skill(libtmux-profiler): Tachyon profiling workflows for libtmux/tmuxp

[tony]

Summary

Adds a Claude Code skill at .claude/skills/libtmux-profiler/ that documents and automates Python 3.15's stdlib profiling.sampling (Tachyon) for libtmux/tmuxp work. Pure tooling addition — no runtime code changes, no public API impact.

What's in here

The skill itself

SKILL.md documents seven workflows for libtmux/tmuxp performance investigation:

• pstats top-N — terminal-only summary of dominant wall-time consumers
• Automated pstats diff — structured before/after comparison via the bundled diff-pstats.py helper (replaces Python 3.15.0a8's broken --diff-flamegraph)
• Single-call microbenchmark — tight-loop per-call timing of one libtmux API
• Heatmap — line-level attribution within a hot function
• Live TUI — top(1)-style real-time profiler
• Attach to a running pytest — diagnose hangs without restart
• Sampling mode selection guide (wall / cpu / gil / exception)

reference/flamegraph-reading.md is a visual-signature cheat sheet for plain and diff flamegraphs.

Helper scripts

Each script in .claude/skills/libtmux-profiler/scripts/ solves a single discrete problem:

setup-tachyon-venv.sh — idempotent bootstrap that creates a .venv-3.15 venv with libtmux installed editable plus the testing dep group. Skips the dev group because watchfiles has no Rust wheel for 3.15a8 yet.

init-profile-session.sh — builds the per-session output directory at /tmp/py-profiling/<TS>/<project>/<branch>/<name>/ and writes a README with HEAD short-sha so the artifact set stays reconstructable after the branch moves.

bench-libtmux-call.py — tight-loop microbench against a real tmux server. Uses a registry-of-callables design (BENCH_TARGETS dict) so the script can never execute caller-supplied Python expressions. Built-in targets: has_session, list_sessions, list_windows, session_name, show_options, list_panes.

diff-pstats.py — markdown table of cumtime deltas between two .pstats files, sorted by abs(Δ) so the biggest movers surface first. Paste-ready for PRs, issues, or chat.

Toolchain bump

.tool-versions adds Python 3.15.0a8 alongside the existing 3.10-3.14. The project's pinned Python (3.14) is unchanged — 3.15 is only used by the skill's .venv-3.15 venv.

.gitignore overrides the global .claude/ ignore so this skill ships with the repo.

What this is not

• No runtime code changes. Nothing in src/libtmux/ is touched.
• No new dependencies. The venv is created on demand by the bootstrap script.
• No CI changes. The skill is invoked manually by Claude Code agents working on the repo.

Verification

The skill is non-runtime, so verification is mostly "the bootstrap is idempotent and the helpers don't crash."

The setup script is idempotent — safe to re-run, skips work when the venv already exists:

$ bash .claude/skills/libtmux-profiler/scripts/setup-tachyon-venv.sh

Smoke-test the diff helper against any two .pstats files:

$ ./.venv-3.15/bin/python .claude/skills/libtmux-profiler/scripts/diff-pstats.py path/to/baseline.pstats path/to/current.pstats --top 10

Smoke-test the microbench against the default has_session target:

$ BENCH_ITERS=100 ./.venv-3.15/bin/python .claude/skills/libtmux-profiler/scripts/bench-libtmux-call.py

Test plan

• Bootstrap runs cleanly on a fresh checkout (requires Python 3.15.0a8 via mise)
• Each helper script runs without errors
• pytest is unaffected — same 885 + 1 skipped as on master

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 46.58%. Comparing base (f0ca2d7) to head (b6b40f1).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #663   +/-   ##
=======================================
  Coverage   46.58%   46.58%           
=======================================
  Files          22       22           
  Lines        2372     2372           
  Branches      390      390           
=======================================
  Hits         1105     1105           
  Misses       1098     1098           
  Partials      169      169           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.