feat(skills): add Agent Skills support with orchestrator-driven activation

cli/planoai/config_generator.py

@@ -1,6 +1,7 @@
 import json
 import os
 import uuid
+from pathlib import Path
 from planoai.utils import convert_legacy_listeners
 from jinja2 import Environment, FileSystemLoader
 import yaml
@@ -8,6 +9,14 @@ from jsonschema import validate, ValidationError
 from urllib.parse import urlparse
 from copy import deepcopy
 from planoai.consts import DEFAULT_OTEL_TRACING_GRPC_ENDPOINT
+from planoai.skills import (
+    MAX_CATALOG_BYTES,
+    Skill,
+    discover_skills,
+    find_project_root,
+    is_project_trusted,
+    total_catalog_size,
+)
 
 SUPPORTED_PROVIDERS_WITH_BASE_URL = [
     "azure_openai",
@@ -162,6 +171,127 @@ def _version_tuple(version_string):
     return tuple(out)
 
 
+def materialize_skills_in_config(config_yaml: dict, project_root: Path) -> None:
+    """Discover and inline Agent Skills referenced by `config_yaml`.
+
+    Mutates `config_yaml` in place. The user's source config may declare
+    `skills:` as a list of strings (skill names) or omit it entirely. After
+    this call, `config_yaml["skills"]` is either absent or a list of fully
+    materialized objects with `name`, `description`, `path`, `body`, etc.
+
+    Project-scope skills under `<project_root>/.plano/skills/` are only loaded
+    when the project has been marked trusted via `planoai skills trust`.
+
+    Per-route `routing_preferences[].skills` allow-lists are preserved as-is
+    so brightstaff can scope the catalog when that route is selected.
+    """
+    requested = config_yaml.get("skills")
+    user_only = not is_project_trusted(project_root)
+
+    discovered, diagnostics = discover_skills(
+        project_root=project_root, include_user_scope=True
+    )
+    for diag in diagnostics:
+        prefix = "error" if diag.severity == "error" else "warning"
+        print(f"[skills] {prefix}: {diag.path}: {diag.message}")
+
+    if user_only:
+        project_skills = [s for s in discovered if s.scope == "project"]
+        if project_skills:
+            print(
+                "[skills] note: project-scope skills are present but the project is "
+                "not trusted yet; run `planoai skills trust` to enable them."
+            )
+        # Keep all non-project scopes (user + agents) — both are user-tier and
+        # auto-trusted, so they always load regardless of project trust state.
+        discovered = [s for s in discovered if s.scope != "project"]
+
+    skills_by_name: dict[str, Skill] = {s.name: s for s in discovered}
+
+    if requested is None:
+        # Default: auto-include every discovered skill.
+        selected: list[Skill] = list(discovered)
+    else:
+        if not isinstance(requested, list):
+            raise Exception("`skills:` must be a list of strings or skill objects")
+        selected = []
+        seen: set[str] = set()
+        for entry in requested:
+            if isinstance(entry, str):
+                name = entry
+            elif isinstance(entry, dict):
+                name = entry.get("name")
+                if not isinstance(name, str):
+                    raise Exception(
+                        "skill entries with object form must include a string `name`"
+                    )
+            else:
+                raise Exception(
+                    f"unsupported entry in `skills:` (expected str or mapping, got {type(entry).__name__})"
+                )
+            if name in seen:
+                continue
+            seen.add(name)
+            skill = skills_by_name.get(name)
+            if skill is None:
+                print(
+                    f"[skills] warning: skill '{name}' is declared in config but no "
+                    f"SKILL.md was discovered under .plano/skills/ or ~/.plano/skills/"
+                )
+                continue
+            selected.append(skill)
+
+    if not selected:
+        config_yaml.pop("skills", None)
+        _strip_unknown_route_skills(config_yaml, set())
+        return
+
+    catalog_bytes = total_catalog_size(selected)
+    if catalog_bytes > MAX_CATALOG_BYTES:
+        print(
+            f"[skills] warning: skill catalog size is {catalog_bytes} bytes, "
+            f"above the recommended cap of {MAX_CATALOG_BYTES}. Consider trimming "
+            f"`routing_preferences[].skills` to the smallest useful set per route."
+        )
+
+    config_yaml["skills"] = [s.to_dict() for s in selected]
+    _strip_unknown_route_skills(config_yaml, {s.name for s in selected})
+
+
+def _strip_unknown_route_skills(config_yaml: dict, known: set) -> None:
+    """Drop unknown skill names from `routing_preferences[*].skills` allow-lists.
+
+    The orchestrator only ever sees skills referenced under some
+    `routing_preferences[].skills`; an unknown name there would render the
+    `<skills>` block with a stale entry the runtime can't resolve, so filter
+    them out here with a warning instead.
+    """
+    routes = config_yaml.get("routing_preferences")
+    if not isinstance(routes, list):
+        return
+    for route in routes:
+        if not isinstance(route, dict):
+            continue
+        allow = route.get("skills")
+        if not isinstance(allow, list):
+            continue
+        filtered = []
+        for name in allow:
+            if not isinstance(name, str):
+                continue
+            if name in known:
+                filtered.append(name)
+            else:
+                print(
+                    f"[skills] warning: routing_preference '{route.get('name')}' "
+                    f"references unknown skill '{name}'; dropping from allow-list."
+                )
+        if filtered:
+            route["skills"] = filtered
+        else:
+            route.pop("skills", None)
+
+
 def validate_and_render_schema():
     ENVOY_CONFIG_TEMPLATE_FILE = os.getenv(
         "ENVOY_CONFIG_TEMPLATE_FILE", "envoy.template.yaml"
@@ -196,6 +326,13 @@ def validate_and_render_schema():
     _ = yaml.safe_load(plano_config_schema)
     inferred_clusters = {}
 
+    # Materialize Agent Skills before further processing so the rest of the
+    # pipeline (Jinja2 envoy template, dump to plano_config_rendered.yaml) sees
+    # the inlined body / description / path.
+    plano_config_path = Path(PLANO_CONFIG_FILE).resolve()
+    project_root = find_project_root(plano_config_path.parent)
+    materialize_skills_in_config(config_yaml, project_root)
+
     # Convert legacy llm_providers to model_providers
     if "llm_providers" in config_yaml:
         if "model_providers" in config_yaml:

cli/planoai/init_cmd.py

@@ -64,6 +64,12 @@ BUILTIN_TEMPLATES: list[Template] = [
         description="stateful responses with memory-backed storage",
         yaml_text=_load_template_yaml("conversational_state_v1_responses.yaml"),
     ),
+    Template(
+        id="skills_routing",
+        title="Agent Skills Routing",
+        description="install Agent Skills (agentskills.io) and let Plano-Orchestrator route to them",
+        yaml_text=_load_template_yaml("skills_routing.yaml"),
+    ),
 ]
 
 

cli/planoai/main.py

@@ -39,6 +39,7 @@ from planoai.init_cmd import init as init_cmd
 from planoai.trace_cmd import trace as trace_cmd, start_trace_listener_background
 from planoai.chatgpt_cmd import chatgpt as chatgpt_cmd
 from planoai.obs_cmd import obs as obs_cmd
+from planoai.skills_cmd import skills as skills_cmd
 from planoai.consts import (
     DEFAULT_OTEL_TRACING_GRPC_ENDPOINT,
     DEFAULT_NATIVE_OTEL_TRACING_GRPC_ENDPOINT,
@@ -746,6 +747,7 @@ main.add_command(init_cmd, name="init")
 main.add_command(trace_cmd, name="trace")
 main.add_command(chatgpt_cmd, name="chatgpt")
 main.add_command(obs_cmd, name="obs")
+main.add_command(skills_cmd, name="skills")
 
 if __name__ == "__main__":
     main()

