chore(context): per-role on-demand token in the routing agents column (stub diet T1) — stacked on #378

CHANGELOG.md

@@ -19,6 +19,8 @@ Sections dated before 2026-09-19 predate the cycle and stay as they are. Named m
 
 ## Unreleased
 
+- chore(context): **a routing row can be on demand for one role: `mdl?` in the agents column — mdl at Stage 5 reads 43,599 → 30,647 words, review 32,181 → 30,265.** The tier column is per row, so a row one role needs every session and another needs once per module was baseline for both. The `<role>?` token keeps a baseline row in that role's stub, among the on-demand rows (trigger only, still findable), and out of its always-read set and word count; the Role(s) column names it (`review (mdl: on demand)`, `every role except mdl, review (on demand for those)`), and the agent audit refuses a `?` on a non-baseline row or beside its own plain token. Applied for mdl to module-review, ui-loop, teamserver-alignment, learned-css-that-never-applied and skills-over-scripts; for review to teamserver-alignment and skills-over-scripts. Baseline total unchanged (73,233). Behaviour A/B (5 reps × 2 tasks + 5 probes × 3 reps, sonnet): trimmed pack no worse than master on every scored metric — planted STOP refused 10/10 each arm, mxcli-check parse failures page 2/5 vs 3/5, retrieval probes 15/15. — context-cost research (r-stub-diet, T1)
+- chore(context): **helpers read the runbook slice for their stage, and the runbook no longer carries a copy of the baseline table — mdl at Stage 5 reads 56,236 → 43,599 words; baseline 75,028 → 73,233.** `skills/conversion-runbook.md` held a fourth rendered copy of the baseline routing table (README, ROUTING.md and every project's CLAUDE.local.md carry it too); it is out of the render surfaces and replaced by a one-line pointer. New `bin/runbook-slice.sh <stage>` prints §1b plus that stage's section, using the span logic gate-check already had (now `bin/lib/runbook-span.sh`, shared). The runbook routing row is split: lead, ba and architect keep the runbook; mdl, gate, test and review get the slice, and `render-routing.sh --check` counts the slice's printed words for them, so the per-role numbers stay honest. New `tests/bench/trim-noloss.sh` checks that every removed line is found again or is a still-routed table row. Behaviour A/B (5 reps × 2 tasks + 5 probes × 3 reps, sonnet): trimmed pack no worse than master on every scored metric — planted STOP refused 10/10 each arm, mxcli-check parse failures page 2/5 vs 3/5, retrieval probes 15/15. — context-cost research (r-stub-diet, T2+T3)
 - process(build): **Stage 5 drafts the ready set in parallel and applies serially.** The runbook dispatch row said "sequential within a module", which serialised drafting even for independent scripts; it now points to `iterative-build-loop.md` → Gate cost, the single statement of the rule. The unit is the build-plan row: every `BUILD` row whose "Depends on" rows are all applied, minus overlapping writers (two rows that replace, alter or drop+create the same document never share a ready set, and no row is drafted while it overlaps a queued unapplied one; `brd-to-build-plan.md` has Produces name every document a row writes), runs at once, its scripts in order, one `mdl-agent` per script. Applies stay one at a time through `bin/exec.sh`, and the ready set is recomputed after each. A queued draft made before a sibling apply landed is re-checked with `mxcli diff --format struct` (preflight STOP rule 23); any removal line (checked against `describe` per BUG-146) or STOP rule 21 target means re-draft, never patch, because a stale restatement of a shared document is exactly rule 21's silent seed-call loss and passes `check --references`. Report `ready set N, drafted N, applied k`; the `mdl-agent` template says a drafter never applies and never depends on an unapplied draft. Field evidence: none yet. — MendixMau
