2026-08-29-steps-in-code-order.md 29 KB

Steps, in the code's order — plan

Status: BUILT 2026-08-29 on feature/steps-servers, P0–P6. Written the same day (at tip 5797d7e) as a plan for a fresh session; what follows is that plan, kept as written, with the notes below on where the build differs from it. The reading is &view=order on the Steps tab, src/ui-server/api/program.ts + ui/src/lib/program-model.ts, and spec §3.13.1 is the description of what was built.

Where the build differs from the plan:

  • A guard carries the decision it belongs to (BranchGuard.branch), not just its text and line — §4.1's "same line, same text" pairing does not tell one switch case from another, nor two try/catch blocks apart. It also carries how the arm it is in leaves (armExit) and, for an early exit, how the arm not taken leaves (exit), which is where WireArm.ends comes from.
  • A function is read ONCE per rail (again on the item), not redrawn at every call: expanding per path turned an 87-step screen into 3,849 items and 618 KB. Once-only is 476 items and 94 KB (+3% wall clock on that picture).
  • A step the walk entered reads on into its own body under its box, so the rail holds the same steps the tree does (§4.1.7 only said this for boundaries under through).
  • WireItem's blocks are one kind with a discriminator (block: 'inline' | 'loop' | 'later' | 'together') rather than four item kinds, and they carry facts (by, via, loop) rather than words — the viewer says them.
  • The reading is drawn on the CANVAS, not as a nested document. §4.2's rail — a column of boxes with forks as rows of arm columns — was built first and rejected on sight by the maintainer: "this is very hard to read… go back to the way it looked, but since the 200 and 401 come after the signing of the jwt those should branch out of it." The right picture is the same Svelte Flow canvas with the same boxes, laid out by WHEN things happen: a line means and then, a row down is one more thing already done, and the fork's condition rides on the line (drawn at rest — on this picture the conditions are the content). ui/src/lib/program-model.ts turns the block tree into that graph; StepsRail/RailBlock are gone.
  • Loops needed their own reading (loopsForFile), and loops and forks nest by which construct BEGINS first, since neither reading knows about the other.
  • The open questions of §7 were answered: order for functions/endpoints and the tree for screens (1); read each function once rather than capping the fold depth (2); a loop's body once, marked (3); the fork carries its condition and the arms say WHEN / WHEN NOT (4).

The Steps tab as it stood (spec §3.13, src/ui-server/api/steps.ts, ui/src/views/StepsView.svelte) is the base; everything here is a second reading of the same walk, not a new walk.

The ask, in the maintainer's words: on proshop's POST /api/users/login the picture draws User.findOne · jwt.sign · 200 · 401 in one row under the handler. That is true — all four are one step from authUser — and it is not what a person wants from a handler. They want the flow of the code: first the lookup, then if the password matches, sign a token and answer 200, else answer 401. "I think this should show the flow/order of the code. That would be way more useful to me." Build that.

Read first, in this order: spec §3.13 and §3.14 (docs/design/codegraph-ui-design-spec.md), then docs/plans/2026-08-28-steps-and-screens-for-apis-and-web.md §2 and §6 (how the pictures work and the gotchas), then the auto-memory note steps-for-apis-2026-08-28 (the two facts that made Steps work, the screenshot harness), then the files in §3 below. CLAUDE.md for tests, docs and the no-kernel-work rule (this plan is all request-time: no extractor changes).


1. The reading to build

For authUser in bradtraversy/proshop_mern (backend/controllers/userController.js):

const user = await User.findOne({ email })
if (user && (await user.matchPassword(password))) {
  res.json({ _id: user._id, …, token: generateToken(user._id) })
} else {
  res.status(401)
  throw new Error('Invalid email or password')
}

the picture reads top to bottom, as a rail that forks where the code forks:

