@axiom-lattice/protocols 4.10.0 → 5.0.0

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @axiom-lattice/protocols@4.10.0 build /home/runner/work/agentic/agentic/packages/protocols
+> @axiom-lattice/protocols@5.0.0 build /home/runner/work/agentic/agentic/packages/protocols
 > tsup src/index.ts --format cjs,esm --dts --sourcemap
 
 CLI Building entry: src/index.ts
@@ -8,13 +8,13 @@
 CLI Target: es2020
 CJS Build start
 ESM Build start
-CJS dist/index.js     48.46 KB
-CJS dist/index.js.map 197.00 KB
-CJS ⚡️ Build success in 718ms
-ESM dist/index.mjs     43.36 KB
-ESM dist/index.mjs.map 193.03 KB
-ESM ⚡️ Build success in 719ms
+ESM dist/index.mjs     48.07 KB
+ESM dist/index.mjs.map 211.72 KB
+ESM ⚡️ Build success in 424ms
+CJS dist/index.js     54.04 KB
+CJS dist/index.js.map 215.96 KB
+CJS ⚡️ Build success in 424ms
 DTS Build start
-DTS ⚡️ Build success in 16819ms
-DTS dist/index.d.ts  259.36 KB
-DTS dist/index.d.mts 259.36 KB
+DTS ⚡️ Build success in 12155ms
+DTS dist/index.d.ts  277.82 KB
+DTS dist/index.d.mts 277.82 KB

