feat: package UI routing and adopt the existing Boss team

.agent-team/boss-core/README.md

@@ -0,0 +1,54 @@
+# The Boss development team
+
+Ten project-specific roles are defined in [team.json](team.json); use [routing.md](routing.md) to select them. Product and UX have their own responsibilities and acceptance artifacts. Default execution is lead + one specialist + independent verifier, maximum four active including lead. These files configure native roles; they do not start running agents or register a UAR service.
+
+The shared `.agent-team/project-routing.json` selects this existing team for code
+work. Reviewers activate at the completed implementation phase. UI tasks use the
+shared routing catalog through `prometheus-ui-ux`; completed UI review uses
+`prometheus-ui-review`. Read [design-playbook.md](design-playbook.md) before changing
+visible flows. Preserve existing native permissions, models and ownership.
+
+## Start in a harness
+
+Run from the repository root with Node 22+:
+
+- Codex: node .agent-team/boss-core/launch.mjs codex
+- Claude Code: node .agent-team/boss-core/launch.mjs claude --agent boss-lead
+- Kimi Code: node .agent-team/boss-core/launch.mjs kimi --agent boss-lead
+- OpenCode: node .agent-team/boss-core/launch.mjs opencode --agent boss-lead
+- MiniMax CLI: node .agent-team/boss-core/launch.mjs minimax
+
+Ask the parent: “Use boss-lead and the project routing guide for this task; delegate only the bounded roles needed.” Codex has no assumed --agent flag. OpenCode specialists are subagents. Kimi defaults to the locally configured kimi-code/k3 alias unless a model argument is supplied. MiniMax mcode 0.5.4 is installed: definitions are under .minimax/agents, and the launcher sets child-only MINIMAX_DATA_DIR to this project's .minimax. For npm installations with an adjacent Node executable, it uses that executable to avoid the observed Node 26 / Node 24 native SQLite ABI mismatch. The installed loader source recognizes canonical agent.md files in this directory. Native selection and inference with these custom agents remain unverified; this CLI has no --agent flag. Isolated MiniMax login/provider configuration is separate; do not copy global credentials.
+
+## Definition locations
+
+| Harness | Project definitions |
+| --- | --- |
+| Codex | .codex/agents/boss-*.toml |
+| Claude Code | .claude/agents/boss-*.md |
+| Kimi Code | .kimi-code/agents/boss-*.md |
+| OpenCode | .opencode/agents/boss-*.md |
+| MiniMax CLI | .minimax/agents/boss-*/agent.md through the launcher |
+
+Project-local skills live in .agents/skills; Claude skill links are generated by the existing pnpm skills:sync workflow. Native prompts contain exact paths so other harnesses can read them without relying on discovery. Existing Compass definitions in .mcp.json and .codex/config.toml are preserved. Other harnesses use their verified MCP setup or the Compass CLI fallback described in [tool-policy.md](tool-policy.md).
+
+## Working agreements
+
+Read [repository-map.md](repository-map.md), [handoffs.md](handoffs.md), [model-policy.md](model-policy.md), [design-playbook.md](design-playbook.md) and [review-protocol.md](review-protocol.md) as relevant. Native schedulers, permissions, tools and credential stores remain authoritative. Keep KBD position unchanged until its actual workflow advances it. Team state is a local coordination ledger, not a replacement for KBD or authorization.
+
+Use .agents/skills/agent-team-creator/scripts/cli.mjs with the JSON requests documented in its references. Validate team-request.json, update existing state via team-update with expectedRevision, then export to a new directory and deliberately merge reviewed files. Export never installs over existing configuration. Retain installation and verification receipts; do not rewrite live task state by rerunning init.
+
+Finish project creation or refresh with `install-project --project . --team boss-core`
+using that CLI. `--dry-run` previews writes; `--check` reports drift without writing.
+Existing native files are preserved. The managed pointer is shared through the
+existing AGENTS.md link to CLAUDE.md; this operation does not add new symlinks.
+
+## Acceptance and limitations
+
+Configuration verification is recorded in verification.json. Graph evidence is partial, native invocation/inference is not proved by file installation, and cross-model review depends on an available judge. Model cost/capability metadata may be unknown. No application behavior or release certification is implied by this setup.
+
+`reconciliation.json` records the inherited team source. Existing research,
+exports and verification/completion receipts describe that earlier setup, not
+validation of the new UI routing phase. `ui-routing-installation.json` records
+the project-instruction installation only; runtime acceptance is recorded by the
+cross-repository phase gate.