● POST /api/users/login · authUser          FIRES FROM POST /api/users/login
│
├─ User.findOne({ email })                  database · User · read
│
╞═ user && (await user.matchPassword(password))
│    ┌─ WHEN ──────────────────────────┐   ┌─ WHEN NOT ─────────────────────────┐
│    │ jwt.sign({ id }, …)  auth        │   │ 401 · res.status(401)   response    │
│    │   via generateToken · inside     │   │   then throw new Error('Invalid …') │
│    │   res.json(…)                    │   └─────────────────────────────────────┘
│    │ 200 · res.json({ _id, …, token })│
│    └─────────────────────────────────┘

Every reading the Steps tab has today survives on each box — kind, FIRES FROM, via, WHEN, what it passes, the status, inside res.json(…) — but the arrangement is the code's: sequence down the rail, branches as forks, an arm that replies (or returns / throws) ends there, arms that fall through rejoin. A helper the walk folded (generateToken) is drawn in place: its own steps appear where it is called, marked via generateToken. A boundary (another endpoint the handler calls, another screen) sits in the sequence where the call is, still not entered unless through asks.

The reading is the outline of a function, not a control-flow graph: statements in source order, if / else / switch / try / early exits as forks, loops as a marked block (the body once, "for each …"), callbacks that run later (.then, setTimeout, useEffect) as a marked block ("later"), Promise.all([a(), b()]) as siblings marked "together". Source order is execution order for straight-line code and for arguments before their call; where it is not (callbacks, concurrency) the block says so rather than pretending.

What stays: the existing tree ("what it sets in motion", rows = distance) remains as the other reading, one toggle away — it is the right picture for a screen, where handlers fire on events and have no order. Default: in order when the anchor's root is a function / method (a handler, an endpoint, any symbol); the tree when the anchor is a screen. Both read the same URL (&view=order / &view=tree), the same panel, the same double-click and Start here →.


2. Why the current picture cannot say it

buildSteps is breadth-first from the anchor's root: a node is a step when it is a screen, a handler, a bridge, an event, a store action or an effect; everything else folds into the link's via. Rows are distance from the anchor. Since 2026-08-29 a row is ordered by the position of the hop that first reached each step, with a hop written inside another call's arguments before that call (WireStep.order, hopCompare in steps.ts) — that is how jwt.sign sits left of 200. The link carries the whole guard chain as one when string and the enclosing call as within.

So the walk already knows, per step, where it is reached (line, column, span) and under what (the guard chain). What it throws away is the structure between siblings: which guards are the same if, which arm is the else of which, what ends an arm. The tree cannot draw authUser → 200 under jwt.sign because that would be a lie (a jwt.sign → 200 edge); the rail can, because a rail is not a graph of causes but the order of statements.


3. Facts to build on (verified 2026-08-29; re-verify, these files move)

Need Where it is today Notes
The steps, links, sites, via, when, trigger, status, order, within src/ui-server/api/steps.ts buildStepsWireStepsPayload (mirrored in ui/src/lib/wire.ts) one box per (function, category), per status for replies; boundaries cut: 'screen'; through=1
Per-site position + span + enclosing call graph/branch-guards.ts callSitesForFileCallSiteText { callee, args, argList, span, within, status } span = the call's start/end; within = the call whose arguments contain it (stops at a function / block)
Per-site guard chain, structured guardsForFileBranchGuard[] (text, negated, form, line), joined by guardLabel into the when string the plan needs the array, not the stringSiteReader.when returns the string; add guards(caller, site): BranchGuard[] beside it (api/when.ts)
The rules that produce guards, per language RULES_BY_LANGUAGE / interface Rules in branch-guards.ts: boundaries, inlineFunctions, bindingParents, blocks, enclosing, earlyExits JS/TS, Swift, Python, Java, Kotlin, C#, Go, C; an early return / throw / raise before a site is already a negated guard in that site's chain (earlyExits)
What fires a site (later-running callbacks) triggersForFileSiteTrigger { kind: 'prop' \| 'option' \| 'callback', name, of }; LATER_CALLEES kind: 'callback' of then / setTimeout / useEffect / addListener = runs later
The hop that first reached a step HopSite in steps.ts (file, line, column, end, within), carried on Fold.first, sorted by hopCompare / hopInside this is the position the rail places a step at
Words for conditions ui/src/lib/conditions.ts: clauses, splitTop, conditionTokens, whenTokens, commonTokens, scenarios WHEN / AND / OR / NOT as words; scenario rows per site
Box words and looks ui/src/lib/steps-model.ts (kindWord, stepLabel, stepSub, triggerWords), ui/src/components/steps/StepNode.svelte reuse the box verbatim; only the arrangement is new
The view, URL, panel ui/src/views/StepsView.svelte (rewrite, navigate, the panel with Start here →, the summary's through checkbox), ui/src/lib/navigation.ts stepsHref / StepsHrefOptions, ui/src/lib/router.svelte.ts, App.svelte add view to the options and the route
Tests to copy the shape of __tests__/ui-steps-api-servers.test.ts (four frameworks in one fixture), ui-steps-cross-tier.test.ts, nextjs.test.ts, ui-steps-model.test.ts (pure model), ui-conditions.test.ts
Headless pictures the auto-memory note: shoot.mjs (Playwright over codegraph ui --no-open, one shot per hash URL; never npm run build while a codegraph ui runs); scripts/try-repo.sh <preset> to clone + index + open proshop, express-realworld, nest-boilerplate, next-saas-starter, fastapi-template are the presets