cli/planoai/skills.py

@@ -0,0 +1,420 @@
+"""Agent Skills discovery for Plano.
+
+Parses SKILL.md files from .plano/skills/ (project scope) and ~/.plano/skills/
+(user scope) following the Agent Skills specification:
+https://agentskills.io/specification.md
+
+The parser is intentionally lenient (per the "Adding skills support" guide):
+warn on cosmetic issues but only skip a skill when its YAML is unparseable or
+its required `description` field is missing.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Iterable
+
+import yaml
+
+from planoai.utils import getLogger
+
+log = getLogger(__name__)
+
+PROJECT_SKILLS_DIR = Path(".plano") / "skills"
+USER_SKILLS_DIR = Path(os.path.expanduser("~/.plano/skills"))
+# Universal Agent Skills install location used by `npx skills add` (vercel-labs/add-skill).
+# Auto-trusted: same security posture as ~/.plano/skills, no project trust needed.
+AGENTS_SKILLS_DIR = Path(os.path.expanduser("~/.agents/skills"))
+
+MAX_CATALOG_BYTES = 5 * 1024
+
+MAX_DIRS_SCANNED = 2000
+
+_NAME_PATTERN = re.compile(r"^[a-z0-9](?:[a-z0-9-]*[a-z0-9])?$")
+
+
+def trusted_projects_file() -> Path:
+    """Resolve `~/.plano/trusted_projects.json` at call time.
+
+    Lazy so tests can override $HOME and have the new path picked up; module
+    import time would freeze it to the developer's actual home directory.
+    """
+    return Path(os.path.expanduser("~/.plano/trusted_projects.json"))
+
+
+def is_project_trusted(project_root: Path) -> bool:
+    """Return True if `project_root` is listed in `~/.plano/trusted_projects.json`.
+
+    Project-scope skills come from arbitrary repos and are gated on this trust
+    decision (set with `planoai skills trust`). Single source of truth, shared
+    between the `skills_cmd` CLI surface and the render pipeline.
+    """
+    path = trusted_projects_file()
+    if not path.exists():
+        return False
+    try:
+        with path.open("r", encoding="utf-8") as fh:
+            data = json.load(fh)
+    except (OSError, json.JSONDecodeError):
+        return False
+    trusted = data.get("trusted_projects", []) if isinstance(data, dict) else []
+    resolved = str(project_root.resolve())
+    return resolved in {str(Path(p).resolve()) for p in trusted}
+
+
+@dataclass(frozen=True)
+class SkillDiagnostic:
+    severity: str  # "warn" or "error"
+    message: str
+    path: Path
+
+
+@dataclass
+class Skill:
+    name: str
+    description: str
+    location: Path
+    base_dir: Path
+    body: str
+    scope: str
+    compatibility: str | None = None
+    license: str | None = None
+    metadata: dict = field(default_factory=dict)
+    allowed_tools: str | None = None
+
+    def to_dict(self) -> dict:
+        """Serialize to a YAML-friendly dict for embedding in rendered config."""
+        return {
+            "name": self.name,
+            "description": self.description,
+            "path": str(self.location),
+            "base_dir": str(self.base_dir),
+            "scope": self.scope,
+            "body": self.body,
+            "compatibility": self.compatibility,
+            "license": self.license,
+            "metadata": dict(self.metadata) if self.metadata else None,
+            "allowed_tools": self.allowed_tools,
+        }
+
+
+def find_project_root(start: Path | None = None) -> Path:
+    """Walk up from `start` looking for `.plano/`, then `.git/`.
+
+    Falls back to `start` (or cwd) if nothing is found. This matches how
+    `npx skills add` chooses a project root.
+    """
+    base = Path(start or Path.cwd()).resolve()
+    cur = base
+    while cur != cur.parent:
+        if (cur / ".plano").exists():
+            return cur
+        cur = cur.parent
+
+    cur = base
+    while cur != cur.parent:
+        if (cur / ".git").exists():
+            return cur
+        cur = cur.parent
+
+    return base
+
+
+def parse_skill_md(path: Path) -> tuple[Skill | None, list[SkillDiagnostic]]:
+    """Parse a single SKILL.md file leniently."""
+    diagnostics: list[SkillDiagnostic] = []
+    try:
+        text = path.read_text(encoding="utf-8")
+    except OSError as exc:
+        diagnostics.append(
+            SkillDiagnostic("error", f"failed to read SKILL.md: {exc}", path)
+        )
+        return None, diagnostics
+
+    frontmatter, body = _split_frontmatter(text)
+    if frontmatter is None:
+        diagnostics.append(SkillDiagnostic("error", "missing YAML frontmatter", path))
+        return None, diagnostics
+
+    data = _parse_yaml_lenient(frontmatter, path, diagnostics)
+    if data is None:
+        return None, diagnostics
+
+    description = data.get("description")
+    if not isinstance(description, str) or not description.strip():
+        diagnostics.append(
+            SkillDiagnostic(
+                "error", "skill is missing required 'description' field", path
+            )
+        )
+        return None, diagnostics
+
+    parent_name = path.parent.name
+    name = data.get("name")
+    if not isinstance(name, str) or not name.strip():
+        diagnostics.append(
+            SkillDiagnostic(
+                "warn",
+                f"missing 'name' field; falling back to parent directory '{parent_name}'",
+                path,
+            )
+        )
+        name = parent_name
+
+    name = name.strip()
+
+    if len(name) > 64:
+        diagnostics.append(
+            SkillDiagnostic("warn", "skill name exceeds 64 characters", path)
+        )
+
+    if not _NAME_PATTERN.match(name):
+        diagnostics.append(
+            SkillDiagnostic(
+                "warn",
+                f"skill name '{name}' violates spec naming rules "
+                "(lowercase alphanumeric + hyphens, no leading/trailing/double hyphens)",
+                path,
+            )
+        )
+
+    if name != parent_name:
+        diagnostics.append(
+            SkillDiagnostic(
+                "warn",
+                f"skill name '{name}' does not match parent directory '{parent_name}'",
+                path,
+            )
+        )
+
+    metadata_raw = data.get("metadata")
+    metadata = {}
+    if isinstance(metadata_raw, dict):
+        metadata = {str(k): str(v) for k, v in metadata_raw.items()}
+
+    skill = Skill(
+        name=name,
+        description=description.strip(),
+        location=path.resolve(),
+        base_dir=path.parent.resolve(),
+        body=body,
+        scope="project",  # may be overridden by caller
+        compatibility=_string_field(data.get("compatibility")),
+        license=_string_field(data.get("license")),
+        metadata=metadata,
+        allowed_tools=_string_field(data.get("allowed-tools")),
+    )
+    return skill, diagnostics
+
+
+def _split_frontmatter(text: str) -> tuple[str | None, str]:
+    if not text.startswith("---"):
+        return None, text
+
+    m = re.match(r"^---\s*\r?\n(.*?)\r?\n---\s*(?:\r?\n)?(.*)$", text, re.DOTALL)
+    if not m:
+        return None, text
+    return m.group(1), m.group(2).strip("\n")
+
+
+def _parse_yaml_lenient(
+    frontmatter: str, path: Path, diagnostics: list[SkillDiagnostic]
+) -> dict | None:
+    try:
+        data = yaml.safe_load(frontmatter)
+    except yaml.YAMLError as exc:
+        retried = _retry_quote_problem_fields(frontmatter)
+        if retried is None:
+            diagnostics.append(
+                SkillDiagnostic("error", f"YAML parse error: {exc}", path)
+            )
+            return None
+        try:
+            data = yaml.safe_load(retried)
+        except yaml.YAMLError as exc2:
+            diagnostics.append(
+                SkillDiagnostic(
+                    "error", f"YAML parse error (after retry): {exc2}", path
+                )
+            )
+            return None
+
+    if not isinstance(data, dict):
+        diagnostics.append(
+            SkillDiagnostic("error", "frontmatter is not a YAML mapping", path)
+        )
+        return None
+    return data
+
+
+_PROBLEM_FIELDS = ("description", "compatibility")
+
+
+def _retry_quote_problem_fields(frontmatter: str) -> str | None:
+    """Wrap unquoted values for fields prone to YAML colon-collisions in quotes."""
+    lines = frontmatter.splitlines()
+    out: list[str] = []
+    changed = False
+    for line in lines:
+        m = re.match(r"^(\w[\w-]*)\s*:\s*(.*)$", line)
+        if m and m.group(1) in _PROBLEM_FIELDS:
+            key = m.group(1)
+            value = m.group(2).rstrip()
+            if value and not (
+                (value.startswith("'") and value.endswith("'"))
+                or (value.startswith('"') and value.endswith('"'))
+            ):
+                escaped = value.replace("\\", "\\\\").replace('"', '\\"')
+                out.append(f'{key}: "{escaped}"')
+                changed = True
+                continue
+        out.append(line)
+    if not changed:
+        return None
+    return "\n".join(out)
+
+
+def _string_field(value) -> str | None:
+    if value is None:
+        return None
+    s = str(value).strip()
+    return s or None
+
+
+def _iter_skill_dirs(root: Path) -> Iterable[Path]:
+    if not root.exists() or not root.is_dir():
+        return
+
+    try:
+        children = sorted(root.iterdir(), key=lambda p: p.name)
+    except OSError:
+        return
+
+    count = 0
+    for child in children:
+        count += 1
+        if count > MAX_DIRS_SCANNED:
+            log.warning(
+                "exceeded max scan budget (%d) while looking for skills in %s",
+                MAX_DIRS_SCANNED,
+                root,
+            )
+            break
+        if not child.is_dir():
+            continue
+        if child.name.startswith("."):
+            continue
+        yield child
+
+
+def discover_skills(
+    project_root: Path | None = None,
+    include_user_scope: bool = True,
+) -> tuple[list[Skill], list[SkillDiagnostic]]:
+    """Discover all skills available to the current project.
+
+    Precedence (highest first): project > user > agents. Project-scope
+    skills shadow lower tiers with the same name; user-scope shadows
+    agents-scope. Both ``~/.plano/skills/`` (Plano-native) and
+    ``~/.agents/skills/`` (the universal Agent Skills install location used
+    by ``npx skills add``) are treated as auto-trusted user-tier scopes.
+
+    Returns ``(skills, diagnostics)`` sorted by name.
+    """
+    project_root = find_project_root(project_root)
+    project_dir = project_root / PROJECT_SKILLS_DIR
+
+    skills_by_name: dict[str, Skill] = {}
+    diagnostics: list[SkillDiagnostic] = []
+
+    if include_user_scope:
+        # Load lowest precedence first so higher tiers shadow.
+        for skill_dir in _iter_skill_dirs(AGENTS_SKILLS_DIR):
+            skill_md = skill_dir / "SKILL.md"
+            if not skill_md.exists():
+                continue
+            skill, diags = parse_skill_md(skill_md)
+            diagnostics.extend(diags)
+            if skill is not None:
+                skill = _set_scope(skill, "agents")
+                skills_by_name[skill.name] = skill
+
+        for skill_dir in _iter_skill_dirs(USER_SKILLS_DIR):
+            skill_md = skill_dir / "SKILL.md"
+            if not skill_md.exists():
+                continue
+            skill, diags = parse_skill_md(skill_md)
+            diagnostics.extend(diags)
+            if skill is None:
+                continue
+            skill = _set_scope(skill, "user")
+            existing = skills_by_name.get(skill.name)
+            if existing is not None and existing.scope == "agents":
+                diagnostics.append(
+                    SkillDiagnostic(
+                        "warn",
+                        f"user-scope skill '{skill.name}' shadows ~/.agents/skills entry at {existing.location}",
+                        skill.location,
+                    )
+                )
+            skills_by_name[skill.name] = skill
+
+    for skill_dir in _iter_skill_dirs(project_dir):
+        skill_md = skill_dir / "SKILL.md"
+        if not skill_md.exists():
+            continue
+        skill, diags = parse_skill_md(skill_md)
+        diagnostics.extend(diags)
+        if skill is None:
+            continue
+        skill = _set_scope(skill, "project")
+        existing = skills_by_name.get(skill.name)
+        if existing is not None and existing.scope in ("user", "agents"):
+            diagnostics.append(
+                SkillDiagnostic(
+                    "warn",
+                    f"project-scope skill '{skill.name}' shadows {existing.scope}-scope skill at {existing.location}",
+                    skill.location,
+                )
+            )
+        skills_by_name[skill.name] = skill
+
+    return sorted(skills_by_name.values(), key=lambda s: s.name), diagnostics
+
+
+def _set_scope(skill: Skill, scope: str) -> Skill:
+    return Skill(
+        name=skill.name,
+        description=skill.description,
+        location=skill.location,
+        base_dir=skill.base_dir,
+        body=skill.body,
+        scope=scope,
+        compatibility=skill.compatibility,
+        license=skill.license,
+        metadata=skill.metadata,
+        allowed_tools=skill.allowed_tools,
+    )
+
+
+def total_catalog_size(skills: Iterable[Skill]) -> int:
+    """Approximate byte size of the catalog the orchestrator will receive."""
+    return sum(len(s.name) + len(s.description) for s in skills)
+
+
+def filter_skills_by_allow_list(
+    skills: Iterable[Skill], allow_list: Iterable[str] | None
+) -> list[Skill]:
+    """Filter skills to those whose `name` appears in `allow_list`.
+
+    If `allow_list` is None, returns all skills. Unknown names are silently
+    dropped — callers warn at config-validation time.
+    """
+    if allow_list is None:
+        return list(skills)
+    allowed = set(allow_list)
+    return [s for s in skills if s.name in allowed]

