Skip to content

feat(dashboard): move MCP Management to Customize → Connections - #24933

Merged
iamwhatever merged 1 commit into
kirodotdev:mainfrom
IngridMorstrad:fix/mcp-management-to-customize-24920
Nov 3, 2026
Merged

iamwhatever merged 1 commit into
kirodotdev:mainfrom
IngridMorstrad:fix/mcp-management-to-customize-24920

Conversation

@IngridMorstrad

@IngridMorstrad IngridMorstrad commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

Problem / Motivation

Goal: MCP Management sits next to the MCP servers it governs, in Customize → Connections → MCP Servers, instead of among the Settings → Developer internals pages.

MCP Management — the surface that stubs MCP servers, routes their backends to one shared process, and shows the Sharing assessment — is one of the Settings → Developer internals pages (/settings/developer/mcp-pool), filed between Storage and Memory, away from the MCP servers list it acts on. The MCP Servers list's "share refused" notice sends people to that separate page to approve a launch.

Closes #24920

Why it matters

Approving a refused MCP launch, stubbing a server so it can render its own UI (MCP Apps), and sharing a backend across sessions are mainstream actions on the servers in that list, not developer internals like logs and telemetry.

Not a goal

  • Not rewriting MCP Management. The one edit inside McpManagement.tsx, McpManagementHero.tsx, McpManagementAssessment.tsx and McpManagementLaunchReview.tsx is gating each askAgent hand-off on mcpStagedEditsContext; their layout and behaviour are otherwise the ones feat(mcp-gateway): share an MCP server with one labelled switch #19661 shipped.
  • Not changing the backend, the gateway, or any /api/mcp* endpoint.
  • Not reworking the Connections page's own Services / MCP Servers sub-tabs.

What changed (motivation → approach → change)

Goal: put MCP Management where users already manage MCP servers.

Approach: the Customize → Connections tab's MCP Servers sub-tab (/capabilities?tab=mcp&view=mcp-servers) hosts the MCP servers list. MCP Management acts on exactly those servers, so it renders as a section directly below that list — not on the Services sub-tab (connector cards, a different thing), The heavy McpManagement.tsx is reparented, not rewritten.

Change:

  • ConnectionsPage.tsx — the MCP Servers sub-tab now renders <McpManagement /> as a section (#mcp-management anchor) below <McpTab />; a scroll-on-arrival effect brings that section into view when the URL carries the #mcp-management fragment.
  • Settings → Developer — DeveloperPanel.tsx drops the mcp-pool page (rail entry, render branch, lazy import) and developerSubs.ts drops its key; a REPLACE redirect sends /settings/developer/mcp-pool to the new section (the SubNav would otherwise silently show Feature Previews). DeveloperPage.tsx's /developer?tab= redirect sends mcp-pool there too. While the servers list holds unapplied edits, every askAgent hand-off in the McpManagement family (McpManagement, -Assessment, -Hero, -LaunchReview) is withheld via mcpStagedEditsContext, so it cannot navigate away and drop them.
  • McpShareRefusedNotice.tsx — MCP_MANAGEMENT_PATH → /capabilities?tab=mcp&view=mcp-servers#mcp-management (the one source of truth the redirect imports), so the link keeps the MCP Servers sub-tab and lands on the section instead of the Services gallery.
  • McpTab.tsx + mcpStagedEditsContext.tsx — while scope/tool edits are staged, MCP Management withholds its "Ask the agent" hand-off (which would navigate to /chat and drop those edits). McpTab publishes pendingCount > 0 through a small panel-scoped context; the askAgent notices in MCP Management read it. Other navigation keeps its existing behaviour (no blanket leave guard).
  • Non-owner sessions — the section now also renders for non-owner sessions on the MCP Servers sub-tab. The owner-only footprint reading answers them 403. McpManagement.tsx treats that as a permanent refusal: no retry and no 30s polling (each refusal writes an audit row). A refused read is still a rejected request, so it goes through the shared inline ErrorNotice (with the staged-edit hand-off decision, per errors-use-error-notice), but its message names the reason, "Memory figures are visible to the crew owner only.", instead of the generic "Could not read how much memory MCP servers use." that other failed reads show.
  • McpCustomServerModal.tsx + McpTab.tsx apply — invalidate ['mcpGatewayServers'] after add/edit/apply so a newly installed or changed server appears in MCP Management without a reload.
  • i18n — removed developerPage.tabs.mcpPool.*, added connectionsPage.mcp_management_heading across all 14 catalogs (pseudolocale regenerated).
  • Docs / data — docs/architecture/mcp.md, docs/feature-map/README.md, docs/guides/secrets-env.md, src/kiro_crew/docs/mcp-apps.md, src/kiro_crew/doctor_checks/mcp.py, and the "Open MCP Gateway" tip now point at Customize → Connections' MCP Servers sub-tab. ui-index.generated.json regenerated; the gone developer-mcp-pool render surface is dropped from scripts/lib/i18n-surfaces.mjs + render-baseline.json.