The walk's caps and budgets (MAX_CALL_SITES, MAX_WHEN_SITES, fold depth, fan-out, steps per picture) apply unchanged; the rail draws the same steps the tree does, so it costs one more structured read per site, nothing more.


4. Design

4.1 The program: a block tree derived from the walk

Add to the payload (same request, view=order asks for it; the tree's steps / links stay so the panel works):

/** The anchor's body as the code reads: items in order, forks where it forks. */
interface WireProgram { root: WireBlock; truncated: number }
type WireBlock = WireItem[];
type WireItem =
  | { kind: 'step'; step: string /* WireStep.id */; link: string /* WireStepLink.id */; site: WireStepSite; via: WireNodeRef[]; within?: string }
  | { kind: 'fork'; on: string /* the condition as written */; form: 'if' | 'switch' | 'ternary' | 'try' | 'exit' | 'guard'; arms: WireArm[] }
  | { kind: 'loop'; on: string /* `item of items` */; body: WireBlock }
  | { kind: 'later'; by: string /* `then` / `useEffect` */; body: WireBlock }
  | { kind: 'together'; body: WireBlock }            // Promise.all / Promise.allSettled arguments
  | { kind: 'inline'; via: WireNodeRef; when: string; body: WireBlock }   // a folded helper, drawn where it is called
  | { kind: 'cut'; why: WireStep['cut'] };           // a cap or a boundary, where the walk stopped
interface WireArm { when: string /* this arm's own clause: `…` or `NOT …`, or `case 'x'` */; ends: 'reply' | 'return' | 'throw' | null; body: WireBlock }

How it is built (server, src/ui-server/api/program.ts, called from buildSteps when asked):

  1. Collect the sites per function. For the anchor's root and every function the walk folded into (the via chain members) or entered (a step with a root), take every hop and effect site the walk made from that function: the step it reached (or the helper it folded into), the site's position (span.start, column-exact), and its structured guard chain (guards(caller, site) — the new SiteReader method — outermost first). One structure per function, keyed by the function's node id; the walk already visits each function once, so record while walking rather than re-walking.
  2. Order each function's sites by hopCompare (source position; a site inside another site's arguments first).
  3. Fold the guard chains into forks. Walk the ordered sites keeping a stack of open forks. For a site with guard chain g1…gn: the longest prefix shared with the open stack stays open; deeper open forks close; each remaining guard opens a fork (for an ifform: 'if', on: g.text) with the arm this site is in (`when: g.negated ? NOT g.text

    : g.text). A later site whose chain has the *same* guard (sameline, sametext) with the oppositenegated joins the **same fork as its other arm** — that is howelseand the||-arm are found without a CFG; aswitch guard (form: 'switch') opens one arm percasetext. A negated guard thatearlyExitsproduced (areturn/throw before the site) is a fork whose *other* arm is the exit:arms: [{ when: g.text, ends: 'return' | 'throw', body: [] }, { when: NOT g.text, body: […] }]` — draw the exit as a terminal, not a box.

