菜单
Run the screenshot regression test suite (33 cases) and report pass/fail/skip triplet against the baseline. Use when the user says "run the regression", "regression tests", "screenshot tests", or asks to confirm no rendering regressions before committing/merging.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Create a fresh agent worktree under .claude/worktrees/ off an up-to-date main, with the project's hygiene rules baked in. Use when the user says "spin up a worktree", "create a worktree for agent X", "set up a worktree for branch Y", or when about to dispatch an agent that needs an isolated work area.
Compare jnext vs CSpect at a specific Z80 PC using DZRP — captures regs, SP, MEM[SP..+15], and slot mapping at the same PC in both emulators and diffs the result. Use when the user says "compare jnext vs CSpect at PC X", "DZRP probe at X", "run a DZRP comparison", or describes a boot-time divergence that needs jnext-vs-CSpect diff at a specific code site.
Add an env-gated diagnostic probe at a specific code site in jnext (the JNEXT_G46B_* pattern — zero cost when env unset, logs to cpu-inst-log channel when on). Use when the user says "add a probe at X", "instrument X", "add a G46B probe", or describes wanting to log emulator state at a specific PC/port/NEXTREG/memory event.
Comprehensive CSpect debugging via DZRP for jnext-vs-CSpect differential investigations. Covers hardened launch lifecycle (timeout-wrapped mono + bracketing pkill), BP-spray PC-stream capture, DZRP API quirks, and when to escalate to a custom plugin. Use whenever invoking CSpect from a script, doing BP-spray differential debugging, or any G46(b) workflow that touches CSpect — including "launch cspect", "run cspect", "BP-spray", "find divergent PC", "diff jnext vs CSpect PC stream", "cspect dzrp".
Run one pass of a Task-2-style subsystem audit. Dispatches a `subsystem-auditor` agent per subsystem in scope (in parallel where possible), then dispatches independent `subsystem-reviewer` agents on each result. Use when the user says "run an audit pass" or "audit subsystem X".
Coalesce commits since the last git tag into a ChangeLog entry following the project's strict rules (4 sections, terse, no commit IDs, no trivial fixes). Use when the user says "update the ChangeLog" or as a step of /version-bump.
| name | regression |
| description | Run the screenshot regression test suite (33 cases) and report pass/fail/skip triplet against the baseline. Use when the user says "run the regression", "regression tests", "screenshot tests", or asks to confirm no rendering regressions before committing/merging. |
Run the jnext screenshot regression suite. 33 end-to-end pixel-comparison tests in test/00regression/.
Optional:
/home/jorgegv/src/spectrum/jnext).build/ — or build/gui-release/ if the change touched GUI code; see feedback_clean_gui_release_for_regression).Build if needed:
cmake --build $TARGET/build -j$(nproc) 2>&1 | tail -5
Run regression with the canonical entry, teeing the log (per feedback_make_regression_canonical + feedback_regression_log_to_file):
short=$(git -C $TARGET rev-parse --short HEAD)
log=/tmp/regression-$short.log
bash $TARGET/test/00regression/regression.sh 2>&1 | tee $log
Parse the tail for Pass / Fail / Skip counts.
Cache "P/F/S @ HEAD on " to $TARGET/.claude/last-test-triplet.txt. The SessionStart hook reads this.
## Regression
regression P/F/S on branch <branch> @ <short-sha>
Log: <path>
New failures vs baseline (33/0/0): <list or "none">
make regression is canonical (feedback_make_regression_canonical).feedback_regression_log_to_file).feedback_pixel_equivalence_for_ref_regen).build/gui-release/ (feedback_clean_gui_release_for_regression).JNEXT_TEST_JOBS=4 on every regression invocation (feedback_jnext_test_jobs). Never raise it for speed — audio-underrun-func and screenshot-paused-func are real-time-bounded and fail under CPU contention.feedback_regression_in_branches).regression-runner subagentIf the user wants the full triplet (ctest + FUSE + regression) rather than just the regression layer, use the test-triplet skill or dispatch the regression-runner agent. This skill runs only the regression layer.