…rkspace

In a non-workspace layout the member IS its own invocation root (standalone
package), where honoring defaultPackage is by design; the finding is about
workspace members, so the case moves to a fixture with a workspaces field
and asserts the pack runs in place.

- rewrite is_bare to mirror cac/mri parsing: a non-flag token after any
  non-boolean flag is that flag's value (required and optional alike, so
  vp dev --host 0.0.0.0 and --open /foo elicit again), and only tokens no
  flag consumes are positionals; the per-tool required-value tables are
  replaced by smaller boolean tables derived from the shipped --help
- pack workspace selectors (-W/--workspace, -F/--filter) define their own
  target set and always bypass elicitation
- document in the RFC that implicit -C runs inside the already-chosen CLI
  (single-version workspace model); full re-delegation for defaultPackage
  targets with their own install/runtime pin is tracked in #2057

The test module imports items explicitly; a restack conflict resolution
kept the tests but dropped the import, which only all-targets compilation
catches (workspace clippy now runs clean with CI's exact invocation).

- pack workspace selectors also bypass elicitation in inline form
  (--filter=ui, --workspace=packages/ui)
- honor the -- option terminator: a token after it is an explicit
  positional, and bare -- stays bare
- global vpr rejects a bare -C with no value instead of delegating it as
  a run argument (mirrors the local wrapper fix)
- local bin.ts catches process.chdir failures (e.g. permission denied on
  an existing directory) as a normal CLI error instead of an uncaught
  stack trace
- render the spread-config build case through the PTY: non-TTY steps
  record raw bytes and the Windows byte stream of a vite build differs
  invisibly from the recorded macOS one

vp pack --root packages/lib at a workspace root was a valid targeted
invocation before this PR; like -W/--workspace and -F/--filter it defines
pack's own target, so it bypasses elicitation (spaced and inline forms)
instead of hitting the picker or the non-TTY listing.

Match the documented convention in packages/prompts/src/milestone.ts: only
the PTY snapshot harness sets the env, so real terminals and piped output
never receive the OSC 8 marker bytes. The picker fixtures keep passing
because the harness sets the gate for spawned processes.

fs.statSync can itself throw (EACCES on a parent, ELOOP) before the
chdir try/catch; chdir reports every failure mode through one catchable
path, with ENOENT/ENOTDIR keeping the directory-not-found message.

- one apply_chdir helper (join/clean/validate/chdir/PWD) replaces the four
  drifting copies in vite_global_cli: main's normal path, the completion
  peek, vpr, and the clap-path fallback in cli.rs — which previously
  skipped set_current_dir and the PWD sync, silently weakening the cd
  equivalence on orderings like a second -C
- one classify() core (bare -> defaultPackage-at-root -> workspace root)
  now feeds both resolve_app_target and needs_elicitation, so the RFC's
  resolution order is written once and the two entry points cannot drift;
  the extracted defaultPackage value rides the enum, removing a duplicate
  config parse
- bin/vpr is a grammar-free shim: bin.ts inserts 'run' itself via argv0
  after its existing -C consumption, deleting the hand-synchronized second
  JS copy of the -C token grammar
- pack runnable check stats src/index.ts before parsing the config (runs
  per workspace package); PWD in the direct-subcommand env is only
  overridden when elicitation retargeted (plain runs keep the caller's
  possibly-symlinked PWD verbatim); vpr parses -C once; identical
  FieldMapInner arms collapse via or-patterns; handler's -C predicate
  simplifies to starts_with

The helper was inserted between normalize_args's bullet-list doc comment
and the function, merging the docs; clippy reads the helper's prose as
unindented list continuations under --deny warnings.

The vp-config e2e broke: a single package with a settings-only
pnpm-workspace.yaml (catalogs, minimumReleaseAge) is a workspace root, and
the runnable-root picker row made the candidate list non-empty, so
non-TTY CI got the target listing and exit 1 for a bare vp pack that
packed fine before this branch.

Deeper rule instead of the row: when the workspace root itself looks
runnable for the command, classify() never elicits and the command runs
in place, exactly the pre-elicitation behavior. The root is never a
picker/listing row; elicitation only engages when the root is not a
plausible target, which is what keeps the feature purely additive.

New settings_only_workspace fixture pins the vp-config shape non-TTY;
app_root_runnable_root now asserts in-place runs for both TTY modes.

The tty=false vite-build capture differs invisibly on Windows (same class
as the round-9 unanalyzable_config_ignored fix); settings_only_workspace
already covers the non-TTY in-place guarantee with Windows-stable tsdown
output.

…3 rebase

The rebase onto main (v0.2.3, vite-task 417b6aae) conflicted on the pin:
417b6aae ships vite_select and encoded_milestone but not the prompt param
the picker needs (#510 is still open). ca55bb232aa8 is #510's current head
(prompt param + main merged in), missing only two dependency-bump chores
from 417b6aae that don't touch the crates vp consumes. Final pin moves to
the merged SHA once #510 lands.

A bare app command walked find_workspace_root twice: once in classify to
decide whether to elicit, once in execute_direct_subcommand to set up the
session. classify now returns the root it found (Classification::RunInPlace)
whenever the command runs in the unchanged cwd, and the caller reuses it.

The hot path is a bare vp dev/build deep in a large monorepo: it now does
one upward walk instead of two, and still never enumerates packages (that
only happens at a root that actually elicits). The hint is handed back only
for the run-in-place cases, where the effective cwd equals the original, so
it always matches a fresh lookup; retargeted (-C, picker, defaultPackage)
paths still resolve the target's own root.