  4. End an arm. An arm whose last item is a reply step (effect.category === 'response') or whose guard came from a throw / return ends (ends); the rail does not rejoin below it.
  5. Inline the folds. A hop that folded into a helper becomes { kind: 'inline', via: helper, body: program(helper) } at the hop's position, recursively, to the walk's fold depth; the helper's own guards are relative to its body (the fold's outer whens are already on the position in the caller). A step reached through a helper appears only in the inline block (never also at the top level).
  6. Later and together. A site whose trigger is kind: 'callback' of a LATER_CALLEES callee, or that sits inside an inline function (the Rules.inlineFunctions climb) passed to .then / setTimeout / useEffect, goes under a later item at the position of the registering call; sites inside Promise.all([...]) arguments go under together. Loops: a site inside a for / while / forEach body goes under loop (the guard reader knows the loop node types per language — Rules.blocks includes loop bodies; add the loop header text as on).
  7. Boundaries and cuts are step items (a screen / endpoint step, cut: 'screen') — the rail shows them where the call is; through enters them by appending their program as an inline block, as the tree enters them today.

This is a derivation from data the walk has; the only new reading is the structured guard array. No CFG, no extractor, no kernel work. It is honest by construction: a fork exists only where a guard was read; a site with an unreadable chain (a language without rules, a drifted file) sits in sequence with no fork, which is the same silence the tree shows today.

4.2 The view

ui/src/views/StepsView.svelte gains the mode; the rail is a new component ui/src/components/steps/StepsRail.svelte (or SequenceView) rendered instead of the Svelte Flow canvas when view === 'order'. No layout engine: the rail is flex / grid — a column of boxes; a fork is a row of arm columns under a condition pill, each arm a column; a rejoin is a hairline back to the rail; an arm that ends stops at its terminal (the reply box, or a small return / throw mark in --ink-3). loop, later, together, inline are bracketed blocks with a left rule and a one-word label in --ink-3 (for each item, later · then, together, via generateToken). Boxes are StepNode's markup and styles verbatim (kind classes, accent rule for bridge / event, dashed effect, / ), sized to content — reuse the component with a rail prop or extract its inner markup; the click / double-click contract is identical (select → panel; double-click → Start here). Keep the pill words: the fork's condition in the conditions vocabulary (conditionTokens), WHEN / WHEN NOT on the arms, ≤ 36 chars with and the full text on hover, as §3.12 pills.

Words (kindWord stays): the toggle reads "Read as: in order · what it sets in motion"; the legend for the rail: "Boxes read top to bottom in the code's order; a fork is an if, switch, try or an early exit, its arms side by side; an arm that replies, returns or throws ends there; via x is a helper drawn where it is called; later runs after the handler returns." Sentence case, no tracking (spec §2).

The panel is unchanged — the same WireStep / WireStepLink shapes, Start here →, Open as a flow →, the scenario rows. The summary keeps depth, through, the counts; Continue through enters boundaries in the rail too.

4.3 URL and defaults

StepsHrefOptions.view?: 'order' | 'tree'; the router parses view; rewrite() carries it; the default when absent: order for a function / method / endpoint anchor, tree for a screen anchor (the anchor's root kind decides, server side, in the payload as defaultView so the viewer never guesses). The Screens tab's double-click and the chooser open the default.


5. Work plan, in order

P0 — structured guards on the reader. api/when.ts SiteReader.guards(caller, site): Promise<BranchGuard[]> (the array guardsForFile already computes; when keeps returning the joined string). Test: branch-guards.test.ts already covers the chains; add one assertion on the array for an if / else pair sharing line and text with negated flipped, and for an early return.

P1 — record per-function sites during the walk. In buildSteps, alongside link(), push { fn, stepId, linkId, hop: HopSite, guards: BranchGuard[], trigger } into a per-function list (programs: Map<fnId, SiteRecord[]>), for hops out of every fold node (steps and folds — a fold into a helper is a record in the caller and the helper's own records are the helper's). Cost: one guards read per hop, within MAX_WHEN_SITES.

