V0.1.3/m2/cursor windsurf adapters

[Vedansi18]

Summary

The integration branch. Adds the per-agent CLI adapters for Cursor and
Windsurf (dev plan §3 M2 §2.2 M9 + M10), wires the production
chat-history watcher to better-sqlite3 to fix the WAL-mode blocker from
dev plan §2.5, fills in the injectFn contract from the project memory
project_b4_prompt_injection_contract, and routes the extension's
view-provider selection through the new injector with clipboard
fallback. Also wires installAction / uninstallAction to actually
call the registered adapters — previously they self-registered but
nothing invoked them.

Stacked on B3 (v0.1.3/m2/webview-ui) with B2 (chat-history-capture)
and M1 (agents framework + Claude Code adapter) merged in —
B4 depends on B1+B2+B3+M1 per dev plan §3.0 dependency table.

Modules covered (per dev plan §3 M2 §2.2)

Module	File(s)	What it does
M9 — Cursor adapter	src/agents/adapters/cursor.ts
VSCodeExtensionAdapter. detect() checks for ~/.config/Cursor /
Library/Application Support/Cursor / %APPDATA%/Cursor. install() prints
deep-link instructions (Open VSX + VS Code Marketplace + `cursor
--install-extensionCLI).chatHistoryPaths()` returns the
User/workspaceStorage base. extractPrompt() is a documented stub — see
"Architectural notes" below. Self-registers via src/agents/index.ts.
M10 — Windsurf adapter	src/agents/adapters/windsurf.ts	Same shape
as cursor.ts. Detection checks both ~/.config/Windsurf/ AND the legacy
~/.codeium/windsurf/ Cascade dir (Windsurf may populate either depending on
version). chatHistoryPaths() returns both paths so the watcher can track
each.

B4 follow-up additions (architecturally required, not in M9/M10 spec)

Added	Why	Tests
WAL-mode fix — chat-history-watcher.ts's default reader swapped from
sql.js to better-sqlite3 with copy-to-staging-dir + `PRAGMA
wal_checkpoint(TRUNCATE)` strategy	Cursor 3.4.20 keeps state.vscdb in WAL
mode — main file is ~4 KB, all writes go to .vscdb-wal. sql.js operates on a
buffer and CANNOT see WAL siblings. Closes dev plan §2.5 blocker for the
production watcher.	+4 tests covering happy-path read, no-ItemTable
defensive return, WAL-mode round-trip, source-file untouched after 3 reads
src/ext-vscode/src/host-detector.ts	The extension needs to know which
host it's inside (vscode.env.appName) to pick the right state.vscdb base
path at activation time. Returns 'cursor' / 'windsurf' /
'vscode-generic'.	11
src/ext-vscode/src/chat-input-injector.ts	Fills the injectFn contract
from project memory project_b4_prompt_injection_contract.md. Tries
host-specific candidate commands via vscode.commands.executeCommand; first
one that runs without throwing wins; otherwise returns false →
handleOptionSelection falls through to clipboard. **Candidate command IDs
are heuristic guesses — verification deferred to B5** against a live Cursor.
8
extension.ts wiring	View provider constructed with `onSelect = (text) =>
handleOptionSelection(text, { injectFn: (t) => chatInputInject(t, {host})
})`.	extension.test.ts updated to 11
Registry-iteration in installAction + uninstallAction	Before this
branch, cursor/windsurf adapters self-registered but nothing called them.
Added a 17-line block after the legacy detectAgents for-loop that iterates
await detectAll(ctx) and runs each non-claude-code adapter's install /
uninstall. Snapshot invariant preserved (cursor/windsurf detect() returns
false in test tmpHome).	+8 (4 happy-path install + 2 uninstall + 2
catch-block via vi.spyOn on real adapter)

