Interactive startup hangs forever (blank TUI, no timeout) when a plugin marketplace dir contains macOS File Provider dataless placeholders; plus silent infinite retry on rejected API key

Claude Code bug report — interactive startup hangs forever on a plugin-marketplace directory containing macOS File Provider dataless placeholders

_Prepared 2026-08-13 (recovery session). File via /bug when convenient. All findings below are
empirical, two-sided-proven (removal cures + re-addition reproduces), from a multi-day diagnosis._

Summary

If settings.json registers a directory-source marketplace (extraKnownMarketplaces with
source.path) whose install location lives on a macOS File Provider volume (Dropbox in
CloudStorage mode), and any files in that tree are dataless placeholders while the File
Provider daemon is unhealthy, interactive claude hangs forever at startup with a fully blank
TUI — no spinner, no timeout, no error, no log line. claude -p (non-interactive) does NOT
hang, which misdirects diagnosis toward TUI/terminal config for days.

Environment

• Claude Code v2.1.228–v2.1.232, macOS Darwin 25.5.0 (Apple Silicon)
• Auth: Claude Max (keyless)
• Marketplace: directory source at

/Users/leeromsegal/Dropbox/2026 Coding/Genome/klick-ai-playbooks
(through the ~/Dropbox symlink → ~/Library/CloudStorage/Dropbox-LeeromSegal/…)

• Tree state: 273 files, of which 57 were dataless placeholders (ls -lO shows dataless),

while fileproviderd + the Dropbox File Provider extension were wedged (~100–200% CPU,
network counters frozen — materialization requests never serviced).

Repro

1. Put a directory marketplace on a File Provider volume; evict part of it (or let the

provider keep it online-only).

1. Wedge or suspend the File Provider daemon (in the wild: the Dropbox FP extension spinning

in sqlite change-enumeration; simulatable with kill -STOP on the provider extension).

1. Register it in settings.json under extraKnownMarketplaces + enable a plugin from it

in enabledPlugins.

1. Launch interactive claude → permanent blank screen. Remove the two keys → instant paint.

Two-sided proof (bisect over full config)

• Full config → BLANK. Config minus env/tui/hooks/statusLine (each) → still BLANK.
• Config minus enabledPlugins+extraKnownMarketplaces → PAINTED.
• Minimal config + only those two keys → BLANK. Same, with only the Dropbox-resident

marketplace entries stripped (other marketplaces kept) → PAINTED.

• Mechanism: a bare cat of one 11KB dataless file in that tree blocks >30s (unbounded);

Claude Code startup evidently performs equivalent blocking reads on the marketplace dir.

• Repointing the marketplace source.path to a fully-material local copy

(~/dev/klick-ai-playbooks), all other config byte-identical → PAINTED.

Why this is worth fixing

• The failure is silent and unbounded: no timeout, no stderr, no entry in claude doctor

territory. A user whose cloud-sync daemon is unhealthy loses Claude Code entirely, with
zero pointer to the cause.

• -p mode completing normally actively misleads diagnosis (looks like a TUI/terminal bug).
• Dropbox/OneDrive/iCloud CloudStorage homes are common; any .claude asset (marketplace,

plugin, hook script) on such a volume can be a placeholder at any time.

Suggested fixes (any one would have prevented days of outage)

1. Bounded reads at startup: wrap marketplace/plugin directory scans in a timeout

(even 10s), then degrade: skip the marketplace, print one warning line, continue.

1. Dataless-awareness on macOS: check for SF_DATALESS/dataless flag before open;

skip or async-hydrate rather than blocking the startup path.

1. Startup phase logging: a CLAUDE_DEBUG=startup breadcrumb ("loading marketplace X…")

would have named the culprit in one run instead of a multi-round settings bisect.

1. claude doctor: flag any configured path that resolves into ~/Library/CloudStorage.

Related earlier finding (same saga, distinct bug — already known internally?)

With ANTHROPIC_API_KEY exported but invalid/unfunded, claude -p silently retried forever
(rc never returned) instead of failing fast with an auth error. Keyless Max auth worked.
Fail-fast + a clear "API key auth failed" message would fix that leg.