cli/planoai/skills_cmd.py

@@ -0,0 +1,471 @@
+"""`planoai skills` command group.
+
+Installs Agent Skills (https://agentskills.io) and surfaces them to Plano.
+
+Three discovery scopes are supported, in descending precedence:
+
+* ``<project>/.plano/skills/`` -- repo-pinned skills. Loaded only when the
+  project has been marked trusted via ``planoai skills trust`` (skill content
+  is injected into the orchestrator prompt, so we gate on trust).
+* ``~/.plano/skills/`` -- Plano-native user-scope. Always trusted.
+* ``~/.agents/skills/`` -- universal Agent Skills location used by
+  ``npx skills add``. Always trusted; lets the upstream skills CLI work
+  out of the box without any Plano-specific awareness.
+
+``planoai skills add`` tries ``npx skills add`` first (the upstream CLI from
+https://github.com/vercel-labs/add-skill), which writes to
+``~/.agents/skills/<name>`` and is picked up automatically thanks to the
+agents-scope above. Falls back to ``git clone`` into ``.plano/skills/`` when
+``npx`` is unavailable.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+import shutil
+import subprocess
+import sys
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from pathlib import Path
+
+import rich_click as click
+from rich.console import Console
+from rich.table import Table
+
+from planoai.consts import PLANO_COLOR
+from planoai.skills import (
+    PROJECT_SKILLS_DIR,
+    Skill,
+    discover_skills,
+    find_project_root,
+    is_project_trusted,
+    trusted_projects_file,
+)
+from planoai.utils import getLogger
+
+log = getLogger(__name__)
+
+_OWNER_REPO_PATTERN = re.compile(r"^[A-Za-z0-9_.-]+/[A-Za-z0-9_.-]+$")
+
+
+@dataclass
+class _InstallTarget:
+    owner: str
+    repo: str
+    ref: str | None = None  # optional branch / tag / commit (e.g. "owner/repo@v1")
+
+    @property
+    def slug(self) -> str:
+        return f"{self.owner}/{self.repo}"
+
+    @property
+    def url(self) -> str:
+        return f"https://github.com/{self.owner}/{self.repo}.git"
+
+
+def _console() -> Console:
+    return Console()
+
+
+def _ensure_skills_dir(project_root: Path) -> Path:
+    skills_dir = project_root / PROJECT_SKILLS_DIR
+    skills_dir.mkdir(parents=True, exist_ok=True)
+    return skills_dir
+
+
+def _parse_install_target(raw: str) -> _InstallTarget:
+    spec = raw.strip()
+    ref: str | None = None
+    if "@" in spec:
+        spec, _, ref_value = spec.partition("@")
+        ref = ref_value.strip() or None
+    if not _OWNER_REPO_PATTERN.match(spec):
+        raise click.BadParameter(
+            f"expected '<owner>/<repo>' (optionally suffixed with '@<ref>'), got '{raw}'"
+        )
+    owner, repo = spec.split("/", 1)
+    return _InstallTarget(owner=owner, repo=repo, ref=ref)
+
+
+def _has_npx() -> bool:
+    return shutil.which("npx") is not None
+
+
+def _has_git() -> bool:
+    return shutil.which("git") is not None
+
+
+def _mark_project_trusted(project_root: Path) -> None:
+    path = trusted_projects_file()
+    path.parent.mkdir(parents=True, exist_ok=True)
+    existing: dict = {}
+    if path.exists():
+        try:
+            with path.open("r", encoding="utf-8") as fh:
+                existing = json.load(fh) or {}
+        except (OSError, json.JSONDecodeError):
+            existing = {}
+    trusted = set(existing.get("trusted_projects", []) or [])
+    trusted.add(str(project_root.resolve()))
+    existing["trusted_projects"] = sorted(trusted)
+    with path.open("w", encoding="utf-8") as fh:
+        json.dump(existing, fh, indent=2)
+
+
+def _read_manifest(skills_dir: Path) -> dict:
+    manifest_path = skills_dir / ".skills.json"
+    if not manifest_path.exists():
+        return {"skills": {}}
+    try:
+        with manifest_path.open("r", encoding="utf-8") as fh:
+            data = json.load(fh)
+    except (OSError, json.JSONDecodeError):
+        return {"skills": {}}
+    if not isinstance(data, dict):
+        return {"skills": {}}
+    data.setdefault("skills", {})
+    return data
+
+
+def _write_manifest(skills_dir: Path, manifest: dict) -> None:
+    manifest_path = skills_dir / ".skills.json"
+    with manifest_path.open("w", encoding="utf-8") as fh:
+        json.dump(manifest, fh, indent=2, sort_keys=True)
+
+
+def _record_install(
+    skills_dir: Path, name: str, target: _InstallTarget, source: str
+) -> None:
+    manifest = _read_manifest(skills_dir)
+    manifest["skills"][name] = {
+        "source": source,
+        "repo": target.slug,
+        "ref": target.ref,
+        "installed_at": datetime.now(timezone.utc).isoformat(),
+    }
+    _write_manifest(skills_dir, manifest)
+
+
+def _remove_from_manifest(skills_dir: Path, name: str) -> None:
+    manifest = _read_manifest(skills_dir)
+    manifest["skills"].pop(name, None)
+    _write_manifest(skills_dir, manifest)
+
+
+def _install_via_npx(
+    target: _InstallTarget, project_root: Path, console: Console
+) -> bool:
+    """Try to install with `npx skills add`. Returns True on success."""
+    env = os.environ.copy()
+    env.setdefault("SKILLS_NO_TELEMETRY", "1")
+    arg = target.slug if target.ref is None else f"{target.slug}@{target.ref}"
+    cmd = ["npx", "--yes", "skills", "add", arg]
+    console.print(
+        f"[dim]Running:[/dim] [cyan]{' '.join(cmd)}[/cyan] [dim](cwd={project_root})[/dim]"
+    )
+    try:
+        result = subprocess.run(
+            cmd,
+            cwd=project_root,
+            env=env,
+            check=False,
+        )
+    except FileNotFoundError:
+        return False
+    return result.returncode == 0
+
+
+def _install_via_git(
+    target: _InstallTarget,
+    project_root: Path,
+    skills_dir: Path,
+    console: Console,
+) -> bool:
+    if not _has_git():
+        console.print("[red]X[/red] git is not installed; cannot fall back from npx")
+        return False
+
+    dest = skills_dir / target.repo
+    if dest.exists():
+        console.print(
+            f"[yellow]![/yellow] {dest} already exists. "
+            "Remove it first with [cyan]planoai skills remove[/cyan] before reinstalling."
+        )
+        return False
+
+    cmd = ["git", "clone", "--depth", "1"]
+    if target.ref:
+        cmd.extend(["--branch", target.ref])
+    cmd.extend([target.url, str(dest)])
+    console.print(
+        f"[dim]Running:[/dim] [cyan]{' '.join(cmd)}[/cyan] [dim](cwd={project_root})[/dim]"
+    )
+    try:
+        result = subprocess.run(cmd, cwd=project_root, check=False)
+    except FileNotFoundError:
+        console.print("[red]X[/red] git binary not found")
+        return False
+    if result.returncode != 0:
+        return False
+
+    shutil.rmtree(dest / ".git", ignore_errors=True)
+
+    if not (dest / "SKILL.md").exists():
+        console.print(
+            f"[red]X[/red] {target.slug} does not contain a SKILL.md at its repo root; "
+            "this does not appear to be a valid Agent Skill."
+        )
+        shutil.rmtree(dest, ignore_errors=True)
+        return False
+    return True
+
+
+def _print_skills_table(console: Console, skills: list[Skill]) -> None:
+    if not skills:
+        console.print(
+            f"[dim]No skills installed.[/dim] Try [cyan]planoai skills add owner/repo[/cyan]."
+        )
+        return
+    table = Table(title="Installed Agent Skills", border_style="dim")
+    table.add_column("Name", style=f"bold {PLANO_COLOR}")
+    table.add_column("Scope")
+    table.add_column("Description")
+    table.add_column("Path", style="dim")
+    for s in skills:
+        desc = s.description.splitlines()[0]
+        if len(desc) > 80:
+            desc = desc[:77] + "..."
+        table.add_row(s.name, s.scope, desc, str(s.location))
+    console.print(table)
+
+
+@click.group(name="skills")
+def skills():
+    """Manage Agent Skills (agentskills.io) for this Plano project."""
+
+
+@skills.command(name="add")
+@click.argument("target", required=True)
+@click.option(
+    "--path",
+    default=".",
+    help="Project directory (defaults to the directory containing .plano/ or .git/).",
+)
+def add_cmd(target: str, path: str):
+    """Install an Agent Skill from a GitHub repo into .plano/skills/.
+
+    TARGET should be `owner/repo` (optionally suffixed with `@ref` for a branch
+    or tag).
+    """
+    console = _console()
+    install_target = _parse_install_target(target)
+    project_root = find_project_root(Path(path).resolve())
+    skills_dir = _ensure_skills_dir(project_root)
+
+    console.print(
+        f"[bold {PLANO_COLOR}]Installing skill[/bold {PLANO_COLOR}] "
+        f"[cyan]{install_target.slug}[/cyan] -> [dim]{skills_dir}[/dim]"
+    )
+
+    # Snapshot what's already discoverable so we can diff after install and
+    # surface every newly-added skill regardless of which scope it landed in
+    # (project for git fallback, agents for `npx skills add`, etc.) and
+    # regardless of how the installed skill name maps to the repo name (e.g.
+    # multi-skill repos like openai/skills -> ~/.agents/skills/pdf).
+    before, _ = discover_skills(project_root=project_root, include_user_scope=True)
+    before_keys = {(s.name, str(s.base_dir)) for s in before}
+
+    used_source: str
+    success = False
+    if _has_npx():
+        success = _install_via_npx(install_target, project_root, console)
+        used_source = "npx-skills"
+        if not success:
+            console.print(
+                "[yellow]![/yellow] npx skills add did not succeed; "
+                "falling back to direct git clone."
+            )
+    if not success:
+        success = _install_via_git(install_target, project_root, skills_dir, console)
+        used_source = "git"
+
+    if not success:
+        console.print(
+            f"[red]X[/red] Failed to install [cyan]{install_target.slug}[/cyan]"
+        )
+        sys.exit(1)
+
+    discovered, diagnostics = discover_skills(
+        project_root=project_root, include_user_scope=True
+    )
+    for diag in diagnostics:
+        if diag.severity == "error":
+            console.print(f"[red]X[/red] {diag.path}: {diag.message}")
+        else:
+            console.print(f"[yellow]![/yellow] {diag.path}: {diag.message}")
+
+    newly_installed = [
+        s for s in discovered if (s.name, str(s.base_dir)) not in before_keys
+    ]
+    if newly_installed:
+        for s in newly_installed:
+            if s.scope == "project":
+                _record_install(skills_dir, s.name, install_target, used_source)
+            try:
+                display_path = s.location.relative_to(project_root)
+            except ValueError:
+                display_path = s.location
+            console.print(
+                f"[green]+[/green] Installed [bold]{s.name}[/bold] "
+                f"[dim]({display_path}, scope={s.scope})[/dim]"
+            )
+        if any(
+            s.scope == "project" for s in newly_installed
+        ) and not is_project_trusted(project_root):
+            console.print(
+                "\n[dim]Project-scope skills are not auto-loaded until this project is "
+                "trusted. Run[/dim] [cyan]planoai skills trust[/cyan] [dim]to enable them.[/dim]"
+            )
+    else:
+        console.print(
+            "[yellow]![/yellow] Install reported success but no new SKILL.md "
+            "was discovered under .plano/skills, ~/.plano/skills, or "
+            "~/.agents/skills. Check the repo structure or pass a "
+            "single-skill repo."
+        )
+
+
+@skills.command(name="list")
+@click.option(
+    "--path",
+    default=".",
+    help="Project directory (defaults to the directory containing .plano/ or .git/).",
+)
+@click.option(
+    "--no-user-scope",
+    is_flag=True,
+    default=False,
+    help="Skip user-scope skills under ~/.plano/skills and ~/.agents/skills.",
+)
+def list_cmd(path: str, no_user_scope: bool):
+    """List discovered Agent Skills across project / user / agents scopes."""
+    console = _console()
+    project_root = find_project_root(Path(path).resolve())
+    discovered, diagnostics = discover_skills(
+        project_root=project_root, include_user_scope=not no_user_scope
+    )
+    _print_skills_table(console, discovered)
+
+    if diagnostics:
+        console.print()
+        for diag in diagnostics:
+            color = "red" if diag.severity == "error" else "yellow"
+            marker = "X" if diag.severity == "error" else "!"
+            console.print(f"[{color}]{marker}[/{color}] {diag.path}: {diag.message}")
+
+
+@skills.command(name="remove")
+@click.argument("name", required=True)
+@click.option(
+    "--path",
+    default=".",
+    help="Project directory (defaults to the directory containing .plano/ or .git/).",
+)
+def remove_cmd(name: str, path: str):
+    """Remove a project-scope skill from .plano/skills/.
+
+    User-scope skills under ~/.plano/skills or ~/.agents/skills must be
+    removed with their respective installer (`npx skills remove <name>` for
+    the latter); planoai will not touch directories outside the project.
+    """
+    console = _console()
+    project_root = find_project_root(Path(path).resolve())
+    skills_dir = project_root / PROJECT_SKILLS_DIR
+    if not skills_dir.exists():
+        console.print(f"[red]X[/red] no skills directory at {skills_dir}")
+        sys.exit(1)
+
+    target_dir = skills_dir / name
+    if not target_dir.exists():
+        discovered, _ = discover_skills(
+            project_root=project_root, include_user_scope=True
+        )
+        project_match = next(
+            (s for s in discovered if s.name == name and s.scope == "project"), None
+        )
+        if project_match is None:
+            other = next((s for s in discovered if s.name == name), None)
+            if other is not None:
+                console.print(
+                    f"[red]X[/red] '{name}' is installed in {other.scope} scope at "
+                    f"{other.base_dir}; planoai only removes project-scope skills. "
+                    "Use the upstream installer (e.g. `npx skills remove`) for that one."
+                )
+            else:
+                console.print(
+                    f"[red]X[/red] no project-scope skill named '{name}' under {skills_dir}"
+                )
+            sys.exit(1)
+        target_dir = project_match.base_dir
+
+    if not target_dir.resolve().is_relative_to(skills_dir.resolve()):
+        console.print(
+            f"[red]X[/red] refusing to delete {target_dir} (outside {skills_dir})"
+        )
+        sys.exit(1)
+
+    shutil.rmtree(target_dir, ignore_errors=False)
+    _remove_from_manifest(skills_dir, name)
+    console.print(f"[green]+[/green] Removed [bold]{name}[/bold]")
+
+
+@skills.command(name="trust")
+@click.option(
+    "--path",
+    default=".",
+    help="Project directory to mark as trusted.",
+)
+@click.option(
+    "--revoke",
+    is_flag=True,
+    default=False,
+    help="Revoke trust instead of granting it.",
+)
+def trust_cmd(path: str, revoke: bool):
+    """Mark this project's .plano/skills/ as trusted for auto-loading.
+
+    Project-scope skills come from the working directory's repo, which may
+    be untrusted. Plano refuses to inject their contents into the
+    orchestrator prompt until you trust the project.
+    """
+    console = _console()
+    project_root = find_project_root(Path(path).resolve())
+
+    if revoke:
+        path = trusted_projects_file()
+        if not path.exists():
+            console.print("[dim]No trusted projects to revoke.[/dim]")
+            return
+        try:
+            with path.open("r", encoding="utf-8") as fh:
+                data = json.load(fh) or {}
+        except (OSError, json.JSONDecodeError):
+            data = {}
+        trusted = {
+            str(Path(p).resolve()) for p in data.get("trusted_projects", []) or []
+        }
+        trusted.discard(str(project_root.resolve()))
+        data["trusted_projects"] = sorted(trusted)
+        with path.open("w", encoding="utf-8") as fh:
+            json.dump(data, fh, indent=2)
+        console.print(f"[green]+[/green] Revoked trust for [bold]{project_root}[/bold]")
+        return
+
+    _mark_project_trusted(project_root)
+    console.print(
+        f"[green]+[/green] Trusted [bold]{project_root}[/bold].\n"
+        f"[dim]Project-scope skills under .plano/skills/ will now be loaded at startup.[/dim]"
+    )