 - fix(platform-link): **CLI commits now show their changed documents in Studio Pro History.** Studio Pro's History → Model changes pane reads a `ModelChanges` array from the commit's `refs/notes/mx_metadata` note. Trap 6c's hand-written template set it to `[]`, so every mxcli/agent commit showed "Model changes 0" (field project: 270 of 270 notes empty). New `project-bin/mx-metadata-note.py` (stdlib only) computes the array from the commit's changed `.mxunit` files, in Studio Pro's entry shape: Status, UnitID, UnitType, UnitName and Module (walked up the `.mpr` Unit tree). It has `changes` / `write` / `backfill` / `sync <remote>` subcommands. `install-project-hooks.sh` now also installs post-commit and post-rewrite hooks that write the note and never block, and `--check` covers them, so `sync-project.sh` rolls them out to wired projects. Notes are published with `mx-metadata-note.py sync <remote>` after the branch push. Push config is never touched. With mxcli v0.26+ in the project root the hooks call `mxcli git note --write` instead (mendixlabs/mxcli#1337); the Python script is the fallback for older pins, and on 160 field commits both wrote the same 3,594 entries. Trap 6c rewritten. Measured against Studio Pro's own notes, 51 commits in 5 apps: 3387/3389 entries exact on all five fields. Field run on a 357-commit Mendix 11.15.0 app: 158 notes backfilled and pushed, and Studio Pro History lists the documents per revision. Review fixes: an earlier revision added `HEAD` + the notes ref to `remote.<r>.push`. It was dropped because any push refspec replaces `push.default`. A branch tracking a differently named upstream got a new same-named branch on Team Server where `simple` refuses, a missing local notes ref failed the whole push, and nobody knows whether Studio Pro's own push reads that config. The installer now only fetches the remote's notes (non-forced). It removes that exact `HEAD` + notes pair if present, and leaves and reports any other entry. `sync` handles a remote with no notes ref (it used to exit 1) and retries a rejected push once after re-merging. A note that already lists changes (Studio Pro's) is never overwritten by `write`/`backfill` without `--force`, and `sync` keeps a colleague's non-empty note over ours. The mxcli probe greps `git note --help` for `--write`. Hooks and docs call the script through `resolve_py`. Fixture `tests/wave2/test-mx-metadata-note.sh` (40 cases, over real captured template units; single- and two-tree layouts, amend, backfill and `--force`, push exit codes identical to a no-hooks clone, leftover-refspec cleanup, sync to an empty remote and its retry, notes fetch at install, a diverged notes merge; green under `init.defaultBranch` master and main). — field project (Mendix 11.15 app), MendixMau
 - learn(fixture-seeding): **explore the data structure first, choose rows by use case, scan actions for resets before any run.** Two seeding runs on public template apps: seeding raised journey passes 33 -> 58 and 20 -> 51 of the same sets (FAIL 93 -> 68, 163 -> 132; INVALID unchanged), and a "Reset all demodata" journey wiped accounts so 104 of 127 journeys reported `The username or password you entered is incorrect`, identical on empty and seeded databases. Skill gains three questions a seed must answer, a use-case checklist, a destructive-action scan, a generic traps table with symptoms, measured expectations and a denominator checklist; routing trigger now fires on seed/test/demo data and empty pages. Independent review fixes: object ids follow the Mendix bigint pattern with a business-key guard (no uuids), associations are seeded where the page's path stores them, four new traps (enum names, inheritance chain, calculated attributes, view/external entities), and the port-offset and container-label traps moved to `harness-architecture.md`. — MendixMau

