design: dev-dashboard revamp umbrella (analysis + Claude-Design prompts)

[rickylabs]

Dev Dashboard Design Revamp — Umbrella (analysis + design-spec only)

DRAFT umbrella PR — nothing here merges to main until the owner reviews. This run changes NO
product code; it produces analysis artifacts and Claude-Design-ready prompts under
.llm/runs/dashboard-design--orchestrator/.

Refs #400 (dev-dashboard epic; no closing keyword — umbrella).

Mission

Produce the design reference bar for the beta.10 dev-dashboard milestone
(https://github.com/rickylabs/netscript/milestone/12): a locked routing hierarchy, screen-by-screen
catalog of the current Claude Design prototype, multi-model analysis passes, a dynamic
plugin/extension architecture proposal, a beta.10 cross-coverage matrix, and multiple
self-contained Claude Design prompts the owner can hand directly to Claude Design.

Slice plan (each merged into this branch, tracked by a per-slice comment)

1. Screenshot inventory + screen catalog of the current prototype (DesignSync + playwright)
2. Prior-work + reference-routing analysis (routing resort ground truth)
3. GLM 5.2 design/UX critique pass (OpenRouter capability test)
4. Codex GPT-5.6 Sol max adversarial UX/DX pass
5. Dynamic plugin/extension system architecture proposal (Fable 5 low single delegation)
6. Beta.10 cross-coverage matrix + issue context augmentation
7. Final Claude-Design prompts + run eval

Improvement axes covered

Zero future-beta prose · complete routing-hierarchy resort · full feature coverage incl. project
writes · beta.10 cross-coverage · diversified AI surface (not a generic chat) · dynamic
plugin/extension system.

Supervisor session: 0e4ec217 (Fable 5 medium) — see
.llm/runs/dashboard-design--orchestrator/orchestrator-session.md.

[rickylabs]

✅ Run complete — dev-dashboard design revamp (all 7 slices)

One action needed from you to publish the umbrella content (auto-mode classifier blocks me from pushing reference-derived analysis to this public repo):

git -C /home/codex/repos/netscript-547-lffix/.claude/worktrees/dashboard-design-orchestrator push

After that push, this PR's diff contains everything below under .llm/runs/dashboard-design--orchestrator/.

The deliverable: six Claude-Design-ready prompts (design-prompts/)

#	Prompt	Produces
P1	Shell, IA & Home	locked route tree, 4-group sidebar w/ derived badges, pathname breadcrumbs, ⌘K (navigate/act/recent), revamped Home
P2	Investigation spine	/flow/:correlationId journey (flagship) + /runs/:correlationId inspector twin
P3	Capability consoles	Workers (jobs + polyglot tasks), Sagas (compensation hero), Triggers (8 types + action chains + builder), Streams (fan-out) — list→detail→leaf, full writes
P4	Control plane	Runtime-config audit/rollback workspace, labeled-edge topology, Catalog procedure detail (duality), Migrations apply, full DLQ, Auth-as-projection-debugger
P5	AI surface	distributed assists (chips + inline grounded cards + proposed-action → confirm), global Ask, AI-drafted triggers, /ai console + durable runs on the spine
P6	Extension platform	/extensions manager, injection-zone inspector, trust tiers + permission prompts, install/create-from-template with file-diff previews, provenance chips everywhere

All six: final-product framing (zero future-beta prose), the locked hierarchy from analysis/routing-resort.md, one canonical Stripe correlation fixture, confirm-with-CLI on every write, satellite doctrine intact, leak-checked for internal names.

Also on the umbrella

• coverage-matrix.md — 30-issue bidirectional matrix. Headline gaps: [dashboard DDX-14] CLI surface + auto-launch #424's flat URL scheme is contradicted by the locked hierarchy; AI console, Auth Sessions, and /extensions have no owning beta.10 issue; polyglot Jobs/Tasks split absent from [dashboard DDX-18a] workers per-capability dashboard section #428's current scope.
• 20 context-augmentation comments posted on epic: NetScript Dev Dashboard — the Aspire/Scalar satellite that drives the framework (ships as a plugin, beta.6) #400, [dashboard DDX-5] Fresh build-console shell + app-registration + IA #415–[dashboard DDX-19] Codegen-from-UI Add-resource action #432, [dashboard DDX-20] S3: Runtime-Config Monitor & Control (flagship) #551–[dashboard DDX-23] seam-event flow plane: unified envelope + HTTP boundary events (S13 co-req) #557 (each names its new reference bar + PR design: dev-dashboard revamp umbrella (analysis + Claude-Design prompts) #685).
• run-eval.md — incl. the GLM 5.2 capability test: no agentic lane can drive GLM today (codex Responses-wire namespace tool type 400s on OpenRouter; claude-custom-env returns empty completions; runtime Claude adapter is smoke-only); critique delivered via direct API, text-only (GLM has no image endpoint). Three launcher/config bugs documented for follow-up.

