Repository navigation
feat(dashboard): Dashboard edit mode, see where a value came from and ask the Bot to change it - #33265
feat(dashboard): Dashboard edit mode, see where a value came from and ask the Bot to change it#33265chenmingwei23 wants to merge 2 commits into
Conversation
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: the preview is off by default and gates every new surface, the evidence is attached, and the one change that applies with the preview off (larger fold row charges) is reversible. Notes
[DESIGN-REVIEWED] 81c48b8 Description read[DESCRIPTION-READ] 9fa8ba965650889733c1505ed38d10219ee3fe43fd0750f38fefd7fdd31146fd 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 |
UX Review (Opus 5.5) — ✅ PASS (a BLOCK with no Trigger and no Outcome does not block)UX-level review of UX-Verdict: PASS cannot evaluate: missing screenshots of the popover's current Edit button and correction form. The attached images come from an earlier revision, and the only recording appeared blank to the reader. Evidence gaps
Notes
[UX-REVIEWED] 81c48b8 Description read[DESCRIPTION-READ] 62dba26d91a7b5e0263b123b03edfb936aff2ed70ddc20437b1c6e45b0d5a4d2 That sha256 names everything this verdict read: the PR description, plus the 36 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) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Everything new sits behind a preview that is off by default. The one change that applies with it off, the larger fold row charge, keeps the existing upper-bound rule for rows read off disk. Notes
What this change shipsInventory (9 items) — 8 justifiedIntent: Let a person trace a Bot's dashboard value back to its source, ask the Bot to change a component, and correct a value the Bot wrote. ADDITION.
[FIRST-PRINCIPLES-REVIEWED] 81c48b8 Description read[DESCRIPTION-READ] a9d8892d6aa32578e9f0e8fedf22c2c4b27e2ac616fbf59fd59a03e95d8ce776 That sha256 names the PR description this verdict read. Recompute it with |
GPT 6.1 Review — 🔴 changes requested (blocking)GPT 6.1 found at least one blocking issue that must be resolved before merging This comment is updated in place on each push. BLOCKING -- website/src/pages/members/CrewDashboardEditLayer.tsx:638 -- Retained corrections remain enabled after a field becomes non-scalar
[BLOCK-MERGE] 81c48b8 Adjudication (Opus 5.5) — is blocking on each finding proportionate?F1 — LOW harm: the gateway refuses the save ( False positive or not applicable? A repository writer can comment: |
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: |
ca18617 to
bf5480b
Compare
8aa206e to
fbed84b
Compare
1891bba to
10d6fd3
Compare
Sensitive Change ReviewSensitive files changed:
|
Sensitive change review
|
10d6fd3 to
ce4faf6
Compare
Sensitive change review
|
Sensitive change review
|
ce4faf6 to
6c534ec
Compare
Sensitive change review
|
Sensitive change review
|
Sensitive change review
|
Sensitive change review
|
Sensitive change review
|
Sensitive change review
|
Sensitive change review
|
Sensitive change review
|
Sensitive change review
|
Sensitive change review
|
Sensitive change review
|
Sensitive change review
|
Sensitive change review
|
|
Re: review at e33377b, the three regex literals in Fixed in
No regex literal remains in This push also rebases onto main (conflicts were in the preview flag list, the Settings previews list, the settings and UI index generated files, the 13 locale catalogs, and two pinned-digest tests). The generated files were regenerated, and the route-table and entry-types digests were re-pinned to the new values. |
Sensitive change review
|
Sensitive change review
|
|
Disposition for GPT 6.1 on
The branch is still two commits. |
Sensitive change review
|
Sensitive change review
|
Sensitive change review
|
Sensitive change review
|
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
Thanks for working through every round on this. One thing still blocks merge on this head.
Where the earlier items stand:
- Feature-map row says Inspect (
docs/feature-map/README.md:212): fixed. _READING_INT_LIMITis in thecrew_log/projection.pymanifest (line 183): fixed.- The orphan context note for the renamed change keys: fixed.
- Escape and E while the panel is collapsed: fixed.
hiddenFromViewguards both listeners (CrewDashboardEditLayer.tsx:369and:1181). - The regexes named in the change request from bolichen97: gone on this head. That review needs a re-review from them.
-
The UX review could not see the correction flow on this head. Its verdict on c2bdab7 reads PASS, but its body opens "Cannot evaluate: missing the screen after 'Save correction'". It was filed as PASS only because it had no Trigger or Outcome lines. A verdict that cannot be reached for missing evidence counts as a BLOCK. Correcting a value is one of the three goals. The description's correction stills stop at the filled form, and the corrected-value redraw is "left for the narrated end-to-end video recorded in a pod". The description does not link that video. The demo video in the comments was recorded on 3a6ee34, many pushes before this head. Your comment on 377efe8 also says the after-save still is not in yet. These surfaces at head show up in no attachment:
- the tile and card after a successful save, with
summary_person"Set by you ..." (CrewDashboardEditLayer.tsx:1061,:764) - the "Use this page" button and its hint (
:1107), andpaused_newer_page(:637) - the failure notices:
send_failedandsend_unconfirmed(:995,:999),correct_failed(:1100),trace_failed(:1251) and the refused state
Smallest fix: take stills of those states from this head and add them to the description. Your stubbed Playwright harness can do it with a stubbed correct route and a lineage reply that has
actor: "person". The after-save state and the "Use this page" control matter most. The UX lane also says shots 04, 28 and 32 show older wording, so re-shoot those against this head. Then let the UX lane run again. - the tile and card after a successful save, with
Separately, the linked issue #33255 is still parked. A maintainer has to sign off the design before this can merge, even after the evidence is in.