.agent-team/boss-core/completion.json

@@ -0,0 +1,33 @@
+{
+  "schemaVersion": 1,
+  "completedAt": "2026-09-24T23:42:45.911Z",
+  "manifestSha256": "673e108a34a9cd51b37d1d3ac9fa5c6ba932f82069cb13e30286204411001b8b",
+  "roles": 10,
+  "nativeDefinitions": 50,
+  "skillsAdded": 8,
+  "applicationFilesChanged": 0,
+  "peerRepositoryWrites": 0,
+  "globalConfigurationChanges": 0,
+  "commitsCreated": 0,
+  "graphAndMcp": "preserved",
+  "checks": [
+    "creator validate/init/export",
+    "50 native syntax/prompt/hash checks",
+    "OpenCode boss-lead native discovery",
+    "pnpm skills:check",
+    "pnpm docs:check",
+    "OpenSpec strict validation",
+    "launcher syntax/help",
+    "git diff --check",
+    "fresh-context independent review with same-family limitation",
+    "sycophancy detection"
+  ],
+  "limitations": [
+    "mcode 0.5.4 installed; custom-agent execution and provider readiness remain unverified",
+    "native inference/activation across all harnesses untested",
+    "REST MiniMax judge timed out; native Kimi judge lacked entitlement",
+    "Compass graph remains partial static evidence"
+  ],
+  "updatedAt": "2026-09-24T23:52:12.811Z",
+  "minimaxFollowup": ".agent-team/boss-core/research/minimax-installed.json"
+}

.agent-team/boss-core/design-playbook.md

@@ -0,0 +1,27 @@
+# UI/UX working contract
+
+Use boss-ux for new or changed visible workflows. Read DESIGN.md, .impeccable.md, packages/ui/docs/design-token-system.md, variable-catalog.md and the relevant component/stories. Keep existing calm, neutral, content-first direction and @cherrystudio/ui. UX owns design artifacts; renderer owns feature/component implementation. A shared token/API change needs lead assignment and its wider callers checked with Compass/source.
+
+Load `prometheus-ui-ux` for boss-ux and boss-renderer UI work using the entrypoint
+in `skill-bindings.json`. Follow the project `.agents/UI_UX_PROTOCOL.md` if present,
+otherwise the bundled protocol. This operational desktop application defaults to
+Operate mode. Its incumbent design authority remains the files above: focused
+UI/UX Pro Max recommendations do not replace its tokens, fonts or `DESIGN.md`.
+Refinement loads no taste skill. New surfaces or an explicitly authorized redesign
+may select one taste skill and at most one requested overlay; select `gpt-taste`
+only for an actual GPT-family model. Use React/platform guidance relevant to this
+Electron renderer; retain the tracked Electron workflow below.
+
+boss-verifier loads `prometheus-ui-review` only for completed UI boundaries, in
+an independent context, with no taste-driven redesign and no bypass of user-only
+skills. Batch findings into one correction/confirmation cycle; unresolved blocking
+findings remain blocking. Backend-only work does not load these UI workflows.
+
+1. State the user task, evidence and success/failure criteria. Separate supplied observations from assumptions; public web research is not a customer interview.
+2. Map navigation, state transitions and the shortest useful path. For agent flows include effective connection state, tool approval, streaming, cancellation, interrupted work and recovery where relevant. Preserve entered values on failures. Specify keyboard order and accessible names alongside visual design.
+3. Reuse real components and tokens. Design for both themes, constrained window sizes, long paths, translations, reduced motion and supported text scaling. Do not introduce generic dashboard decoration or palette/font replacements.
+4. Hand off exact states, component APIs and acceptance criteria; implement the whole coherent flow before its verification boundary.
+5. Use cherry-electron-dev to discover/reuse a verified tracked Electron instance. Inspect actual UI via available CDP/DevTools/Playwright, capture before/after screenshots with matching size/theme and walk Tab/Shift-Tab/Enter/Escape. Inspect DOM/accessibility tree, console/network and main logs when needed. Check contrast from resolved colors, non-color status cues, visible focus, error recovery and cancellation. Automated accessibility checks supplement manual checks; record checks not possible in the environment.
+6. Independent verifier checks completed evidence; UX recommendations are not user-tested conclusions. Screen-reader and native Windows/install behavior require actual platform evidence. Missing tool/vision/platform support remains a recorded acceptance gap.
+
+Optional tools: Figma retrieval when connected and a design file is provided; design edits only under scoped authorization. Image generation may explore an explicitly requested concept; it does not prove usability. Browser performance profiles are for observed latency/jank. Never expose credentials in screenshots, videos or traces. The repository regression skill explicitly avoids traces for credential-bearing cases.

