feat: agent sessions can spawn and orchestrate other sessions

[roughcoder]

Problem

A session cannot delegate: splitting work across providers means the user manually creating threads and shuttling context between them. There is no way for a long-running "manager" chat to run multiple flows and find out when they finish.

What this does

Adds a sessions toolkit to the per-thread t3-code MCP server, so any agent session can, from inside its own turn:

• list_session_providers — enumerate spawnable provider instances and models
• spawn_session — create a new thread (own git worktree by default, linked via spawnedByThreadId) that starts working immediately
• send_to_session / read_session / stop_session — scoped to threads the caller spawned
• post_report — post a completion report (status, markdown summary, artifacts)

A new SessionSpawnReactor watches domain events: when a spawned thread posts its report (or its provider session errors), it starts a turn on the spawning thread carrying the result — the parent is woken instead of polling. Spawned sessions automatically get report instructions appended to their first message so the contract holds on every provider.

Supporting changes: the thread bootstrap macro (create + worktree + setup script + rollback) is extracted from ws.ts into a shared ThreadTurnBootstrap service; reports persist as a projection (migrations 041/042); web UI renders a "Spawned by" banner on child threads and report cards in the timeline; Settings → General gains a Session orchestration toggle (enableSessionOrchestration, checked per tool call so it applies to running sessions instantly).

Guardrails: mutating tools reach direct children only; max 8 children per thread, 3 levels deep; a child's permission mode is capped at its parent's; spawned threads participate in normal archive/settle session cleanup.

Also: bumps @anthropic-ai/claude-agent-sdk to 0.3.226 (new SDK message types handled explicitly in ClaudeAdapter), and fixes an MCP compatibility bug — an empty tool parameter struct encodes to a typeless anyOf schema, which makes Claude Code silently drop the whole server's toolset.

Model options

list_session_providers returns each model's option descriptors (reasoning effort, context size, boolean flags) with their valid choices, and spawn_session accepts an options array of {id, value} selections — validated against the chosen model's descriptors and carried into the spawned thread's model selection. Verified live: a Claude session spawned a Codex child with reasoningEffort=low, the selection persisted on the child thread, and the completion report woke the parent (cross-provider both directions: Codex→Claude and Claude→Codex).

Verification

• Focused suites green (decider/projector/reactor/persistence/client-runtime, 77+ tests incl. new decider.reports.test.ts); contracts, server, client-runtime, and web all typecheck clean.
• Live end-to-end in the web app: a Codex parent spawned a Codex child into its own worktree; the child did the work, posted a success report with artifacts, and the parent was woken with the report and summarized it. Cross-client checked with Claude (tools load and list_session_providers returns all five providers). Settings toggle verified live: flipping it off made a running session's next tool call fail with disabled_in_settings; flipping back on restored it.
• Contract changes are additive (optional fields + decoding defaults) so existing clients keep decoding; mobile renders spawned threads as ordinary threads (banner/report-card affordances are web-only for now — mobile decision pending per adapter/provider checklist).

Screenshots of the report card, spawned-by banner, parent wake-up, and settings toggle to be attached in comments.

Built with Claude Fable 5 on Claude Code.

🤖 Generated with Claude Code