind internal checks that were not traced.
• Claude Code now reads the output of test runners, including pytest, unittest, jest, vitest, mocha, go, cargo, maven and ctest, and classifies a run as passed, failed or unknown.
• Bash analytics now record which programs a command runs (bash_programs, which is a list, none or incomplete), the scope of a test run (bash_test_scope) and its result (bash_test_outcome).
• Success and failure records also gain output_end_shown, triggered_note_rule, triggered_note_shown, oom_kill_during_command, oom_kill_note_shown and reads_own_output.
• A new analytics record is sent when a background command settles, with CPU and memory-group (cgroup) use, outcome, stop cause, bash_command_kinds, programs, test scope and test outcome.
• The Read tool's analytics also gain reads_own_output.

The Bash tool's description, the text that tells Claude how to use it, is unchanged in the capture taken under this release.

Why

After a failed or killed command, Claude may now be told how the output ended and whether memory ran out, so it has more to go on before trying again. The analytics changes record what commands run and how tests turn out, and do not change what you see on screen.