chore: backport upstream server fixes (2026-09-29)

[TheDarkPhantom]

Backports a batch of upstream server stability and speed fixes from pingdotgg/t3code (upstream/main at d2c9281). Each fix is its own commit. Cherry-picks keep the -x trailer, and hand-ports carry Ported from pingdotgg/t3code#NNNN.

Landed

Cherry-picked (conflicts resolved to keep the fork's structure):

• fix(server): background git status fetches no longer fill the disk with failed repacks pingdotgg/t3code#13812 Background git status fetches pass --no-auto-gc, so a failing auto-gc can no longer leave tmp_pack_* files on every poll
• fix(server): the SQLite WAL file shrinks back after large writes pingdotgg/t3code#13684 PRAGMA journal_size_limit (32 MB) shrinks the WAL back after large writes. Placed after the fork's foreign_keys PRAGMA
• perf(server): stop remapping every thread on each thread event pingdotgg/t3code#13720 Projector patches one thread per event instead of mapping over every thread
• fix(git): exclude SSH ports from provider URLs pingdotgg/t3code#12537 ssh://host:port/... remotes use the hostname only, so the SSH port no longer ends up in the provider's HTTPS URL
• fix(opencode): accept v2 serve ready line when spawning server pingdotgg/t3code#13651 OpenCode v2's "server listening" startup line is now accepted

Hand-ported:

• fix(sqlite): retry failed statement preparations pingdotgg/t3code#10584 A failed SQL prepare is no longer cached for 10 minutes. Cache.makeWith uses an exit-based TTL (in the fork's apps/server/src/persistence/NodeSqliteClient.ts)
• perf(server): build the thread list snapshot without decoding it twice pingdotgg/t3code#13693 The shell and archived shell snapshots are built from rows that are already decoded, without a second whole-snapshot decode. The cherry-pick conflicted heavily, so I made the same change by hand
• perf(server): shutdown no longer rewrites every stopped session row pingdotgg/t3code#13688 Shutdown runStopAll skips rows that are already stopped with no active turn. The analytics property is now stoppedSessionCount. The test is adapted because the fork has no recording analytics helper
• fix(server): installed editors no longer vanish when discovery is slow pingdotgg/t3code#13669 Editor discovery wraps lookups in withPathDirectoryListings: each PATH directory is listed once, relisted when its mtime changes, and only names that exist are probed. The fork has no CommandResolutionCache, so the test drops it
• fix(terminal): settling a thread closes its idle shells pingdotgg/t3code#13673, settle half only. New TerminalManager.closeIdle re-inspects each running terminal and closes only the ones at an idle prompt (no subprocess, and no input or output during the check). ProviderCommandReactor now handles thread.settled and calls it while the thread is still settled
• perf(server): cut idle wakeups from the Connect relay and session reaper pingdotgg/t3code#13774, session-reaper part only. listBindings({ excludeStopped: true }) filters stopped rows in SQL
• fix(acp): keep one answer when a running tool reports progress pingdotgg/t3code#13386, only the shownToolCallIds logic. Only the first time a tool call is shown splits the assistant segment. Later progress and repeated completions for that call no longer split the reply. I confirmed the new test fails without the fix

Skipped or partial, and why

• fix(terminal): settling a thread closes its idle shells pingdotgg/t3code#13673, setup-script half: the fork's ProjectSetupScriptRunner has no completion-sentinel observation to attach the close to. The upstream async-prompt helper detection in deriveSubprocessInspectResult depends on its process-table inspector, which the fork lacks. Without it, a terminal whose prompt theme forks a helper process stays open, which is the safe direction. The upstream settle also stops the session with onlyIfSettled, but that needs a contracts change, so it is not ported.
• perf(server): cut idle wakeups from the Connect relay and session reaper pingdotgg/t3code#13774, AgentAwarenessRelay, cloud and relay parts: out of scope, as the batch instructions required.
• fix(acp): keep one answer when a running tool reports progress pingdotgg/t3code#13386, upstream tool-update throttling: the fork doesn't have it. The upstream test's expected event sequence still passes on the fork.

Verification

• tsgo --noEmit passes for apps/server and packages/shared
• vp test run on the touched suites: Sqlite, NodeSqliteClient, ProviderService, ProviderSessionDirectory, ProviderSessionReaper, projector, projector.settled, ProjectionSnapshotQuery, ProviderCommandReactor, ProjectSetupScriptRunner, terminal Manager, AcpJsonRpcConnection, externalLauncher, opencodeRuntime.cliParsers, shared shell and sourceControl. All pass locally.
• GitVcsDriverCore.test.ts: the new --no-auto-gc test passes. One test fails locally, "preserves newline characters in worktree paths", because Windows can't create paths containing newlines. It is unrelated to this change.
• CursorAdapter and GrokAdapter tests fail locally with spawn EFTYPE, because Windows can't spawn the shebang mock agent directly. This is an environment issue, so CI is the real check.
• parseChangelog.test.ts passes. vp lint and vp fmt are clean on the changed files. vp run check:avicode passes.

🤖 Generated with Claude Code