Real SVG icons, and Kubernetes manifests identified by content

[cjimti]

Closes #38. First of #37's eight children (#39–#45 follow).

The tree drew its file types as ▸ ◆ ▤ ▫ and the editor tabs as coloured CSS squares — four buckets for every file in a Kubernetes manifest repository, which is to say a manifest, a Helm chart and a CI workflow all looked identical. This replaces them with a real SVG set, and decides what a YAML file is by reading it rather than by guessing from its directory.

1. Two icon sources, resolved differently on licensing

lucide-react (ISC) — a real dependency. Zero runtime dependencies, a React peer only, and it tree-shakes: the built bundle contains chevron-right and ellipsis-vertical and does not contain air-vent, zap or sailboat, which was checked against the artifact rather than assumed. Its icons draw in currentColor, which is what UI chrome needs — a close button has to go muted → foreground on hover, and an <img> cannot do that.

material-icon-theme (MIT) — vendored, not depended on. Seventeen SVGs copied verbatim into frontend/src/icons/material/ with the upstream LICENSE and a README recording version and provenance.

Not taking it as a dependency is a deliberate rejection. It is a VS Code extension, we import none of its code, and depending on it drags chroma-js into the production npm graph — whose declared license is the compound expression (BSD-3-Clause AND Apache-2.0). Both halves are on the allowlist, but license-checker-rseidelsohn --onlyAllow matches declared strings, so that is a gate failure waiting to happen on behalf of a package nothing here calls.

vscode-icons was rejected outright, as #38 said it would be: CC BY-SA.

Kustomize has no artwork in the set (kusto in there is Azure Data Explorer, a different product), so kustomization.yaml uses lucide's stacked layers. Worth flagging that DESIGN.md defers kustomize support to v1.x — an icon is not support, and the ticket asked for it.

Both sources sit behind one module. FileIcon takes an IconKind, UiIcon takes an action name ("close", not "an X"), and the tree, both tab strips and the changes rows all call the same two. FILE_SVG is typed Record<Exclude<IconKind, "kustomize">, string>, so adding a bucket to the union without artwork behind it fails type-checking rather than rendering nothing.

2. Kubernetes manifests are identified by content

The first commit answered this from the path: any YAML below the repository root that was not a chart, kustomization or workflow. That was wrong in the way a rule about icons should never be — it asserts something about a file it has not read. docker-compose.yaml two directories down got a Kubernetes wheel, and the only reason codecov.yml did not was a second guess layered on the first about root-level YAML being configuration.

The second commit replaces it. iconKind can no longer return "kubernetes" at all, and there is a test aimed squarely at the rule it replaces. Names now settle only what names can settle:

Rule	Icon
.github/workflows/**	GitHub Actions — that is where GitHub looks
Chart.yaml, Chart.lock, values*, under templates/	Helm — Helm named them
kustomization.y(a)ml	Kustomize
every other .yaml/.yml	plain YAML, until its content says otherwise

The upgrade is lazy, and its shape is the point. A listing renders immediately with the icons names earn; then useFileTree reads the head of that directory's plain-YAML files and the manifests among them change icon. Nothing blocks on a file read to draw a directory, and nothing is read outside the directories actually open.

internal/watch.ReadPrefixes is the read: the first 2 KiB of each named file, batched per directory. It returns evidence rather than a verdict on purpose — deciding in Go would put a second copy of the apiVersion/kind rule beside the frontend's. looksLikeManifest in tree.ts is the single implementation, used by the tree against a prefix and by the editor tab strip against the whole file it already holds, which is what keeps a manifest's tab and its tree row from disagreeing about the same file.

The rule is a column-0 apiVersion: and a column-0 kind:, both with values, anywhere in the head. Column 0 is what makes "top level" checkable without a YAML parser: a kind: under spec: is indented and a list item's never starts a line. Both keys are required, so a chart's values file carrying its own kind: is not promoted. Reading a prefix means this can only be wrong in one direction — a manifest whose two keys both sit past 2 KiB keeps the plain icon — and understating is the right way for an icon to fail.

Defects found by adversarial review, fixed here

Object-literal lookup returned inherited members. The icon tables were object literals, so BY_NAME["constructor"] answers "present" and returns Object — a Function handed back where an IconKind was promised, which TypeScript cannot catch because the index signature declares the value type. A file named constructor or toString would have rendered an icon with no artwork behind it. Both tables are Maps; the test over the five Object.prototype names fails against the previous version.