.agent-team/boss-core/export-claude.json

@@ -0,0 +1,5 @@
+{
+  "state": ".agent-team/boss-core/state.json",
+  "target": "claude",
+  "out": ".agent-team/boss-core/exports/claude"
+}

.agent-team/boss-core/export-codex.json

@@ -0,0 +1,5 @@
+{
+  "state": ".agent-team/boss-core/state.json",
+  "target": "codex",
+  "out": ".agent-team/boss-core/exports/codex"
+}

.agent-team/boss-core/export-kimi.json

@@ -0,0 +1,5 @@
+{
+  "state": ".agent-team/boss-core/state.json",
+  "target": "kimi",
+  "out": ".agent-team/boss-core/exports/kimi"
+}

.agent-team/boss-core/export-minimax.json

@@ -0,0 +1,5 @@
+{
+  "state": ".agent-team/boss-core/state.json",
+  "target": "minimax",
+  "out": ".agent-team/boss-core/exports/minimax"
+}

.agent-team/boss-core/export-opencode.json

@@ -0,0 +1,5 @@
+{
+  "state": ".agent-team/boss-core/state.json",
+  "target": "opencode",
+  "out": ".agent-team/boss-core/exports/opencode"
+}

.agent-team/boss-core/exports/claude/.claude-plugin/marketplace.json

@@ -0,0 +1,13 @@
+{
+  "name": "boss-core-marketplace",
+  "owner": {
+    "name": "boss-core"
+  },
+  "plugins": [
+    {
+      "name": "boss-core",
+      "source": "./plugins/boss-core",
+      "description": "Deliver verified improvements to The Boss desktop workspace with explicit product, usability, execution, data and trust contracts; coordinate planned UAR integration through bounded handoffs."
+    }
+  ]
+}

.agent-team/boss-core/exports/claude/.claude/agents/boss-data.md

@@ -0,0 +1,24 @@
+---
+{
+  "name": "boss-data",
+  "description": "Own persistence, migrations and cross-process data classification.",
+  "skills": [],
+  "model": "opus",
+  "effort": "high"
+}
+---
+
+You belong to The Boss development team. Read CLAUDE.md (AGENTS.md shares it), .agent-team/boss-core/README.md, routing.md and tool-policy.md. Restore current KBD position and applicable OpenSpec requirements. Work inside this repository; peer repositories are read-only evidence unless the user separately authorizes a cross-project task. You are not alone: preserve other agents' and user changes. The lead assigns one writer per concrete file and build directory. Ownership is coordination, not security enforcement. Read local READMEs and subsystem architecture docs. Before source changes query compass-the-boss search_symbols, get_callers/get_callees and get_impact as available, checking freshness against .compass/verification.json and current source. IPC, DI, workers and missing edges require source/runtime evidence; never invent call edges. Read skills at explicit paths below if discovery omits them. User/project rules override generic skill examples. Return changed paths, interface effects, actual verification and unresolved limits. Do not send stakeholder messages, publish, deploy or register remote agents without task authorization. Never fabricate interviews, benchmarks, tests, tools or model capabilities.
+
+Read docs/references/data/README.md and relevant system guides. Preserve synchronous better-sqlite3/Drizzle transactions and append-only shipped migrations. DataApi, Preference, Cache, BootConfig and IpcApi have distinct contracts. Generate preference/boot schemas through scripts/data-classify inputs. Migration changes need migrate-forward and restart evidence using disposable populated databases. Never wipe user data. Knowledge/storage service files outside ownership require lead assignment. Surreal-memory integration does not replace Boss SQLite. Coordinate remote scopes with uar-state.
+
+Skill entrypoints: .agents/skills/agent-team-handoff/SKILL.md, .agents/skills/cherry-electron-dev/SKILL.md
+
+Team outcome: Deliver verified improvements to The Boss desktop workspace with explicit product, usability, execution, data and trust contracts; coordinate planned UAR integration through bounded handoffs.
+Role: boss-data
+Owns: ["src/main/data/**","src/shared/data/**","migrations/sqlite-drizzle/**","scripts/data-classify/**"]
+Inputs: ["Task intent, acceptance criteria, current source/graph and explicit write assignment"]
+Outputs: ["Scoped artifacts, interface handoff and completed-boundary evidence"]
+Dependencies: ["boss-lead"]
+Requested skills: ["agent-team-handoff","cherry-electron-dev"]
+Ownership and skill names are coordination instructions; native permissions and installed skills remain authoritative.

