mutation-night-watch

Mutation Night-Watch

Role: worker. Launches and schedules a report-only measurement run. It never generates a test, never invokes an LLM, and never commits — that is the mutation-kill agent's job, the next morning, reading this skill's output.

Naming note. The skill is named mutation-night-watch; its underlying scripts, output directory (reports/mutation-nightwatch/), and launchd label use mutation_nightwatch/mutation-nightwatch — this is an intentional, stable naming split; do not rename the on-disk paths without a migration plan (#1856).

Why this exists, and how it differs from mutation-kill

mutation-kill already has two modes: interactive (an agent generates tests turn by turn) and --headless (unattended, but every round still shells out to claude --print — LLM cost and API availability are both in the loop). Neither mode is meant to run overnight with nobody watching: a multi-hour unattended claude --print loop is expensive and, on a flaky network, silently stalls.

Night-watch is a third, LLM-free mode: measurement only. It runs the real mutation tool for each detected stack, scores the result with the same mutation_report.py the rest of the mutation-testing skill family uses, and writes the result where mutation-kill (or a human) picks it up the next day. It does not replace mutation-kill's fixing modes — see mutation-testing/SKILL.md for those.

Constraints

• Report-only. Never generation. The script (../mutation-testing/scripts/mutation_nightwatch.py) has no code path that shells to an LLM. If a stack's tool run mutates a survivor count, that is the mutation tool itself running normally — nothing here writes a test.
• Mechanical repair only, never code. Before measuring, the script runs each stack's package-pin-sync step (npm ci when node_modules is missing, dotnet restore) so a stale dependency tree doesn't fail the whole run on a mechanical break. It never edits a production or test file, and never commits — see the script's mechanical_repair.
• Detected-but-unscored stacks are reported, not guessed. pitest (Java) and go-mutesting (Go) are detected but have no mutation_report.py parser yet; they show up in the summary as skipped_no_parser, never a fabricated score.
• Always confirm scope before scheduling a recurring run. A single ad hoc --detach launch needs no extra confirmation beyond what the user already asked for; wiring a scheduler entry (Task Scheduler / launchd / cron) that runs unattended on a recurring cadence is a standing change to the user's machine — confirm the schedule and scope before writing the scheduler config.