Plugin Install Ecosystem Report
Installs cleanly, but only what runs is trusted
12,568 install cleanly (L4) but only 6,448 run (L5); 6,755 of the tested registry never reached runtime verification. 13,325 registry plugins × 3 dsh versions, one ladder.
A Verdict mix per versionstacked to max 13,324
B L4 installed vs L5 runsdecided pass rate
C Registry mix · 0.1.3-alpha.113,325
D Install ladder L1 → L5 · 0.1.3-alpha.1bar = pass rate
E Install channel duel · 0.1.3-alpha.1delta 22.1%pp
F L5 runtime-verified trendpeak rc.1
01 Findings
About half the registry runs — three versions in a row
6,448 runtime-verified (48.4% of the registry), after 5,543 on 0.1.2-alpha.1 and 6,233 on 0.1.2-rc.1; test coverage reached 100%.
Scope. Runtime verdicts state installability/loadability in a sandbox — nothing beyond the smoke sequence. Limitation. Coverage tracks queue/instrumentation reach, not ecosystem quality. Untested ≠ failing.
Failures are only ~22.6%; the gray bucket is plain dependencies, not breakage
3,009 failed; 3,746 installed but never declared dsh.bundle — the dsh CLI installs them as plain dependencies, unmounted as plugins, so the web runtime gate does not apply.
Scope. Record layer keeps unknown separate (L5.4 notFound → unknown, neither pass nor fail). Limitation. “Plain dependency” ≠ unusable in dsh; other profiles are not tested by this report.
Release dips are bookkeeping, not breakage
Runtime-verified moved 5,543 → 6,233 → 6,448 while the registry kept growing; older verdicts are kept as history and re-queued.
Scope. Version scoping is deliberate — verdicts are never silently upgraded or deleted. Limitation. Window is a retrospective; figures are build-time snapshots.
Installed ≠ running — the core signal of this window
12,568 install cleanly (L4) but only 6,448 run (L5); 6,755 of the tested registry (failed + plain deps) never reached runtime verification.
Scope. L5 = sandbox install → web boot → HTTP → plugin inventory active. Limitation. Per-plugin detail: GET /data/install/<id>.json.
02 Methodology & FAQ
- Verdict rule (L5 wins): newest real L5 verdict per version wins; no L5 → newest L4. Only L5 pass counts as “runs / install verified”.
- State layers (record layer unchanged): Failed = install/boot/HTTP failure. Gray = installed but no dsh.bundle — dsh installs it as a plain dependency, unmounted as a plugin (web runtime gate not applicable). Record layer calls it unknown (neither pass nor fail); this report names it by its semantics, never as failure.
- Window/version scope: fixed retrospective over the three versions; verdicts hold only for the version tested; older coverage uses today’s registry as denominator.
- Reproduce: GET /data/install/<id>.json per plugin · /verification/.
Q Is “no dsh.bundle” a failure?
No. It installs, but dsh treats it as a plain dependency and never mounts it as a plugin — the web runtime gate is not applicable. Not “unusable in dsh”, not a failure.
Q Why does 0.1.3-alpha.1 show fewer L5 than rc.1?
Version-scoped bookkeeping plus per-version sweep rhythm: after a release, older verdicts stop counting until the retest queue catches up. A dip is accounting, not breakage.
Cite this report
dsh.so — DeepSeek Harness Plugin Install Ecosystem Report, Vol.1 (published 2026-09-07). Window: 0.1.2-alpha.1 → 0.1.2-rc.1 → 0.1.3-alpha.1 · verification batch started 2026-9-6 14:21:35（UTC+8） Registry: 13,325 · runtime verified (L5) 6,448 · installed (L4) 12,568 · failed 3,009 https://dsh.so/reports/install-vol1/