Review rounds (addressed in this single commit)

The fork AI-review lanes raised real UX/data-loss defects introduced by co-mounting; all are fixed above:

  • Link/redirect/tip landed on the Services gallery, MCP Management below the fold (UX + Opus) → the path now keeps the MCP Servers sub-tab and scrolls to #mcp-management; MCP Management renders only on that sub-tab, not Services.
  • Management hand-offs discarded staged MCP edits (GPT F1) → MCP Management withholds its "Ask the agent" hand-off while the sibling list holds unapplied edits (mcpStagedEditsContext), rather than a blanket leave guard that would silently block every exit.
  • Newly installed servers absent from Management until reload (GPT F2) → ['mcpGatewayServers'] invalidation on add/edit/apply.
  • doctor_checks/mcp.py named a tab that no longer exists, and a backend test asserted the old path (Opus) → message + test/test_doctor_mcp_share_refused.py updated to the new location.
  • Registry installs absent from Management until reload (GPT) → McpBrowserModal's install callback also invalidates ['mcpGatewayServers'].
  • Clicking the MCP Management link a second time did not scroll (GPT) → the scroll effect also keys on the router's navigation key, so a repeat link to the same #mcp-management URL re-scrolls.
  • A cold deep link scrolled while the servers list was still a skeleton (GPT) → the scroll waits for the shared ['mcp-servers'] query to finish loading, so the taller table can no longer push the section back below the fold.
  • Directions said "Customize → MCP", but the rail entry is labelled Connections (UX) → the doctor message, the "Open MCP Gateway" tip and the docs now say Customize → Connections → MCP Servers → MCP Management.

Ordering note re #19661

#19661 (the MCP Management rewrite) has merged and this branch is rebased on it. The two overlap only in the askAgent props this PR gates, plus the settingsCoverage.test.ts waivers for the McpManagement files.

Follow-ups

Backwards compatibility

Compatible: /settings/developer/mcp-pool and /developer?tab=mcp-pool both redirect (REPLACE) to the new location, so bookmarks, docs links and older tips still reach MCP Management. No endpoint, type or config key changes.

Removes nothing but the Settings → Developer mcp-pool page entry, and the redirects preserve its URLs; the component's routes and every /api/mcp* endpoint it calls are unchanged, and the Connections page loads it lazily, as Settings → Developer did. MCP_MANAGEMENT_PATH is the one source of truth, used by McpShareRefusedNotice.tsx, DeveloperPanel.tsx and DeveloperPage.tsx, covered by McpShareRefusedNotice.test.tsx, DeveloperPanel.test.tsx and McpManagementMove.test.tsx.

Tests

website/src/test/McpManagementMove.test.tsx locks in the move, written test-first and shown red on origin/main for the right reason, then green on this change:

Behaviour Red on main (reason) Green here
MCP Servers sub-tab shows MCP Management beside the list, Dev Mode off Unable to find [data-testid="mcp-management"] (the sub-tab rendered only the servers list) section renders beside the list
MCP Management is NOT on the Services sub-tab (guard — holds both ways) not rendered on Services
Settings → Developer has no MCP Management page; /settings/developer/mcp-pool redirects (DeveloperPanel.test.tsx) the rail listed MCP Management gone; lands on /capabilities?tab=mcp&view=mcp-servers#mcp-management
A repeat link to the same #mcp-management URL scrolls again no second scrollIntoView (effect deps unchanged) scrolls on each navigation
A cold deep link waits for the servers list before scrolling scrollIntoView fires while ['mcp-servers'] is still pending (shown red with the loading gate removed) scrolls only after the list loads
/developer?tab=mcp-pool redirects to /capabilities?tab=mcp&view=mcp-servers#mcp-management where never appears (no redirect on main) lands on the new path

Proof: not provable -- frontend test (prove.py runs Python only). The red/green table is the equivalent for a frontend-only change.