P2 — the block builder. src/ui-server/api/program.ts: buildProgram(records, roots) → WireProgram, pure over the records (no graph access), unit-tested on hand-made records: straight line; if / else; early return; nested ifs; switch; try / catch; a helper inlined; a site inside arguments; later; together; a loop; a cap. Then wire it into the payload behind view=order.

P3 — the rail. StepsRail.svelte, the toggle, view in the URL, the legend and summary words, the default per anchor kind. Snapshot-free tests: the pure layout of arms is trivial, so test the model (ui/src/lib/program-model.ts: the block → rows/columns with the pills' words) in __tests__/ui-program-model.test.ts, and drive one real picture with the harness (P5).

P4 — languages beyond JS. Nothing to add for guards (P5 of the previous plan built them for Python, Java, Kotlin, C#, Go, C); check later / loop / together node types per language in Rules and add the loop-header text. Test: extend ui-steps-api-servers.test.ts with an in order assertion per framework (FastAPI's raise HTTPException is an early exit; Spring's if (…) return ResponseEntity.badRequest() too).

P5 — validation pictures. With the harness: proshop POST /api/users/login (the §1 reading), POST /api/products/:id/reviews (three outcomes: !product → 404 early exit, alreadyReviewed → 400, else create + 201), express-realworld POST /api/users/login, nest-boilerplate POST /api/v1/auth/email/login (guards, DI, thrown exceptions), next-saas-starter signIn (a server action: validation → lookup → redirect), fastapi-template POST /login/access-token. Read each against the source; a fork that is not in the code, or an order that is not the code's, is a bug, not a style choice. Re-shoot the mobile app (/capture/review) to confirm the tree mode is untouched.

P6 — docs. Spec §3.13 gains "3.13.1 In order" with the §1 rendering and the rules of §4.1; CHANGELOG.md [Unreleased] gets one user-facing bullet ("The Steps tab reads a handler in the code's order …"); CLAUDE.md's src/ui-server/ note names api/program.ts; the previous plan's status header points here.

Done when: proshop's login reads as §1 in the browser, the reviews endpoint shows its three arms with the 404 as an early exit, and the tree mode is byte-for-byte the picture it is today.


6. Gotchas, known from building the tree

  • Source order is execution order only for straight-line code. Arguments before their call (the nesting rule, already in hopCompare); await does not reorder; a callback registered now runs later (later); Promise.all is concurrent (together). Do not draw an order the code does not fix — say later / together instead.
  • A nested const handleX = async () => … is not a node (only function declarations and useCallback-bound arrows are); its body's sites attribute to the enclosing component, and the guard climb treats the inline arrow as transparent. The rail inherits that: such a handler's sites appear in the component's program. The fix is an extractor rule with the Rust kernel twin — out of scope here; note it where it shows.
  • The reply box is per (function, status) since ca9a7fd; a status set by the statement before the reply (res.status(401); throw …) is that reply's (statusSetBefore), and a bare res.json is a 200 (implicitResponseStatus). In the rail the 401 arm reads 401 · res.status(401) then throw new Error(…) as its terminal — the throw is the arm's ends, read from the early-exit guard on the next site, or from the site's own statement's successor; either way it is a mark, not a step.
  • Site keys. A site is (line, column); synthesized edges carry the call's column (the tier synthesizer sets it); an edge with no column lands on the first non-blank column of the row, which for const x = await fetch(…) is constcallAt climbs to the first call from there. Pass the callee name (want) when you know it.
  • Budgets. MAX_WHEN_SITES / MAX_CALL_SITES bound the tree reads per request; the rail adds one guards read per hop — the same sites the when read already parsed, so it is cheap, but count it. A drifted file yields no guards (hasDriftedOnDisk): the rail then shows sequence without forks, and should say "conditions not read: the file changed since the index".
  • Two passes per fold (spec §3.13): a node that is a step is never also folded through its contains edge; the program's records must come from the same pass that made the link, or a step appears twice.
  • The index keeps only the last segment of a deep member call; every call text in the rail must come from callSitesForFile (as the tree's do), never from the edge's refName.
  • Never npm run build while a codegraph ui server is running (it loads modules lazily and hangs); the harness in the memory note kills its server group on exit for that reason.