Epic #400 gets a cross-link comment next. This PR stays DRAFT for your review.

[rickylabs]

Slice 8 — PR #506 absorbed (partial, deliberate)

Per owner direction, feat/dashboard-design-prototype (#506) is now merged into this umbrella (5d905018, +11,417 lines) so #506 can be closed as superseded when #685 merges (manual close — PR-on-PR has no auto-close keyword).

What was absorbed:

• tools/design-sync/ (57 files) + deno task design:sync + the .gitignore scratch rules — the reusable fresh-ui → Claude Design converter. Scoped deno check on the tool: green (11 files, 0 errors). This is the tool the six revamp prompts will be executed with.
• The design + rescope run artifacts (63 files under .llm/runs/) and resources/design/ prototype screens.

What was deliberately DROPPED (resolved to main): the branch's 42-file packages/fresh-ui delta. That work was extracted and landed via the #547 pixel-polish pass (the branch even records #547's landing); the merge conflict on layouts.css confirmed main has moved past it. Any residual registry improvements belong to the #509 fresh-ui lane with proper review — not riding a design umbrella. Zero packages//plugins/ changes remain in this PR's diff.

Why absorb rather than merge #506 separately: #506's base is beta.5-era (317e4b50) and conflicts with current main; merging it separately would (a) require the same conflict surgery anyway, (b) re-land fresh-ui code already superseded by #547, and (c) auto-close #425 via its Closes #425 body line while the PR no longer represents the current state of that scope. Absorbing keeps one review/merge event and an explicit record of what was kept vs dropped.

Board dispositions at #685 merge time (owner calls):

• plan(design): Dev Dashboard E2E Claude Design prototype + production design-sync — Plan & Design ready for review #506 → close as superseded by design: dev-dashboard revamp umbrella (analysis + Claude-Design prompts) #685 (comment + close).
• [dashboard DDX-15] Claude design-sync artifact + panel prototype #425 (DDX-15) → plan(design): Dev Dashboard E2E Claude Design prototype + production design-sync — Plan & Design ready for review #506 would have auto-closed it; recommend instead reconciling it with feat(design): Dev Dashboard E2E Claude Design prototype + production design-sync system (tools/design-sync) #507 (prototype + production design-sync tracking): design-sync now lands via this umbrella, and the prototype lineage continues through the revamp prompts — either close [dashboard DDX-15] Claude design-sync artifact + panel prototype #425 as superseded by feat(design): Dev Dashboard E2E Claude Design prototype + production design-sync system (tools/design-sync) #507/design: dev-dashboard revamp umbrella (analysis + Claude-Design prompts) #685 or fold its residual scope into feat(design): Dev Dashboard E2E Claude Design prototype + production design-sync system (tools/design-sync) #507.

[rickylabs]

Slice 4b — GLM 5.2 AGENTIC design pass (through the repaired #696 lane) + prompt sharpenings

Now that the OpenRouter lane is fixed, the GLM design pass was re-run as a full agentic session (22 turns, $1.27, claude-print lane, branch design/ddr-s4b-glm-agentic, merged at 4a60cee6). File: analysis/glm-design-pass-agentic.md.

Its conclusion: verdict 7/10; "the locked revamp direction is correct — the six prompts target the three real structural diseases (flat in-memory routing, read-only bias, siloed chat-AI) while preserving the genuine strengths (ns- token system, correlation spine, confirm-with-CLI signature)."* Biggest remaining risk: the six prompts, pasted independently, shared no binding fixture/derived-stats contract — six screens could silently diverge on ids, stat numbers and model labels, undoing the "one id, numbers that reconcile" thesis.

Notable evidence-grounded catches (things the one-shot pass missed or got wrong):

• Polyglot runtime badges are entirely absent in the markup (one-shot said "verify and polish") — it's build-from-scratch work.
• The one-shot had invented route names (/journey/:id, /sagas/instances/:id) that diverge from the locked tree — repo access fixed that.
• Only 3 real netscript … CLI lines exist in the whole prototype vs ~15 writes the prompts add — CLI-transparency must be an enforced invariant, not a convention.
• Two real shell defects named: the Console/Consoles sidebar label collision and the synthetic Console / breadcrumb prefix.

Applied (commit 0946740d): the shared-contract preamble items added to design-prompts/00-README.md (retire-list for ns-waterfall/ns-preview-tag, hard CLI invariant on ns-confirm, numbers-reconcile rule, neutral AI model labels) and P1 now names the sidebar/breadcrumb defects explicitly so a design agent can't preserve them.

GLM capability re-assessment for the run eval: agentic GLM (repo access, xhigh) is a materially better design analyst than one-shot GLM — it cited class names/markup lines, deduped against the six prompts, and self-corrected its earlier pass. Still below the Opus/Codex passes on architectural judgment, but now clearly useful as a cheap evidence-checking critic. Lane cost: ~$1.3/pass.