Agent-authored Slack Block Kit components in respond_to_user

[josephschorr]

What

Lets an agent compose presentational Slack Block Kit layout in its replies — not just a markdown string. The agent passes a blocks array on respond_to_user; the Slack kind renders it as the message body, with the existing text field as the notification preview and degrade fallback.

How

Channel-neutral seam, Slack-native content — no kind == "slack" branch leaks into the agent tool:

• "components" capability on channelkinds.Kind (Slack opts in). Flows to ChannelBinding.Capabilities → respond_to_user, exactly like "markdown".
• Two optional interfaces on the kind (mirroring TextFormatter): ComponentFormatter (model-facing authoring instructions, injected into the blocks schema description) and ComponentValidator (ValidateComponents + ComponentsPlainText). The runner already blank-imports the Slack kind, so both are reachable in-process.
• Validate before publish. respond_to_user calls the kind's validator synchronously; a bad payload returns a precise, model-facing IsError (no envelope published) instead of a silent post-time degrade to plain text.
• Transport. New OutboundUserMessagePayload.Components json.RawMessage; the Slack sender posts it via MsgOptionBlocks with text as fallback, keeping the existing blocks→plaintext degrade net.
• Egress safety. The info-leakage gate measures block text too (via ComponentsPlainText), so content can't bypass the gate by riding in blocks.

Supported blocks (read-only / presentation-only)

header, section (no accessory), divider, context, rich_text, markdown, table, data_visualization (line/bar/area/pie charts), card (no actions), container, plan, task_card.

Rejected with precise, correctable errors: interactive blocks/elements (actions, input, context_actions, data_table, carousel, section accessories, card actions), alert (modals-only), images/icons (deferred — see Scope), and @channel/@here/@everyone broadcasts. Chart and container payload shapes are validated to Slack's actual requirements — verified against the live API — so the agent gets a clean error rather than a degrade (container needs a title/rich_text_title; charts need a title plus per-type segments or series + axis_config.categories).

slack-go upgrade

Bumps github.com/slack-go/slack v0.23.1 → v0.30.1 for typed support of the newer blocks (data_visualization, table, card, container, markdown, plan, task_card, …). The only breaking change was GetUserInfoContext becoming variadic; the Slack wrapper interfaces and test fakes are adapted to match.

Scope

Slice 1 is presentational only. Interactivity (buttons/menus routing clicks back into the session) is out of scope. Images are deferred to Slice 2 (artifact-backed — resolve a session-owned artifact handle rather than an arbitrary URL).

Testing

• Table-driven validator tests: one row per accept/reject rule (each block type, each interactive/image/broadcast rejection, structural limits, chart/container shape requirements).
• respond_to_user tests against the real Slack validator: schema exposes blocks only under the capability; validation failure → IsError, no publish; success carries Components; a leakage-gate test proves block-only content is still measured.
• Slack sender render + degrade tests.
• Exercised end-to-end on a live Slack workspace (charts, tables, cards, containers, rich_text, markdown all render; rejections surface clean errors).

Follow-ups (tracked, not in this PR)

• Surface an upstream Slack invalid_blocks back to the agent as a system note for any future/unknown block that passes our validator but Slack still rejects (channelsd → session feedback), instead of only degrading.
• The channelsd silence-watchdog counts outbound-message silence, not internal tool-call progress, so a long internal run can read as a false "stall."

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
openagentprimitives	Ready	Preview	Oct 9, 2026 11:15pm UTC