README.md

@@ -771,11 +771,12 @@ Then filter by Role(s). A helper dispatched with a role stub (ba, architect, mdl
 
 | Always relevant for | Reference this | Stage(s) | Role(s) |
 |---|---|---|---|
-| Before writing any .js or .sh for a check, gate or report — and before adding a rule to an existing one: judgement goes in a skill, code only fetches facts a reader cannot | `skills/skills-over-scripts.md` | every stage | every role |
+| Before writing any .js or .sh for a check, gate or report — and before adding a rule to an existing one: judgement goes in a skill, code only fetches facts a reader cannot | `skills/skills-over-scripts.md` | every stage | every role except mdl, review (on demand for those) |
 | Any pass whose input is missing, stale or unresolvable — before recording UNMEASURED, N/A or a silent skip: name what was missing, say what you assessed against instead, still deliver a verdict | `skills/degrade-to-judgement.md` | every stage | every role |
 | Any time an exit code, a tool's output or a subagent's report is about to become a stated finding — verify before you conclude | `skills/tool-output-is-not-ground-truth.md` | every stage | every role |
 | Before obeying any learned-* STOP or workaround that costs a detour — probe the binary you actually have, then stamp the verdict back into the rule | `skills/retesting-learned-rules.md` | every stage | every role |
-| Any pipeline work at all — every session, before producing any stage artifact: read §1b plus your own stage's section, not the whole file (gate-check.sh prints the line spans); READMEs and the guide are orientation only | `skills/conversion-runbook.md` | P,0,1,2,3,4,5,6,7 | every role |
+| Any pipeline work at all — every session, before producing any stage artifact: read §1b plus your own stage's section, not the whole file (gate-check.sh prints the line spans); READMEs and the guide are orientation only | `skills/conversion-runbook.md` | P,0,1,2,3,4,5,6,7 | lead, ba, architect |
+| Any pipeline work at all, as a helper — every dispatch, before producing anything: run the toolkit's bin/runbook-slice.sh <stage>, which prints §1b plus your stage's section of the conversion runbook; never open skills/conversion-runbook.md whole | `bin/runbook-slice.sh` | P,0,1,2,3,4,5,6,7 | mdl, gate, test, review |
 | Any question before asking the user or writing anything — query the model, then read the source, then ask the human, in that order | `skills/query-the-model.md` | P,0,1,2,3,4,5,6,7 | every role |
 | Putting a question TO the user — any gate, any stage: ask in chat not in a file, two named options plus your recommendation, one batch per gate then end the turn | `skills/interview-protocol.md` | P,0,1,2,3,4,5,6,7 | lead, ba, architect |
 | The user gives feedback or asks for a change mid-project — "change X", "it also needs Y", a remark on a screenshot or the running app, a stakeholder's ask relayed in chat: log a CR row in PROJECT.md ## Change log BEFORE any edit, size it, then edit the highest artifact it alters (BRD, decision, architecture, wireframe), never the MDL first | `skills/change-intake.md` | 2,3,4,5,6,7 | lead, ba |
@@ -796,20 +797,20 @@ Then filter by Role(s). A helper dispatched with a role stub (ba, architect, mdl
 | Writing a microflow with any loop, a retrieve/commit/call inside a loop, nested or multiple loops, >20 activities counting loop bodies, or a list built from a list — post the checklist before the first MDL line | `skills/microflow-preflight.md` | 5 | mdl |
 | Building any page or snippet — before the first widget. Wireframe, tokens, gallery reuse, cross-check; no wireframe means STOP | `skills/ui-preflight-pages.md` | 5 | mdl |
 | Writing or reviewing any page or snippet — the spacing scale (8/16/24/32/48), section rhythm, and the page-header scaffold every full page starts with; sections at 0px apart and pages with no H1 are the defects it retires | `skills/design-spacing.md` | 5 | mdl, review |
-| After every page-building script, and any time the UI looks wrong — the cheap repeatable look during the build: one page, one screenshot, four questions, scored when a wireframe exists. Feeds Gate: UI, never replaces it | `skills/ui-loop.md` | 5 | mdl, review |
+| After every page-building script, and any time the UI looks wrong — the cheap repeatable look during the build: one page, one screenshot, four questions, scored when a wireframe exists. Feeds Gate: UI, never replaces it | `skills/ui-loop.md` | 5 | review (mdl: on demand) |
 | After the FIRST build that follows any design-system port, and before any page is built on it — reads the BUILT stylesheet and reports how many framework knobs point at a design token, how many tokens arrived, how many component classes arrived, each with its denominator. Measured on a real run: 55 tokens ported correctly into the right file, 0 of 35 knobs bound and 0 of 20 classes present, two build phases shipped in the framework's default blue with mx check, mxcli lint, the MDL suite and two e2e journeys all green | `project-bin/check-design-reaches-app.sh` | 3,5 | architect, mdl, gate, review |