cli/planoai/templates/skills_routing.yaml

@@ -0,0 +1,71 @@
+version: v0.3.0
+
+# This template wires Agent Skills (https://agentskills.io) into Plano so the
+# built-in Plano-Orchestrator can decide *per request* which skill(s) to attach
+# to the selected route.
+#
+# 1. Install skills locally:
+#      planoai skills add owner/pdf-processing
+#      planoai skills add owner/code-review
+#    Skills land under .plano/skills/<name>/SKILL.md.
+#
+# 2. (Required for project-scope skills) Mark the project trusted so its
+#    skills auto-load at startup:
+#      planoai skills trust
+#
+# 3. Reference the installed skills under `routing_preferences[].skills`.
+#    During the intent step Plano-Orchestrator receives both a <routes> and a
+#    <skills> XML block; it picks a route and zero or more skills, and the
+#    SKILL.md bodies of the chosen skills are injected into the upstream
+#    system prompt for that turn.
+
+model_providers:
+  - model: anthropic/claude-sonnet-4-5
+    default: true
+    access_key: $ANTHROPIC_API_KEY
+
+  - model: openai/gpt-4.1-2025-04-14
+    access_key: $OPENAI_API_KEY
+
+# Catalog of skills available to this project. Entries are skill names
+# (resolved against .plano/skills/<name>/) — the CLI inlines each SKILL.md
+# body into the rendered config at `planoai up` time.
+#
+# Omit `skills:` entirely to auto-include every discovered skill.
+skills:
+  - pdf-processing
+  - code-review
+
+listeners:
+  - type: model
+    name: model_listener
+    port: 12000
+
+# Routing preferences double as Plano-Orchestrator's <routes> catalog. Attach
+# `skills:` to a route to make those skills eligible for activation when the
+# orchestrator picks that route.
+routing_preferences:
+  - name: code review
+    description: |
+      Reviewing pull requests, analyzing diffs, and suggesting improvements
+      to existing code.
+    models:
+      - anthropic/claude-sonnet-4-5
+      - openai/gpt-4.1-2025-04-14
+    skills:
+      - code-review
+
+  - name: document understanding
+    description: |
+      Summarizing PDFs and other long-form documents, extracting structured
+      data such as tables, line items, or signatures.
+    models:
+      - anthropic/claude-sonnet-4-5
+      - openai/gpt-4.1-2025-04-14
+    skills:
+      - pdf-processing
+    selection_policy:
+      prefer: cheapest
+
+tracing:
+  random_sampling: 100

cli/planoai/templates/template_sync_map.yaml

@@ -27,3 +27,8 @@ templates:
     template_file: conversational_state_v1_responses.yaml
     demo_configs: []
     transform: none
+
+  - template_id: skills_routing
+    template_file: skills_routing.yaml
+    demo_configs: []
+    transform: none

