doctor --bench, machine health, recommended fix, and the speed-setup skill

CHANGELOG.md

@@ -19,6 +19,7 @@ Sections dated before 2026-09-19 predate the cycle and stay as they are. Named m
 
 ## Unreleased
 
+- new(doctor, speed-setup): **`bin/doctor.sh --bench [<project>]`** times 50 process starts, git status, python/node start, one gate-check, an mxcli check and (on a COPY of the model) mx check against measured reference constants, prints one verdict line and writes `.claude/loop/bench-<timestamp>.tsv` for comparing machines; a **Machine health** section (normal run and `--bench`) warns on leftover mxbuild `deno` helpers and free memory under 2 GB (mxbuild is killed with exit 137 when memory runs out), `--cleanup` stops exactly those helpers after a typed y (no tty: prints the command); both runs end with ONE **Recommended fix** block (Gate mode: check, else WSL, dev container, or Claude in the cloud; antivirus-exclusion advice removed, it needed IT); new `skills/speed-setup.md`. Field run: ab-581/gk on Linux 4 cores, 46 orphan deno helpers found, bench 10.4 s vs 10.9 s reference — Maurits (mauritscgvisser)
 - perf(exec): **~55% fewer forks and ~60% fewer process starts per exec** (Windows: ~100 ms each; follow-up to #584/#585). Measured on a two-module Purchasing/Approval app (Mendix 11.13, `Gate mode: check`), Linux, `strace -f` counting clone+execve for one `./bin/exec.sh`: forks 733 → 335, execs 350-375 → 135, 3 runs each; verdicts, exit codes, output and the BUILD-LOG row are byte-identical. What changed: **exec.sh** resolves the model, platform, Python and toolkit root once and hands them down as exported memos (`MXTK_MPR_MEMO`, `MXTK_MODEL_DIR_MEMO`, `MXTK_PLATFORM`, `MXTK_PY_MEMO`), uses `EPOCHSECONDS` / `printf %(…)T` / `${x##*/}` instead of `date` / `basename` / `dirname`, reads PROJECT.md and CLAUDE.local.md fields in-shell, and parses lint-gate's output with ONE awk instead of ~10 grep/sed/tr/sort/paste; **_common.sh** waits for an unbounded gate with `wait` instead of a `sleep 1` process per second (10 on a warm build, 150+ on a cold Windows one), reads the model version with one Python start instead of two, and no longer starts basename/cut/sed/grep to match a Studio Pro folder to a version; **model-stamp.sh** asks git once (it asked 6 times), keeps the file list in an array instead of mktemp/paste/rm, reads the stamp once, and `clear` answers before any set-up; **snapshot-mpr.sh**, **look-ledger.sh**, **script-owners.sh**, **metrics-ledger.sh**, **row-groups.sh**, **build-plan-status.sh**, **lint-gate.sh**, **exec-approval.sh**, **mdl-scan.sh** drop their dirname/basename/date/ls|wc|tr/find|wc forks; lint-gate skips its rule-receipt Python while no rule file, receipt or disabled-list is newer than the last clean check. `MXTK_PROFILE=2` traces the scripts exec.sh starts as well (=1 counts only exec.sh's own commands, so a started script is one line there), and the profile lines now name the file. Reported by mkr87. — mkr87 (#584, #585)
 - new(doctor, init): **`bin/doctor.sh <project>` measures performance, not just presence**: a "Performance (measured)" section reads the last 20 `steps:` rows of the project's BUILD-LOG and reports the median exec time, the median gate and the slowest step against a normal machine (17-34 s per exec, gate 9-20 s), warning above 60 s per exec or 45 s gate with the lever to pull; on Windows with slow process starts it recommends (never requires) WSL2 or a dev container (Docker Engine in WSL2, Rancher Desktop, Podman Desktop). `Gate mode: check` is the default for every project, existing ones included after sync (`Gate mode: full` in PROJECT.md keeps a full build per exec). Windows field run (ExpenseTracker, Windows 11, Git Bash, 103 ms per process start): check-only execs took 90-103 s against 138-139 s with the full build (gate 60 s → 12-14 s); the 60-70 s of non-gate overhead is the next target. Reported by mkr87. — mkr87 (#584, #585)
 - new(exec): **check-only gate by default (`Gate mode: full` in PROJECT.md, or `MXTK_GATE_MODE=full`, opts out)**: each exec runs `mx check` (beside mxbuild, same version) instead of the full deploy build, converting its JSON into mxbuild's errors shape so the delta gate, rollback and report are unchanged. The full build still runs when Java/JavaScript/theme/widgets changed since the last full build, when the project has Java actions and the script drops or renames something (`drop … if exists`, the re-create idiom, does not count), on `--patch`, before any full build is recorded, and after 10 check-only passes in a row (`MXTK_CHECK_MAX`). The other full builds sit at the work's natural boundaries, as mxcli itself advises (`mx check` after every exec, mxbuild when packaging): a new pre-push hook and `module-close.sh` run `bin/verify-model.sh` when the stamp says the model was only `mx check`ed (the stamp now records `kind: check|full`; `model-stamp.sh check --full`), and `bin/verify-model.sh` closes a batch on demand. Commits stay allowed on a check stamp. Field run (purchase-requests app, Linux, Mendix 11.13.0): gate 20 s → 9-10 s; a push after check-only execs ran the full build (20 s) and re-stamped the model as full; a dropped attribute caught as 6 × CE1613 by `mx check` on the broken model. Golden input: captured `mx check -j` output in `tests/wave2/fixtures/mx-check/`. For #584/#585 (Windows, mkr87). — mkr87 (#584, #585)

