feat(js): add Bun runtime support (bun test, run, bunx)

[HenriqueSchroeder]

Summary

Adds Bun runtime support to RTK with filters for bun test, bun run, bun install, bun build, and bunx.

What's new

• bun test filter: strips passing tests, keeps failures with error context, compact summary. Handles both TTY (✓/✗) and piped ((pass)/(fail)) formats. Buffers lines preceding a fail marker since Bun prints error context before the marker, then flushes them on match.
• bun install filter: strips progress bars, keeps summary.
• bun build filter: compact build output.
• bun run <script> filter: passes through script output with error-focused filtering.
• bunx passthrough with intelligent routing (tsc, eslint, prisma → specialized filters).
• Discovery rules so bunx tsc, bunx eslint, etc. get rewritten to specialized RTK filters.

Token savings

All filters verified >= 60% savings against real bun test output (see tests/fixtures/bun_test_real.txt, 140 lines of real bun output with 79 pass / 2 fail).

Tests

• Snapshot tests for all filter variants
• Token savings assertions with real fixtures (bun_test_real.txt)
• Clap routing tests for bun test/run/install/build subcommands and bunx

Test plan

• cargo test — 1372 tests pass locally
• cargo clippy --all-targets — no new warnings in bun_cmd.rs
• cargo fmt --all --check — clean
• Manual: rtk bun test in a Bun project verifies end-to-end filtering

Notes

Branch is based on an older point in history (pre-0.37.0). Merge conflicts with current develop (notably src/discover/rules.rs where Bun entries need to be combined with the expanded npx/pnpm exec patterns, and src/main.rs PrismaCommands derive) are expected and will need resolution before merge.

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.

[pszymkowiak]

[w] wshm · Automated triage by AI

📊 Automated PR Analysis

✨ Type	feature
🟡 Risk	medium

Summary

Adds comprehensive Bun runtime support to RTK, including filters for bun test, bun install, bun build, bun run, and bunx commands. Implements intelligent routing for bunx to specialized filters (tsc, eslint, prisma) and includes snapshot tests, token savings assertions, and clap routing tests.

Review Checklist

• Tests present
• Breaking change
• Docs updated

---

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

[bZichett]

Bumping. Great to see this coming soon. Much appreciated library btw. Been using for a few weeks with success.

[KuSh]

Some previous comments have not been taken into account or answered, conflicts have to be handled

[pszymkowiak]

Reviewed this locally (build + the 17 bun tests pass, clippy clean) — solid work, and it's the most complete of the open Bun PRs: the dedicated bun test failure extraction (buffering the error context Bun prints before the fail marker) is exactly where the token savings are, and the real-output fixtures back it up. Two mechanical things before merge:

1. Rebase needed — the branch is stale. On the current branch cargo test fails to compile the test binary on unrelated dead-code-as-error (FILTERS_TOML in constants.rs, load in toml_filter.rs) that develop has since resolved. I confirmed it's purely staleness: after merging develop in, the build is clean and all 17 bun tests pass. A rebase onto develop fixes it.

2. Coordinate the vitest_cmd.rs refactor with #2418. This PR decouples run_test from the Commands enum via a new TestFramework enum — #2418 does the same refactor. Whichever merges first, the other will need to drop/reconcile that part to avoid a conflict.

Substance is good — happy to re-review once rebased.

[aeppling]

Thanks for the rebase. The review threads are in good shape: I re-checked the 19 resolved ones against the code and they have all landed (bun x alias and bun run <tool> prefixes, public test/install/build/run, route_bunx coverage, "ok" outputs, LC_ALL=C, lock file untouched).

Two blockers before this can go green, both from develop moving under the merge:

1. lazy_static was removed from the crate (5269df7). bun_cmd.rs:6 fails to import it and all the BUN_*_RE statics are undefined. Use std::sync::LazyLock like playwright_cmd.rs:167:

   static BUN_FAIL_MARKER_RE: LazyLock<Regex> = LazyLock::new(|| Regex::new(r"...").unwrap());

2. RtkRule gained pipeline_final_safe (523c803) and now has a Default. The two new bun rules in rules.rs (around lines 105 and 114) need ..RtkRule::default() like their neighbors.

With those the test binary should build again. Please run cargo fmt --all --check && cargo clippy --all-targets && cargo test on the merged head before pushing; it currently fails with 15 errors.

Two threads are still open with a reply but no resolution; a one-line confirmation each would let them be closed:

• non-English locale: LC_ALL=C on test/install/build without having tested de_DE is fine by me, just say so explicitly in the thread.
• bun run keeping untrimmed lines: agreed, it is passthrough; note it in the thread.

