Repository navigation
feat(dashboard): the Dashboard tab renders the package, with no default page - #18760
chenmingwei23 wants to merge 2 commits into
Conversation
UX Review (Opus 5.5) — 🟡 CONCERNSUX-level review of UX-Verdict: CONCERNS The new empty state tells the user to "Ask this crewmate to build its dashboard" but gives them no way to do it from there, so they have to find the chat on their own. Watch
Suggestions
[UX-REVIEWED] 5810f07 Description read[DESCRIPTION-READ] 14cfa01a969ccc7f1785457293d57068b8231b935c02ecb059d5cbac2e8cda5e That sha256 names everything this verdict read: the PR description, plus the 14 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 |
Design Review (Opus 5.5) — 🟡 CONCERNSDesign-level review of Design-Verdict: CONCERNS The server still writes and renders a default page for crewmates with no dashboard, and only the tab declines to show it, so agent writes succeed against a page nobody sees. Watch
[DESIGN-REVIEWED] 5810f07 Description read[DESCRIPTION-READ] be6e164ffebd52a198539453f1ea0ca2154b0870bfc3bc178a3482733ee955ee 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 |
First Principles Review (Opus 5.5) — 🔴 BLOCKPremise-level review of First-Principles-Verdict: BLOCK product-shape change without accepted RFC: crewmates with no adopted template stop seeing the default page, which the in-progress RFC says they must see. Not justified as shipped
What this change shipsInventory (7 items) — 4 justifiedIntent: a crewmate's Dashboard tab draws the package its agent composed, and makes up no page. This is an ADDITION.
Blockers
[FIRST-PRINCIPLES-REVIEWED] 5810f07 Description read[DESCRIPTION-READ] cfda6cd1e160fe7755b2c8b34ddd78b33ab75e9bf88bacd9dd9cedf168eea30a That sha256 names the PR description this verdict read. Recompute it with |
Opus 5.5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
f90731b to
6e42e0c
Compare
f31db48 to
a3979d5
Compare
6e42e0c to
bf1837b
Compare
a3979d5 to
846c847
Compare
bf1837b to
718f6a4
Compare
4d7bf2f to
42f6ae9
Compare
2f6be4e to
bc1d575
Compare
42f6ae9 to
4ae57d9
Compare
bc1d575 to
a192e65
Compare
Human judgment recorded@chenmingwei23 marked the first-principles AI finding as false positive, not applicable, or explicitly accepted for
This decision applies only to this commit. A new push requires a new judgment. |
4ae57d9 to
05ab809
Compare
a192e65 to
d48ff15
Compare
05ab809 to
1303279
Compare
…lt page The v3 Dashboard tab reads a dashboard artifact's layout package and draws from it. There is no composed default page: a member with no package gets an empty state, scoped to PACKAGE-BOUND members so the shipped templates keep working. An absence is the whole subject here -- no page drawn, no mint, no held page -- so the capture entry carries one row per state, and the row whose body holds a full dashboard while the tab shows the empty state is the one a reviewer should look at hardest. Its header names which parts are real (the component, the providers, the theme tokens, the i18n catalogs) and which are stand-ins (the controller response, the mint, the document), and the two no-fallback captions say they prove the frontend half only. One home per value, on the receive side too. The block patch carries field NAMES in `blocks` and every value once in `read`, redacted and formatted by the renderer's own rules, so `blocks` is read for WHICH blocks moved and never for a value -- a copy there would show a bare 1200000000 where the first paint showed 1.2 GB. The apply path posts `read` whole, under the document's own message type, which the bootstrap listener compares exactly. A fold push is not a full repaint. The patch constant held the full-paint value, and that listener replaces the whole read and re-initialises every block, so a fold advancing gave a canvas-owning block a second canvas with two scenes animating over each other. The constant is renamed to PAGE_FULL_PAINT_MESSAGE_TYPE, a fold push forwards the renderer's own block_patch() output verbatim, and blockPatchMessage() is gone so the narrowing, the formatting and the message name stay in one language. Three pins read Python: each type against the module that owns it, plus an inequality between them, because convergence is the bug. A lower push version is a restart, not a replay. The push counter lives in server memory per live page, so a gateway restart arms a fresh page at 0 and the next frame arrives below what the tab holds; read as a replay it was dropped, and so was every frame after it, until the counter climbed back and the page resumed having missed the gap. The check is strict equality, and `held` is taken from the read rather than from a frame that failed it, because a refetch exists precisely for when what the tab holds cannot be trusted. Mutations pin the halves that a passing suite would not otherwise prove: tolerating a lower version reddens 5 tests, advancing `held` from a refetched frame reddens the case asserting the position comes back from the read, and posting the full-paint type for a fold push makes the capture run refuse to write the R5 shots. Refs #18642
…from The schema is built from the live data-type and block catalogs rather than written out beside them, so a type added to a catalog appears in the schema with no second edit. website/src/types/dashboardPackage.schema.json and the dashboardPackage.ts types it generates are its only consumers, which is why it ships here rather than with the write gate.
d48ff15 to
5810f07
Compare
Problem / Motivation
Goal: a crewmate's Dashboard tab draws the dashboard its agent composed, and shows an honest empty state when there is none.
The parent PR gives a layout package a validated home, but nothing renders it. The tab reads the template registry, so a package-bound member's page cannot be drawn. Worse, a read with no page makes the tab compose a DEFAULT from whatever data is to hand, so a member whose agent composed nothing sees the fallback's guess.
