chore(prometheus-skills): bundle agent team skills

resources/skills/agent-team-creator/SKILL.md

@@ -0,0 +1,115 @@
+---
+name: agent-team-creator
+description: "Create the smallest useful coding agent team for a task, with guided role discovery and native exports for UAR, Codex, Claude Code, Copilot, Kimi Code, MiniMax CLI, OpenCode and DeepSeek. Use when a user asks to create, configure or choose an agent team; use agent-team-manage for existing task state. Do not use for existing task updates (see agent-team-manage)."
+license: MIT
+compatibility: Requires Node.js 22 or newer. Git is optional for handoff snapshots. Model gateways, memory services and native harness CLIs are optional and separately configured.
+metadata:
+  version: "1.0.0"
+  tags: "agents, teams, orchestration, coding"
+---
+
+# Agent Team Creator
+
+Help the user choose a useful team, then create inspectable definitions. A role
+describes responsibility; a skill provides reusable instructions; a model supplies
+capability; a harness owns execution. Do not turn these into one interchangeable
+concept or start a second agent loop.
+
+## Start with the task
+
+Read project instructions and any active KBD work first. Reuse answers already
+given. Ask only for missing outcome, scope, deliverables, budget preference and
+review needs. Use ordinary language: “Is this one isolated change, or does it
+span design, implementation and verification?” The user need not know agent
+terminology. Choose a short team ID and explain it rather than requiring jargon.
+
+Use [assets/intake.json](assets/intake.json) as the JSON request shape. `guide`
+returns missing intake fields when answers are incomplete, or a proposed team
+with role explanations and a single-agent alternative. Once task details are known,
+the guide returns `proposedRoles` and ownership questions until every proposed role
+has output paths in `ownership: {"role-id": ["relative/path/**"]}`. Inspect the
+project, suggest concrete disjoint paths, and reuse known scope. Reviewers can read
+broadly but need a separate findings output path. Only `ready: true` returns `team`:
+
+```text
+node <this-skill>/scripts/cli.mjs guide --input intake.json
+```
+
+Recommend one implementer for a small isolated task. Add specialists only for
+concrete work that can be assigned separately. Independent review costs another
+pass; explain that tradeoff. Let the user refine roles using their existing
+authorization and preferences. Do not require a ritual confirmation for every
+reversible file creation. Experts may supply a manifest directly.
+
+The guide’s skill IDs are suggestions, not assertions that anything is installed.
+Discover actual installed skills or use an available skills directory/search.
+Inspect each chosen skill’s provenance, scope, tools and instructions. Replace
+unavailable suggestions, or propose installation when needed; do not silently
+install external code. Bind discovered skills to each role’s `skills` array.
+
+## Define and export
+
+Save the proposed `team` from guide output. Fill `owns`, inputs, outputs and
+dependencies for each role before parallel edits. Empty ownership is unresolved,
+not permission over the repository. The [manifest reference](references/manifest.md)
+and [JSON schema](schemas/team.schema.json) define the common contract.
+
+Select models with `$agent-team-models`: declared strength, required capabilities,
+and known prices, followed by concrete IDs. Tier labels alone do not configure a
+native model. Every native option can be carried in `native.<target>.options`,
+role-native overrides, or exact `native.<target>.files`. Read
+[native harness contracts](references/native-harnesses.md) for the selected
+target’s mapping, source, plugin support and limitations. Do not load every
+harness manual for a single-target task.
+
+```text
+node <this-skill>/scripts/cli.mjs validate --input team-request.json
+node <this-skill>/scripts/cli.mjs init --input team-request.json
+node <this-skill>/scripts/cli.mjs export --input export-request.json
+```
+
+`team-request.json` contains `{"team": <manifest>, "state": <local-state-path>}`.
+`export-request.json` contains `{"state": <local-state-path>, "target": "codex",
+"out": <new-proposal-directory>}`. Targets are `uar`, `codex`, `claude`,
+`copilot`, `kimi`, `minimax`, `opencode`, `deepseek`, and the separate `bossfang`
+integration. MiniMax means its own `mcode` CLI.
+
+Export never overwrites an existing output directory or native configuration.
+Inspect `team-export.json`, diagnostics, native files and the source/version
+receipt. Preservation of arbitrary options is not semantic validation. Validate
+with the installed harness when available, or report source-only support.
+
+## Install and operate within the requested scope
+
+Apply reviewed proposals only where the user authorized them. Merge existing
+native configuration deliberately; never replace it wholesale. Follow the
+native reference for supported project agents or plugin/marketplace installation.
+Do not invent plugin agent fields where a harness has none. A plugin installation
+does not start an agent team.
+
+For UAR or BossFang, keep registry registration, activation and execution separate.
+Use the exported native payload and verified route, operator-selected instance
+URL and environment credential reference. Record each returned native ID and
+outcome; multi-agent registration is not atomic. BossFang Hands and standalone
+agent/workflow registration are alternative native deployment paths. Choose one.
+Do not auto-activate a Hand or run a workflow merely because definitions exist.
+
+For actual work, the chosen harness owns native spawning, permissions, sessions,
+subagent depth and model overrides. Use its available tools/current CLI contract;
+the team ledger does not schedule processes or enforce Cedar. Resolve missing
+native capabilities explicitly. Use `$agent-team-manage` for tasks and
+`$agent-team-handoff` when changing owners/harnesses.
+
+## Evidence and recovery
+
+Report generated paths, roles and rationale, chosen model policy, native support
+level, unresolved ownership/configuration and the next authorized action. Never
+report export as live execution. Node 22+ runs the compiled package without a
+root checkout or runtime dependencies. Maintainers rebuild the `.mts` source
+with pinned TypeScript 7.0.2 under `runtime/`; full and mini ship identical bytes.
+
+The local state is a coordination record for trusted collaborators, not a
+distributed authorization service. See [task and handoff contracts](references/task-handoff.md)
+for revision/lock recovery and [models and memory](references/models-memory.md)
+for optional shared services and Karpathy boundaries. Existing KBD state stays
+authoritative; never hand-edit its generated projections.