Optional:
- The owning specs still describe the old shapes.
learn-cron-dashboard.mdsays the record body is read fortypeanddataonly.crew-log-projection.mdlists the custom and agentic rows withoutsource,points,provenance,actororprior. The two new owner routes are in no spec. crew_log/entry_types_surfaces.py:221keeps 120points, but the only reader draws the last 40 values and drops the timestamps. Thecustomrow charge increw_log/projection_budget.pyis sized for 120 and applies with the preview off. A limit of 40 and a measured charge would make that change smaller.dashboard/handlers/member_dashboard_lineage.py:544: a timeout on the instance write lock raisesInstanceError, and nothing catches it, so the route returns a bare 500. Returning a coded refusal such aspage_busywould match the other refusals on this route.
…e preview Component ids on view.blocks (data-dashboard-block, checked at the write gate), the frame's edit bridge and kirocrew.pick, the lineage route, the person's correction of an agentic value (actor person, told to the Bot on its next turn), and a script record's declared source/points beside the gateway's job/script/sha stamp. Preview off composes and folds exactly as before.
Sensitive change review
|
|
Disposition for Opus 5.5 on
|
…view Edit toggle on a Bot's Dashboard tab, the capture layer over the sandboxed frame, the lineage popover (connector line, hop-by-hop, Esc, never over the value), the component flip with provider and a change request sent as the person's message, popover Ask/Correct, Sources, the preview card, the strings in every catalog, the feature map row, the config reference and the generated indexes. Part of #33255
|
Disposition for GPT 6.1 on
|
Sensitive change review
|
|
Converting this PR to draft and pausing pushes. GPT 6.1 raised one new blocking finding on |
|
Closing for now. The team is reconsidering how much to invest in Bot dashboards versus the conductor-mode Board panel, and the last push (81c48b8) drew another blocker after the declared final push (GPT 6.1 at CrewDashboardEditLayer.tsx:638, retained correction stays enabled after a field turns non-scalar). The branch stays; reopen if the dashboard direction is confirmed. |
Problem / Motivation
Goal: On a Bot's Dashboard tab, the person can see where any value came from, click a component to ask the Bot to change it, and correct a value the Bot wrote, all behind an opt-in feature preview.
A Dashboard page shows numbers with no way to ask "where did this come from?" from the page itself. To change one card the person has to describe it in chat from memory. A value the Bot wrote cannot be fixed by hand. Pages that draw in script (canvas, WebGL, script-written text) and pages fed by a script cron (a CloudWatch metric, a PR count) are the hardest case: nothing the eye reads is a bound element.
Why it matters
The person trusts a number only when they can trace it, and asks for a change faster when they can point at it. Without lineage, a stale or wrong value looks the same as a current one, and a script-only board gives no hint of which job, script or reading produced it.
Not a goal
seqon fold rows, evidence on verdict entries, andsource/observed_atondashboard_write. Each shows in the lineage as aneeds: <what>hop instead.deriveon a model field (a page's own math).What changed (motivation → approach → change)
Approach: the host draws every word of a lineage, and the page only says which field path is under the pointer, so a page can point at a field but cannot fake where it came from.
dashboard_edit_mode(default off): a Settings card, a gateway flag, and a browser flag. Off, the composed document, every package write and fold row, and the Bot's turn are byte-for-byte as before (pinned by tests).view.blocks[].idbecomesdata-dashboard-block; ids are checked (charset, unique, at most 64) at the package write gate only, so a stored page never stops reading.pointmessages with{path, rect, block}for bound cells, and throughkirocrew.pick(handler)for script-drawn values; keyboard focus reports too. Its capture listener stopsedit/point/wheelbefore the page sees them.GET /api/members/{slug}/dashboard/lineage: owner-only, read-only, path validated; hops from the field spec, the fold value, the agentic row (writer, time, earlier write, and freshness from the field'sstale_after_sthroughdashboard_agentic.stale_after_map, the same time the page's out-of-date mark uses) or the custom record (writer,stale_after, last failure, gateway stamp, script's declared source).E; the popover pins while Edit is open; the first Esc closes Edit, the second the popover). Every value: ask the Bot to change it. A value the Bot writes: correct it throughPOST .../dashboard/correct(owner-only, refused to an agent caller, same type check asdashboard_write, scalars only), recorded withactor: person, the replaced write kept asprior, and a[DASHBOARD]line on the Bot's next turn. A script's record: the job behind it (schedule, hold time, script, last run), Open Schedule, and Run now, which asks first. A folded value: ask the Bot to change how it is shown. Each option that does not apply is shown greyed with its reason.ctx.crew_log_append(..., source={kind, ...}, points=[[t, v], ...])withkindone of cloudwatch / http / sql / prometheus / file / other, at most 2 KiB,console_urlhttps only and never fetched (a link the redactor rewrites is dropped, not the record); at most 120 points. The gateway stamps the job, run time and script path. It records no digest of the code, because the file on disk when the record arrives may not be the code that ran. The scheduler names no run id yet, so the lineage shows aneeds: run id and code digesthop instead. Each hop is marked "recorded by gateway" or "declared by script".Open questions (each took the conservative option; reversible)
Other decisions taken while the owner was away: the page stays one page with named regions; only the person can Edit, ask or correct, never a page script.