Files

Path	Type	Tests
src/agents/adapters/cursor.ts	M9 — VSCodeExtensionAdapter	15
src/agents/adapters/cursor.test.ts	tests	—
src/agents/adapters/windsurf.ts	M10 — VSCodeExtensionAdapter	16
src/agents/adapters/windsurf.test.ts	tests	—
src/agents/index.ts (modified)	registry side-effect imports	—
src/cli/commands/install.ts (modified)	registry-iteration in install +
uninstall	—
src/cli/commands/install.test.ts (modified)	+8 tests	—
src/ext-vscode/src/host-detector.ts	runtime host classification	11
src/ext-vscode/src/host-detector.test.ts	tests	—
src/ext-vscode/src/chat-input-injector.ts	OptionInjector implementation
8
src/ext-vscode/src/chat-input-injector.test.ts	tests	—
src/ext-vscode/src/chat-history-watcher.ts (modified)	WAL-aware default
reader (better-sqlite3); readFileFn removed from API; defaultReadItemTable
exported for tests	14 (was 10, +4)
src/ext-vscode/src/chat-history-watcher.test.ts (modified)	+4 reader
tests	—
src/ext-vscode/src/extension.ts (modified)	view provider constructed
with injectFn-aware onSelect	11 (unchanged)
src/ext-vscode/src/extension.test.ts (modified)	mock setup updated for
new imports	—
src/ext-vscode/package.json (modified)	better-sqlite3 moved devDeps →
deps; sql.js removed from deps	—
src/ext-vscode/esbuild.config.mjs (modified)	`external: ['vscode',
'better-sqlite3']` (was sql.js)	—

Architectural notes worth flagging for review

Decision	Rationale
**WAL fix uses better-sqlite3 (native binding), not sql.js + checkpoint
shim**	Same approach as the dev-only scripts/dump-cursor-state.ts. Native
binding adds .vsix complexity for B6 publish (prebuilt binaries per platform)
but eliminates an entire class of WAL-correctness bugs at the source.
chatInputInject candidate command IDs are educated guesses
Verification deferred to B5 (smoke-test) where the engineer can run
vscode.commands.getCommands(true) against a live Cursor and prune/re-order
the list. The runtime getCommands(true) filter skips unavailable IDs cleanly
→ net effect on unverified hosts = clipboard fallback (safe).
extractPrompt remains a stub on both adapters	Documented
architectural choice. The CLI never decodes rows (the extension's
chat-history-watcher does, via src/ext-vscode/src/extractors/). Returning
null is contract-compliant ("I don't know"). Full migration-path JSDoc on
cursor.ts explains what to do if a CLI caller is ever added.
Extension watcher start-up is NOT wired in activate()	Intentional.
The "stop trigger" timing (when does the extension call nexpath stop?) needs
real-Cursor behaviour to settle. Wiring + smoke-testing belong together in
B5.
Branch is a 4-commit + 2-merge history	B4 depends on B1+B2+B3+M1. B3
was cut off B1 (not B2), so merges of B2 + M1 are required for B4 to have all
four contracts available. Per dev plan §3.0 dependency table.

Test plan

Automated

# Root + sub-package typecheck
npx tsc --noEmit                          # root
(cd src/ext-vscode && npx tsc --noEmit)   # sub-package — both EXIT=0

# Sub-package tests in isolation
(cd src/ext-vscode && npx vitest run)
# expected: 185/185 pass across 16 test files

# Full root test suite
npm test
# expected: 2080 passing + 18 pre-existing TtySelectFn carry-forward

# Extension bundle still builds
(cd src/ext-vscode && npm run build)
# expected: out/extension.js produced (~14.7 KB)

# Install snapshot (the zero-diff gate)
npx vitest run src/cli/commands/install.snapshot.test.ts
# expected: 1 test passes — snapshot byte-identical to pre-B4

Manual

# Verify nexpath install detects Cursor on a machine with Cursor installed
node dist/cli/index.js install --yes
# expected console output includes:
#   ✓ Cursor       — install the Nexpath extension to activate guidance:
#       Open VSX:
https://open-vsx.org/extension/nexpath/nexpath-vscode
#       VS Code Marketplace:
https://marketplace.visualstudio.com/items?itemName=nexpath.nexpath-vscode
#       Or via CLI:          cursor --install-extension nexpath.nexpath-vscode