[feigeCode]

Description

Two commits, both aimed at getting this branch to a state where CI can go green.

1. 8a9b55c — documentation for the two new shell host modules added by feat(shell): 宿主新增 navop.i18n / navop.theme 两条 host 模块 navop#384
   (navop.i18n, navop.theme).
2. b2fb914 — fixes for the three guards that have kept main's Repository checks red since
   2026-10-09, plus one version drift of the same family.

Commit 1 — the docs (why they have to travel with the host PR)

The extension skill is the only place an extension author learns the module surface, and nothing
generates it from the host enum. On its own, the host PR leaves every extension author reading a
table that is missing two rows.

• references/manifest.md — two rows in the shellViews[].modules table + the module
  implementation path list.
• references/script-api.md — the import block, and §6 reworked into 6.1/6.2/6.3.
• SKILL.md — one checklist line.

Two things the docs now state explicitly, because getting either wrong is a silent failure:

• cx.theme() and navop.theme are not one naming scheme. cx.theme().colors keeps serving
  gpui-base's 18 semantic tokens; the component-library names the host actually paints with
  (list_active, table_hover, sidebar_*, chart_*) only exist under navop.theme. The
  component theme has no surface / destructive — it has popover / danger.
• navop.i18n hands over the language, not a catalogue, and the language must be read inside
  render(). A language switch re-runs the existing instance's render(), not init(), so a
  catalogue chosen once in init() stays stale and reports no error.

Commit 2 — main's three pre-existing guard failures

main has been red on Repository checks for five consecutive pushes (last green: 2026-10-08,
run 37758388874). Those three failures block the same check this PR renders on, so they are fixed
here. No version bumps and no releases beyond the one drift repair called out below.

a. Windows x86 backfill matrix resolves every supported published extension (expected 18 / actual 19)

4e3825a added the gemini-acp extension but did not update the hard-coded list inside the guard.
node scripts/backfill-windows-x86.mjs matrix all returns 19 entries; the extra one is
gemini-acp, sorted after duckdb. Count 18 → 19, one line added to the list.

b. IPC driver manifests expose context menu actions for supported object workflows

iotdb and tdengine both do implement schema/dump_ddl (iotdb:
internal/drivers/iotdb/server.go:188; tdengine: extensions/ipc/tdengine/src/handlers.rs:542
dispatched from src/driver.rs:167) but neither declares a single DumpSql* action in its
driver.json. So this is a missing menu, not a false method declaration — removing
schema/dump_ddl would have been the wrong repair.

The other seven drivers that declare schema/dump_ddl (dm, duckdb, gbase8a, gbase8s, kingbase,
opengauss, oscar) all use the same shape: ImportExport.export_* labels,
Database/Schema/Table targets, ContextMenu. Both drivers now declare the three actions in
that shape. The label keys come from the host
(crates/db_view/locales/db_view.yml:1818/1822/1826), so no driver locale needed a new key.

c. repository manifest is maintained as a lightweight marketplace index

gemini-acp's entry in manifest.json (also hand-added by 4e3825a) carried two fields no other
entry has:

• artifacts — R2 plugin-manifest data; the lightweight index explicitly forbids it (the guard
  asserts artifacts / asset_urls / fallback_asset_urls / sha256s are absent).
• engines — not present in gemini-acp's own source manifest at all, and the 42 other entries omit
  it even where the source manifest declares it (dm, duckdb, the composite extensions). It was
  fabricated by the hand edit.

Both removed, restoring the uniform 8-field entry shape.

d. Version drift in the same family

fba18c4 release(acp-agent) bumped claude-acp / codex-acp to 0.1.1 in the index only. The
normal release pattern (4e435b5) updates the source manifest and the index together, and the
remote tags claude-acp-v0.1.1 / codex-acp-v0.1.1 both exist — so the release happened and the
source manifests were simply left behind. Both acp_agent.json files bumped to 0.1.1.

Observation, deliberately out of scope

oceanbase's driver.json labels its dump actions with ImportExport.dump_sql_structure /
dump_sql_data / dump_sql_structure_and_data. Those keys do not exist in the host locale
(only the export_* family does), so its three export menu entries cannot resolve a label. That is
a separate defect on a published extension, with a user-visible consequence, and it needs its own
decision — it is not fixed here.

Related

• Host side: feat(shell): 宿主新增 navop.i18n / navop.theme 两条 host 模块 navop#384
• Host enum these docs mirror:
  ../navop/crates/extension-runtime/src/extension/manifest/contributes/shell.rs

How to Test

node --test tests/scripts.test.mjs
# before this branch: 139 tests, 136 pass, 3 fail
# after:              139 tests, 139 pass, 0 fail

node scripts/changed-extensions.mjs <base-sha> <head-sha>
# -> 4 entries: iotdb, tdengine, claude-acp, codex-acp

Because the fix commit touches real extension directories, CI now runs more than Repository checks on this PR — the previous push had an empty matrix:
Test changed extensions for the four extensions, and Build Windows x86 for iotdb, claude-acp
and codex-acp (windows_x86: true). That path is known-good: run 37893458082 shows the acp test
and Windows-x86 packaging jobs passing, with only Repository checks failing.

Local runs of the two extension test suites whose files this PR edits:

CGO_ENABLED=0 go test ./internal/drivers/iotdb/...              # ok, 0.381s
cargo test --manifest-path extensions/ipc/tdengine/Cargo.toml   # 61 passed + 6 passed, 0 failed

iotdb's test only asserts that certain actions/forms are present, and tdengine's
declared_methods_cover_driver_json_set compares methods — which this PR does not touch — so
neither can regress from added actions. Verified anyway rather than assumed.

Note on the Go run: a plain go test fails to build on this machine with
unknown architecture ... arm64e ... tapi error: malformed file against the local
MacOSX27.0.sdk; that is a local SDK/linker mismatch, not code. CGO_ENABLED=0 builds and runs
the test, and CI runs it on ubuntu-latest regardless.

AI Assistance

AI-generated, human-reviewed. AI drafted the three markdown files and the four data-file repairs;
I reviewed each against the host implementation and against the existing repo conventions. No
generated Rust, Go, Java or JS was merged.

Specifics, since the fix commit reasons about facts that are easy to get wrong:

• The direction of each repair was checked against evidence, not inferred: the iotdb/tdengine
  implementations were located before deciding to add menus rather than drop method declarations,
  and the acp version direction was confirmed by the presence of the remote release tags.
• The label family to copy was chosen after enumerating all ten drivers that declare
  schema/dump_ddl — three shapes exist, and the dominant one (7 of 10) was used. The
  dump_sql_* family was rejected because the host locale does not define those keys.
• The remaining out-of-scope defect (oceanbase) is reported rather than silently fixed.
• Every number and file:line in this description was re-derived from the tree or from the CI log
  at the time of writing.