Skip to content

plat 538

github-actions[bot] edited this page Oct 5, 2026 · 2 revisions

PLAT-538: move RTS knowledge to Brain, then retire per-workflow KB sharing

State: open. P2. PR 268 (Brain, product id knowledgebase) is on main (8221d81e4), not deployed. The old sharing path (knowledgebase_sources on a workflow manifest) still exists.

Why a ticket: deploying Brain migrates nothing. Retiring the old path before RTS is migrated would silently remove knowledge from running workflows.

RTS today (2026-10-05, read through the MCP): each workflow has its own knowledgebase/ folder; attachments, all read-only: rtsaws reads rtslatency; automationtesting reads rtslatency and rtsaws; rtsprreviweer reads rtslatency, automationtesting and rtssprinttracking. rtslatency and rtssprinttracking read none. Backups already cover knowledgebase (rtslatency to GitHub, automationtesting to S3).

Order: deploy with the product on for the owner only; migrate (preview, import, rebind consumers, cutover; pause writers and schedules first) rtslatency and rtssprinttracking, then rtsaws, then automationtesting, then rtsprreviweer; check that no manifest lists knowledgebase_sources; only then remove the old path.

Removal change must: log or refuse loudly when a workflow still has knowledgebase_sources; never drop it silently. Scan Excellence, Confida and Dominion for knowledgebase_sources before the removal.

Known follow-ups from the PR 268 review (not blockers): default MCP consent includes Brain write (role-bounded); the backup token is stored with the host key, not in Vault; a Builder chat getting the Brain tools has no test.

Left: everything above.

Pilot, 2026-10-05 (rtslatency): Brain folders RTS and RTS/Latency created; grants set (admin Owner, yoav and laxmi Reader, the workflow's owner and readers). migration_preview then migration_import ran through the MCP: 6 files imported (context/context.md and five notes, all read back through Brain), notes/_index.json skipped (unsupported), source files untouched. Not done: cutover. Three workflows still read this one through the old path and must be rebound first: automationtesting (alias rtslatency), rtsaws (rtslatency), rtsprreviweer (latency); the schedules must be paused around cutover. Next: do the same for rtssprinttracking, then rtsaws, automationtesting, rtsprreviweer.

Clone this wiki locally