Migrate to 1.1

Tantra 1.1 keeps the 1.0 journals and store schemas. Existing JSON, filesystem, SQLite, and PostgreSQL data loads without a migration or historical backfill.

Child actors

Child human asks are no longer supported. Move approval-gated work to the root. When a child needs input or authority, have it send() a message to its direct parent. Runtime construction now rejects static child permissions whose effective verdict is ask, and dynamic child escalation fails without emitting AskRaised.

Plain child completion is not a result. It leaves the child reusable and sends the direct parent a status-only lifecycle input. A completed assignment must call finish(result) exactly once to close the child and deliver its result. Parents can use the injected status(agent_id) tool for direct children; applications can use Runtime.status() and Runtime.tree_status() without reading journals.

Applications should poll actor status and subscribe to child journals only when detailed output is visible or otherwise needed. Keep a separate cursor for every subscribed actor. ActorStatus.active is process-local, not distributed ownership.

spawn() accepts an optional durable display name. The registered agent remains the reconstruction type. Existing records have no name and display their actor type; no backfill is required.

Sarathi subagents and skills

Sarathi now registers sarathi and one general-purpose subagent. Legacy researcher journals and headers remain readable, but those actors cannot resume or accept new work because the type is no longer registered. Start new delegated work with subagent.

Research specialization moved into the packaged research skill. Sarathi works inline by default and delegates only when explicitly requested or independently useful. Put shallow, normal, or deep in the task when a particular research level matters; omission means normal. Subagents have non-interactive read/search capabilities and cannot use approval-gated memory writes or ask a human.

Compaction and providers

Compaction remains opt-in. Install PruneThenSummarize on the Runtime to enable complete-request budgeting, token-bounded recent history, durable summaries, and one-shot confirmed-overflow recovery. Summarization is lossy, while prompts, execution-environment blocks, tool schemas, and current input remain exact.

Provider.limits(model) may now be synchronous or awaitable. OpenAICompatible uses explicit per-model limits first, then lazy catalogue metadata, then 128_000 context and 4_096 output tokens. Configure explicit limits for endpoints whose catalogue omits or misreports them.