Coordination: #2418 carries the same TestFramework refactor of vitest_cmd.rs. Whichever merges first, the other rebases and drops that hunk; worth a note on #2418 so it is not a surprise.

Once the build is green I am happy to re-review. The bun test failure extraction is the valuable part here and the real-output fixtures back it.

[aeppling]

Final pass — CONTRIBUTING / CODING_PRACTICES + fuzz testing

Thanks for addressing the previous review — public fns, route_bunx, bun x handling, LazyLock regexes and the rewrite coverage are all in good shape now.

Reviewed head c3c2257 one last time against the guidelines and fuzz-tested the binary against a real Bun 1.3.14 project (pass/fail/todo/skip, 250 failures, timeouts, snapshots, SIGKILL, --bail, --coverage, bare bun run/bunx, build/install). fmt, clippy, tests green. Routing, rewrite rules and refactors are solid.

Blockers

1. Shared infrastructure not used. bun_cmd.rs bypasses the core building blocks every filter relies on: no with_tee on any run_filtered, no CAP_* cap, no force_tee_hint. Result: data is dropped with no way to recover it (250 failures → 2996 lines uncapped; build output file name, coverage table, $-prefixed script lines, Bailed out, snapshot/todo counts all silently gone). CONTRIBUTING "Never Block": hidden data requires a recovery hint. See pytest_cmd.rs:286-294 and tsc_cmd.rs:40 for the pattern.

2. Bare bun run / bunx / bun x become rtk errors. Real commands list scripts / print usage; rtk bails. Pass through like rtk npm run does.

3. ok printed on crash. SIGKILL in a test → Bun prints header only, exit 137 → rtk prints ok. Filter never sees the exit code (run_filtered vs run_filtered_with_exit).

Must fix

4. bun_build.txt / bun_install.txt fixtures are not real Bun output; build regex never matches the real name N bytes (entry point) line. Replace with real captures.
5. Summary regexes miss 1 snapshots, 6 expect() calls, 1 todo, snapshots: +1 added, Bailed out after N failure.
6. filter_bun_test trims lines: caret ^ loses column, toEqual diff loses indentation.
7. filter_bun_run drops every $-prefixed line; Bun echoes one $ cmd line, strip only that.
8. --coverage requested → table dropped.

Requesting changes. Blocker 1 gates this on its own. Happy to re-run the same fuzz set on the next revision.

[HenriqueSchroeder]

All eight are in, pushed as 5804705. Ran the same kind of pass you did: fmt, clippy and the 2757 tests are green, and I drove the built binary against a real bun 1.3.14 project.

Blocker 1. Every run goes through run_filtered_with_exit with a tee label now. Failures are grouped into blocks (context + marker) and capped at CAP_WARNINGS, with the remainder behind a tee hint; install caps its package list and build its artifact list at CAP_LIST. A 40-failure run gives 10 blocks, the overflow line and the hint, with the summary intact.

Blocker 2. Bare bun run, bunx and bun x pass through, so they print bun's own script list and usage instead of an rtk error.

Blocker 3. finish() reports the exit code when the filter kept nothing and the command failed, for all four filters. A SIGKILLed run prints bun test exited with code 137 plus the tee hint, with 137 propagated. One nuance: when the raw output is just the banner, never_worse keeps bun's raw line instead, which is fine, it still isn't ok and the code still propagates.

4. Replaced every bun fixture with a real capture from 1.3.14, not just build and install. The build regex now matches index.js 72 bytes (entry point). Worth recording: 1.3.14 no longer prints a line per passing test, so the old 1.3.10 fixture with its 79 (pass) lines is kept to cover that format.

5. The summary parser picks up 1 todo, snapshots: +1 added, the N snapshots, N expect() calls variant and Bailed out after N failure.

6. Context lines are kept verbatim. There is a test that measures the caret indentation, so a future trim will not slip back in.

7. Only the single $ cmd line bun echoes is dropped; a script printing its own $-prefixed lines keeps them.

8. The coverage table survives.

While I was in there I answered the locale question with actual data: on 1.3.14 the output of bun test, bun install and bun build is byte identical under LC_ALL=C, pt_BR.utf8 and de_DE.UTF-8, decimals dot separated in all three. de_DE is not generated on my box so it falls back to C; pt_BR is the real one. Keeping LC_ALL=C as cheap insurance.

Note dropped on #2418.