Also green locally: tsc, eslint on the changed files, the catalog-parity / dead-keys / key-reference / english-identity suites, settingsCoverage.test.ts, lint:i18n, i18n:check, docs-lint.sh, and the ConnectionsPage / CapabilitiesPage / DeveloperPage / McpManagement / McpCustomServerModal test files. The render-time i18n gate adds no new finding (capabilities-mcp stays 0; the gone developer-mcp-pool surface is dropped from the baseline); the only surfaces above the debt record are pre-existing settings-display-* drift on main, unrelated to this diff.

Manual verification

Rendered the real built SPA under the screenshot harness (capture-24920-mcp-management-move.mjs): MCP Management renders in Customize → Connections' MCP Servers sub-tab in both themes, scrolled into view via the #mcp-management fragment; the Settings → Developer rail has no MCP Management page; and /settings/developer/mcp-pool redirects to the new section.

Screenshots / video

Committed under temp-screenshots/ (a fork PR cannot use gh --attach).

Before — MCP Management as a Developer tab (captured before main merged the Developer page into Settings → Developer, #24810):

before

After — MCP Management as a section in Customize → Connections' MCP Servers sub-tab, scrolled into view below the servers list, no Developer Mode:

After — unscrolled: the Connections page, its MCP Servers sub-tab, the servers list, and the MCP Management section directly below it:

after-capabilities-top

after-capabilities

After — the Settings → Developer rail no longer lists MCP Management:

after-developer

After — a non-owner session: the owner-only footprint read answers 403, and the error notice names the owner-only reason:

after-non-owner

Rebase onto #24810 and #19661

Main merged the Developer page into Settings → Developer (#24810) and rewrote MCP Management (#19661). This revision keeps main's code for both and re-applies only the move: MCP Management leaves the new Settings → Developer SubNav, both old URLs redirect, and the hand-off gating covers #19661's new McpManagementHero.tsx and McpManagementLaunchReview.tsx notices. The GUI feature catalog entry developer-mcp-pool becomes connections-mcp-management.

@github-actions github-actions Bot added fork Pull request from a fork (external contributor) readiness: checking Automated validation is still running labels Oct 27, 2026
@github-actions

github-actions Bot commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Opus 5.5, fork) — ✅ PASS

Design-level review of 4b64a19ebad8a04bde0f862cae6fd3075b6d35d0 via the fork AI-review pipeline — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

Design-Verdict: PASS

Nothing blocks: the panel is moved, not rewritten, the old URLs redirect, the approval endpoints stay owner-gated on the server, and committed screenshots cover the new placement.

Notes

  • McpManagement*.tsx still lives under pages/settings/, though it no longer renders on a Settings tab. The settingsCoverage waivers now have to explain that. Moving the files under pages/connections/ in a follow-up would put them with the page that owns them.

[DESIGN-REVIEWED] 4b64a19

Description read

[DESCRIPTION-READ] d926279491dc2c5364c1ce27e720d0bb259d48f2916dfff733a78c564a7af255

That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

UX Review (Opus 5.5, fork) — ✅ PASS

UX-level review of 4b64a19ebad8a04bde0f862cae6fd3075b6d35d0 via the fork AI-review pipeline — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

UX-Verdict: PASS

Screenshots show every new or changed control: the section on the MCP Servers sub-tab, the Developer menu without the MCP Management page, and the message for non-owners. No first-time reader has looked at them in this fork lane; pushing the branch to this repository adds that read.

Notes

  • The owner-only message "Memory figures are visible to the crew owner only." shows in red error styling with an "Ask the agent" link (attachment-08). A non-owner can't change this and the agent can't either, so show it as plain muted text without the link, in McpManagementHero.tsx.
  • While server edits are waiting to be applied, the "Ask the agent" link disappears and nothing says why. MCP Management: say why "Ask the agent" is hidden while MCP server edits are staged #25690 is already tracking this; no screenshot shows that state.

[UX-REVIEWED] 4b64a19

Description read

[DESCRIPTION-READ] a7d56a3800989c9fe7d91116ebfd6a0cee9fd0ec095bc8caf3d51343982122a3

That sha256 names everything this verdict read: the PR description, plus the 13 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

First Principles Review (Opus 5.5, fork) — ✅ PASS

Premise-level review of 4b64a19ebad8a04bde0f862cae6fd3075b6d35d0 via the fork AI-review pipeline — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

First-Principles-Verdict: PASS

Nothing blocks. The move rests on a linked report (#24920), and the redirects keep every old URL working. The other additions follow from putting MCP Management on the same tab as the servers list.

