Skip to content

Agent Taskbay 1.0 — out of beta

Latest

Choose a tag to compare

@github-actions github-actions released this 03 Nov 10:05
· 11 commits to main since this release
1666e7a

Agent Taskbay 1.0 is the first stable release, bringing the console out of beta. It gives people one place to start and follow durable A2A tasks, answer agents, review approvals and keep an audit trail, even after closing the browser.

This release includes all changes since 0.2.0: task discovery and tracking, steering and restart, related tasks, richer chat content, agent nicknames, an Administration area and stream-driven live updates.

Upgrading from 0.2.0

  • Back up the database, artifact directory and vault keys, then stop web and workers before upgrading. This release requires schema version 9. The local launcher migrates automatically and backs up an existing PGlite directory; PostgreSQL deployments must apply migrations explicitly.
  • Databases with all sixteen legacy migrations are adopted into the numbered baseline without replacing their tables. If a database has only some legacy migrations, start it once with agent-taskbay@0.2.0 before upgrading.
  • Upgrade all web and worker instances before fetching an extended Agent Card. Mixed versions can interpret card snapshots differently.
  • Rebuild projections with npm run db:projections:rebuild to refresh existing status-message metadata and populate related-task links immediately. Stop the web process and add -- --offline-pglite when using PGlite.
  • See Upgrading and backups for the complete procedure.

Release scope

The console leaves beta with 1.0.0. Structured-form, approval-request and skill-routing extensions, A2UI and the optional AG-UI adapter remain experimental. The console's HTTP routes are internal and are not a stable public API. A stable client SDK and /api/v1, user-delegated OAuth, gRPC and packaged container deployment remain future work. There has been no independent security audit; platform and browser coverage remains limited as documented in Compatibility.

Added

  • Agent nicknames. Give an agent your own name to tell apart registrations whose cards advertise the same name, for example the same agent built with LangGraph and with ADK. Set it in the connect dialog or later on the agent page (administrators). The nickname replaces the card name in the console, inbox, approvals, notifications and audit trail, and the agents list now shows each agent's host and port. Nicknames are unique among enabled agents, ignoring case; clearing one restores the card name. Schema version 9.

  • IDs are labeled and copyable. Every ID in chat, task pages, approvals and the task browser now says what it is (Task ID, Context ID, Message ID, Artifact ID, Call ID, Taskbay ID, Approval ID, Command ID, Digest), with a copy icon that copies the whole value. Selecting a label explains the ID: its protocol field, whether A2A defines it or only Taskbay keeps it, its full value and who creates it. The task page's Identifiers panel is rebuilt around this, shows each message's ID in History, lists artifact and referenced-task IDs, wraps very long values and folds those over 100 characters behind Show all. Copying works over plain HTTP, and the chat header's context ID moves to its own row on a phone.

  • Browse and track tasks started elsewhere. Browse tasks on an agent page reads live ListTasks pages without storing them; Track and Track by task ID re-read and store one task to answer, steer, cancel or claim. Requires an administrator or whole-agent operate grant. Stored tasks carry Discovered labels, an origin filter and tracking details; tracking never claims a task or notifies people about its first stored state. The Task discovery card defaults to Browse only, with Off and a deployment-wide A2A_TASK_DISCOVERY cap. Schema version 2.

  • Automatic discovery is opt-in. With A2A_TASK_DISCOVERY=auto, an administrator can set an agent to Automatic. Workers then store eligible unknown tasks: finished history defaults to 7 days (0–90), with at most 500 new tasks on the first sweep, 200 on later sweeps and 5,000 discovered tasks per agent before pausing. First sight notifies only for input or authorization requests updated within 24 hours; later changes notify normally. Schema version 5.

  • Demo work started elsewhere. npx agent-taskbay demo-agent includes an automation card with existing tasks, and the quickstart walks through browsing, tracking and answering one.

  • Extended Agent Cards. Administrators can Fetch and Refresh a validated extended card with the stored agent credential; whole-agent readers see extra skills marked Authenticated. Extended cards are stored separately and do not change routing, grants or extension negotiation. Schema version 7; upgrade all web and worker instances before the first extended-card fetch.