[aeppling]

Thanks, re-ran the same fuzz set on 5804705. Blockers 1-3 and points 4-6, 8 verified fixed: tee hints, caps, run_filtered_with_exit, SIGKILL propagates 137 without ok, bare bun run/bunx pass through, caret/diff context intact, todo/snapshot/bail/coverage kept, real fixtures.

One must-fix left, #7: the "drop first $ line" heuristic drops the wrong line. Bun writes the $ cmd echo to stderr and the script to stdout, and stdout lands first in the capture. 8/8 runs dropped the script's $HOME is not set and kept the echo. Inject --silent before the script (only suppresses the echo, script output verified identical) and remove the $ heuristic.

Approve once that's in.

[HenriqueSchroeder]

Good catch on the ordering, that is exactly it. Reproduced before touching anything: bun writes the echo to stderr and the script to stdout, stdout lands first in the capture, and rtk bun run dropped the script line 5/5.

Fixed in 3d30c06 your way: bun run --silent before the script name, and the $ heuristic is gone, so the filter no longer looks at $-prefixed lines at all. run_args builds the argv and a test pins --silent ahead of the script, since after it bun would forward the flag to the script instead.

Verified 8/8 with a script printing $HOME is not set: output is now byte identical to bun run --silent run directly. fmt, clippy and 2758 tests green.

[aeppling]

All blockers and must-fixes resolved. LGTM.

@KuSh still pending changes

[KuSh]

Another pass over the current head. Most of the earlier threads are genuinely addressed — the bun x alias, bun tsc / bun run tsc, the bun x special case in main, c.arg("--").arg(tool), the finish() consolidation, trim-upfront, the public test/install/build/run fns, LC_ALL=C and the Cargo.lock revert all check out. I ran the rewrite engine over the bun command forms and all four spellings (bun X, bun run X, bunx X, bun x X) resolve to the right specialized filter, so the jest/vitest/prisma/tsc threads are correctly closed.

Two things below are blocking (the biome arm never runs biome, and is_coverage_row eats failure context on any line containing ||). The rest are smaller.

Two more that I can't attach inline because the files aren't in the diff:

• src/core/tracking.rs, categorize_command — not extended with "bun" | "bunx", so every rtk bun … / rtk bunx … invocation is bucketed as "other" instead of "js" in telemetry and rtk gain.
• src/core/toml_filter.rs, RUST_HANDLED_COMMANDS — not extended with bun/bunx. A project TOML filter whose match_command matches bun will compile with no "already handled in Rust" shadow warning and then silently never fire, since clap now routes rtk bun first.

[KuSh]

Re-raising the factorization point — my earlier comment on this is now collapsed as outdated, and the duplication has since drifted.

route_bunx and the Commands::Npx arm (~line 2425) are the same tool→filter table, and they no longer agree:

	route_bunx	Commands::Npx
jest / vitest / biome	present	missing
prisma fallthrough	bun_cmd::run_tool	~30 lines of inlined passthrough
final fallback	bun_cmd::run_tool	npm_cmd::exec

The drift is the point: the missing arms are exactly what I flagged earlier, and the &args[1..] slicing bug on the eslint/biome arm below exists identically in the npx copy — one bug, copied.

Something like route_js_tool(args, verbose, fallback) with npx passing npm_cmd::exec and bunx passing bun_cmd::run_tool fixes both call sites at once and stops the next tool from being registered in only one of them. The npx prisma branch should call that fallback too rather than hand-rolling its own passthrough.

[KuSh]

This strips the linter name, but lint_cmd::run expects it as args[0] — detect_linter reads args[0], and rtk lint biome is the documented form.

So rtk bunx biome check . calls lint_cmd::run(["check", "."]), detect_linter treats check as the linter, and rtk shells out to bunx -- check .. Verified end-to-end: bunx answers could not determine executable to run for package check, and rtk then reports Lint: 1 errors, 0 warnings. biome never runs and the failure is masked.

Suggested change

	"eslint" | "biome" => lint_cmd::run(&args[1..], verbose),
	"eslint" | "biome" => lint_cmd::run(args, verbose),

Note bunx eslint src misroutes the same way (bunx -- src), mirroring the same slicing in the Commands::Npx arm — see the factorization thread.

[KuSh]

This misfires on bun's own source frames. bun prints failure context as NN | <source>, so any frame whose source contains || (two pipes plus the gutter pipe = 3) or a TS union type is classified as a --coverage row: buffer.clear() throws away every preceding context line of that failure, and the line is then re-emitted after the FAILURES block.