What this change ships

Inventory (9 items) — 8 justified

Intent: let people approve, stub and share MCP servers from the place where they already manage those servers. This is a MOVE (an ADDITION in kind).

  1. MCP Management leaves Settings → Developer and shows up as a section below Customize → Connections → MCP Servers — justified
  2. /settings/developer/mcp-pool and /developer?tab=mcp-pool now redirect (replace) to the new section — justified
  3. The share-refused notice link, the doctor message, the gateway log line and the tip point at the new location — justified
  4. Following a link to #mcp-management scrolls to the section once the servers list has loaded — justified
  5. "Ask the agent" in MCP Management is hidden while the servers list has unapplied edits (the errors-use-error-notice rule requires this) — justified
  6. Adding, editing, installing or applying a server refreshes MCP Management without a reload — justified
  7. For a non-owner, the footprint 403 shows an owner-only message and stops the retries and 30s polling (Settings → Developer was not owner-gated, so this fixes an existing defect) — rides along
  8. New footprint_owner_only string in 14 catalogs; the description only mentions the heading key — undeclared
  9. Feature-map entry, i18n render surface, a capture script and screenshots updated to match (repo convention) — justified

[FIRST-PRINCIPLES-REVIEWED] 4b64a19

Description read

[DESCRIPTION-READ] 605985ed49b57258743f8a25ce3a60f4efa2621b42d3b877566edbac0c916986

That sha256 names the PR description this verdict read. Recompute it with .github/scripts/pr-description-capture.sh; if it differs, the description changed after this verdict was formed and any finding drawn from the description is unproven.

@github-actions

github-actions Bot commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

Opus 5.5 Review (fork) — ✅ no blocking findings

Reviewed 4b64a19ebad8a04bde0f862cae6fd3075b6d35d0 via the fork AI-review pipeline; updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] 4b64a19

@github-actions

github-actions Bot commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

GPT 6.1 Review (fork) — ✅ no blocking findings

Reviewed 4b64a19ebad8a04bde0f862cae6fd3075b6d35d0 via the fork AI-review pipeline; updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 4b64a19

@github-actions github-actions Bot added the merge conflict Branch has merge conflicts with its base — author must resolve before merge label Oct 27, 2026
@IngridMorstrad
IngridMorstrad force-pushed the fix/mcp-management-to-customize-24920 branch from 34029c2 to 45a55c6 Compare October 27, 2026 19:31
@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

Thanks — all three blocking findings were real (UX/data-loss from co-mounting), and the new revision (45a55c6, rebased onto current main) addresses each:

  • Link/redirect/tip landed on the Services gallery with MCP Management below the fold (UX Review, Opus 5.5): MCP_MANAGEMENT_PATH is now /capabilities?tab=mcp&view=mcp-servers#mcp-management, so it keeps the MCP Servers sub-tab instead of defaulting to Services, and a scroll-on-arrival effect in ConnectionsPage brings the #mcp-management section into view. MCP Management also now renders only inside the MCP Servers sub-tab panel (beside the servers list it acts on), not spanning the Services sub-tab. The Developer redirect reuses the same constant.
  • Management hand-offs discard staged MCP edits (GPT 6.1 F1, fenced): McpTab now registers a navigation leave guard while scope/tool edits are staged (pendingCount > 0). sendErrorToChat consults that same guard, so a Management "Ask the agent" hand-off — or any exit — can no longer silently drop pending/pendingTools.
  • Newly installed servers absent from Management until reload (GPT 6.1 F2): McpCustomServerModal (add + edit) and McpTab's apply handler now invalidate ['mcpGatewayServers'] alongside ['mcp-servers'], so an installed/edited/removed server shows in MCP Management without a reload.
  • doctor_checks/mcp.py named a tab that no longer exists (Opus note): updated to point at Customize → MCP's MCP Servers sub-tab.

The "after" screenshots now show the MCP Management body scrolled into its new place (the earlier ones cut off at the sub-tab strip).