resources/skills/agent-team-creator/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Agent Team Creator"
+  short_description: "Choose and create a useful coding agent team"
+  default_prompt: "Use $agent-team-creator to help me choose and create a team for my task."

resources/skills/agent-team-creator/assets/intake.json

@@ -0,0 +1,11 @@
+{
+  "id": "checkout-team",
+  "outcome": "Implement an accessible checkout with verified payment error handling",
+  "complexity": "complex",
+  "areas": ["code", "design", "security", "docs"],
+  "deliverables": ["Checkout implementation", "Integration evidence", "Operator documentation"],
+  "budget": "balanced",
+  "review": true,
+  "harness": "codex",
+  "scope": "project"
+}

resources/skills/agent-team-creator/references/manifest.md

@@ -0,0 +1,146 @@
+# Team manifest and CLI
+
+The portable manifest describes intent and coordination. It is not a native team
+API, a scheduler, an authentication credential, or a Cedar policy. The native
+harness remains the execution and permission authority.
+
+## Minimum expert request
+
+Save this as `team-request.json` in the target project. Paths in requests resolve
+against the CLI working directory. Keep state in a local, access-controlled
+directory; it contains prompts, evidence and optional memory content.
+
+```json
+{
+  "state": ".agent-team/state.json",
+  "team": {
+    "schemaVersion": 1,
+    "id": "feature-team",
+    "outcome": "Implement the agreed feature and supply integration evidence",
+    "scope": "project",
+    "harness": "codex",
+    "roles": [
+      {
+        "id": "implementer",
+        "description": "Implement the agreed feature",
+        "prompt": "Read the project instructions and acceptance criteria. Implement the complete behavior before verification. Report changed files and evidence.",
+        "skills": [],
+        "owns": ["src/feature/"],
+        "inputs": ["Acceptance criteria"],
+        "outputs": ["Implementation", "Integration evidence"],
+        "dependsOn": [],
+        "modelPolicy": { "tier": "medium", "capabilities": ["function_calling"] }
+      }
+    ]
+  }
+}
+```
+
+`scope` chooses project, UAR or BossFang administration intent. It does not select
+a network instance or transfer execution authority. `harness` is one of the eight
+execution targets. BossFang is a separate export target; it may coordinate UAR or
+its own native runtime as configured outside this skill.
+
+Role IDs are stable portable lowercase identifiers. Role inputs/outputs describe
+the handoff contract; `dependsOn` is validated for references and cycles. Native
+runtime dependency behavior varies; portable dependencies also gate local tasks.
+`owns` is a planning agreement, not an enforced filesystem permission. Assign
+disjoint write ownership before parallel work, and let reviewers read broadly.
+
+## Model binding
+
+Optional `modelPolicy` exists on the team and each role. `skillPolicies` maps
+installed skill IDs to the same policy shape. Tasks can provide `modelPolicy`.
+Fields: `model`, `tier` (`low`, `medium`, `hard`), `capabilities` (string array),
+`maxInputPerMillion`, `maxOutputPerMillion` (nonnegative USD estimates).
+
+Use `models-select` before invocation; tier/cost/capability intent is not itself
+a native model ID. The exporter translates explicit team/role `model` values
+where supported. Skill/task policy selection must be applied through the native
+invocation controls, or by updating an appropriate role and re-exporting.
+
+## Preserving every native option
+
+The common schema deliberately does not enumerate every native harness setting.
+Use the complete current native documentation linked from `native-harnesses.md`:
+
+- `roles[].native.<target>` overrides that role's generated native fields.
+- `native.<target>.options` supplies target-specific settings or defaults. Its
+  meaning is explicit in the adapter reference; some targets preserve it in
