chore: take upstream pi's toolchain split for install, checks, and tests

[flora131]

Takes upstream earendil-works/pi's toolchain split task for task, not just its test runner. npm installs, builds, checks, and runs the suites; vitest replaces bun test for the three root suites; Bun keeps exactly the two jobs pi also gives it — compiling release binaries and running scripts/*.ts.

The gate: does npm/Node preserve raw-TypeScript extension loading?

Yes for the shipped loader and for tests; no for bare node. The boundary, not the yes/no, is the design constraint:

.js→.ts specifier rewriting for raw-TypeScript packages survives under Node only through a transpiling loader. jiti provides it in production; Vite provides it under vitest. Node's own resolver and type-stripping do not. Therefore no entrypoint invoked by bare node may depend on it.

Verified on a real npm ci tree: workspace layout is identical to Bun's hoisted linker (all seven node_modules/@bastani/* are symlinks into packages/*), jiti + Node loads packages/workflows/src/durable/factory.ts out of a 316-file graph, and the Bun runtime loads the same tree unchanged. Changing the installer does not change the runtime — which is exactly why pi can install with npm and still compile binaries with Bun. loader-virtual-modules.ts is untouched. scripts/*.ts therefore stay on bun run; the one script covered by node --test is .mjs.

Scope migrated

	Before	After
Install	bun install --frozen-lockfile, bun.lock	npm ci --ignore-scripts, package-lock.json
Build / check	bun run build, bun run typecheck	npm run build, npm run check
Root suites	bun test --timeout 30000 test/{unit,integration,ci}	vitest --run --project {unit,integration,ci}
coding-agent suite	bun run --cwd packages/coding-agent --bun test	npm run test --workspace=@bastani/atomic (Node)
Script tests	none	node --test scripts/*.test.mjs
Binaries	bun build --compile	unchanged, Bun pinned 1.3.14

Also migrated in the same PR, so nothing is left contradictory: every .github/workflows/test.yml job and publish.yml (which would otherwise install from a deleted bun.lock), prek.toml hooks renamed off the bun- prefix, engines (node >=22.19.0, bun >=1.3.14), packageManager: bun@1.3.14 removed so Corepack cannot drive npm through Bun, and AGENTS.md / CONTRIBUTING.md / DEV_SETUP.md / docs/ci.md / packages/coding-agent/docs/development.md.

The package-manager decision, and what happens to the security control

No security control is dropped. bunfig's [install] block ports 1:1 to a committed .npmrc:

bunfig	.npmrc	Note
minimumReleaseAge = 259200	min-release-age=3	seconds → days; verified present in npm 11.17.0
minimumReleaseAgeExcludes	min-release-age-exclude[]	same three pi packages, same semantics
exact = true	save-exact=true
linker = "hoisted"	—	hoisted is npm's only layout; the resulting tree is identical, symlinks included
saveTextLockfile = true	—	moot; package-lock.json is already reviewable JSON

Because npm ci installs the lockfile verbatim, the age gate only binds when the lockfile is updated. Two reinforcements: the .npmrc is committed, so it binds every contributor's npm install rather than only CI, and .github/dependabot.yml gains cooldown: 3 on all three ecosystems so automated bumps cannot outrun it.

This is also a defect fix. package-lock.json was already tracked and already the input to the published shrinkwrap, and npm ci failed on main (prek@0.4.5 vs 0.4.10, vitest@4.1.9 vs 4.1.10, tsx, @napi-rs/cli). The repository maintained two lockfiles and verified one, while the unverified one fed a published artifact. Regenerating it changes 62 already-implied devDependency versions — the notable ones are prek 0.4.5→0.4.10, vitest 4.1.9→4.1.10, tsx 4.23.0→4.23.1, @napi-rs/cli 3.7.2→3.7.3, vite 8.1.0→8.1.5 — called out here rather than left to ride inside a large lockfile diff.

Distinct test counts

Diffed on test name, not file + name: Bun attributes barrel-re-exported tests to the importing file, so file attribution legitimately changes. scripts/compare-test-inventory.mjs is committed and gates on an empty symmetric difference plus a floor.