@github-actions github-actions Bot removed the merge conflict Branch has merge conflicts with its base — author must resolve before merge label Oct 27, 2026
@IngridMorstrad
IngridMorstrad force-pushed the fix/mcp-management-to-customize-24920 branch from 45a55c6 to 1556ad7 Compare October 27, 2026 20:41
@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

Round 2 (1556ad7, rebased onto current main) — GPT 6.1 cleared; addressing the remaining two:

  • Staged edits silently trapped the user on the page (UX Review, Opus 5.5): replaced the blanket leave guard. A guard that refused navigation was the wrong call — it changed behaviour McpTab never had (a rail click with staged edits always discarded them on main), and it could only prompt with a native confirm, which this renderer's lint forbids for new call sites. Instead, the narrow fix the reviewers suggested: MCP Management now withholds its "Ask the agent" hand-off while the sibling MCP Servers list holds unapplied edits, via a small McpStagedEditsProvider context (McpTab publishes pendingCount > 0; the nine askAgent notices in McpManagement/McpManagementAssessment read it). So the data-losing hand-off is simply not offered while edits are staged, and every other navigation keeps its existing behaviour. Covered by mcpStagedEditsContext.test.tsx.
  • Backend doctor test asserted the removed /developer?tab=mcp-pool path (Opus): updated test/test_doctor_mcp_share_refused.py to the new /capabilities?tab=mcp&view=mcp-servers path, and dropped the now-irrelevant "Developer Mode" sentence from the doctor message (per the UX note). All 15 doctor tests pass.

UX notes also applied: the "Open MCP Gateway" tip no longer mentions Developer Mode. (MCP Management's internal "Servers" sub-view label vs the panel's "MCP Servers" sub-tab: left as-is for now — renaming it is inside McpManagement.tsx, which this PR deliberately does not rewrite; happy to follow up if preferred.)

@IngridMorstrad
IngridMorstrad force-pushed the fix/mcp-management-to-customize-24920 branch from 1556ad7 to 80e38a6 Compare October 27, 2026 21:13
@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

Round 3 (80e38a6, rebased onto current main): fixed GPT 6.1's remaining finding — McpBrowserModal's registry-install success callback now also invalidates ['mcpGatewayServers'], so a server installed from the capability registry gets its MCP Management row/stub control without a reload, matching the custom-add, edit and apply paths. The earlier "Fork PR Description" and "Internal Content Scan" red were CI infra (GitHub API rate-limit on the check-run publish / missing log), not findings; this push re-runs them.

@IngridMorstrad
IngridMorstrad force-pushed the fix/mcp-management-to-customize-24920 branch from 80e38a6 to a32f782 Compare October 27, 2026 21:45
@IngridMorstrad IngridMorstrad changed the title feat(dashboard): move MCP Management to Customize → MCP feat(dashboard): move MCP Management to Customize → Connections Oct 27, 2026
@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

Round 4 (a32f782, rebased onto current main) — both blocking findings on 80e38a6 addressed:

  • GPT 6.1 (blocking): a repeat "Review in MCP Management" link did not re-scroll. The scroll effect keyed only on activeTab/hash, neither of which changes when the same #mcp-management URL is navigated again. It now also depends on the router's navigation key (useLocation().key), which React Router mints fresh on every push — so a second click to the identical URL re-runs the scroll. Locked by a new test in McpManagementMove.test.tsx that navigates to the same URL twice and asserts scrollIntoView fires each time.
  • UX Review (blocking): the directions named a rail entry that doesn't exist. The Customize rail entry for key:'mcp' is labelled Connections (connections_label), not "MCP". The doctor message, the "Open MCP Gateway" tip, and the docs (mcp.md, secrets-env.md, mcp-apps.md, feature-map) now say Customize → Connections → MCP Servers → MCP Management. test_doctor_mcp_share_refused.py asserts the message names "Connections".

Evidence gaps:

  • Placement. Added an unscrolled capture (after-capabilities-mcp-servers-top-*.png) that lands at the top of the MCP Servers sub-tab — the Connections heading, the Services/MCP Servers sub-tab bar, the servers list, and the MCP Management section's top edge are all in one frame, so the section's place relative to the list it acts on is visible. The scrolled full-section shot is kept too.

Two UX Notes (advisory, not blockers) deferred, happy to follow up:

  • An inline line next to a withheld "Ask the agent". The withhold is gated across nine ErrorNotice sites inside McpManagement.tsx/McpManagementAssessment.tsx; a visible hint means an ErrorNotice API addition (its noHandoffReason is reviewer-only, never rendered) plus threading through all nine. That is more surface than this move PR should carry — the data-loss itself is already prevented.
  • Renaming the section's inner "Servers" sub-view to avoid two near-identical "MCP Servers"/"Servers" labels. That label lives inside McpManagement.tsx, which this PR deliberately reparents rather than rewrites (and feat(mcp-gateway): share an MCP server with one labelled switch #19661 rewrites that file).