README.md

@@ -782,6 +782,7 @@ Every mxcli project has a `.ai-context/skills/` directory (bundled by `mxcli ini
 | Needing Studio Pro load evidence without a human at the GUI — direct-binary launch and log capture; a capture technique, NOT a validated pass/fail oracle | `skills/scriptable-sp-verification.md` |
 | A runtime test or seed reads/writes data that then is not there (the test wrote a row, the DB query shows nothing), or vice versa — three local Postgres instances can answer on this box; resolve the real port from the project's own compose file first | `skills/learned-local-db-confusion.md` |
 | gate-check.sh reports a Stage 0 file not found that plainly exists — ANALYSIS_BASE falls back to project root until Stage 1; move the file, do not debug the script | `skills/gate-check-file-locations.md` |
+| The user says the toolkit, an exec or the gate is slow, or asks to update the toolkit, run doctor and check whether it can be faster on their machine (werk de toolkit bij, draai doctor, controleer of het sneller kan) | `skills/speed-setup.md` |
 
 **Reference — lookup tables, not method**
 

ROUTING.md

@@ -270,6 +270,7 @@ picks the row up. That is the whole procedure — there is no second list to rem
 | gate-check.sh reports a Stage 0 file not found that plainly exists — ANALYSIS_BASE falls back to project root until Stage 1; move the file, do not debug the script | `skills/gate-check-file-locations.md` | ba,gate | 0 | ondemand |
 | A CE error (mx check or mxbuild: CE0066, CE0117…), a cited BUG-n, or behavior that looks like a known mxcli quirk rather than a modeling mistake — `bin/bug-lookup.sh CE0117` / `BUG-102` / "keyword" / the whole pasted error message (ranked top 3) prints the matching ledger entries, so the session reads one entry, not the whole ledger | `bin/bug-lookup.sh` | mdl,gate | 5,6 | baseline |
 | You hit, or worked around, a bug in mxcli or the toolkit — `bin/file-bugs.sh` (the `--add` line is in learned-detection-gaps.md) keeps docs/bugs.jsonl, redacts, shows duplicates, is filed only by a human with `!` (`--file PBUG-NN`, y/N, denied to agents), or prints new-issue links (`--links`, `--json`) | `bin/file-bugs.sh` | mdl,gate | 5,6 | baseline |
+| The user says the toolkit, an exec or the gate is slow, or asks to update the toolkit, run doctor and check whether it can be faster on their machine (werk de toolkit bij, draai doctor, controleer of het sneller kan) | `skills/speed-setup.md` | lead | - | ondemand |
 
 #### Reference — lookup tables, not method