7. Open questions for the maintainer

  1. Default reading per anchor — the plan says in order for functions / endpoints, tree for screens. Or in order everywhere, with a screen's program being its render body plus each handler's program under later · onPress?
  2. How far to inline. A helper drawn in place at every call is faithful but wide; the tree's fold depth (8) may be too deep for a rail. Proposal: inline to depth 3, then a cut item with Start here →.
  3. Loops. Body once with for each …, or unrolled never. Proposal: once, marked.
  4. Should the rail replace the pills' "→ …x" placement rule? The rail's forks carry the condition once, on the fork; the arm's WHEN / WHEN NOT is the pill. The scenario rows in the panel stay as they are.

8. What the validation pictures found (2026-08-29, P5)

Read against the source, endpoint by endpoint. Every reading below is the one the rail draws today.

Repo Anchor Reads as Verdict
bradtraversy/proshop_mern POST /api/users/login User.findOne · fork on the password check · [via generateTokenjwt.sign, then 200] | [401], both arms answering §1 exactly
" POST /api/products/:id/reviews Product.findById · fork on product · [fork on alreadyReviewed400 | 201] | [404] three outcomes, right
gothinkster/node-express-realworld-example-app POST /users/login via login → two 422 guards, prisma.user.findUnique, if userbcrypt.compareif matchvia generateToken, then 403; then the handler's own 200 right, after the span fix
brocoders/nestjs-boilerplate POST /auth/email/login via validateLoginfindByEmail, !user422, two throwing guards, bcrypt.compare, then sessionRepository.create and via getTokensDatatogether Promise.all of two jwtService.signAsync right, after dropping the update name-match
leerob/next-saas-starter signIn (server action) drizzle select, length === 0 → return, !isPasswordValid → return, together Promise.all of setSession (→ signTokenSignJWT) and logActivity (→ db.insert), then redirectTo === 'checkout' → the whole checkout session | /dashboard right, after the lending fix
fastapi/full-stack-fastapi-template POST /login/access-token via authenticatesession.exec, not db_user → return, not verified → return, updated_password_hashsession.add/commit/refresh; then not user400, elif not user.is_active400; then via create_access_tokenjwt.encode right, after the elif fix
amniservices-mobile-app /capture/review unchanged: 87 steps, 160 links, the tree, defaultView: 'tree' the regression check

Each of those was read first as the rejected rail and then as the canvas graph; the readings are the same, the picture is not. What the canvas gives that the rail could not: proshop's login fits the sketch the maintainer drew (User.findOne → the fork → jwt.sign200 | 401), and nest-boilerplate's login reads down the page — findByEmailWHEN NOT user → 422 | WHEN user AND user.provider === …bcrypt.compareWHEN isValidPasswordsessionRepository.createtogether · Promise.alljwtService.signAsync.

Three defects the pictures caught, all in the walk and both readings — a hop's span taken from a call that was not the one asked for (an inline Express handler's edges carry the route's line, so the registration's span swallowed the body); a name-match the call as written disproves (crypto.createHash('sha256').update(…) followed into the caller's own AuthService.update); and a value lent nothing because one plain references edge counted as a body of its own (const signIn = validatedAction(schema, async (data) => { … }) drew one call out of nine). Plus an elif whose body raises being read as ending the arm it is written in.

Left open, deliberately:

  • A mongoose document's product.save() is not in the effects table (api/effects.ts), so proshop's review endpoint draws its 201 but not the write before it. A JS rule for <lowercase receiver>.save would catch it and would also catch canvas.save() / ctx.save() / sharp(...).toFile-adjacent idioms in any web app — a call for the maintainer, not a silent widening.
  • A nested const handleX = async () => … is still not a node (the plan's §6): its sites belong to the enclosing component. Fixing it is an extractor change with a Rust kernel twin.
  • The via name of an inlined helper is its bare name (via create), which is ambiguous when two classes have a create. The panel disambiguates; the rail could say the class.