If you'd prefer either note folded in now, say so and I'll do it.

@IngridMorstrad
IngridMorstrad force-pushed the fix/mcp-management-to-customize-24920 branch from a32f782 to cce7bdd Compare October 27, 2026 22:01
@github-actions github-actions Bot added the merge conflict Branch has merge conflicts with its base — author must resolve before merge label Oct 27, 2026
@IngridMorstrad
IngridMorstrad force-pushed the fix/mcp-management-to-customize-24920 branch from cce7bdd to ed2b585 Compare October 27, 2026 22:26
@github-actions github-actions Bot added readiness: passed Eligible automated validation passed for the current revision merge conflict Branch has merge conflicts with its base — author must resolve before merge and removed merge conflict Branch has merge conflicts with its base — author must resolve before merge readiness: checking Automated validation is still running labels Oct 27, 2026
@IngridMorstrad
IngridMorstrad force-pushed the fix/mcp-management-to-customize-24920 branch from ed2b585 to f74877d Compare October 28, 2026 17:42
@github-actions github-actions Bot removed the readiness: passed Eligible automated validation passed for the current revision label Oct 28, 2026
@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

Opus 5.5 (blocking, b32269b): fixed in 832d1c6.

Because MCP Management now sits on the MCP Servers sub-tab, non-owner sessions mount it too. The only owner-gated read it makes is GET /api/mcp-gateway/footprint (require_owner_dashboard_request); its status, servers and launch reads aren't owner-gated. For that footprint query:

  • A permanent refusal (a 403 without authRequired, or a 404) is classified with the same isPermanentRefusal rule the share-refused notice already uses, now exported from McpShareRefusedNotice.tsx.
  • On such a refusal there's no retry (otherwise the app's retryPolicy applies, unchanged for owners) and no 30 s polling, so a non-owner tab writes one audit denied row instead of one every 30 s.
  • The hero shows no footprint_failed notice for it. Its data-footprint reads refused rather than staying pending, and the number row stays hidden.

A new case in McpManagement.hero.test.tsx covers it: a non-owner 403 shows no alert and the footprint is requested exactly once. It fails with the old query options and passes with the new ones. The branch is also rebased onto current main, because the PR had gone DIRTY.

@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention and removed merge conflict Branch has merge conflicts with its base — author must resolve before merge readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running labels Oct 29, 2026
@IngridMorstrad
IngridMorstrad force-pushed the fix/mcp-management-to-customize-24920 branch from 832d1c6 to fadb5a4 Compare October 29, 2026 04:49
@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

Opus 5.5 (blocking, 832d1c6, feature-map-correctness): fixed in fadb5a4. The feature map's Developer intro now says eight pages after Feature Previews; the table has eight rows: logs, system, telemetry, storage, memory, config, debug-tools, archive. It also names the one exception to the same-page redirect: /developer?tab=mcp-pool and /settings/developer/mcp-pool both go to Customize → Connections → MCP Servers (/capabilities?tab=mcp&view=mcp-servers#mcp-management). I swept the other docs for page counts or redirect claims; only the historical RFC still says "nine", and I left it as written. check_feature_map.py and docs lint pass.

@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: passed Eligible automated validation passed for the current revision and removed readiness: action required A blocking check or review needs attention readiness: checking Automated validation is still running labels Oct 29, 2026

@iamwhatever iamwhatever left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Recommendation: REQUEST_CHANGES (issue required) at fadb5a4079a3752a96d339b85ff545ffb351508a.

The move itself is sound. MCP Management mounts lazily as a #mcp-management section under the MCP Servers list (ConnectionsPage), and both old URLs redirect. Its "Ask the agent" hand-offs are withheld while the list holds unapplied edits. No backend handler changes, and the approve/enable endpoints stay owner-gated. All lanes pass on this head, every check is green, and the branch merges cleanly with current main.

Blocker: two follow-ups you deferred have no linked issue.

  • The inline reason when "Ask the agent" is withheld while edits are staged (McpManagement.tsx). Your comment says "Still deferred".
  • Renaming MCP Management's inner "Servers" sub-view so it does not read like the "MCP Servers" sub-tab (deferred here).

Please fix them now, or open one issue per item and link both in the description.

AI findings judged non-blocking

  • UX Review: the tip still says "Shared MCP gateway" toggle, mixes ›/→, and its button reads "Open MCP Gateway". These are wording nits (tips_curated.json).
  • Design Review: the description does not mention the non-owner footprint 403 handling. The code is correct and tested (footprint query); adding a line to the body would help.
  • Design Review: other exits still drop staged edits. McpTab behaves the same way on main.
  • First Principles: the capture script and PNGs are committed. A fork PR cannot attach evidence, and they have no runtime effect.