 | Before exec'ing ANY page script — compares the drafted MDL's shell against the wireframe's: page column, layout/nav shell, one H1. Measured 0/10 pages on a real first build, repaired wholesale 47 scripts later | `project-bin/check-page-shell.sh` | 5 | mdl, gate, review |
 | After drafting and again after exec'ing any page script — scores the page MDL (or `mxcli describe` output on stdin) against its wireframe: headings/actions/content/classes, weighted. The scored companion to check-page-shell's binary gate; 32% median measured without it, 90% first-draft with it. Every run is appended to the project's docs/PAGE-FIDELITY.tsv — first non-stub row per page = first-build score of record vs the ≥80% target (forward-reference stubs score with --stub, exempt) | `project-bin/page-fidelity.js` | 5 | mdl, gate, review |
-| Reviewing any module before calling it done — the ONE pass: build, gate, prove, LOOK (is it logical, does it look right, does it match our design, over every page not just the tested ones), confirm with the denominator stated | `skills/module-review.md` | 5,6 | mdl, review, test |
+| Reviewing any module before calling it done — the ONE pass: build, gate, prove, LOOK (is it logical, does it look right, does it match our design, over every page not just the tested ones), confirm with the denominator stated | `skills/module-review.md` | 5,6 | review, test (mdl: on demand) |
 | Before any mxcli exec / exec.sh / --mcp write — ask or run? the knob decides | `bin/exec-approval.sh` | 5,6 | mdl, gate |
 | Before calling any module tested, or before re-running tests after a change — what testing a module means, which re-runs a change owes (UI-only vs logic lane), and the false-green register of confirmed ways a test reports green over a broken feature | `skills/testing-shape.md` | 5,6 | test, gate, review |
 | Finishing any module — before calling it done. One command that runs every instrument and keeps "instrument faulted" apart from "feature failed"; in a wired project run the installed copy at bin/verify-module.sh | `project-bin/verify-module.sh` | 5,6 | mdl, gate, test, review |
-| A style change that appears to have done nothing, a focus ring that paints on every page landing, or an app still grey after a design port every instrument called green — the three ways a correct rule paints nothing (matches nothing / matches chrome / loses the cascade), the two reads that tell them apart, and the class that arrived in the stylesheet and is bound to no widget | `skills/learned-css-that-never-applied.md` | 5,6 | mdl, review, gate |
+| A style change that appears to have done nothing, a focus ring that paints on every page landing, or an app still grey after a design port every instrument called green — the three ways a correct rule paints nothing (matches nothing / matches chrome / loses the cascade), the two reads that tell them apart, and the class that arrived in the stylesheet and is bound to no widget | `skills/learned-css-that-never-applied.md` | 5,6 | review, gate (mdl: on demand) |
 | Before trusting a green check/exec/DESCRIBE result as proof, or when a runtime symptom appears over a fully green model — the register of constructs that pass early rungs and fail later ones | `skills/learned-detection-gaps.md` | 5,6 | mdl, gate, review |
 | Creating any entity, or calling a module security-ready — entity and grants land in one script, and ready means SHOW SECURITY MATRIX proves it | `skills/security-is-not-a-later-script.md` | 5 | mdl, gate, review |
 | A CE error or behavior that looks like a known mxcli quirk rather than a modeling mistake — `bin/bug-lookup.sh CE0117` / `BUG-102` / "keyword" prints the matching ledger entries, so the session reads one entry, not the 32k-word ledger | `bin/bug-lookup.sh` | 5,6 | mdl, gate |
 | The first command of every session, and any time someone asks "where are we" or "what next" — one screen: stage, done/overdue, the ONE next action, from the instruments, never from memory | `bin/status.sh` | P,0,1,2,3,4,5,6,7 | every role |
-| Any session that will push a model to Mendix Team Server, and BEFORE telling the user a Team Server push is blocked — which remote is authoritative, settle-then-push order, and the four checks that have to fail first | `skills/teamserver-alignment.md` | 5,6,7 | mdl, gate, review |
+| Any session that will push a model to Mendix Team Server, and BEFORE telling the user a Team Server push is blocked — which remote is authoritative, settle-then-push order, and the four checks that have to fail first | `skills/teamserver-alignment.md` | 5,6,7 | gate (mdl, review: on demand) |
 <!-- ROUTING:END -->
 
 **Why this has to be explicit instead of implicit:** a project's own skill files are usually written before a given toolkit learning exists, or before a new one is added later — they never grow a cross-reference to it on their own. When you `git pull` this toolkit and it brings in a new baseline-worthy skill (most often a new `learned-*.md`), update every consuming project's routing to match — don't assume the next session will find it by chance.