Suite	Before	After	Symmetric difference
test/unit	4417 (579 files)	4417 (579 files)	empty (8 reviewed renames)
test/integration	289	289	empty
test/ci	32	33	empty; +1 new contract
packages/coding-agent	2928	2928	—

The brief's 4426/581 does not reproduce; the merge-base measures 4417/579, and that frozen number is the contract. Both skipped tests are still skipped, asserted by --expect-skipped.

Vitest reports 5398 records for those 4417 distinct unit names: Bun's single process caches a shared module so a test defined there registers once, while vitest isolates files so it registers per importer. That is more execution, not more coverage — hence the name-level diff.

The eight reviewed renames (all in the two rewritten guard suites)

Old	New
duration parsing attributes each sample to its file and ignores non-test lines	…and skips tests that never ran
GitHub Actions log groups do not become part of the test file path	reported absolute file paths are normalised to repository-relative posix paths
an explicit budget still resolves from output wrapped in Actions log groups	…from the file path the report names
a run that prints no per-test durations is reported blind, not clean	a run that reports no per-test durations…
the spawned suite never inherits Bun's per-test record suppression	every reported test is scored, including ones that printed no duration under Bun
a passing no-retry file header does not suppress an unrelated flake retry	a passing no-retry file does not…
the wrapper scores real Bun output under the agent-quiet and Actions reporters	the wrapper scores a real vitest run through the JSON reporter it requests
the wrapper scores real Bun output behind the bun run <script> form CI uses	…behind the npm run <script> form CI uses

Each old name described Bun-stdout scraping that no longer exists; each new one asserts the equivalent property of the JSON reporter. No assertion was deleted.

Wall clock

Measured on the same machine, /usr/bin/time -p, natives built.

Suite	Before (Bun)	After (vitest/Node)
test/unit	118.21 s	67.65 s
test/integration	10.79 s	8.42 s
test/ci	1.80 s	1.98 s
Root suites total	130.80 s	78.05 s
packages/coding-agent	33.91 s	32.57 s

Each suite was run twice back to back to expose order dependence; counts and results were identical.

Per-API Bun replacements

A lexer distinguished real code from Bun API names appearing inside child-process source strings. That materially narrowed the work: all four Bun.serve calls are inside bun --eval scripts and need no replacement at all, and so are 7 Bun.sleep, 4 Bun.file, 3 Bun.write, 3 Bun.spawnSync and 3 Bun.spawn occurrences.

API	Real calls	Replacement	Why a helper rather than a direct port
Bun.sleep	169	sleep → node:timers/promises	—
import.meta.dir	60	moduleDir(import.meta.url)	undefined under Node
Bun.file(p).text/json/exists	36	readText / readJson / fileExists	Bun.file is lazy, helpers eager; readJson<T> returns unknown, not any
process.execPath (spawning Bun)	29	bunExecutable()	under vitest it is Node, so a .ts child silently ran under the wrong runtime
Bun.spawnSync	15	spawnSyncCollect	Node returns status; a direct port makes every assert.equal(r.exitCode, 0) compare against undefined and pass on a failed child
Bun.spawn	14	spawnProcess	no .exited promise, no web-stream stdio, no stdin.flush(), and ENOENT reported asynchronously so a caller's catch loses the code
Bun.write	10	writeFileEnsuringDir	Bun.write creates parent directories; fs.writeFile throws ENOENT
Bun.Glob	2	tinyglobby	already installed
Bun.Archive	2	tar-stream (new devDependency)	restores symmetry with the JSZip half of the same file
Bun.YAML.parse	1	yaml, promoted to an explicit devDependency
import { $ } from "bun"	1 file	spawnSyncCollect	the only bun module import; the commands, bun pm pack included, are unchanged

Two Bun surfaces were left alone on purpose. test/unit/fixtures/blocking-tool-extension.ts is only ever passed as a path to a Bun-spawned CLI, and the seven packages/ files behind isBunBinary/isBundledBuild are shipped runtime code.

One of those shipped files needed handling rather than editing: packages/web-access/subprocess.ts calls Bun.spawn/Bun.sleep unguarded, and its five unit tests import it in-process. Re-executing that file under Bun would have collapsed five distinct test names into one wrapper assertion — a real coverage loss to work around a runtime detail — so installBunGlobal() supplies the two primitives instead. All five names and all five assertions survive; what is no longer covered is Bun's own spawn implementation, and the product file is untouched.