Fast-forward over the previous ca55bb23 pin (contains it), so no divergence.
The new head merged main, so it now carries main's v0.2.3 release commit
(417b6aae, the two dep-bump chores my earlier pin was missing) plus
ef49b4ac (fix: clarify task process wait errors, #515) and the v0.2.4 dep
bump. The prompt param the picker needs is unchanged; snapshots are
byte-stable. Final pin still moves to the merged SHA once #510 lands.

An interactive PTY case can wedge unbounded on the Namespace Linux runner
(PTY spawn/teardown flakiness under the parallel harness, the Linux analog
of the ConPTY blocking that keeps the Windows snapshot job on a
GitHub-hosted runner). The harness's 50s per-step timeout does not cover
TestTerminal::spawn, and the RwLock execution gate then stalls the whole
suite. With no job timeout that burned the 6h default and never flushed the
step log. A 10-min step bound (healthy runs finish in under 2 min) fails
fast and preserves the partial log so the wedged case is identifiable.

vpr had no snapshot coverage: add vpr -C <dir> <task> and its cd-form
equivalent (both print the target cwd, proving the global vpr shim and
local bin/vpr consume -C) plus bare vpr -C (missing-value error) to cwd_flag.

The dev/build/preview positional had only is_bare unit coverage: add
vp build apps/web to app_root_listing, proving a positional forwards to
Vite as [root] (output at apps/web/dist, cwd unchanged) and skips
elicitation entirely (no Selected/Tip lines), unlike the -C forms.

vite-task #510 (configurable selector prompt) merged, so move the pin off
the now-defunct PR-head commit (28310dc6) to the permanent main commit that
merged it. Content-identical for the crates vp consumes (empty tree diff on
vite_select/pty_terminal_test), so this is purely a stable-pin move: the
picker's prompt param and snapshots are unchanged.

Two regressions surfaced in review, each reproduced by a new snapshot fixture:

- defaultPackage declared through a TypeScript wrapper (satisfies UserConfig,
  as const) was ignored: extract_config_from_expr only stripped parentheses,
  so the wrapper made the config unanalyzable. Unwrap TS-only wrappers via
  get_inner_expression. Fixture default_package_ts_wrapper + extractor unit
  tests.
- a workspace root that builds via vite.config.ts (build.lib/SSR, no root
  index.html) was filtered out of elicitation once a member existed, so
  non-interactive vp build listed members instead of building the root.
  The root is now runnable when it declares a build block, not only when it
  has index.html (a shared lint/fmt config still declares neither). Fixture
  config_only_root_build.

The recurring Linux snapshot hang survives #2102's spawn bound (a run that
contained the bound still wedged for the full 10-min step timeout), so the
wedge is at a different, still-unbounded point (or cumulative create-test
timeouts around the ctrl-c case). Add snap_trace (gated on VP_SNAP_TRACE=1),
writing START/phase/END lines straight to fd 2 so they survive libtest
capture and reach the CI log even when the step is killed by its timeout. A
step that logs START without END is the wedged one, and the phase markers
(spawn dispatched / spawned / waiting for exit) localize within it. Enable
it on the Linux PTY CI step; off by default everywhere else.

…eview

Three review follow-ups, each reproduced first:

- unwrap TS wrappers INSIDE defineConfig too: defineConfig({...} satisfies
  UserConfig) and factory returns now analyze, not just the outer export.
  (default_package_define_config_wrapper + extractor unit tests)
- the declared-build root signal is build-only: a lib/SSR root is a target
  for vp build, but vp dev/preview still need index.html, so dev/preview at
  a lib root elicits instead of serving the root. (dev_at_lib_root_elicits)
- an explicit -c/--config file is an explicit target: vp build -c <cfg>
  forwards to Vite instead of eliciting. (explicit_config_forwards + is_bare
  unit tests)

The last hang confirmed the trace works (the mac job logged clean
START/END for every step) but GitHub discards a timed-out step's own log,
so the wedged Linux job's trace was unrecoverable. Write the trace to
VP_SNAP_TRACE_FILE instead and print it from a following if: always() step,
which runs even after the PTY step is killed by its timeout. The next
Linux wedge will show the START-without-END case in the Print PTY trace
step. VP_SNAP_TRACE=1 still goes to stderr for local runs.

config_only_root_build and explicit_config_forwards captured a vite build
with tty=false; a piped vite build's raw byte stream differs invisibly on
Windows (the round-9 unanalyzable_config_ignored class), so both failed the
Windows snapshot job. Render them through the PTY grid instead, which
normalizes the bytes; the assertions (root builds in place, -c config
forwards) are unchanged. dev_at_lib_root_elicits stays tty=false: it
captures the plain-text elicitation error, which is Windows-stable.

- vpr: restore dynamic import so it runs after enableCompileCache (a static
  import is hoisted, defeating the compile cache for the CLI bundle)
- static config: unwrap TS-only wrappers (as const, satisfies) on property
  values so defaultPackage: './x' as const stays a static string literal;
  add extractor unit tests and a value_wrapper snapshot fixture
- global -C: deliver PWD to delegated children via the spawn env instead of
  mutating our own PWD, which is unsound once the Tokio runtime has threads

The new expr_to_json TS-wrapper assertions exceeded the line width;
cargo fmt --check gates CI Lint.