package/CHANGELOG.md

@@ -1,5 +1,118 @@
 # @axiom-lattice/protocols
 
+## 5.0.0
+
+### Major Changes
+
+- 8f9dd0d: Remove the deprecated TeamAgent/teammate feature and its UI surfaces.
+
+### Minor Changes
+
+- f1f8840: Durable cross-replica public-channel replies (Phase B)
+
+  - Add a durable `ChannelReplyJob` protocol, in-memory/PostgreSQL stores, and a
+    `lattice_channel_reply_jobs` migration (v200) with lease-based claim, bounded
+    reschedule, and terminal delivery state.
+  - Route non-Room public ingress for adapters declaring `durableReply: true`
+    through a reply job plus a pending `source='run_api'` run (job written first,
+    with compensation on run-create failure). Lark opts in; WeCom/WeChat/Room stay
+    legacy because their reply transport is connection-bound/process-local.
+  - Add `ChannelReplyDispatcher`, which polls due jobs on any replica, waits for
+    the run ledger to become terminal, reads the final assistant text from the
+    thread checkpoint, sends it through the adapter, and marks the job delivered.
+    Empty output and loader/send failures retry with capped exponential backoff;
+    sends have a 15s deadline; missing runs converge after a 30s grace window;
+    cancelled runs do not send a waiting notice. Terminal jobs are pruned after
+    24h.
+  - Keep Room/trusted dispatch and `ChannelConnectionLeader` unchanged. Durable
+    replies are at-least-once; outbound metadata carries `inputMessageId` for
+    channel-side deduplication. Ingress falls back to legacy when the dispatcher
+    runtime is not registered. Opt out with `CHANNEL_DURABLE_RUNS=false`.
+
+- ea90717: `keep_latest` checkpoint pruning (M2-B2b)
+
+  - Add `CheckpointMaintenance.pruneThread` and transactional per-namespace latest-N checkpoint retention with blob retention derived from `channel_versions`.
+  - Add `POST .../prune { strategy: "keep_latest", keep?: 1..100 }`; `delete` is unchanged.
+  - Reuse strict ownership, thread-run lease, and empty-inbox guards; keep the thread row and chunk stream.
+
+- 351b0e9: Multi-replica run cancel and queue-first HITL
+
+  - Add ledger-backed run cancel intent plus Redis owner wake-up and heartbeat fallback; abort/recover can be accepted by any replica without ingress stickiness.
+  - Make `input.respond` queue-first with `expected_interrupt_id` worker validation and the official LangGraph resume map; reserve answers across replicas and release in-flight reservations after enqueue.
+  - Unify Room and Web App native HITL resume mapping with the Console command path.
+  - Expose run-cancel readiness/metrics and document the A2A boundary: A2A remains a deferred adapter over the Agent runtime, not a second runtime.
+
+- 1fde314: Pending run task queue (Phase 1)
+
+  - Make `lattice_agent_runs` the durable task queue for the new single-run API: create writes `pending`, a worker claims/executes through `Agent.executeRun`, and recovery reclaims expired rows in place and completes durable rollback work.
+  - Add run lifecycle `join` (bounded terminal wait), `cancel?action=interrupt|rollback`, and HITL command runs; Web App runtime mirrors the same handlers.
+  - Add `client.runs.join` / cancel `action` / create `command`, and single-run HITL with `ThreadStream.openRun`. `RunCreateResult` is now `{ runId, messageId?, status: "pending" }`: the old queue-first `status: "queued"` / `queued: true` shape and the required `messageId` are gone (command runs have no user message).
+  - Keep the legacy message queue, Room, Project Task, A2A, and legacy HITL on `source='legacy'`; the two executors share the `thread-run` lease.
+
+- da5294d: Tenant-scoped namespaced KV store (M2-C)
+
+  - Add `StoreProtocol` (put/get/delete/search/listNamespaces/deleteExpired) with deep JSONB filters, substring query, and TTL.
+  - Add InMemory, Local SQLite, and PostgreSQL implementations (PG migration v195 with GIN and partial expiry indexes).
+  - Add `/api/store/*` routes and a single-flight store TTL sweeper.
+  - `index` hints are accepted but vector indexing and semantic search remain deferred.
+
+- 03e9984: Thread commands and explicit interrupts (M3-B)
+
+  - Add normalized `AgentInterrupt` plus `Agent.getInterrupts()`/`normalizeAgentInterrupts()` for clarify, task_review, and custom interrupts.
+  - Expose `interrupts` in thread status using one state read.
+  - Add `POST /api/assistants/:assistantId/threads/:threadId/commands` with `run.start`, `input.respond`, and `agent.getTree`.
+  - Map multitask strategies to queue modes, add process-local 5-minute answer dedupe, and return Mermaid from `get_draw_graph()`.
+
+- 2609f3b: Thread copy and prune-delete (M2-B1)
+
+  - Add `CheckpointMaintenance` and transactional `copyThread` / `deleteThread` to `PostgreSQLCheckpointStorage`, with a typed `CHECKPOINT_TARGET_NOT_EMPTY` conflict and metadata `thread_id` rewrite.
+  - Add owner-scoped, busy-guarded copy and prune-delete routes; prune clears the thread's chunk stream best-effort and retains run-ledger rows.
+  - Share the thread ownership resolver between read and maintenance controllers.
+
+- 12ce057: Thread read model (M2-A)
+
+  - Add `Agent.getStateHistory` / `Agent.getStateAtCheckpoint` (bounded to 50 checkpoints).
+  - Add `CheckpointStorage` / `ThreadCheckpointStorage` protocols and a read-only `PostgreSQLCheckpointStorage` over the built-in `PostgresSaver` tables.
+  - Add tenant/assistant-scoped read-only routes for thread status, checkpoint history, checkpoint state, and checkpoint storage usage; missing checkpoints and foreign threads return 404 and unexpected errors are sanitized.
+
+- 23902af: Per-thread TTL and expiry sweeper (M2-B2a)
+
+  - Add optional `expiresAt` to `Thread`/`CreateThreadRequest` and `listExpiredThreads` to `ThreadStore`.
+  - Persist `expires_at` in InMemory, Local SQLite, and PostgreSQL stores (PG migration v194 + partial index).
+  - Accept `ttlSeconds`/`expiresAt` when creating a thread (invalid values return 400).
+  - Add a lease/inbox/maintenance-guarded TTL sweeper with `THREAD_TTL_SWEEP_*` configuration.
+  - Copy does not inherit TTL; run-ledger rows are retained on sweep.
+
+## 4.11.0
+
+### Minor Changes
+
+- b0d43d4: Tenant model providers can now offer chat, embedding and speech-to-text models from one catalog, and every model-provider or collection change takes effect without a gateway restart.
+
+  `ModelKind` (`chat` / `embedding` / `asr`) is stored on the model row rather than the provider, because one provider is one credential and one base URL and a single OpenAI-compatible endpoint commonly serves all three. Each kind keeps its own runtime lattice and lazily materialises its client from the shared store, so a tenant model resolves on every replica without replica-to-replica synchronisation. A loader guard keeps a chat consumer from ever resolving an embedding or ASR row, and the tenant default model remains chat-only.
+
+  The provider panel exposes the kind per model, offers a per-model thinking toggle (stored as `extra.disableThinking`, translated to the upstream kwargs by core), and can run one embedding inference to prove a model works and report its vector dimension. Collection and voice pickers now read the tenant catalog and fall back to the deployment's boot-registered models.
+
+  Invalidation is now uniform across kinds: a provider change evicts chat, embedding and ASR clients on every replica, and a collection change — including a new `embeddingKey` — drops the cached vector stores it invalidated, locally and across replicas through the outbox.
+
+- c57f417: WeCom: tool-driven template cards and interactive card events.
+
+  - `wecom_card` is the only card producer: it validates/formatting the card and
+    registers it for the current reply instead of returning JSON that the model
+    has to copy. The channel takes the registered card on completion and sends it
+    as a separate `template_card` reply; it no longer parses card JSON from model
+    text, so no tool call means no card.
+  - `template_card_event` is forwarded to the agent. Clicks update the original
+    card in place and the agent's follow-up goes through `sendMessage` (the event
+    `req_id` is reserved for the card update). Same-button clicks are deduped;
+    different buttons on the same card still work.
+  - Event-triggered runs that produce a card send **only** the card (no extra
+    text). `OutboundStreamChunk` gains an optional `reason`; failed/interrupted/
+    timeout runs discard a pending card.
+  - Streaming now respects WeCom's 30 messages/minute limit (2500ms interval),
+    uses an 8-minute hard timeout, and closes a completed run after a short grace
+    window even when the final AI message is empty.
+
 ## 4.10.0
 
 ### Minor Changes