Duration guard: rewritten, not retired

The brief expected vitest output to trip the guard's blind state. It would not — and the truth is worse. resolveDefaultTimeoutMs accepted only a bun/bunx binary whose subcommand was test, so under vitest it returned undefined, enabled became false, and blind was unreachable: the guard silently became a no-op. That was already live, because agent-suite wrapped bun run --cwd packages/coding-agent --bun test, which resolves to vitest. That job's gate has been scoring nothing for as long as it has existed.

Rewriting it strictly improves it:

	Bun stdout parser	vitest JSON reporter
Unit tests scored	4288 of 4417 (97.1 %)	5396 of 5398 (the 2 skipped carry no duration)
File attribution	wrong for ~67 barrel-re-export files	correct
Budget source	regex over bun test --timeout	resolved testTimeout from the config the command selects
blind	unreachable	meaningful for the first time

Deleted: the four stdout regexes, ranTestCount, perTestReportingEnv/QUIET_REPORTER_ENV, the ::group:: handling, and the bun/bunx budget allowlist. Kept unchanged: WARN_RATIO 0.4, FAIL_RATIO 0.7, the sort, renderDurationTable, .ci-diagnostics/<suite>-durations.md written on green runs too, and the declaredTimeouts source scan with its scope > name qualification. scripts/run-flaky-test-suite.ts keeps retry-once, --no-retry-file, per-attempt logs, the debug dump, the annotations, and all-attempt scoring; only its input contract changed. It requests --reporter=default --reporter=json so the step log stays diagnosable while the gate gets a machine-readable report, and it adds those flags itself so the CI command stays byte-identical to what a developer runs.

End-to-end on this branch: Samples: 33 of 33 test(s) run, budget 30000 ms resolved through npm run test:ci-contracts → config. Against the full unit suite: 0 failures, 0 warnings, worst test at 33.7 % of budget.

Both guard suites were rewritten alongside, including two tests that drive a real vitest run through the real wrapper — one invoking vitest directly, one through npm run <script>.

Contracts re-pointed

• ci-workflow-contracts: the --timeout 30000 assertion becomes "the three projects resolve to one identical testTimeout in [30000, 120000]", by importing the config. The intent — one platform-neutral budget, declared once, identical locally and in CI — is preserved verbatim, and it now also rejects a --timeout flag reappearing in any script.
• test-workflow-topology: updated for the new child commands, keeping the --parallel|--shard|--concurrent|--max-concurrency prohibition, which matters more now that vitest parallelises by default and sharding is the tempting way to hide a load-sensitive test. Adds a new contract that every work job installs with npm ci --ignore-scripts and pins both runtimes it uses.
• The two required contexts are untouched. The test gate job's id, name:, matrix values and if: always() are unchanged; it runs no install, build, or test.

Tests fixed for load sensitivity

1. subagents-attempt-watchdog.test.ts — three sites set a 600/700 ms wall-clock cap and asserted the literal. The fragility was never the stalled attempt (that child stalls forever, so detection is deterministic at any cap) but the healthy fallback attempt, which must spawn a real CLI and print inside the same window. Replaced with a named STALLED_ATTEMPT_CAP_MS = 2_500, with each assertion derived from the constant. Strictly more robust, not weaker: ~4× headroom on the healthy path, detection unchanged. No skip, no fileParallelism: false, no serialization.

2. subagents-async-event-journal.test.ts — surfaced a genuine crash vector. runPiStreaming creates its transcript WriteStream with no error listener, so any failure to open or write it becomes an uncaught exception that kills the host process, asynchronously and long after the call returned. Under load the open landed after the test's scratch directory was removed and failed the whole run. Fixed at the source, in packages/subagents/src/runs/background/subagent-runner-streaming.ts: the transcript is diagnostic, so it now degrades with a logged warning exactly as the child-event journal beside it already does. This is the only shipped file changed for behaviour.

3. slash-dispatch-resume.ts had a duplicate import { InMemoryDurableBackend }. Bun tolerated the repeated binding; every standards-conforming parser rejects it. Latent bug, removed.

4. web-access-subprocess.test.ts and subagents-pi-spawn.test.ts needed the Bun runtime named explicitly rather than inherited from process.execPath — the shipped resolver only treats a .ts CLI as runnable when the exec path is Bun, which is precisely what those tests assert.

No test was skipped, disabled, weakened, or deleted, and the assertion style stays node:assert/strict throughout.