cli/test/test_skills.py

@@ -0,0 +1,394 @@
+"""Tests for cli/planoai/skills.py and the config-rendering hooks that
+materialize SKILL.md bodies into the rendered plano config.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+from unittest import mock
+
+import pytest
+
+from planoai.skills import (
+    AGENTS_SKILLS_DIR,
+    PROJECT_SKILLS_DIR,
+    USER_SKILLS_DIR,
+    Skill,
+    discover_skills,
+    parse_skill_md,
+    total_catalog_size,
+)
+
+
+@pytest.fixture(autouse=True)
+def _isolate_user_scopes(tmp_path, monkeypatch):
+    """Default both user-tier scopes to non-existent dirs so the dev's real
+    ~/.plano/skills and ~/.agents/skills cannot bleed into tests.
+    """
+    monkeypatch.setattr("planoai.skills.USER_SKILLS_DIR", tmp_path / "_no_user_skills")
+    monkeypatch.setattr(
+        "planoai.skills.AGENTS_SKILLS_DIR", tmp_path / "_no_agents_skills"
+    )
+
+
+def _write_skill(
+    base: Path,
+    name: str,
+    description: str = "Process PDFs. Use when handling PDF files.",
+    body: str = "# Body\n\nDo the thing.",
+    extra_frontmatter: str = "",
+) -> Path:
+    skill_dir = base / name
+    skill_dir.mkdir(parents=True, exist_ok=True)
+    frontmatter = f"name: {name}\ndescription: {description}\n{extra_frontmatter}"
+    (skill_dir / "SKILL.md").write_text(
+        f"---\n{frontmatter}---\n\n{body}",
+        encoding="utf-8",
+    )
+    return skill_dir / "SKILL.md"
+
+
+def test_parse_skill_md_minimal(tmp_path):
+    skill_md = _write_skill(tmp_path, "pdf-processing")
+
+    skill, diagnostics = parse_skill_md(skill_md)
+
+    assert skill is not None
+    assert skill.name == "pdf-processing"
+    assert skill.description.startswith("Process PDFs")
+    assert "Do the thing." in skill.body
+    assert skill.location == skill_md.resolve()
+    assert skill.base_dir == skill_md.parent.resolve()
+    # No warnings or errors for a well-formed skill.
+    assert diagnostics == []
+
+
+def test_parse_skill_md_lenient_when_name_mismatches_directory(tmp_path):
+    skill_dir = tmp_path / "some-dir"
+    skill_dir.mkdir()
+    (skill_dir / "SKILL.md").write_text(
+        "---\nname: different-name\ndescription: ok\n---\n\nbody",
+        encoding="utf-8",
+    )
+
+    skill, diagnostics = parse_skill_md(skill_dir / "SKILL.md")
+
+    assert skill is not None
+    assert skill.name == "different-name"
+    assert any(
+        "does not match parent directory" in d.message for d in diagnostics
+    ), diagnostics
+
+
+def test_parse_skill_md_warns_on_invalid_name(tmp_path):
+    skill_dir = tmp_path / "Bad-Name"
+    skill_dir.mkdir()
+    (skill_dir / "SKILL.md").write_text(
+        "---\nname: Bad-Name\ndescription: ok\n---\n\nbody",
+        encoding="utf-8",
+    )
+
+    skill, diagnostics = parse_skill_md(skill_dir / "SKILL.md")
+    assert skill is not None
+    assert any("violates spec naming rules" in d.message for d in diagnostics)
+
+
+def test_parse_skill_md_recovers_from_unquoted_colons(tmp_path):
+    skill_dir = tmp_path / "code-review"
+    skill_dir.mkdir()
+    (skill_dir / "SKILL.md").write_text(
+        "---\nname: code-review\n"
+        "description: Use this skill when: the user asks about code review\n"
+        "---\n\nbody",
+        encoding="utf-8",
+    )
+
+    skill, diagnostics = parse_skill_md(skill_dir / "SKILL.md")
+    # Lenient parse retries with quoted values and succeeds.
+    assert skill is not None
+    assert skill.name == "code-review"
+    assert "Use this skill when:" in skill.description
+
+
+def test_parse_skill_md_rejects_when_description_missing(tmp_path):
+    skill_dir = tmp_path / "broken"
+    skill_dir.mkdir()
+    (skill_dir / "SKILL.md").write_text(
+        "---\nname: broken\n---\n\nbody",
+        encoding="utf-8",
+    )
+
+    skill, diagnostics = parse_skill_md(skill_dir / "SKILL.md")
+    assert skill is None
+    assert any(d.severity == "error" for d in diagnostics)
+
+
+def test_parse_skill_md_rejects_when_frontmatter_missing(tmp_path):
+    skill_dir = tmp_path / "no-frontmatter"
+    skill_dir.mkdir()
+    (skill_dir / "SKILL.md").write_text("just markdown", encoding="utf-8")
+
+    skill, diagnostics = parse_skill_md(skill_dir / "SKILL.md")
+    assert skill is None
+    assert any("frontmatter" in d.message for d in diagnostics)
+
+
+def test_discover_skills_project_only(tmp_path, monkeypatch):
+    (tmp_path / ".plano").mkdir()
+    project_skills_dir = tmp_path / ".plano" / "skills"
+    project_skills_dir.mkdir()
+    _write_skill(project_skills_dir, "pdf-processing")
+    _write_skill(project_skills_dir, "code-review")
+
+    skills, diagnostics = discover_skills(
+        project_root=tmp_path, include_user_scope=False
+    )
+
+    names = sorted(s.name for s in skills)
+    assert names == ["code-review", "pdf-processing"]
+    assert all(s.scope == "project" for s in skills)
+
+
+def test_discover_skills_picks_up_agents_scope(tmp_path, monkeypatch):
+    """`npx skills add` writes into ~/.agents/skills/<name>. That directory
+    must be discovered as a user-tier (auto-trusted) scope so the upstream
+    CLI works without Plano-specific awareness.
+    """
+    agents_dir = tmp_path / "fake-home" / ".agents" / "skills"
+    agents_dir.mkdir(parents=True)
+    _write_skill(agents_dir, "pdf", description="agents-scope description")
+
+    monkeypatch.setattr("planoai.skills.AGENTS_SKILLS_DIR", agents_dir)
+
+    (tmp_path / ".plano").mkdir()
+    (tmp_path / ".plano" / "skills").mkdir()
+
+    skills, _ = discover_skills(project_root=tmp_path, include_user_scope=True)
+    by_name = {s.name: s for s in skills}
+    assert "pdf" in by_name
+    assert by_name["pdf"].scope == "agents"
+    assert by_name["pdf"].description == "agents-scope description"
+
+
+def test_discover_skills_user_scope_shadows_agents_scope(tmp_path, monkeypatch):
+    """When the same skill name exists in both ~/.plano/skills and
+    ~/.agents/skills, the Plano-native one wins and a diagnostic is emitted.
+
+    Project root lives in its own subtree (with no .plano/ ancestor) so it
+    cannot collide with the patched user-scope dir.
+    """
+    home = tmp_path / "fake-home"
+    home.mkdir()
+    agents_dir = home / ".agents" / "skills"
+    agents_dir.mkdir(parents=True)
+    _write_skill(agents_dir, "pdf", description="agents copy")
+    monkeypatch.setattr("planoai.skills.AGENTS_SKILLS_DIR", agents_dir)
+
+    user_dir = home / ".plano" / "skills"
+    user_dir.mkdir(parents=True)
+    _write_skill(user_dir, "pdf", description="user copy")
+    monkeypatch.setattr("planoai.skills.USER_SKILLS_DIR", user_dir)
+
+    project_root = tmp_path / "elsewhere" / "proj"
+    project_root.mkdir(parents=True)
+    skills, diagnostics = discover_skills(
+        project_root=project_root, include_user_scope=True
+    )
+    by_name = {s.name: s for s in skills}
+    assert by_name["pdf"].scope == "user"
+    assert by_name["pdf"].description == "user copy"
+    assert any("shadows ~/.agents/skills" in d.message for d in diagnostics)
+
+
+def test_discover_skills_project_overrides_user_scope(tmp_path, monkeypatch):
+    user_skills_dir = tmp_path / "fake-home" / ".plano" / "skills"
+    user_skills_dir.mkdir(parents=True)
+    _write_skill(user_skills_dir, "shared", description="user-scope description")
+
+    monkeypatch.setattr("planoai.skills.USER_SKILLS_DIR", user_skills_dir)
+
+    (tmp_path / ".plano").mkdir()
+    project_skills_dir = tmp_path / ".plano" / "skills"
+    project_skills_dir.mkdir()
+    _write_skill(project_skills_dir, "shared", description="project-scope description")
+    _write_skill(project_skills_dir, "only-project")
+
+    skills, diagnostics = discover_skills(
+        project_root=tmp_path, include_user_scope=True
+    )
+
+    by_name = {s.name: s for s in skills}
+    assert by_name["shared"].scope == "project"
+    assert by_name["shared"].description == "project-scope description"
+    assert by_name["only-project"].scope == "project"
+    assert any("shadows user-scope skill" in d.message for d in diagnostics)
+
+
+def test_total_catalog_size_counts_name_and_description():
+    skills = [
+        Skill(
+            name="a",
+            description="d1",
+            location=Path("/x"),
+            base_dir=Path("/"),
+            body="b",
+            scope="project",
+        ),
+        Skill(
+            name="bb",
+            description="dd",
+            location=Path("/y"),
+            base_dir=Path("/"),
+            body="b",
+            scope="project",
+        ),
+    ]
+    assert total_catalog_size(skills) == (1 + 2) + (2 + 2)
+
+
+def test_materialize_skills_in_config_default_inlines_bodies(tmp_path, monkeypatch):
+    project_root = tmp_path
+    (project_root / ".plano").mkdir()
+    project_skills_dir = project_root / ".plano" / "skills"
+    project_skills_dir.mkdir()
+    _write_skill(
+        project_skills_dir,
+        "pdf-processing",
+        description="Process PDFs.",
+        body="# Body\nfollow these steps.",
+    )
+
+    fake_home = tmp_path / "home"
+    fake_home.mkdir()
+    monkeypatch.setenv("HOME", str(fake_home))
+    # Mark the project trusted so .plano/skills is loaded.
+    trusted = fake_home / ".plano" / "trusted_projects.json"
+    trusted.parent.mkdir(parents=True, exist_ok=True)
+    trusted.write_text(
+        json.dumps({"trusted_projects": [str(project_root.resolve())]}),
+        encoding="utf-8",
+    )
+    monkeypatch.setattr(
+        "planoai.skills.USER_SKILLS_DIR",
+        fake_home / ".plano" / "skills",
+    )
+
+    from planoai.config_generator import materialize_skills_in_config
+
+    config_yaml = {"version": "v0.4.0"}
+    materialize_skills_in_config(config_yaml, project_root)
+
+    assert "skills" in config_yaml
+    materialized = config_yaml["skills"]
+    assert len(materialized) == 1
+    entry = materialized[0]
+    assert entry["name"] == "pdf-processing"
+    assert "follow these steps." in entry["body"]
+    assert entry["scope"] == "project"
+
+
+def test_materialize_skills_in_config_skips_untrusted_project_skills(
+    tmp_path, monkeypatch
+):
+    project_root = tmp_path
+    (project_root / ".plano").mkdir()
+    project_skills_dir = project_root / ".plano" / "skills"
+    project_skills_dir.mkdir()
+    _write_skill(project_skills_dir, "pdf-processing")
+
+    fake_home = tmp_path / "home"
+    fake_home.mkdir()
+    monkeypatch.setenv("HOME", str(fake_home))
+    monkeypatch.setattr(
+        "planoai.skills.USER_SKILLS_DIR",
+        fake_home / ".plano" / "skills",
+    )
+
+    from planoai.config_generator import materialize_skills_in_config
+
+    config_yaml = {"version": "v0.4.0"}
+    materialize_skills_in_config(config_yaml, project_root)
+    # Untrusted -> project skills are not loaded.
+    assert "skills" not in config_yaml
+
+
+def test_materialize_skills_in_config_loads_agents_scope_without_trust(
+    tmp_path, monkeypatch
+):
+    """Even with no project trust, skills installed by `npx skills add` into
+    ~/.agents/skills/<name> must materialize into the rendered config — that
+    directory is the universal Agent Skills install location and is
+    user-tier, not project-tier.
+    """
+    project_root = tmp_path / "proj"
+    (project_root / ".plano").mkdir(parents=True)
+
+    fake_home = tmp_path / "home"
+    fake_home.mkdir()
+    monkeypatch.setenv("HOME", str(fake_home))
+    monkeypatch.setattr(
+        "planoai.skills.USER_SKILLS_DIR",
+        fake_home / ".plano" / "skills",
+    )
+
+    agents_dir = fake_home / ".agents" / "skills"
+    agents_dir.mkdir(parents=True)
+    _write_skill(agents_dir, "pdf", body="# Body\nhandle the pdf.")
+    monkeypatch.setattr("planoai.skills.AGENTS_SKILLS_DIR", agents_dir)
+
+    from planoai.config_generator import materialize_skills_in_config
+
+    config_yaml = {"version": "v0.4.0"}
+    materialize_skills_in_config(config_yaml, project_root)
+
+    assert "skills" in config_yaml
+    materialized = config_yaml["skills"]
+    assert len(materialized) == 1
+    assert materialized[0]["name"] == "pdf"
+    assert materialized[0]["scope"] == "agents"
+    assert "handle the pdf." in materialized[0]["body"]
+
+
+def test_materialize_skills_in_config_respects_allow_list(tmp_path, monkeypatch):
+    project_root = tmp_path
+    (project_root / ".plano").mkdir()
+    skills_dir = project_root / ".plano" / "skills"
+    skills_dir.mkdir()
+    _write_skill(skills_dir, "skill-a")
+    _write_skill(skills_dir, "skill-b")
+
+    fake_home = tmp_path / "home"
+    fake_home.mkdir()
+    monkeypatch.setenv("HOME", str(fake_home))
+    trusted = fake_home / ".plano" / "trusted_projects.json"
+    trusted.parent.mkdir(parents=True, exist_ok=True)
+    trusted.write_text(
+        json.dumps({"trusted_projects": [str(project_root.resolve())]}),
+        encoding="utf-8",
+    )
+    monkeypatch.setattr(
+        "planoai.skills.USER_SKILLS_DIR",
+        fake_home / ".plano" / "skills",
+    )
+
+    from planoai.config_generator import materialize_skills_in_config
+
+    config_yaml = {
+        "version": "v0.4.0",
+        "skills": ["skill-a"],
+        "routing_preferences": [
+            {
+                "name": "demo route",
+                "description": "demo",
+                "models": ["openai/gpt-4o"],
+                "skills": ["skill-a", "does-not-exist"],
+            }
+        ],
+    }
+    materialize_skills_in_config(config_yaml, project_root)
+
+    assert [s["name"] for s in config_yaml["skills"]] == ["skill-a"]
+    # Unknown allow-list entries are pruned but the known one is kept.
+    assert config_yaml["routing_preferences"][0]["skills"] == ["skill-a"]

