feat(ng): compact Angular build tables safely

[yeahjack]

Summary

Refs #4239. This PR covers the first, bounded part of the request: native rtk ng build and a narrow rewrite of bare ng build. It does not close the broader issue or overlap with the Karma work in #2541.

• Compact recognized Angular build-table padding while retaining every chunk, filename, raw/transfer size, initial total, build status, duration and output location
• Remove only the exact redundant leading progress pair when a valid table is present; keep unfamiliar output and compiler/template context
• Preserve the ng executable resolved from PATH, argv, exit status and complete stderr, including budget warnings; no package-manager fallback or injected Angular flags
• Pass through explicit help/verbose/watch modes and uncertain workspaces. This includes configuration-based watch, JSONC/unreadable configuration, unknown builders and missing local workspaces
• Reuse guard/recovery/tracking, with a small opt-in capture helper that relays cancellation without normalizing line endings or capping stderr

Test, serve, wrappers and path-based executable rewrites remain outside this initial filter.

Real-output measurements

Generated and built an actual local Angular application with the official Angular CLI/framework/build 22.2.1, Node 24.19.0, npm 11.9.0 and TypeScript 6.0.3. The checked-in fixtures are raw captures, including separate stdout/stderr and recorded build help.

Using RTK's production ceil(bytes / 4) estimator on native-vs-RTK builds:

Case	Native estimate	RTK estimate	Reduction
Default successful build	130	99	23.85%
Six lazy routes	260	193	25.77%
Budget warning	161	130	19.25%
Budget failure	165	134	18.79%
TypeScript error	104	104	0%
Template error	160	160	0%

These are command-output estimates, not actual tokenizer counts or billing savings. Angular's current output is already compact, so this change deliberately makes modest savings and leaves diagnostic-only failures unchanged.

The successful fixtures satisfy the current contribution guide's 20% threshold. The CLI-testing rules still mention a 60% blanket release floor; this implementation does not meet the 60% target. Maintainer feedback on that discrepancy and the value of this limited filter would be useful during review.

Verification

• cargo fmt --all --check
• cargo clippy --all-targets -- -D warnings
• cargo test --all: 4,112 passed, 8 ignored
• cargo test --all -- --ignored: 8 passed
• Hostile-environment run with German locale variables, foreign GIT_DIR, RTK_NO_TOML=1, RTK_TEE=0 and --include-ignored: 4,120 passed
• Actual native-vs-RTK executions of all six Angular build cases: matching exit codes, byte-identical stderr and preserved stdout information; real help output also matches byte-for-byte
• Real configuration-based watch remains live and emits through RTK before completion
• Process regressions cover argv/path spaces, missing-value --watch, passthrough streaming, SIGTERM flushing/no orphan, recovery, uncapped warnings, exact output accounting, malformed output and narrow rewrite boundaries
• Independent review completed; its watch-parsing and cancellation findings were fixed and covered by regressions
• Release build passed; a 60-sample local finite-command probe measured 3.32 ms median RTK invocation, including workspace lookup and tracking

Linux was exercised here. Windows/macOS runtime validation and broader Angular version coverage are not claimed. Native fixture EOF blank lines are intentional and preserved byte-for-byte.

[rtk-wshm-sync-bot]

wshm · Automated triage by AI

📊 Automated PR Analysis

✨ Type	feature
🟡 Risk	medium

Summary

Adds a new rtk ng build command that compacts Angular CLI build-table output (padding, redundant progress lines) while preserving bundle sizes, diagnostics, exit codes, and stderr. It conservatively falls back to passthrough for help/verbose/watch modes, unknown builders, JSONC configs, or uncertain workspaces, and includes real Angular-generated fixtures and extensive regression/hostile-environment testing.

Review Checklist

• Tests present
• Breaking change
• Docs updated

Linked issues: #4239

---

Analyzed automatically by wshm · This is an automated analysis, not a human review.

[yeahjack]

@KuSh When you have a chance, could you take a look at #4405, #4409 and #4410? A gentle follow-up on these pending reviews. Thank you!