ReadPrefixes opened before it stat'd. Opening a named pipe blocks until something writes to it, and a repository is allowed to contain one — the call would have hung holding the batch, and the tree's icons with it. It Lstats first and takes regular files only, which also rejects symlinks (os.Root refuses to follow one out of the root but will follow one within it).

A UTF-8 BOM defeated the match. It sits in front of the first key, on the line most likely to carry apiVersion. Stripped.

The batch cap was reachable. 1024, so this binding cannot be turned into a repository walk — but a flat directory of two thousand manifests would have failed the call and kept the plain icon forever. The frontend batches at 256 rather than discovering the cap.

A project switch could cross-contaminate. An in-flight classification would have written one project's verdicts into another's tree, where paths are relative and deploy.yaml means different files. Checked against the current root before the state write.

Two pins raised, both by exactly one — these want your opinion

Per CLAUDE.md these need maintainer sign-off, so they are the thing to review first:

• internal/watch 21 → 22 exported. One name, ReadPrefixes. The other two identifiers it needed (prefix length, oversized-batch sentinel) are unexported because nothing outside names them. ReadFile could not serve this: it answers "give me this file to edit" and enforces an editing size limit for it, which is the wrong contract for "tell me what these thirty files are".
• maxAppMethods 23 → 24. One binding. It passes the test that ceiling's own comment sets — a request with an answer and no throughput, not a stream wearing a binding's clothes. The alternative was folding heads into ListDirectory's reply, which would make every expansion pay for reads it usually does not need, which is exactly the cost being avoided.

Both are argued at the pin, not only here.

Smaller decisions worth naming

• assetsInlineLimit: 8192 in vite.config.ts. helm.svg is 4143 bytes and the 4096 default emitted it as a separate asset fetched through the Wails asset server on first paint, for a 4 KiB string that belongs in the bundle. icons.test.ts gates that every vendored icon stays under whatever the limit is, reading both numbers off disk — the same approach style.test.ts uses, and for the same reason: neither fact is visible to a rendering test, since vitest's dev transform serves assets as URLs and inlining is a property of vite build.
• kindFromIcon was the trap. It maps IconKind to the bucket that picks the CodeMirror language. Splitting YAML five ways while leaving it matching only "yaml" would have silently opened every manifest and workflow as plain text, no highlighting and no folding. There is a YAML_ICONS set now and a comment saying what it holds up.
• FileTree.tsx crossed the 600-line cap, so CreateRow/InlineField moved to their own module — the part of that file that knows nothing about trees. Shaving comments to fit a gate was the other option and is not one.
• Buttons that lost their + gained an aria-label, since "file" alone is a worse accessible name than "+ file" was. Each label contains its visible word (WCAG 2.5.3).

Known consequence, stated rather than hidden

The changes rows show YAML as YAML. They have no content behind them, and wiring them to the tree's classifications would make a changed file's icon depend on whether its directory happened to be expanded — a panel inconsistent with itself is worse than one uniformly generic. #40 replaces that panel entirely.

Verification

make verify green, including bindings-check with the regenerated bindings committed.

• 671 frontend tests, 0 lint issues (frontend and Go)
• Go coverage: internal/app 93.0%, internal/watch 94.4%; patch floor clear
• Licenses clean with the new dependency; real Wails smoke build; no dead code

Checked in the running app against a real manifest repository, not only in jsdom:

• Seven .yml files under a manifests directory render as plain YAML and then upgrade to the Kubernetes wheel as their heads arrive — the async wiring unit tests cannot show
• Folders open and close with the chevron; README.md opens in preview with the pencil offering edit; the dirty dot fills
• A Kubernetes tab keeps YAML highlighting — kindFromIcon confirmed end to end
• Both themes crisp at the 22px row height

One gap: patch-coverage is a Go-only gate, so the frontend half of this diff is covered by frontend-test and frontend-lint rather than by a measured patch percentage.

[codecov]

Codecov Report

❌ Patch coverage is 88.57143% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 93.43%. Comparing base (7ee400e) to head (b4e2a67).

Files with missing lines	Patch %	Lines
internal/watch/prefix.go	86.66%	2 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #46      +/-   ##
==========================================
- Coverage   93.54%   93.43%   -0.12%     
==========================================
  Files          31       32       +1     
  Lines        1518     1553      +35     
==========================================
+ Hits         1420     1451      +31     
- Misses         62       64       +2     
- Partials       36       38       +2     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.