cli/test/test_skills_cmd.py

@@ -0,0 +1,235 @@
+"""CLI tests for the `planoai skills` command group."""
+
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+from unittest import mock
+
+import pytest
+from click.testing import CliRunner
+
+from planoai.skills_cmd import skills
+
+
+def _seed_project(tmp_path: Path) -> Path:
+    """Create a project that find_project_root will pick up via .plano/."""
+    project = tmp_path / "project"
+    project.mkdir()
+    (project / ".plano").mkdir()
+    return project
+
+
+@pytest.fixture(autouse=True)
+def _isolate_user_scopes(tmp_path, monkeypatch):
+    """Default both user-tier scopes to non-existent dirs so the dev's real
+    ~/.plano/skills and ~/.agents/skills cannot bleed into the test sandbox.
+    Individual tests can override these via further monkeypatching.
+    """
+    monkeypatch.setattr("planoai.skills.USER_SKILLS_DIR", tmp_path / "no-user-skills")
+    monkeypatch.setattr(
+        "planoai.skills.AGENTS_SKILLS_DIR", tmp_path / "no-agents-skills"
+    )
+
+
+def _write_skill(base: Path, name: str, description: str = "demo skill") -> None:
+    skill_dir = base / name
+    skill_dir.mkdir(parents=True, exist_ok=True)
+    (skill_dir / "SKILL.md").write_text(
+        f"---\nname: {name}\ndescription: {description}\n---\n\nbody",
+        encoding="utf-8",
+    )
+
+
+def test_list_empty(tmp_path, monkeypatch):
+    project = _seed_project(tmp_path)
+    monkeypatch.chdir(project)
+    # Isolate user-scope skills dir.
+    monkeypatch.setattr("planoai.skills.USER_SKILLS_DIR", tmp_path / "no-such-home")
+
+    runner = CliRunner()
+    result = runner.invoke(skills, ["list"])
+
+    assert result.exit_code == 0, result.output
+    assert "No skills installed" in result.output
+
+
+def test_list_shows_project_skills(tmp_path, monkeypatch):
+    project = _seed_project(tmp_path)
+    monkeypatch.chdir(project)
+    monkeypatch.setattr("planoai.skills.USER_SKILLS_DIR", tmp_path / "no-such-home")
+
+    _write_skill(project / ".plano" / "skills", "pdf-processing")
+    _write_skill(project / ".plano" / "skills", "code-review")
+
+    runner = CliRunner()
+    result = runner.invoke(skills, ["list", "--no-user-scope"])
+
+    assert result.exit_code == 0, result.output
+    assert "pdf-processing" in result.output
+    assert "code-review" in result.output
+
+
+def test_remove_deletes_skill_dir(tmp_path, monkeypatch):
+    project = _seed_project(tmp_path)
+    monkeypatch.chdir(project)
+    monkeypatch.setattr("planoai.skills.USER_SKILLS_DIR", tmp_path / "no-such-home")
+
+    skills_dir = project / ".plano" / "skills"
+    _write_skill(skills_dir, "pdf-processing")
+    (skills_dir / ".skills.json").write_text(
+        json.dumps(
+            {
+                "skills": {
+                    "pdf-processing": {
+                        "source": "git",
+                        "repo": "owner/pdf-processing",
+                    }
+                }
+            }
+        ),
+        encoding="utf-8",
+    )
+
+    runner = CliRunner()
+    result = runner.invoke(skills, ["remove", "pdf-processing"])
+
+    assert result.exit_code == 0, result.output
+    assert not (skills_dir / "pdf-processing").exists()
+    manifest = json.loads((skills_dir / ".skills.json").read_text(encoding="utf-8"))
+    assert "pdf-processing" not in manifest["skills"]
+
+
+def test_remove_unknown_skill_errors(tmp_path, monkeypatch):
+    project = _seed_project(tmp_path)
+    monkeypatch.chdir(project)
+    monkeypatch.setattr("planoai.skills.USER_SKILLS_DIR", tmp_path / "no-such-home")
+    (project / ".plano" / "skills").mkdir()
+
+    runner = CliRunner()
+    result = runner.invoke(skills, ["remove", "nope"])
+    assert result.exit_code != 0
+
+
+def test_add_falls_back_to_git_when_no_npx(tmp_path, monkeypatch):
+    project = _seed_project(tmp_path)
+    monkeypatch.chdir(project)
+
+    # Force the npx branch off and stub git clone to create a SKILL.md.
+    monkeypatch.setattr("planoai.skills_cmd._has_npx", lambda: False)
+    monkeypatch.setattr("planoai.skills_cmd._has_git", lambda: True)
+
+    def fake_subprocess_run(cmd, **kwargs):
+        # cmd is like ["git", "clone", ..., url, dest]
+        dest = Path(cmd[-1])
+        dest.mkdir(parents=True, exist_ok=True)
+        (dest / "SKILL.md").write_text(
+            "---\nname: my-skill\ndescription: example\n---\n\nbody",
+            encoding="utf-8",
+        )
+        (dest / ".git").mkdir()
+        return mock.Mock(returncode=0)
+
+    monkeypatch.setattr("planoai.skills_cmd.subprocess.run", fake_subprocess_run)
+
+    runner = CliRunner()
+    result = runner.invoke(skills, ["add", "owner/my-skill"])
+    assert result.exit_code == 0, result.output
+    assert (project / ".plano" / "skills" / "my-skill" / "SKILL.md").exists()
+    # Trust hint should be shown for untrusted projects with project-scope installs.
+    assert "planoai skills trust" in result.output
+
+
+def test_add_discovers_skill_installed_into_agents_scope_by_npx(tmp_path, monkeypatch):
+    """`npx skills add` writes to ~/.agents/skills/<name>; planoai must
+    pick it up from that universal scope and *not* nag about trust.
+    """
+    project = _seed_project(tmp_path)
+    monkeypatch.chdir(project)
+
+    agents_dir = tmp_path / "agents" / "skills"
+    agents_dir.mkdir(parents=True)
+    monkeypatch.setattr("planoai.skills.AGENTS_SKILLS_DIR", agents_dir)
+
+    # Pretend npx is on $PATH and succeeds, dropping the skill in ~/.agents/skills
+    # rather than .plano/skills (which is what the upstream CLI actually does).
+    monkeypatch.setattr("planoai.skills_cmd._has_npx", lambda: True)
+
+    def fake_install_via_npx(target, project_root, console):
+        skill_dir = agents_dir / "pdf"
+        skill_dir.mkdir(parents=True, exist_ok=True)
+        (skill_dir / "SKILL.md").write_text(
+            "---\nname: pdf\ndescription: process pdfs\n---\n\nbody",
+            encoding="utf-8",
+        )
+        return True
+
+    monkeypatch.setattr("planoai.skills_cmd._install_via_npx", fake_install_via_npx)
+
+    runner = CliRunner()
+    result = runner.invoke(skills, ["add", "openai/skills"])
+
+    assert result.exit_code == 0, result.output
+    assert "scope=agents" in result.output
+    assert "planoai skills trust" not in result.output
+
+
+def test_list_includes_agents_scope_entries(tmp_path, monkeypatch):
+    project = _seed_project(tmp_path)
+    monkeypatch.chdir(project)
+
+    agents_dir = tmp_path / "agents-skills"
+    agents_dir.mkdir()
+    monkeypatch.setattr("planoai.skills.AGENTS_SKILLS_DIR", agents_dir)
+    _write_skill(agents_dir, "pdf")
+
+    runner = CliRunner()
+    result = runner.invoke(skills, ["list"])
+
+    assert result.exit_code == 0, result.output
+    assert "pdf" in result.output
+    assert "agents" in result.output
+
+
+def test_remove_rejects_agents_scope_skill(tmp_path, monkeypatch):
+    project = _seed_project(tmp_path)
+    monkeypatch.chdir(project)
+    (project / ".plano" / "skills").mkdir()
+
+    agents_dir = tmp_path / "agents-skills"
+    agents_dir.mkdir()
+    monkeypatch.setattr("planoai.skills.AGENTS_SKILLS_DIR", agents_dir)
+    _write_skill(agents_dir, "pdf")
+
+    runner = CliRunner()
+    result = runner.invoke(skills, ["remove", "pdf"])
+
+    assert result.exit_code != 0
+    assert "npx skills remove" in result.output
+
+
+def test_trust_marks_project(tmp_path, monkeypatch):
+    project = _seed_project(tmp_path)
+    monkeypatch.chdir(project)
+
+    fake_home = tmp_path / "home"
+    fake_home.mkdir()
+    monkeypatch.setenv("HOME", str(fake_home))
+
+    runner = CliRunner()
+    result = runner.invoke(skills, ["trust"])
+
+    assert result.exit_code == 0, result.output
+    trusted_file = fake_home / ".plano" / "trusted_projects.json"
+    assert trusted_file.exists()
+    data = json.loads(trusted_file.read_text(encoding="utf-8"))
+    assert str(project.resolve()) in data["trusted_projects"]
+
+
+def test_add_rejects_invalid_target(tmp_path, monkeypatch):
+    project = _seed_project(tmp_path)
+    monkeypatch.chdir(project)
+    runner = CliRunner()
+    result = runner.invoke(skills, ["add", "not-a-spec"])
+    assert result.exit_code != 0

