Summary

Operator-facing view of recent Kora ↔ Joshua DM exchanges in the cockpit. Pairs with CC#3's KR-FEAT-SLACK-DM bucket (Feature 5 backend) — ST2 will swap the stub /api/slack-dm/recent body for a real read of ${HERMES_HOME}/slack_dm_log.jsonl. Joshua picked Slack first as his preferred ops channel.

Shape is pinned by the new test suite so the FE shipping in this PR keeps rendering correctly during the cut-over.

What's in here

• Backend stub: GET /api/slack-dm/recent returning 4 representative messages deliberately spanning inbound (received) + outbound (sent_ok) + filtered_non_joshua so the operator's first look surfaces the filtering posture as well as the happy path. stub: true keeps the FE banner visible.
• SlackDMPanel.tsx (new): title + stub banner + stats strip (total + by-direction + filtered/error counts) + filter pills (all / inbound / outbound / filtered / errors) + newest-first timeline with direction arrows (←/→) + user-label badge + relative time + status badge (only for non-default statuses — happy-path rows stay clean) + expandable detail view (channel_id, thread_ts, full timestamp, full text, id).
• Dashboard card feat(KR-3 ST1): iso_node_* typed-graph tool family #9 ("Slack DM ↔ Joshua") with MessageCircle icon. Layout keeps lg:grid-cols-3; row 2 becomes 3+3+3 — the cleanest possible balance, no orphan card. Headline goes destructive when filtered_non_joshua > 0 OR sent_failed > 0 OR handler_error > 0 per spec §2(c).
• Route /slack-dm + nav entry between /agent-activity and /mcp-clients.

4-layer security contract

This iteration's specific risks: real DM text is arbitrary user-typed content; raw Slack IDs / tokens / signing secrets must never reach the operator's browser.

1. user_id_label is a LABEL — joshua / kora_bot / unknown_user, never a raw Slack user ID (U[A-Z0-9]{8,} shape). Backend test pins per-field AND walk-whole-payload — a future leak via a raw_user_id diagnostic or in the message text gets caught.
2. channel_id is a STUB label (D_<label>) in v1. Backend test pins the stub format AND explicitly rejects real Slack channel shapes (C/G/D[A-Z0-9]{8,}) so a future flip without hashing/truncation gets caught at the API edge.
3. text rendered as PLAIN TEXT via React's default child escaping. Frontend pin: SlackDMPanel.tsx never uses dangerouslySetInnerHTML for message bodies (comment-stripped grep so the warning comment doesn't self-trigger).
4. Walk-the-whole-payload guards: xoxb-/xoxp-/xapp-/xoxa-/xoxr- Slack token shapes, AND 32-char hex runs (Slack signing secret shape) — backend bug or future log-entry edit that leaks creds gets caught before reaching the operator's browser (where it could end up in diag bundles or screenshots).

Test plan

• tests/kora_cli/test_web_server_slack_dm.py — 15 tests: shape, 4-message stub pin, direction+status diversity, per-entry schema, all 4 security guards, by_direction_24h reconciliation, FE source-pin against dangerouslySetInnerHTML, cron-regression sanity.
• Full admin-panel regression: 219/219 across 20 suites (skipping test_web_server_mcp_clients.py — PM tracking as task Fix nous refresh token rotation failure in case where api key mint/retrieval fails NousResearch/hermes-agent#269).
• pnpm tsc --noEmit clean.
• pnpm build clean.
• Manual smoke: load /slack-dm, exercise filters, expand rows, verify destructive tone on the dashboard card.

Notes

• Layout choice: row 2 now has 9 cards → clean 3+3+3 with lg:grid-cols-3. No need to bump columns or push to row 3 — landed the symmetry sweet spot.
• Convention divergence from spec: spec hints /admin/slack-dm; using flat /slack-dm to match every prior panel in this branch (established by KR-HB-PANEL onward).
• Newest-first ordering: spec offered the choice; defaulted to newest-first to match every other panel in the branch (recent-activity log semantic). Stable secondary sort on id breaks ties.

🤖 Generated with Claude Code