feat(server): add /compact context compaction on every agent

[TheDarkPhantom]

The fork only observed compaction. Nothing could trigger it, so a long thread kept growing until the provider compacted it on its own terms. This ports upstream's /compact command (server, contracts, shared only; the web UI comes later).

How it works

• A user message of exactly /compact with no attachments is intercepted by ProviderCommandReactor and routed to the new ProviderService.compactThread instead of being sent as a prompt. It is rejected when the thread has no conversation yet, or while a turn is running.
• Every adapter now has to declare compaction (a required field on ProviderAdapterShape, following refactor(server): let adapters declare context compaction pingdotgg/t3code#10112):
  • Codex runs its native thread/compact/start, and a finished contextCompaction item now emits thread.state.changed: compacted.
  • OpenCode runs its native session.summarize({ auto: false }) and maps session.compacted.
  • Claude and Grok send /compact as a turn, and Cursor sends /compress.
  • The test adapter declares unsupported, which fails with a validation error.
• The bookkeeping lives in a new ProviderCompaction.ts, so ProviderService.ts only gains a few call sites. It allows one compaction per thread and uses a 10-minute timeout. A native run that times out blocks retries until it reports back or the session restarts. A slash-command run waits for its own turn to settle, and if that turn completes without reporting a compaction, one is synthesized. The compacted event carries the /compact message id as requestId.
• Messages sent during compaction queue and replay in order, each with the interaction mode and model it was sent with (fix(server): queue messages during context compaction pingdotgg/t3code#11107). If compaction fails, is interrupted, or the session stops, they are reported as "Queued message was not sent".
• The projection keeps the /compact pending row while messages queue behind it. It clears that row on a context-compaction or provider.turn.start.failed activity whose requestId matches, or when the reactor sets the session back to ready.
• Ingestion attaches beforeTokens/afterTokens, taken from the provider when it reports them (Claude's compact_boundary) and otherwise from the last two context-window readings. The activity summary reads "Compacted context 899K → 0 tokens".
• Claude (feat(claude): compact old threads before they burn through usage pingdotgg/t3code#8144):
  • New autoCompactWindow setting (100000 to 1000000). A settings patch with any other value is rejected, and a bad value already stored decodes to empty.
  • ThreadTokenUsageSnapshot.autoCompactThreshold.
  • onUserDialog with supportedDialogKinds: ["resume_return"], surfaced through the existing AskUserQuestion user-input path with "Compact and continue" / "Keep full history" / "Don't ask again".
  • New packages/shared/src/claudeCompaction.ts.
  • The ultrathink prefix is no longer added to slash commands, so they still run.
• All five providers advertise /compact as a slash command.

No DB migrations. The web-only follow-ups (pingdotgg#11103, pingdotgg#9623, pingdotgg#9430, and the web half of pingdotgg#8144) are listed in FUTURE_ENHANCEMENTS.md.

Ported from pingdotgg#9293, pingdotgg#10112, pingdotgg#11107 and pingdotgg#8144.

Tests cover reactor /compact routing, the empty-thread rejection, queue and replay order, queued-message cancellation on failure, per-adapter compaction (Codex, OpenCode, native and slash-command in ProviderService, unsupported, timeout quarantine), the projection's pending-row handling, ingestion requestId and token sizes, the Claude setting and dialog, the settings decode, and the shared copy.

Done by Claude Opus 5.5 (1M context) in Claude Code.

🤖 Generated with Claude Code