sdsrss/agentsmd/agentsmd-perf-baseline

agentsmd-perf-baseline

Use the selected SKILL.md absolute path from the live skills list; never infer it from the process cwd. Define and call the adjacent launcher in the same shell. The launcher preserves selected bundle → manifest-owned standalone → versioned CLI identity checks and exports plugin context only for a verified selected bundle.

SKILL_MD="<selected SKILL.md absolute path from the live skills list>"
agentsmd_skill_run() {
  node "$(dirname "$SKILL_MD")/scripts/agentsmd-run.js" "$SKILL_MD" "$@"
}

Turns "the hooks add ~200–400 ms" (a guess) into a measured per-hook table. For each hook it times, over N runs, the median of OFF (DISABLE_AGENTSMD_HOOKS=1 → the hook exits at its kill-switch line: bash-spawn + startup floor) vs ON (the hook does its real work); delta = ON − OFF is the hook's own logic cost.

agentsmd_skill_run                    # all 19 hooks, 10 runs
agentsmd_skill_run --event=PreToolUse # all registered PreToolUse hooks
agentsmd_skill_run --runs=3 --json

• Report concurrent event wall time separately from aggregate process cost (the sum of separately measured ON medians). The registry has 5 PreToolUse:Bash hooks and 8 Stop hooks; the event-wide benchmark also includes the PreToolUse edit-journal entry, so it is not an exact Bash-matcher measurement.
• Hooks run against a synthetic non-triggering echo Bash event (the common per-call case, not the block path), in an isolated sandbox CODEX_HOME — measuring never writes to the live ~/.codex (§8.V3) and cleans up after itself (§8.V4).
• Synthetic measurements exclude Codex's harness IPC, model time, and real long-transcript or blocking paths. Do not turn aggregate cost or an event-wide group into measured end-to-end user latency.

Operator/dev tool, read-only on the live env. From the repo instead of an install: node scripts/perf-baseline.js.