Deliberately declined

pi's CI is one ubuntu-latest job with no matrix and no timeout-minutes. Copying it would delete Windows coverage, orphan the two required check contexts, and break the per-job budgets test-workflow-topology.test.ts asserts. Parity is a toolchain goal, not a CI-topology goal, and AGENTS.md now says so.

One deviation from the plan, on evidence: oven-sh/setup-bun stays in static-checks. The plan had it removed, but that job runs npm run docs:check, which is bun run scripts/validate-docs-links.ts. Every work job now sets up both runtimes, and the new topology contract asserts it.

Left for follow-up

• Codemod from "bun:test" → from "vitest" across 629 files and delete the shim alias. Deliberately separate: bundling a 629-file mechanical rewrite with a package-manager swap, a CI rewrite and a guard rewrite would make behavioural regressions indistinguishable from churn. scripts/compare-test-inventory.mjs is committed so the same proof can be re-run when the shim goes.
• Adopt Biome for lint and format, as pi does, and fold it into npm run check. It would reformat effectively every file and drown this diff. Until then lint is an alias for check rather than pretending to be a linter.
• The 11 packages/coding-agent tests that fail without built native bindings fail identically under Bun and Node — pre-existing, unrelated, and green once packages/natives is built (as CI does).

Validation

bun run typecheck                              clean
bun run lint                                   clean
vitest --project unit          x2   579 files  5396 passed | 2 skipped   67.65 s / 66.35 s
vitest --project integration   x2    34 files   469 passed | 1 skipped    8.42 s
vitest --project ci            x2     4 files    33 passed                1.98 s
npm run test --workspace=@bastani/atomic  x2   2898 passed | 30 skipped  32.57 s
npm run test:scripts                            3 passed
npm ci --ignore-scripts (clean checkout)        2.9 s, workspace symlinks intact
run-flaky-test-suite.ts -- npm run test:ci-contracts   33/33 scored, budget 30000 ms

---

Repair pass — independent verification findings

Reviewed against a full independent reproduction. Every finding is fixed in substance; nothing was argued away.

Blocking: the bun:sqlite loss was eleven tests, not one

src/core/tools/resource-selectors.ts loads bun:sqlite through createRequire and throws "SQLite selectors require Atomic's Bun runtime" without it. Moving packages/coding-agent off bun --bun test therefore did not fail its SQLite suites — it emptied them:

File	Shape	Effect under Node
resource-selector-tools.test.ts	(loadBunSqlite() ? it : it.skip)	1 test skipped (the 29→30 skip the reviewer found)
resource-selector-tools.test.ts	if (sqlite) { … } around 2 expects	test green, 2 assertions dead
read-sqlite-search-meta-parity.test.ts	const mod = sqlite(); if (!mod) return;	5 tests green, 0 assertions
resource-write-parity-edges.test.ts	same	2 tests green, 0 assertions
resource-selector-hardening.test.ts	same	2 tests green, 0 assertions

The last three files were invisible to the skip count and to the name diff, because an early return is not a skip. Running all four under Node with the guards made hard fails exactly 11 tests, which is the size of the hole.

Fix. Those four files now form a Bun-hosted vitest project:

npm run test --workspace=@bastani/atomic       # vitest --run --project agent      (Node)
npm run test:bun --workspace=@bastani/atomic   # bun --bun vitest --project agent-bun

agent excludes exactly the files agent-bun collects, so the new agent-suite CI step is coverage, not a repeat. The guards are hard requires again through test/helpers/bun-sqlite.ts, which throws and names the right command. test/ci/ci-workflow-contracts.test.ts makes the split structural rather than a convention: every test file naming bun:sqlite must be in BUN_HOSTED_TESTS, collected by agent-bun, excluded from agent, sharing one testTimeout, carrying no if (!sqlite) return and no ? it : it.skip, and CI must run the step.

scripts/test-duration-guard.ts now steps over a leading bun/bunx so the new step is scored like every other suite — resolved at 30000 ms, 47 of 47 samples, no blindness. Only the runtime's own leading flags are dropped: an earlier draft filtered every - argument, which swallowed --project and made the guard fall back to whichever projects happened to agree. That bug has its own test.