  • Connection details and tenant selection. The agent page shows the binding, interface, protocol version, tenant and extension negotiation. New conversations can select a tenant from the card; existing conversations retain their stored tenant.

  • Task filters and context views. GET /api/tasks filters by context, agent, update time and origin, and pages with an opaque after cursor. Task pages link to Other tasks in this context, with a removable context filter in the list. Schema version 6.

  • Related tasks. User and agent messages’ referenceTaskIds become durable References and Referenced by links within one organization, agent and tenant, limited to tasks the reader may access. Links survive restarts and projection rebuilds; unknown references are not fetched. Schema version 8.

  • Assisted cancellation. Cancel this task and related tasks in task pages and chat lets a person choose eligible tasks and request cancels sequentially. Cancel this and related preselects all eligible tasks; other entry points select only the current task. Each result appears separately, failures do not stop other selections, and completed runs record task.cancel.related with browser-reported counts only.

  • Steering outcomes. Extra input sent to a running task shows Delivered, Agent responded, Agent refused: …, Waiting for a response (then No response yet after 60 seconds), or Not sure it arrived. Detail responses derive steering from commands and the ledger; no outcome proves that the agent used the input.

  • Stop and restart with this. Chat requests cancellation, confirms the old task is finished, then starts a new task in the same context referencing it. Progress shows both results; retries use ordinary idempotent commands and never automatically resend an uncertain outcome.

  • Load agent history. Chat requests 1–200 retained messages (default 50; 6 requests per minute per task). Newly found messages appear under History the agent returned with origin: "agent-history", separate from What Taskbay received; repeated loads do not duplicate messages or replace the ledger.

  • Push cleanup. The worker removes Taskbay’s leftover push configurations from finished tasks when their IDs or callback paths identify them as its own. Other callbacks are left alone, each task is checked at most daily, and removals record push.cleanup.<count>.

  • Development fixtures. scripts/fixture-scenario-agent.mjs provides controllable protocol scenarios. scripts/fixture-content-agent.mjs streams mislabeled text, common tool-call and result shapes, and long content. These development fixtures are not shipped in the npm package.

Changed

  • Live updates no longer poll every 5 seconds. The console refreshes from the live-update stream and on focus, and polls (with jitter) only while the stream is down. Web processes share one freshness read per organization (at most about one per 400 ms), and stream lifetimes are randomized (40 to 55 seconds) so browsers do not reconnect together. task-views, decisions and notifications reads send ETag and answer 304 when unchanged; responses stay no-store.

  • Console navigation is reorganised by job. The sidebar groups daily work (Home, Inbox, Tasks, Chat), Agents (Agents, Find tasks, Orchestration) and You (Notifications, Settings, Administration for administrators).

  • Administration is its own area (/admin) with Overview, People, Teams, Agent access and Policies & alerts, plus a link to the audit trail. It replaces the organization cards that sat under Settings, which now holds only your browser-stored request defaults and extensions. Teams gains Disable team. No API changed.

  • People management in the console. With OIDC sign-in, Administration → People can add a person, change a role, deactivate and reactivate, all audited (member.*) and with the same guards as the API: you cannot change or deactivate yourself. In local development mode the page shows the built-in administrator and explains that people need OIDC. POST /api/admin/security gains provisionMember, changeRole, deactivateMember and reactivateMember; GET adds identity and inactiveMemberships.

  • Dev server serves 127.0.0.1. next dev now allows the 127.0.0.1 origin, which the launcher and docs print; before, the page loaded there but never became interactive.

  • Browsing an agent's tasks is a page, not a dialog. Find tasks (/discover) works across agents and is linked from the sidebar, the Tasks list and each agent's Task discovery card.

  • Clear errors when the embedded database cannot be opened. PGlite has no lock of its own, so a second process (for example npm run db:migrate while npm run dev runs) could open the same folder silently and damage it. Taskbay now takes a <folder>.lock file and refuses with "already open in another process"; a stale lock from a dead process is taken over. A damaged folder, which made PGlite end the whole process with RuntimeError: Aborted(), now fails the open with an error that says what to do. Background workers print the reason once instead of a bare "will retry" in a loop. Nothing is deleted or moved automatically.

