langstage

Fixed

• Custom non-streaming agents replied with nothing in chat (gh #-dogfood). A CompiledGraph whose node returns a finished AIMessage (rule-based / router / retrieval agents, or any LLM call outside a token-streaming node) rendered an empty assistant turn. Root cause was in the shared core; bumped langgraph-stream-parser to >=0.6.4, which emits such content as a fallback.
• Canonical LANGSTAGE_WORKSPACE_ROOT was ignored when computing the default agent's workspace — default_agent.py read only the deprecated DEEPAGENT_WORKSPACE_ROOT. It now reads the canonical name first and warns on the legacy one.

Fixed

• Declare langchain as a dependency. langstage.middleware (canvas) imports langchain.agents.middleware and is loaded on hot paths (app, the check CLI command, the default agent), but langchain wasn't declared — it only arrived transitively via the [deepagents] extra. A clean pip install langstage therefore failed with ModuleNotFoundError: No module named 'langchain' on langstage check / langstage run. Now a hard dependency. (Found in production by the daily QA routine.)

CI

• Added a minimal-install job: installs with no extras (pip install .) and runs an import + CLI smoke — the lane that matches a real pip install langstage. The other jobs install [deepagents], which masked the missing langchain by pulling it transitively. This guards against undeclared dependencies going forward.

Changed

• Durable agent checkpointer. The auto-attached checkpointer is now upgraded to a SQLite-backed AsyncSqliteSaver (<workspace>/.langstage/checkpoints.db) at server startup, so conversation + interrupt state and the task review gate survive restarts, and orphaned tasks resume from their last checkpoint instead of from scratch. Only checkpointers LangStage auto-attached are upgraded — a user-supplied checkpointer is never replaced. Falls back to in-memory if the SQLite saver can't initialize.
• Picks up langgraph-stream-parser 0.6.1 (sharper async-delegation tool descriptions, so the agent reaches for the task board for long/background work).

Added

• Dependency: langgraph-checkpoint-sqlite.

Bring-your-own-agent integration — make pointing --agent at any LangGraph graph "just work," and tell users exactly what lights up.

Added

• langstage check --agent <spec> — a preflight doctor that loads your agent and reports which features will light up (checkpointer, Canvas, Plan/write_todos, async-delegation tools) and which need a convention or tool to unlock.
• LANGSTAGE_TOOLS — a one-import bundle (from langstage import LANGSTAGE_TOOLS) of the host's scheduling + async task-delegation tools, so a BYO agent unlocks agent self-delegation and agent-created schedules in one line.
• README "Bring your own agent" section documenting the integration contract.

Changed

• Checkpointer is now auto-attached when a loaded agent has none (was: a console warning, then silently degraded). Conversation memory, human-in-the-loop interrupts, and the task review gate now work for any BYO graph out of the box; supply your own checkpointer for durability. (Matches the AG-UI bridge's behavior.)
• __version__ now reads from package metadata (was a stale hardcoded constant).

Fixed

• Task board columns now expand to fill the panel width (they were fixed-width, leaving dead space when the right panel was widened).

Docs

• README: documented the task board (delegate, live-tail detail pop-up, review gate, agent self-delegation, REST API).

Async task board — delegate tasks to background copies of the agent, track them on a Kanban board, and interact with each run.

Added

• Task board (new Board tab): delegate a task from the UI and watch it move queued → ongoing → review_needed → done (cancel/retry per card). Backed by the core 0.6 task engine (TaskRunner + a durable SQLite store) — non-blocking, single-process, no extra infra.
• /api/tasks REST: list/get/create/cancel/retry, plus GET /{id}/events, POST /{id}/resume (HITL approve/reject), POST /{id}/message (talk-back).
• Task detail pop-up: click a card to replay/live-tail its agent's full event stream, approve/reject a paused task, and send follow-ups.
• Agent self-delegation: the default agent now carries start/check/list/update/cancel_async_task tools, so it can spawn async sub-tasks (tracked on the board with parent links).
• The cron scheduler now enqueues onto the task board (durable) and its timezone handling is fixed (UTC throughout).

Changed

• Parser pinned >=0.6,<0.7 (+ [agui]); added aiosqlite.
• The "Tasks" tab (the todo/plan list) is renamed "Plan" (the async tasks live on the new Board tab).

Notes

• Runs on the existing in-memory checkpointer; task rows + transcripts persist in SQLite (the board survives a restart). Durable mid-run checkpoint resume is a later pass.
• Single-process: run one server worker.

Adopt AG-UI: widen the langgraph-stream-parser ceiling to <0.5 and add an [agui] extra so this surface's agent can be served over AG-UI via langstage-agui. Additive; no runtime changes.

cowork-dash is now langstage — the web stage (and namesake) of the LangStage family ("every stage for your LangGraph agent"). The rename also clears the collision with Anthropic's Claude Cowork.

Changed

• Distribution cowork-dash → langstage; module cowork_dash → langstage; command cowork-dash → langstage (the old command remains as a deprecated alias, and a deprecated alias package keeps import cowork_dash working with a warning).
• Canonical config vocabulary via langgraph-stream-parser 0.3: LANGSTAGE_* env vars, project langstage.toml, global ~/.langstage/config.toml. The full legacy DEEPAGENT_* / deepagents.toml vocabulary still resolves as a deprecated fallback.
• Parser pinned >=0.3,<0.4.