Blocking (process): the migration's own guard, pointed at all four suites

scripts/compare-test-inventory.mjs was committed unreferenced and was never run against the suite that regressed. It now:

• takes a repeatable --candidate and unions the parts, because a suite split across runtimes must be compared whole — pointing it at one half of packages/coding-agent is precisely how the loss stayed hidden (that half alone reports 47 missing names);
• diffs the skip set as well as the name set, so a test that ran before and skips now fails the comparison instead of keeping its name and its green tick;
• auto-detects a Bun stdout log or a vitest JSON report as the baseline.

Its rules are covered by scripts/compare-test-inventory.test.mjs, which CI executes in static-checks via npm run test:scripts. The comparison itself stays a migration-time gate — it needs a baseline captured from the runner being replaced, and after this merges there is none — and AGENTS.md records the four invocations and that reason.

Re-run against a baseline captured at HEAD~1 under Bun:

Suite	Baseline	Candidate	Missing	Newly skipped	Skipped
test/unit	4417	4423	0	0	2 → 2
test/integration	289	289	0	0	1 → 1
test/ci	32	34	0	0	0 → 0
packages/coding-agent	2893	2893 (agent ∪ agent-bun)	0	0	29 → 29

29, not the 30 the Node-only run produced. The eight unit renames are the allowlist already documented above; test/ci carries one more (every Bun test suite entry point declares one shared per-test timeout → every test suite entry point resolves to one shared per-test timeout). The six new unit names are the tests added by this repair.

Everything else

• Named timeout constants. test/unit/flaky-test-suite-runner.test.ts uses REAL_VITEST_SUITE_TIMEOUT_MS at both structural call sites. The PR required this in AGENTS.md and then broke its own rule with two bare 120_000 literals.
• Aliased declarations have a fixture again. declarationPattern still reads const runTest = built ? test : test.skip; the fixture covering it was dropped in the rewrite. Restored, in both argument shapes, plus an it alias.
• Both uncovered guard paths. run-flaky-test-suite.ts now treats an unreadable report as blind, not only a missing one — a corrupt report measures exactly as much as no report, and would otherwise have thrown inside the scorer. Two fixture modes cover what nothing did: a suite that writes no report at all, and a corrupt report whose deterministic failure is found by the log scan in findFailedDeterministicFile.
• setDefaultTimeout does what it says. It clamps with Math.min against the declared budget instead of an unconditional vi.setConfig. TEST_TIMEOUT_MS moved to a leaf module so the shim does not pull vitest/config into 629 files' workers for one number; the clamp is unit-tested.
• bumpNpmLock is exercised. test/unit/bump-version-script.test.ts's fixture root has a package-lock.json, so the branch is entered: workspace entries and first-party ranges stamped, third-party pins and link entries untouched, 3 workspace entries reported. The no-lockfile case is kept as an explicit test rather than the accidental default.
• WARN_RATIO is unchanged, and AGENTS.md now says why it must stay so: vitest's transform cost put the slowest unit test just under the 40 % warn line, and a warning there is the gate working.
• The Bun.spawn fidelity gap is stated for reviewers, in AGENTS.md as well as the helper: installBunGlobal() keeps all five web-access-subprocess names and assertions and everything subprocess.ts itself does, but Bun's own spawn implementation — what the shipped binary runs on — is no longer covered by those tests.
• Docs follow: docs/ci.md, DEV_SETUP.md, packages/coding-agent/docs/development.md. A stale paragraph describing the removed 500-line file-length gate is gone.

Validation after the repair

npm run check                                        clean (tsc + shrinkwrap)
npm run test:unit                        580 files   5402 passed | 2 skipped
npm run test:integration                  34 files    469 passed | 1 skipped
npm run test:ci-contracts                  4 files     34 passed
npm run test --workspace=@bastani/atomic  370 files   2852 passed | 29 skipped
npm run test:bun --workspace=@bastani/atomic         47 passed | 0 skipped
npm run test:scripts                                 13 passed
npm ci --ignore-scripts --dry-run                    clean
run-flaky-test-suite.ts -- npm run test:bun ...      47/47 scored, budget 30000 ms
compare-test-inventory.mjs, all four suites          0 missing, 0 newly skipped

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[greptile-apps]

Too many files changed for review. (1799 files found, 100 file limit)

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
bastani	🟢 Ready	View Preview	Jul 30, 2026, 6:22 PM

💡 Tip: Enable Workflows to automatically generate PRs for you.