config/plano_config_schema.yaml

@@ -316,6 +316,52 @@ properties:
         description: "Maximum token length for the orchestrator/routing model context window. Default is 8192."
   system_prompt:
     type: string
+  skills:
+    type: array
+    description: |
+      Agent Skills (https://agentskills.io) registered for this Plano project.
+      Each entry may be a string (the skill name, resolved against
+      .plano/skills/<name>/ or ~/.plano/skills/<name>/) or an object with name
+      + optional inline metadata. The Python CLI auto-populates body/path
+      during config rendering. Skills are attached to routes via
+      `routing_preferences[].skills`; when omitted there, the orchestrator
+      sees every entry declared (or auto-discovered) here.
+    items:
+      oneOf:
+        - type: string
+        - type: object
+          properties:
+            name:
+              type: string
+            description:
+              type: string
+            path:
+              type: string
+              description: "Absolute path to the SKILL.md file (set by the CLI at render time)."
+            base_dir:
+              type: string
+              description: "Absolute path to the skill directory (set by the CLI at render time)."
+            body:
+              type: string
+              description: "Markdown body of SKILL.md (inlined at render time so WASM does not need filesystem access)."
+            scope:
+              type: string
+              enum:
+                - project
+                - user
+            compatibility:
+              type: string
+            license:
+              type: string
+            metadata:
+              type: object
+              additionalProperties:
+                type: string
+            allowed_tools:
+              type: string
+          additionalProperties: false
+          required:
+            - name
   prompt_targets:
     type: array
     items:
@@ -549,6 +595,17 @@ properties:
           items:
             type: string
           minItems: 1
+        skills:
+          type: array
+          description: |
+            Agent Skills associated with this routing preference. When
+            Plano-Orchestrator selects this route, the listed skills are also
+            considered for activation and their SKILL.md bodies are injected
+            into the upstream system prompt. Skill names must match an entry
+            in the top-level `skills:` catalog or be discoverable under
+            `.plano/skills/` or `~/.plano/skills/`.
+          items:
+            type: string
         selection_policy:
           type: object
           properties:

crates/brightstaff/src/handlers/agents/selector.rs

@@ -132,11 +132,15 @@ impl AgentSelector {
             .determine_orchestration(messages, Some(usage_preferences), request_id)
             .await
         {
-            Ok(Some(routes)) => {
+            Ok(Some(selection)) => {
-                debug!(count = routes.len(), "determined agents via orchestration");
+                debug!(
+                    count = selection.routes.len(),
+                    skill_count = selection.skills.len(),
+                    "determined agents via orchestration"
+                );
                 let mut selected_agents = Vec::new();
 
-                for (route_name, agent_name) in routes {
+                for (route_name, agent_name) in selection.routes {
                     debug!(route = %route_name, agent = %agent_name, "processing route");
                     let selected_agent = agents
                         .iter()

crates/brightstaff/src/handlers/llm/mod.rs

@@ -37,7 +37,7 @@ use crate::tracing::{
     collect_custom_trace_attributes, llm as tracing_llm, operation_component,
     plano as tracing_plano, set_service_name,
 };
-use model_selection::router_chat_get_upstream_model;
+use model_selection::{inject_activated_skills_into_request, router_chat_get_upstream_model};
 
 const PERPLEXITY_PROVIDER_PREFIX: &str = "perplexity/";
 
@@ -282,26 +282,16 @@ async fn llm_chat_inner(
         Err(response) => return Ok(response),
     };
 
-    // Serialize request for upstream BEFORE router consumes it
+    // Route the request (or use pinned model from session cache) BEFORE
-    let client_request_bytes_for_upstream: Bytes =
+    // serializing for upstream — skill body injection happens here, so the
-        match ProviderRequestType::to_bytes(&client_request) {
+    // upstream bytes must be produced after routing returns.
-            Ok(bytes) => bytes.into(),
+    let (resolved_model, activated_skills) = if let Some(cached_model) = pinned_model {
-            Err(err) => {
-                warn!(error = %err, "failed to serialize request for upstream");
-                let mut r = Response::new(full(format!("Failed to serialize request: {}", err)));
-                *r.status_mut() = StatusCode::INTERNAL_SERVER_ERROR;
-                return Ok(r);
-            }
-        };
-
-    // --- Phase 3: Route the request (or use pinned model from session cache) ---
-    let resolved_model = if let Some(cached_model) = pinned_model {
         info!(
             session_id = %session_id.as_deref().unwrap_or(""),
             model = %cached_model,
             "using pinned routing decision from cache"
         );
-        cached_model
+        (cached_model, Vec::new())
     } else {
         let routing_span = info_span!(
             "routing",
@@ -313,11 +303,16 @@ async fn llm_chat_inner(
             route.selected_model = tracing::field::Empty,
             routing.determination_ms = tracing::field::Empty,
         );
+        // The router consumes the request (it converts it to OpenAI format
+        // internally to extract conversation messages). Clone so we can
+        // still mutate the original below when Plano-Orchestrator activates
+        // any Agent Skills.
+        let request_for_routing = client_request.clone();
         let routing_result = match async {
             set_service_name(operation_component::ROUTING);
             router_chat_get_upstream_model(
                 Arc::clone(&state.orchestrator_service),
-                client_request,
+                request_for_routing,
                 &request_path,
                 &request_id,
                 inline_routing_preferences,
@@ -335,8 +330,11 @@ async fn llm_chat_inner(
             }
         };
 
-        let (router_selected_model, route_name) =
+        let (router_selected_model, route_name, activated) = (
-            (routing_result.model_name, routing_result.route_name);
+            routing_result.model_name,
+            routing_result.route_name,
+            routing_result.activated_skills,
+        );
         let model = if router_selected_model != "none" {
             router_selected_model
         } else {
@@ -362,10 +360,34 @@ async fn llm_chat_inner(
                 .await;
         }
 
-        model
+        (model, activated)
     };
     tracing::Span::current().record(tracing_llm::MODEL_NAME, resolved_model.as_str());
 
+    // If Plano-Orchestrator activated any Agent Skills for this route, inject
+    // their SKILL.md bodies into the system prompt before we hand the bytes
+    // off to the upstream provider.
+    if !activated_skills.is_empty() {
+        info!(
+            count = activated_skills.len(),
+            skills = ?activated_skills.iter().map(|s| s.name.as_str()).collect::<Vec<_>>(),
+            "injecting activated Agent Skills into upstream system prompt"
+        );
+        inject_activated_skills_into_request(&mut client_request, &activated_skills);
+    }
+
+    // Serialize request for upstream AFTER potential skill injection.
+    let client_request_bytes_for_upstream: Bytes =
+        match ProviderRequestType::to_bytes(&client_request) {
+            Ok(bytes) => bytes.into(),
+            Err(err) => {
+                warn!(error = %err, "failed to serialize request for upstream");
+                let mut r = Response::new(full(format!("Failed to serialize request: {}", err)));
+                *r.status_mut() = StatusCode::INTERNAL_SERVER_ERROR;
+                return Ok(r);
+            }
+        };
+
     // --- Phase 4: Forward to upstream and stream back ---
     send_upstream(
         &state.http_client,