Two live-update faults sit alongside: a fold push uses the FULL-PAINT type, so a canvas-owning block gets a second canvas, and the gap check drops any version at or below the one held, so a restart arming a fresh page at 0 freezes the tab.
Why it matters
A dashboard a reader cannot trust is worse than none: the composed default cannot be told apart from a real page, so "my report says X" stops being a statement about the project. Both live-update faults are silent: the canvas duplication corrupts the page while every assertion passes, and the restart fault shows stale numbers, then resumes having missed the gap entirely.
Not a goal
What changed (motivation -> approach -> change)
Goal: render the package, invent no page.
blocks, values once inread. A copy would show a bare1200000000where the paint showed1.2 GB.PAGE_FULL_PAINT_MESSAGE_TYPEand the push forwards the renderer'sblock_patch()output verbatim.blockPatchMessage()is deleted, so this side builds no payload.heldadvances only from a read.revision, not the package version, which is 0 for every template page, so two compared equal and a template apply swapped onto the screen without proving it loads.postMessagecarries an inlinenosemgrepfor the wildcard-target rule, matching four existing sites: sandboxed withoutallow-same-origin, the document's origin cannot be named. Its controls: no network, no agent script, a sender-is-parent check.aimbadge in the Artifacts library andArtifact.kindwidens todashboard; no picker offers it, the backend keeps it out ofUSER_SELECTABLE_KINDS.package_json_schema()(draft 2020-12) is built FROM the live catalogs, so a type added to one needs no second edit.dashboardPackage.schema.jsonis its output anddashboardPackage.tsis generated from that; its enum is the renderers' ten block types.The no-template design comes from #18642's design doc.
Backwards compatibility
Breaking: a crewmate who has adopted no template stops seeing the registry default page and gets the empty state.
The fix sits at SYMPTOM level.
default_instance()returns the registry default with its state left atSTATE_EMPTYandapi_member_dashboardrenders it, so an unadopted crewmate arrives withstate: 'empty'AND a page attached, and the gate declines to draw it. Scoping the gate to package-bound members was tried and abandoned: it makes five tests here assert the opposite of their intent.Writer sweep on
origin/main: the only writer of theempty-with-default contract isapi_member_dashboard, unmodified.Reader:
website/src/pages/members/CrewDynamicDashboard.tsx-- crew page -- declines the attached page for ANY member whose read saysempty--CrewDynamicDashboard.test.tsxReader:
src/kiro_crew/dashboard/handlers/member_dashboard.py-- crew page, unmodified -- still substitutes and renders the default, now declined rather than unsent --test_member_dashboard_routes.pyReader:
src/kiro_crew/dashboard/handlers/agent_panel.pyread_instance-- the agent's panel tool -- resolvesemptyto the DEFAULT's manifest, so an unadopted crewmate's write is accepted against a page this tab does not draw. FIXED with the template line (drop-templates) --test_agent_panel_routes.pyReader:
src/kiro_crew/dashboard_package.py:322_from_template_instance(from #18707) -- the WRITE route -- substitutes the default at:362, so a write lands on a Model this tab will not render. LEFT with the other three: all four go with the template line (drop-templates) --test_dynamic_dashboard.pyReader:
website/src/pages/members/dashboardBlockPush.ts-- crew page -- readsblocksfor which moved,readfor values -- its own testReader:
website/src/hooks/useWebSocket.ts-- app bundle -- forwards the patch frame; an emptypatchre-reads --useWebSocket.frameTrace.test.tsReader:
website/src/api/client/agents.ts-- app bundle -- carries the package, template and push-version wire keys, shape unchanged --ApiClient.refactor.surface.test.tsThe rest is compatible:
package_json_schema()is additive, the renamed constant internal.Tests
Zero skips, and no
skipIfleft in this PR's test files. Three were cross-language pins on modules the later PRs own, moved out with them. The fourth pinneddashboard_frame.py, which IS here: it is unconditional, andreadFileSyncthrowing IS the assertion.Five mutations, each reverted after its run: a lower push version tolerated fails 5 cases;
heldadvanced from a refetched frame fails position-from-read; the full-paint type for a fold push makes the capture refuse the R5 shots; probation onlayoutfails the changed-template case; the schema's block enum bound to the field catalog fails enums-follow. The last two restores are sha256-verified.Manual verification
The component, providers, theme tokens and i18n catalogs are real; the controller response, the mint and the host document are stand-ins, and each capture header says which. R6 mounts the real
LibraryTableover two stand-in records.Screenshots / video
R1 the tab draws the package. The HOST stays dark in both shots: the light one shows the component's own light surfaces inside it, so what these evidence is the cards' tokens and not a light page.











R2 a member with no package gets the empty state.
R3 the tab declines a page it was sent, and the same on the PACKAGE-BOUND shape. Frontend half only: the route is stubbed, so this is the tab refusing, not the server declining.
R4 the body carries a full dashboard and the tab still shows the empty state.
R5 before/after a fold push, both card palettes on the same dark host: only subscribing blocks moved, values formatted by the renderer's rules, nothing re-initialised.
R6 the Artifacts library lists a dashboard artifact, its kind badged in the same
aimvariant as thewebapprow beside it, in the raw kind word.Related Issues
Refs #18642. Depends on #18705 and #18784; the T3 one-pager is approved.
Checklist
Contribution License Agreement
Made under Apache-2.0, per CONTRIBUTING.md.