[KuSh]

Round 1 of 3 — CHANGES REQUESTED

Claim: rtk ng build (and a rewrite of bare ng build) compacts Angular build-table padding and the leading progress pair, keeps every bundle and diagnostic, preserves argv, exit code and stderr, and passes help, verbose, watch and uncertain workspaces through.
Scope: accept (round 1, frozen)
Ran: real Angular 22.2.3 app built in Docker (Node 22): default production build, 4 lazy routes, budget warning, budget error, TypeScript error, template error, -c development, --progress=false, --stats-json, --source-map, ng build app, --help, --watch=false, app + ng g library workspace, watch: true in a configuration, outside a workspace, ng missing from PATH, SIGTERM mid-build, NG_FORCE_TTY=1 with closed stdin. 20 rows, native ng build against rtk rewrite "<cmd>" and then the rewritten command, stdout and stderr captured separately. Invariants: under the hook | split streams | every decision (mutation run below) | signal relay. Savings 28.2% on the default build, 28.7% with lazy routes, 22.7% with --progress=false, against the 20% floor (CONTRIBUTING.md).
Pins: 50 mutations over the decisions the PR adds: 29 killed, 21 survived the full suite. 3 have no observable effect (-h and -v, which Angular rejects as Unknown argument, and watch inside an array). The other 18 are grouped below.
Matrix: run this round. Pin check: run this round.
CI @48b0a4a1b: all green. GitHub reports a merge conflict with develop.

Blocks merge (3, frozen at round 1)

The behaviour is correct on every real build I ran: every stdout token is kept apart from the progress pair, stderr is byte-identical, exit codes match native, rtk recall returns the full output, and a configured watch stays live. What blocks is that 18 of the decisions behind that result can be changed with the whole suite still green. Each item below lists its surviving mutations and the rows a test has to assert. Each mutation must turn the new test red.

1. src/cmds/js/ng_cmd.rs:112 — unpinned behaviour: which workspaces are captured (8 mutations). This guard decides between buffering the build and streaming it live, so it is the decision that matters most if it is wrong.

   • .all( → .any( in known_builders. Row: a workspace with an app plus a library generated by ng g library (@angular/build:ng-packagr). ng build app must pass through byte-identical; it does today.
   • Remove @angular-devkit/build-angular:application, :browser-esbuild or :browser from the list (3 mutations). Row: one workspace per builder name, each compacted. :application is the default builder for Angular 17–19 projects.
   • Drop .or_else(|| project.get("targets")). Row: a project using targets instead of architect, compacted.
   • Drop !projects.is_empty() &&. Row: "projects": {} (what ng new --create-application=false writes), passed through.
   • Err(_) => return true → false. Row: an angular.json that exists but cannot be read (for example, a directory), passed through.
   • current_dir().map_or(true, …) → map_or(false, …). Row: rtk started in a directory that has since been removed, passed through.

2. src/cmds/js/ng_cmd.rs:185 — unpinned behaviour: which lines count as table rows (6 mutations). Each guard below can be removed without any test failing. Each needs one input that only this guard rejects (or accepts), with the exact output asserted:

   • The header lookahead in filter_build made always true. Row: a header line followed by a non-row line; both must stay byte-identical.
   • Drop || cells[0].ends_with(".mjs"). Row: a .mjs row, compacted.
   • Drop && !cells[1].is_empty(). Row: a row with an empty Names cell, kept verbatim.
   • Size check made always true. Row: main.js | main | huge after a header, kept verbatim.
   • Drop |- from SIZE. Row: a row with a - size cell, compacted.
   • Remove the CRLF progress prefix. Row: CRLF output, progress pair stripped.

   None of my real builds produced a .mjs row, a - size or CRLF output. If Angular never prints them, removing those arms closes the item as well as a test would. You know the output better than I do, so your call.

3. unpinned behaviour: process and registration (4 mutations).

   • src/core/stream.rs:727: cmd.stdin(Stdio::inherit()) → Stdio::null(). Row: a fake ng that echoes one line it reads from stdin; printf 'y\n' | rtk ng build must show y.
   • ng_cmd.rs:52: the filtered != result.stdout guard on the hint and the never_worse call at :58 each back up the other, so neither can be killed alone. Row: a failing compiler-only build with tee enabled (the integration tests set RTK_TEE=0); stdout must be byte-identical with no [full output line. The test must fail when both are removed together.
   • src/core/tracking.rs:550: remove "ng" from SUBCOMMAND_ROUTERS. Row: the stored label for rtk ng build app keeps build, as the sibling router tests check.

