chore(server): port upstream Claude and stability fixes (2026-10-08)

[TheDarkPhantom]

The fork hasn't merged upstream since 2026-07-30, so it was missing several Claude and server stability fixes. This PR ports them by hand into the fork's V1 orchestration and ClaudeAdapter, one commit per fix, along with the upstream tests for each.

Ported

• fix(server): Claude /compact no longer ends early and leaves the thread busy pingdotgg/t3code#14497 Claude /compact no longer ends early and leaves the thread busy. A result from a turn Claude started itself (a leftover background task after resume, a peer wake-up) no longer completes the user's queued turn. New turns now send their turn id as the prompt uuid, so newer CLIs can echo it back in user_message_uuids. The pinned SDK 0.3.170 types don't have the echo fields yet, so the code reads them through a narrow cast.
• fix(server): stop a second server from resending Claude turns pingdotgg/t3code#13295 A second server sharing a state directory no longer resends Claude turns. After a dispatch fails, reconcile republishes only the events that dispatch appended. The test helper gained a file-backed DB option. The test subscribes with a startImmediately fork because the fork has streamDomainEvents, not upstream's subscribeDomainEvents.
• fix(server): recover claude skill scalar frontmatter pingdotgg/t3code#15452 Claude skills whose frontmatter has : inside a plain scalar (e.g. description: ... via kane-cli: run ...) appear in the $ picker again. The test drops upstream's invocation-flag fields, which the fork's skill model doesn't have.
• fix(shared): DrainableWorker keeps running after a failed item pingdotgg/t3code#16223 DrainableWorker keeps running after an item fails, drain resolves after the scope closes, and a refused offer is no longer counted.
• feat(models): add Claude Haiku 5.5 and retire Sonnet 5 and Opus 5 to legacy pingdotgg/t3code#16903 (partial) Adds Claude Haiku 5.5 to the fork's hardcoded model list in ClaudeProvider.ts. Upstream made this change in model-manifest.json, which the fork doesn't have. Haiku 5.5 offers effort low to max (default medium) plus ultrathink, keeps Extra High, and defaults to the 1M context window ([1m]) with 200k as an option. It needs Claude Code 2.1.293, with an upgrade nudge at the top of the ladder. It also gets the haiku-5.5, claude-haiku-5.5 and claude-haiku-5-5 aliases and the New badge.

Skipped

• fix(server): Claude subagents with their own model no longer show the parent's model pingdotgg/t3code#14540 (subagents with their own model show the parent's model): this depends on upstream's subagent observability work (feat: native subagent & workflow observability pingdotgg/t3code#5219 / fix(server): keep the authoritative subagent model when snapshots race task_started pingdotgg/t3code#7583): the taskAgents registry, pendingTaskModels, the task.updated runtime event, and model/agentId on task payloads. The fork has none of these, so the bug can't happen here.
• feat(models): add Claude Haiku 5.5 and retire Sonnet 5 and Opus 5 to legacy pingdotgg/t3code#16903, the rest: Sonnet 5 and Opus 5 stay current (Sonnet 5 is still the fork's default model). The bare sonnet and haiku shortcuts are unchanged. haiku still points to Haiku 4.5, which has no minimum CLI version and is the default text generation model.

Verification

• vp test run on ClaudeAdapter, OrchestrationEngine, ClaudeSkills, DrainableWorker and shared model.test.ts: 121/121 pass.
• Confirmed that the new fix(server): stop a second server from resending Claude turns pingdotgg/t3code#13295 test fails without the engine fix.
• ProviderRegistry: all 26 Claude tests pass. The non-Claude failures (codex binaryPath re-probe, OpenCode removals, cursor merged snapshot) are flaky and also fail on unmodified main on Windows.
• DrainableWorker consumers (5 reactor suites plus AgentAwarenessRelay): all pass except one CheckpointReactor test, which fails the same way on the old worker because of a Windows CRLF ('v2\r\n') mismatch.
• tsgo --noEmit for server, contracts, shared and web: no errors. vp lint and vp fmt are clean (one existing warning in ClaudeAdapter.test.ts). vp run check:avicode and vp run test:avicode (7/7) pass.

🤖 Generated with Claude Code