+  `native-options.json` for manual application rather than inventing a config path.
+- `native.<target>.files` maps relative paths to exact UTF-8 text for native hooks,
+  MCP settings, plugin source, profiles, configuration or other options.
+- Each team-native wrapper includes `source` and `version` recording the contract
+  the user chose. Native role overrides and the built-in source contract are
+  retained in the export receipt too.
+
+Example extension inside `team`:
+
+```json
+{
+  "native": {
+    "codex": {
+      "version": "installed version recorded by operator",
+      "source": "https://developers.openai.com/codex/config-reference/",
+      "options": { "agents": { "max_threads": 4 } },
+      "files": { ".codex/team-notes.md": "Project-specific native configuration notes\n" }
+    }
+  }
+}
+```
+
+Unknown common fields fail to prevent silently misspelled configuration. Native
+options are preserved without claiming their semantic validity. Objects merge
+recursively and arrays replace. JSON `null` cannot be serialized to TOML; omit
+the field or supply a separate native file. Native file paths cannot escape the
+proposal root, use Windows-reserved names, or collide with generated files,
+case variants or directories. Generated files cannot be silently replaced by
+opaque files. Change supported fields through role/options overrides, or author
+an independently reviewed native file outside the automatic export path.
+
+The generated receipt identifies source-only versus live verification. Validate
+against the installed version before applying options. Do not commit real
+credentials in any native file; use the target's environment/credential facility.
+This escape hatch preserves future options; it does not assert future support.
+
+## CLI protocol
+
+All commands use one UTF-8 JSON request and return JSON. Malformed requests and
+conflicts exit nonzero. No shell quoting or shell-specific redirection is needed:
+
+```text
+node <skill>/scripts/cli.mjs <command> --input request.json
+```
+
+| Command | Request |
+|---|---|
+| `guide` | Intake fields from `assets/intake.json`, followed by `ownership` mapping role IDs to nonempty arrays of project-relative output paths/globs. Missing scope returns questions and `proposedRoles`; only `ready: true` returns `team` |
+| `validate` | `team` manifest |
+| `init` | `team`, new `state` filename |
+| `status` | `state` |
+| `team-update` | `state`, `expectedRevision`, replacement same-ID `team` |
+| `export` | `team` or `state`, `target`, new `out` directory |
+| `task`, `complete-kbd` | See `task-handoff.md` |
+| `handoff-create`, `handoff-accept` | See `task-handoff.md` |
+| `models-discover`, `models-select` | See `models-memory.md` |
+| `memory-queue`, `memory-publish` | See `models-memory.md` |
+
+Exports are proposals, never in-place installation. If a write fails partway,
+the incomplete directory remains inspectable and lacks its final receipt. Choose
+a new output directory for a retry. No global tool installation or remote agent
+registration occurs through this CLI. Skills guide native installation and
+execution using the user's authorized target and the actual harness contract.
+
+## Building and distributing
+
+`runtime/package.json` pins TypeScript 7.0.2 and Node type declarations. Run
+`npm ci --prefix <skill>/runtime` then `npm run build --prefix <skill>/runtime`
+when maintaining source. Runtime consumers need only Node.js 22+ and the copied
+skill files; they do not need npm, TypeScript or the repository checkout. Full
+and mini distribute identical source and emitted `.mjs` files. Each sibling skill
+declares its dependency on this creator companion in its instructions.
+
+The four SKILL.md frontmatters follow the
+[AgentSkills specification](https://agentskills.io/specification): version and
+comma-separated tags live as string-valued metadata. Existing pack extensions
+remain accepted for backward compatibility but are not required by these skills.