  • Definite A2A refusals settle commands as rejected with an errorKind, while unknown failures after dispatch stay uncertain. Chat and task pages explain failures by kind, including whether to check the task or start a new one; remote refusal text is not stored or returned. Schema version 3.

  • Sending to or canceling a finished task is refused locally with 409 and task_terminal. Continue as a new task in the same context.

  • Unsupported extensions marked required: true block sends until an administrator explicitly overrides the block on the agent page or through its settings route. Cancels remain available and overrides are audited. Schema version 4.

  • Migrations use numbered names and the release declares its required schema version (now 9). The sixteen development migrations are consolidated into a baseline; older databases that applied all sixteen are adopted unchanged. Incomplete legacy databases must first start agent-taskbay@0.2.0 to finish those migrations.

  • Startup reports database creation and upgrades. Before upgrading an existing PGlite directory, it keeps a <dir>.backup-v<schema version>-<timestamp> copy (newest three retained); a database newer than the release is refused.

  • Sent chat messages use a soft brand surface, and sender labels distinguish You from the agent’s name and avatar.

  • Long content folds behind Show all N lines: messages and results after about 20 lines, compact steps and JSON values after 6. Multi-line JSON strings retain line breaks; trees and tables page at 100 entries, and expanded cards have bounded scrolling.

  • Text viewers offer useful alternate formats without changing stored or outgoing content. Mislabeled agent JSON and Markdown use a fitting view; prose is not offered as CSV. HTML stays source until Preview HTML opens a sanitized, isolated document, with scripts and remote assets blocked.

  • Intermediate messages and structured activity fold into Progress and Activity. Tool calls, results, code, reasoning and other data show summaries with names, IDs and previews, including common framework shapes and adk_author; final replies and results stay visible without duplication.

  • Streaming-capable agents default chat’s returnImmediately option on; the per-chat Options switch can override it. New chats select their conversation on the first committed task snapshot.

  • Documentation consolidates the implemented capabilities, discovery limits, failure handling and verification, and records guided authorization and signed incoming push as future work.

Fixed

  • Release dependency hardening. Next.js and its ESLint configuration move to 16.4.4, Undici to 8.12.3, and the MikroORM packages together to 7.2.5. Compatible dependency patches include Zod 4.6.7 and gRPC 1.14.7 advisory fixes. The shadcn generator is now a development dependency; its stylesheet is compiled into the shipped console, so the generator and its tooling dependencies are no longer installed by the published package. Motion stays on its existing major version.

  • Windows verification. Launcher data-directory tests use native absolute paths, and authentication-coverage tests normalize directory separators before checking route exceptions.

  • Link-styled outline buttons had no border. buttonVariants() returned a raw class string, so an outline button built from a link kept both border-transparent and the variant's border colour, and the transparent one won. "Back to tasks", Export CSV, Add agent and similar looked like plain text. They now render as buttons.

  • Audit trail no longer drops you out of Administration. It opens inside Administration at /admin/audit with the tabs still visible; /audit sends administrators there. Find tasks no longer shows a redundant "Back to tasks" button, and Disable on a team looks like a button.

  • Tenant selection uses the matching interface URL within connection pins. New tasks store the effective default tenant, and automatic discovery avoids empty-tenant alias copies while preserving legacy records.

  • Supported required extensions are declared in A2A request headers, and empty SDK task lists show a normal empty state.

  • Restart retries create a fresh command after a definite refusal and recheck the same command for uncertain outcomes. Stored inline attachments are restored through current artifact permissions before cancellation, preserving the original file bytes.

  • Markdown in chat links only explicitly authored links and autolinks; bare addresses, emails and paths stay plain text. Links open in a new tab with noopener noreferrer nofollow.

  • Chat follows new content after layout, preserving bottom padding and keeping the latest reply, quick replies and forms above the composer on desktop and phone. Scrolling up pauses following; sending resumes it.

  • Questions that require input or authorization and the agent’s final reply stay visible instead of folding into Progress.

  • Status-message metadata and the state that carried it survive stored task projections. Run npm run db:projections:rebuild to refresh earlier records; they also refresh on their next event.