— crew-pr-reviewer (r134)

@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

Thanks. I went with the issues route; both are linked under Follow-ups in the description:

  1. Inline reason when "Ask the agent" is withheld while edits are staged → MCP Management: say why "Ask the agent" is hidden while MCP server edits are staged #25690
  2. Rename MCP Management's inner "Servers" sub-view → MCP Management: rename the inner "Servers" sub-view so it doesn't read like the "MCP Servers" sub-tab #25691

I also added a line to the description about the non-owner footprint 403 handling, per the Design Review note. There's no code change, so the head stays at fadb5a4.

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: passed Eligible automated validation passed for the current revision labels Oct 29, 2026
@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

GPT 6.1 (blocking, 4883fb7, errors-use-error-notice): fixed in 1e2825f.

McpManagement.tsx now passes footprintFailed={footprintQ.isError}, so a non-owner's 403 shows the existing inline failure notice instead of a silent gap. The refusal still suppresses retry and the 30s polling (isPermanentRefusal in retry/refetchInterval is unchanged). The hero test for the 403 case now asserts the notice is shown and the read happens once. The PR description's non-owner line is updated to match.

@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

UX Review (Opus 5.5, blocking, d1d952e, non-owner refusal shown as a fault): fixed in af3d59c.

  • McpManagementHero.tsx: when the footprint read is refused (footprintRefused, a non-owner's 403), the hero now shows a plain muted line, "Memory figures are visible to the crew owner only.", with no error styling and no "Ask the agent". Any other failed read still uses the shared inline failure notice. Still one request, no retry, no polling.
  • New key pages.mcpManagement.footprint_owner_only in all locales (pseudolocale regenerated).
  • The hero test for the 403 case now checks for the owner-only line and asserts no alert and no failure text.
  • Evidence gap closed: the capture script has a non-owner frame (after-non-owner-footprint-refused-{dark,light}.png), now in the description.

The second evidence gap (an error notice with edits staged, "Ask the agent" withheld) is the behaviour tracked by follow-up #25690.

@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

GPT 6.1 (2 blocking, af3d59c): both fixed in ffeb500.

  1. Deep-link scroll ran before the server list settled (ConnectionsPage.tsx): the scroll effect now returns early while the shared ['mcp-servers'] query is loading and has isLoading in its deps, so it scrolls once the real table has replaced the skeleton. New test in McpManagementMove.test.tsx holds the list pending, asserts no scroll, resolves it, asserts the scroll; it fails with the gate removed.
  2. Permission refusal bypassed ErrorNotice (McpManagementHero.tsx): the refused read goes through the shared inline ErrorNotice again, with the same staged-edit hand-off decision, retry and polling still off. Its message is the owner-only reason ("Memory figures are visible to the crew owner only.") rather than the generic read failure, which keeps the honest wording the UX review asked for. The non-owner screenshot in the description shows it.

On the UX review's "no error styling, no Ask the agent" suggestion for that refusal: errors-use-error-notice (blocking) counts a rejected request as an error whatever its status and says a refusal whose agent can still run keeps askAgent, so I followed the rule and kept only the wording change.

@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

Opus 5.5 (blocking, ffeb500, dashboard.md § Developer): fixed in f62e860. "MCP pooling" is gone from the Developer page list, and the section now says /developer?tab=mcp-pool and /settings/developer/mcp-pool redirect to Customize → Connections → MCP Servers, where MCP Management is a section below the list. Docs lint passes.

UX Review (ffeb500, non-owner refusal): this conflicts with GPT 6.1's upheld blocking finding on af3d59c, which applied errors-use-error-notice: a rejected request (the 403) must render through ErrorNotice, and a refusal whose agent can still run keeps askAgent. The dashboard has no client-side owner signal, so the read can't be skipped for non-owners. I've left it on ErrorNotice with the owner-only wording and am asking a maintainer which way this surface should go.

@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

UX Review (non-owner owner-only notice shown as an error with "Ask the agent"): keeping the current code, deliberately.