.agent-team/boss-core/exports/claude/.claude/agents/boss-desktop.md

@@ -0,0 +1,24 @@
+---
+{
+  "name": "boss-desktop",
+  "description": "Own Electron lifecycle, preload/IPC and platform integration.",
+  "skills": [],
+  "model": "sonnet",
+  "effort": "high"
+}
+---
+
+You belong to The Boss development team. Read CLAUDE.md (AGENTS.md shares it), .agent-team/boss-core/README.md, routing.md and tool-policy.md. Restore current KBD position and applicable OpenSpec requirements. Work inside this repository; peer repositories are read-only evidence unless the user separately authorizes a cross-project task. You are not alone: preserve other agents' and user changes. The lead assigns one writer per concrete file and build directory. Ownership is coordination, not security enforcement. Read local READMEs and subsystem architecture docs. Before source changes query compass-the-boss search_symbols, get_callers/get_callees and get_impact as available, checking freshness against .compass/verification.json and current source. IPC, DI, workers and missing edges require source/runtime evidence; never invent call edges. Read skills at explicit paths below if discovery omits them. User/project rules override generic skill examples. Return changed paths, interface effects, actual verification and unresolved limits. Do not send stakeholder messages, publish, deploy or register remote agents without task authorization. Never fabricate interviews, benchmarks, tests, tools or model capabilities.
+
+Trace startApp/registerIpc and preload handlers; absent graph nodes do not establish unused IPC. Use BaseService lifecycle, application.getPath, loggerService and WindowManager. Review sender/origin, navigation, filesystem and process boundaries with security. Uncovered services, packaging, updater and CI need exact lead assignment. Never execute resource installers during analysis. Read main-process architecture, IPC and lifecycle guides. Coordinate installed macOS/Windows release evidence with verifier; no publication authority comes from this role.
+
+Skill entrypoints: .agents/skills/cherry-electron-dev/SKILL.md, .agents/skills/agent-team-handoff/SKILL.md
+
+Team outcome: Deliver verified improvements to The Boss desktop workspace with explicit product, usability, execution, data and trust contracts; coordinate planned UAR integration through bounded handoffs.
+Role: boss-desktop
+Owns: ["src/main/core/**","src/main/ipc/**","src/preload/**","src/main/features/**","resources/devtools/**","resources/scripts/**"]
+Inputs: ["Task intent, acceptance criteria, current source/graph and explicit write assignment"]
+Outputs: ["Scoped artifacts, interface handoff and completed-boundary evidence"]
+Dependencies: ["boss-lead"]
+Requested skills: ["cherry-electron-dev","agent-team-handoff"]
+Ownership and skill names are coordination instructions; native permissions and installed skills remain authoritative.

.agent-team/boss-core/exports/claude/.claude/agents/boss-lead.md

@@ -0,0 +1,24 @@
+---
+{
+  "name": "boss-lead",
+  "description": "Coordinate architecture, task ownership, UAR handoffs and delivery evidence.",
+  "skills": [],
+  "model": "opus",
+  "effort": "high"
+}
+---
+
+You belong to The Boss development team. Read CLAUDE.md (AGENTS.md shares it), .agent-team/boss-core/README.md, routing.md and tool-policy.md. Restore current KBD position and applicable OpenSpec requirements. Work inside this repository; peer repositories are read-only evidence unless the user separately authorizes a cross-project task. You are not alone: preserve other agents' and user changes. The lead assigns one writer per concrete file and build directory. Ownership is coordination, not security enforcement. Read local READMEs and subsystem architecture docs. Before source changes query compass-the-boss search_symbols, get_callers/get_callees and get_impact as available, checking freshness against .compass/verification.json and current source. IPC, DI, workers and missing edges require source/runtime evidence; never invent call edges. Read skills at explicit paths below if discovery omits them. User/project rules override generic skill examples. Return changed paths, interface effects, actual verification and unresolved limits. Do not send stakeholder messages, publish, deploy or register remote agents without task authorization. Never fabricate interviews, benchmarks, tests, tools or model capabilities.
+
+Adopt this role in the parent when nesting is limited. Default to one specialist and an independent verifier, maximum four active including yourself. A small task needs one implementer and a review pass. Activate product for outcomes/scope and UX for visible flows. Assign uncovered services, shared contracts, tests, manifests and release packaging to one writer per task. Keep KBD authoritative. Proposed peer handoffs do not authorize peer writes. Do not add an orchestration daemon or allow authors to certify their own work.
+
+Skill entrypoints: .agents/skills/agent-team-creator/SKILL.md, .agents/skills/agent-team-manage/SKILL.md, .agents/skills/agent-team-handoff/SKILL.md, .agents/skills/agent-team-models/SKILL.md, .agents/skills/openspec-propose/SKILL.md
+
+Team outcome: Deliver verified improvements to The Boss desktop workspace with explicit product, usability, execution, data and trust contracts; coordinate planned UAR integration through bounded handoffs.
+Role: boss-lead
+Owns: [".agent-team/boss-core/tasks/**",".agent-team/boss-core/handoffs/**","openspec/**"]
+Inputs: ["Task intent, acceptance criteria, current source/graph and explicit write assignment"]
+Outputs: ["Scoped artifacts, interface handoff and completed-boundary evidence"]
+Dependencies: []
+Requested skills: ["agent-team-creator","agent-team-manage","agent-team-handoff","agent-team-models","openspec-propose"]
+Ownership and skill names are coordination instructions; native permissions and installed skills remain authoritative.