Optional — will not hold merge

• ng build -c development gets 0% (349 → 349 bytes). Angular ends each row of the development table with an empty cell (main.js | main | 1.39 MB | ), so cells.len() != columns rejects every row and the header lookahead fails. Accepting one empty trailing cell would compact development builds too.
• A workspace that contains a library passes ng build app through untouched (455 → 455 bytes), because known_builders checks every project rather than the one being built. This is safe, but common monorepos get nothing from the filter.
• tee_and_hint writes a recovery entry on every failing build, including compiler-only failures where the hint is then thrown away. Calling it only when filtered != result.stdout avoids that.
• ng is not in RUST_HANDLED_COMMANDS (src/core/toml_filter.rs:266), so a user TOML filter on ng gets no shadow warning. Several existing commands are missing from that list too; #3770 tracks it.
• The h and v short arms in should_filter: Angular rejects ng build -h and ng build -v (Error: Unknown argument: h, exit 1), and that error prints the same captured or not.
• The new sentence in the main README intro reads as a one-feature announcement; the what-rtk-covers.md row already covers it.

Checked and correct — no need to re-verify

• Every stdout token kept apart from the progress pair, and stderr byte-identical, on 12 real builds (success, lazy, development, budget warning and error, TypeScript and template errors, --progress=false, --stats-json, --source-map, ng build app).
• Exit codes match native on every row, including 1 on budget and compiler errors.
• The recovery hint on the budget-error build (rtk recall <id>) is left alone by rtk rewrite and returns the full stdout and stderr, identical to native apart from build timing.
• --help, --watch=false, a configured watch: true (live output under timeout), and running from outside a workspace all pass through.
• SIGTERM mid-build: rtk prints what it captured, exits like native, and leaves no node or esbuild process behind.
• NG_FORCE_TTY=1 with stdin closed: same 418 bytes and exit 127 as native, no hang.
• Rewrite: ng build positive, plus cd … && ng build, CI=1 ng build, ng build 2>&1, --configuration=production; negative for ng b, ng build | tail, ng build>out.txt.
• Merged onto current develop (the stream.rs conflict is two adjacent additions; keeping both resolves it): fmt and clippy clean, cargo test --all 4995 passed (one machine-local failure that also fails on develop).

Your questions

• "The CLI-testing rules still mention a 60% blanket release floor … Maintainer feedback on that discrepancy and the value of this limited filter would be useful" → The floor is 20% of bash output (CONTRIBUTING.md); the 60% in .claude/rules/cli-testing.md is stale and does not apply. The filter is worth merging as claimed: 28% on real default builds, and diagnostic-only failures left alone is the right call.

Next

Merge develop (in the src/core/stream.rs conflict keep both functions), then add tests that kill the 18 surviving mutations above, or remove the arms in item 2 that Angular never reaches. After that, this merges.
Rounds: 1/3. Threads: 3 open (waiting on the author), 0 resolved this round.

[KuSh]

Blocks merge, item 1: which workspaces get captured. Changing .all( to .any( here, removing any one of the three @angular-devkit/build-angular:* names, dropping the targets fallback, dropping !projects.is_empty(), Err(_) => return false in workspace_may_watch, or current_dir().map_or(false, …) in run: each change leaves the whole suite green. The review body lists the row each test must assert. The first is a real ng g library workspace, where ng build app must pass through byte-identical.

[KuSh]

Blocks merge, item 2: which lines count as rows. The header lookahead in filter_build, the .mjs suffix, !cells[1].is_empty(), the SIZE check, the - alternative in SIZE, and the CRLF progress prefix can each be removed with the suite still green. Each needs an input only that guard decides. Removing an arm that Angular never reaches closes it as well as a test would.

[KuSh]

Blocks merge, item 3: Stdio::inherit() → Stdio::null() here passes every test. A fake ng that echoes one line read from stdin would pin it. The same item covers the hint guard plus never_worse pair (ng_cmd.rs:52/58), which only fails when both are removed, and "ng" in SUBCOMMAND_ROUTERS (tracking.rs:550).