Reproduced with a test failing on const name = user.name || user.nick || "anon"; — frame lines 10-13 were dropped, the failure block started at line 14, and line 13 reappeared under the coverage section. || is ubiquitous in JS, so this fires often.

The coverage table is a contiguous block with a ---|--- separator and fixed headers; keying on that (or on being inside a detected coverage section) would be much safer than a bare pipe count.

[KuSh]

BUN_TEST_COUNT_RE was deliberately anchored with $ so app output isn't read as the N pass summary, but this prefix check (and "Bailed out after " below) is still open.

Reproduced: a test logging Ran 5 migrations before assert before failing had that line hoisted into the summary section, and buffer.clear() then dropped the failure's code frame and caret. Worth anchoring the same way, e.g. ^Ran \d+ tests?\b.

[KuSh]

A build failure loses its location. Only lines containing error/warn/Error/Warn, artifact lines and summary lines survive; bun's code frame, the caret, and the at /path/file.ts:1:15 line match none of those and are dropped.

Verified: frame + caret + error: Could not resolve: "missing" + at /home/u/p/index.ts:1:15 filters down to just the error: line. The file and line number are exactly what the LLM needs to act, so keeping the at <path>:<line>:<col> line (and ideally the frame) would be worth the tokens.

[KuSh]

This drops all blank lines, which contradicts the reasoning used to keep line rather than trimmed a few lines down ("this is passthrough of arbitrary output where indentation is content"). If indentation is content here, so are the paragraph breaks.

Section A\n\n item 1\n item 2\n\nSection B collapses into four consecutive lines, destroying the structure of any report or diff a script prints.

[KuSh]

Missing the terminator that every sibling rule has (^npm\s+(exec|run|…)(\s|$), ^(bun\s+x|bunx)(\s|$), ^ls(\s|$)), so this matches on a prefix.

I ran rewrite_command over a few forms and these all rewrite:

bun build.ts     -> Some("rtk bun build.ts")
bun test.ts      -> Some("rtk bun test.ts")
bun install.ts   -> Some("rtk bun install.ts")
bun removeall    -> Some("rtk bun removeall")

bun <file.ts> is a normal invocation since bun executes files directly, so bun build.ts isn't contrived. Execution stays correct (those land in BunCommands::Other → run_passthrough), but the hook rewrites commands that gain nothing and discover reports 80% savings for them.

Suggested change

	pattern: r"^bun\s+(test|run|install|build|add|remove)",
	pattern: r"^bun\s+(test|run|install|build|add|remove)(\s|$)",

[KuSh]

Separately from the anchoring: savings_pct: 80.0 over-reports for part of this rule's own alternation. add and remove land in BunCommands::Other → run_passthrough (0% reduction), and bun run <script> is near-passthrough. bun add zod currently classifies as 80% estimated savings, so rtk discover will over-promise. The subcmd_savings list already carries the real numbers for test/install/build — the base rate is the one that's off.

[KuSh]

This duplicates the pre-existing line 228 (rtk uv run pytest # Preserve uv env, keep program output) with a contradictory description, three lines apart. Looks like a rebase artifact from the pre-0.37.0 base — probably wants dropping.

[KuSh]

Thanks for this — solid work, especially the fixture coverage.

Closing in favour of #988, which is further along in review and covers both runtimes (bun and deno) rather than bun alone. Keeping two overlapping bun implementations in flight isn't serving anyone, and #988 is the one about to land.

That said, this PR has things #988 doesn't: bun test --coverage and --bail handling, and a much richer fixture set. Those would be very welcome as a focused follow-up on top of #988 once it merges — please do open one if you're up for it.

[HenriqueSchroeder]

Fair call, no argument from me. Two overlapping implementations is worse than one.

I'll open the follow-up once #988 lands: --coverage and --bail handling, plus the fixtures. The fixture set is real captures from bun 1.3.14 and covers the cases #988's don't: todo/skip/snapshot counts, a SIGKILLed run for exit code propagation, and a 40-failure run for the cap. Happy to hand those over earlier if @geofflittle wants them before the merge, they are on my branch and independent of the rest.

One thing from your last pass is worth rescuing, since it outlives this PR: the &args[1..] slicing you flagged on the eslint/biome arm exists in the Commands::Npx arm on develop today, so rtk npx eslint src misroutes to npx -- src and masks the failure the same way. Not mine to fix here anymore, but it is live.

[KuSh]

Thanks for the reminder, @HenriqueSchroeder. I’ve opened #3812 for that. I’ll ping you here if and when #988 is merged so we can follow up. Thanks again!