Backwards compatibility
Compatible: every new behaviour but the row charges below is behind the preview, which is off by default; new crew-log fields are optional and absent with the preview off.
Removes nothing: the change adds a preview, two routes and optional fields; no path, key or validator that works today changes.
One change applies with the preview off: the warm-fold memory charge per row rises for the
customfold (8,900 to 18,100 bytes) and theagenticfold (8,400 to 18,700 bytes), so the charge covers the widest row those folds can keep. Under a fixedslot_fold_cache_bytesceiling, fewer of those cells are kept warm, and one that is not warm is folded again on its next read.Tests
test/test_dashboard_edit_mode.py(46): preview off keeps the document, blocks and fold rows unchanged; block ids on/refused/duplicate at the write gate only; bridge order and thepickanchor; field-path grammar; lineage for fold, agentic (person write over a Bot write) and script records; the folds keep actor/prior and source/points/provenance; the record door's checks and the gateway stamp; both routes 404 with the preview off; the Bot's turn line.website/src/pages/members/dashboardEdit.test.ts(6): manifest shapes, picked-message sanitizing, refusing undeclared paths, popover placement, correction parsing, change-request text.website/src/pages/members/CrewDashboardEditLayer.test.tsx(6): Edit is posted to the frame, an undeclared path is refused with no gateway call, the gateway's lineage is drawn and a folded value's Edit greys Correct, Edit pins the popover and Esc closes Edit before the popover, Run now asks before it runs, and a change request is sent only on Send.test_dashboard_frame.py,test_artifact_dashboard_package.py,test_feature_previews_config.py,test_config_baseline.py,test_crew_log_types.py,test_agent_panel_routes_record.py, the configuration doc tests, and vitestcatalogParity,previewSurfaces,previewFlagServerSync,featurePreviewsSectionNotice,CrewDynamicDashboard.Proof: not provable -- other: feature PR, every new test fails on main because the code it calls is new
Manual verification
Screenshots below come from an isolated pod of this branch with the preview on.
Screenshots / video
These frames come from the built SPA (
website/distat this head), driven by headless Playwright. No server or socket was used. Every request to a fake origin is answered bypage.route: static files come from the build, and the gateway's/apiroutes are stubbed. The page payload and every lineage answer are generated by the gateway's own code (dashboard_frame.composewith the edit bridge, andmember_dashboard_lineage.build_lineageover seeded fold rows), so their shapes are the real ones. The Bot is "Atlas" with the preview on. Light and dark are shown for every step.The stubs cannot show steps that need a real Bot turn: the change request arriving in the Bot's chat and the Bot answering it, a corrected value being written and redrawn, and a script job actually running. Those are left for the narrated end-to-end video recorded in a pod.
The change card, recorded: point at p99, click its component, and a card titled "Change latency" turns over from that box. Closing it shows the page again.
The Inspect button is shown at rest, with no hover needed:
Inspect mode on:
Point at a counted value: one plain sentence, with the steps closed under Details:
Details open on the same value:
A script-written value (CloudWatch p99), with its record, freshness, script run, what the script read and its readings:
A value past its stale limit says so under the summary and on its Freshness step ("Out of date: went stale 23h ago."), never as a raw timestamp:
The Edit button is named for what its menu offers on each value: "Correct it or ask Atlas..." on a value the Bot writes, "Ask Atlas or see the job..." on a script value, "Ask Atlas to change this..." on a counted one. On a script value, correcting it by hand is unavailable, with the reason shown, and the job behind it is offered:
The job panel (job, schedule, how long the value counts as current, shown as "2h", script, last run, Open Schedule, Run now). The actions stay pinned at the popover's bottom while the steps above scroll:
Run now asks first ("Run the job once now, outside its schedule?"):
After the confirm, the run is started (the run route is stubbed), and Run now stays off so it is not started twice:
A value the Bot wrote, showing who wrote it, the person's earlier correction with its reason, and freshness:
Correcting that value. The new value, the optional reason and "Save correction" stay fully in view at the bottom:
Click a component: the card titled "Change latency", which shows the provider and the context chips:
A request typed, with Send enabled:
After Send, the request shows in Atlas's open chat at once as the person's own message, and the card stays on "Sent to Atlas." until the person presses Done:
Sources: every field the page reads, with how each is fed:
The preview's card in Settings, under Developer and then Feature Previews:
Related Issues
Part of #33255
Checklist