The 403 is a rejected request, and errors-use-error-notice (blocking, website/AUTOSDE.yaml) says a rejected request renders through ErrorNotice whatever its status, and that a refusal whose agent can still run keeps askAgent. GPT 6.1 upheld that as blocking on af3d59c, and the written rule outranks a lane preference, so the notice stays on ErrorNotice with the owner-only wording ("Memory figures are visible to the crew owner only.") and the staged-edit hand-off decision. Background is in my earlier comment: #24933 (comment)

Skipping the read for non-owners would avoid the error entirely, but the dashboard has no client-side owner signal, and adding one is out of scope here. A maintainer can override this if they prefer muted status text for this surface.

@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

Opus 5.5 (blocking, f62e860, doctor fix-it text split kirocrew restart): fixed in 3c3099a. _doctor_mcp_share_refused now passes keep=("kirocrew restart",) to render._print_wrapped, and test/test_doctor_mcp_share_refused.py is back to asserting "kirocrew restart" in out (unflattened), which fails without the keep= and passes with it.

On the note about the refused footprint read re-fetching on remount/focus: non-blocking, and the same pattern McpShareRefusedNotice already uses, so I've left it out of this PR.

Frontend Tests (2) on f62e860 failed on a 15s timeout in i18nAllLanguagesEntry.test.ts (./integration/setup.ts module-graph walk took 17s); this PR doesn't touch that setup, and it re-runs on the new head.

iamwhatever
iamwhatever previously approved these changes Nov 1, 2026

@iamwhatever iamwhatever left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approve at 092b8bb.

Our r134 change request is resolved: both deferred items are now tracked in #25690 and #25691 and linked under Follow-ups. The move is sound. MCP Management mounts lazily as the #mcp-management section under the MCP Servers list (ConnectionsPage.tsx#L816-L830), both old URLs redirect, and the backend changes are text only. All lanes pass on this head and every check is green.

On the non-owner notice: the author kept it on ErrorNotice because the blocking errors-use-error-notice rule requires it, and the wording now says the figures are owner-only. That reply holds; a maintainer can still choose muted text.

AI findings judged non-blocking:

— crew-pr-reviewer (r139)

@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

Rebased onto main to clear a conflict; no change to the approved diff. Could you re-approve?

MCP Management (stub MCP servers, share their backends, Sharing assessment)
was one of the Settings → Developer internals pages, among logs, telemetry
and storage, away from the MCP servers it governs. The share-refused notice on
the MCP Servers list sent people to that separate page to approve a launch.

Render McpManagement as a section below the servers list in Customize →
Connections → MCP Servers (/capabilities?tab=mcp&view=mcp-servers); the
component is reparented, not rewritten. Drop the mcp-pool page from Settings
→ Developer and redirect both old links, /settings/developer/mcp-pool and
/developer?tab=mcp-pool, to the new section so bookmarks, docs and tips keep
working. While the servers list holds unapplied edits, MCP Management's
"Ask the agent" hand-offs are withheld so they cannot navigate away and drop
them. Point MCP_MANAGEMENT_PATH, the doctor message, the curated tip, docs, the
feature map and the GUI feature catalog at the new location, remove the
now-unused tab label from every locale, and add tests for the redirects, the
placement and the repeat-link scroll.

Closes kirodotdev#24920
@IngridMorstrad

Copy link
Copy Markdown
Contributor Author

Rebased onto main to pick up the Windows E2E fix (#28792); no change to the approved diff. Could you re-approve?

@iamwhatever iamwhatever left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approve at 4b64a19.

This head is our r139-approved 092b8bb rebased onto main: git range-diff shows only main-side context lines and regenerated digests changing, with the same 61 files. The move is sound. MCP Management mounts lazily as the #mcp-management section under the MCP Servers list (ConnectionsPage.tsx#L818-L843), both old URLs redirect, launch and stub POSTs stay owner-gated, and the Python changes are text only. The two deferred items are tracked in #25690 and #25691. All lanes pass on this head and every check is green.

AI findings judged non-blocking:

  • Design: McpManagement*.tsx still lives under pages/settings/. File placement only, no behaviour (McpManagement.tsx).
  • UX: the owner-only notice uses error styling with "Ask the agent". The author kept ErrorNotice because the blocking errors-use-error-notice rule requires it, and the text names the owner-only reason (McpManagementHero.tsx#L165-L177). A maintainer can still choose muted text.
  • UX: no reason shown when "Ask the agent" is withheld. Tracked in #25690.
  • First Principles: the footprint_owner_only key is not named in the description. Description wording only; the string ships in all catalogs.

— crew-pr-reviewer (r144)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

fork Pull request from a fork (external contributor)

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Move MCP Management from Developer to Customize → MCP

2 participants