crates/brightstaff/src/handlers/llm/model_selection.rs

@@ -1,5 +1,9 @@
-use common::configuration::TopLevelRoutingPreference;
+use common::configuration::{SkillRef, TopLevelRoutingPreference};
+use common::skills_runtime::augment_system_prompt_with_skills;
+use hermesllm::apis::openai::{Message, MessageContent, Role};
 use hermesllm::clients::endpoints::SupportedUpstreamAPIs;
+use hermesllm::providers::request::ProviderRequest;
+use hermesllm::transforms::lib::ExtractText;
 use hermesllm::ProviderRequestType;
 use hyper::StatusCode;
 use std::sync::Arc;
@@ -29,6 +33,10 @@ pub struct RoutingResult {
     /// Full ranked list — use subsequent entries as fallbacks on 429/5xx.
     pub models: Vec<String>,
     pub route_name: Option<String>,
+    /// Agent Skills activated by Plano-Orchestrator for this request.
+    /// Their `body` field (the SKILL.md content) is prepended to the
+    /// upstream system prompt by the caller in `send_upstream`.
+    pub activated_skills: Vec<SkillRef>,
 }
 
 pub struct RoutingError {
@@ -128,8 +136,8 @@ pub async fn router_chat_get_upstream_model(
 
     match routing_result {
         Ok(route) => match route {
-            Some((route_name, ranked_models)) => {
+            Some(decision) => {
-                let model_name = ranked_models.first().cloned().unwrap_or_default();
+                let model_name = decision.models.first().cloned().unwrap_or_default();
                 current_span.record("route.selected_model", model_name.as_str());
                 bs_metrics::record_router_decision(
                     route_label,
@@ -139,8 +147,9 @@ pub async fn router_chat_get_upstream_model(
                 );
                 Ok(RoutingResult {
                     model_name,
-                    models: ranked_models,
+                    models: decision.models,
-                    route_name: Some(route_name),
+                    route_name: Some(decision.route_name),
+                    activated_skills: decision.activated_skills,
                 })
             }
             None => {
@@ -159,6 +168,7 @@ pub async fn router_chat_get_upstream_model(
                     model_name: "none".to_string(),
                     models: vec!["none".to_string()],
                     route_name: None,
+                    activated_skills: Vec::new(),
                 })
             }
         },
@@ -172,3 +182,61 @@ pub async fn router_chat_get_upstream_model(
         }
     }
 }
+
+/// Prepend the bodies of `activated_skills` to the system prompt of the
+/// upstream request so the chosen LLM has access to each skill's instructions.
+/// Works across every provider variant by going through the OpenAI message
+/// shape (`get_messages`/`set_messages`).
+///
+/// When there is already a leading system message we augment it in place;
+/// otherwise a new system message is inserted at position 0. No-op when
+/// `activated_skills` is empty.
+pub fn inject_activated_skills_into_request(
+    client_request: &mut ProviderRequestType,
+    activated_skills: &[SkillRef],
+) {
+    if activated_skills.is_empty() {
+        return;
+    }
+
+    let skill_refs: Vec<&SkillRef> = activated_skills.iter().collect();
+
+    let mut messages = client_request.get_messages();
+
+    let (system_idx, base_text) = match messages.iter().position(|m| m.role == Role::System) {
+        Some(idx) => {
+            let text = messages[idx]
+                .content
+                .as_ref()
+                .map(|c| c.extract_text())
+                .unwrap_or_default();
+            (Some(idx), Some(text))
+        }
+        None => (None, None),
+    };
+
+    let augmented = augment_system_prompt_with_skills(base_text, &skill_refs);
+    let Some(augmented_text) = augmented else {
+        return;
+    };
+
+    match system_idx {
+        Some(idx) => {
+            messages[idx].content = Some(MessageContent::Text(augmented_text));
+        }
+        None => {
+            messages.insert(
+                0,
+                Message {
+                    role: Role::System,
+                    content: Some(MessageContent::Text(augmented_text)),
+                    name: None,
+                    tool_calls: None,
+                    tool_call_id: None,
+                },
+            );
+        }
+    }
+
+    client_request.set_messages(&messages);
+}

crates/brightstaff/src/main.rs

@@ -314,11 +314,12 @@ async fn init_app_state(
         .orchestrator_model_context_length
         .unwrap_or(brightstaff::router::orchestrator_model_v1::MAX_TOKEN_LEN);
 
-    let orchestrator_service = Arc::new(OrchestratorService::with_routing(
+    let orchestrator_service = Arc::new(OrchestratorService::with_routing_and_skills(
         format!("{llm_provider_url}{CHAT_COMPLETIONS_PATH}"),
         orchestrator_model_name,
         orchestrator_llm_provider,
         config.routing_preferences.clone(),
+        config.skills.clone(),
         metrics_service,
         session_ttl_seconds,
         session_cache,

crates/brightstaff/src/router/orchestrator.rs

@@ -1,7 +1,9 @@
 use std::{borrow::Cow, collections::HashMap, sync::Arc, time::Duration};
 
 use common::{
-    configuration::{AgentUsagePreference, OrchestrationPreference, TopLevelRoutingPreference},
+    configuration::{
+        AgentUsagePreference, OrchestrationPreference, SkillRef, TopLevelRoutingPreference,
+    },
     consts::{ARCH_PROVIDER_HINT_HEADER, REQUEST_ID_HEADER},
 };
 use hermesllm::apis::openai::Message;
@@ -13,7 +15,7 @@ use tracing::{debug, info};
 
 use super::http::{self, post_and_extract_content};
 use super::model_metrics::ModelMetricsService;
-use super::orchestrator_model::OrchestratorModel;
+use super::orchestrator_model::{OrchestratorModel, OrchestratorSelection};
 
 use crate::metrics as bs_metrics;
 use crate::metrics::labels as metric_labels;
@@ -30,12 +32,27 @@ pub struct OrchestratorService {
     orchestrator_model: Arc<dyn OrchestratorModel>,
     orchestrator_provider_name: String,
     top_level_preferences: HashMap<String, TopLevelRoutingPreference>,
+    /// Agent Skills catalog (deduplicated by name) attached to any
+    /// `routing_preferences[].skills` list. Empty when no route has skills.
+    skills_catalog: Vec<SkillRef>,
     metrics_service: Option<Arc<ModelMetricsService>>,
     session_cache: Option<Arc<dyn SessionCache>>,
     session_ttl: Duration,
     tenant_header: Option<String>,
 }
 
+/// Result of `determine_route`: which route was picked, the ranked candidate
+/// models for that route, and the Agent Skill bodies the orchestrator chose
+/// to activate alongside it. Skills are resolved against
+/// `routing_preferences[<route>].skills`, so unknown / cross-route names are
+/// silently dropped.
+#[derive(Debug, Clone, Default)]
+pub struct RouteDecision {
+    pub route_name: String,
+    pub models: Vec<String>,
+    pub activated_skills: Vec<SkillRef>,
+}
+
 #[derive(Debug, Error)]
 pub enum OrchestrationError {
     #[error(transparent)]
@@ -66,6 +83,7 @@ impl OrchestratorService {
             orchestrator_model,
             orchestrator_provider_name,
             top_level_preferences: HashMap::new(),
+            skills_catalog: Vec::new(),
             metrics_service: None,
             session_cache: None,
             session_ttl: Duration::from_secs(DEFAULT_SESSION_TTL_SECONDS),
@@ -84,14 +102,53 @@ impl OrchestratorService {
         session_cache: Arc<dyn SessionCache>,
         tenant_header: Option<String>,
         max_token_length: usize,
+    ) -> Self {
+        Self::with_routing_and_skills(
+            orchestrator_url,
+            orchestration_model_name,
+            orchestrator_provider_name,
+            top_level_prefs,
+            None,
+            metrics_service,
+            session_ttl_seconds,
+            session_cache,
+            tenant_header,
+            max_token_length,
+        )
+    }
+
+    /// Like `with_routing`, but also seeds the orchestrator with a catalog of
+    /// Agent Skills referenced by `routing_preferences[].skills`. The
+    /// orchestrator gets a `<skills>` block in its system prompt and may
+    /// select zero or more skills alongside the picked route; this enables
+    /// the LLM handler to inject the chosen SKILL.md bodies into the
+    /// upstream request.
+    #[allow(clippy::too_many_arguments)]
+    pub fn with_routing_and_skills(
+        orchestrator_url: String,
+        orchestration_model_name: String,
+        orchestrator_provider_name: String,
+        top_level_prefs: Option<Vec<TopLevelRoutingPreference>>,
+        skills_catalog: Option<Vec<SkillRef>>,
+        metrics_service: Option<Arc<ModelMetricsService>>,
+        session_ttl_seconds: Option<u64>,
+        session_cache: Arc<dyn SessionCache>,
+        tenant_header: Option<String>,
+        max_token_length: usize,
     ) -> Self {
         let top_level_preferences: HashMap<String, TopLevelRoutingPreference> = top_level_prefs
             .map_or_else(HashMap::new, |prefs| {
                 prefs.into_iter().map(|p| (p.name.clone(), p)).collect()
             });
 
-        let orchestrator_model = Arc::new(orchestrator_model_v1::OrchestratorModelV1::new(
+        let skills_catalog = build_skills_catalog_for_routes(
+            skills_catalog.as_deref().unwrap_or(&[]),
+            &top_level_preferences,
+        );
+
+        let orchestrator_model = Arc::new(orchestrator_model_v1::OrchestratorModelV1::with_skills(
             HashMap::new(),
+            skills_catalog.clone(),
             orchestration_model_name,
             max_token_length,
         ));
@@ -105,6 +162,7 @@ impl OrchestratorService {
             orchestrator_model,
             orchestrator_provider_name,
             top_level_preferences,
+            skills_catalog,
             metrics_service,
             session_cache: Some(session_cache),
             session_ttl,
@@ -170,7 +228,7 @@ impl OrchestratorService {
         messages: &[Message],
         inline_routing_preferences: Option<Vec<TopLevelRoutingPreference>>,
         request_id: &str,
-    ) -> Result<Option<(String, Vec<String>)>> {
+    ) -> Result<Option<RouteDecision>> {
         if messages.is_empty() {
             return Ok(None);
         }
@@ -206,9 +264,13 @@ impl OrchestratorService {
             )
             .await?;
 
-        let result = if let Some(ref routes) = orchestration_result {
+        let result = if let Some(ref selection) = orchestration_result {
-            if routes.len() > 1 {
+            if selection.routes.len() > 1 {
-                let all_routes: Vec<&str> = routes.iter().map(|(name, _)| name.as_str()).collect();
+                let all_routes: Vec<&str> = selection
+                    .routes
+                    .iter()
+                    .map(|(name, _)| name.as_str())
+                    .collect();
                 info!(
                     routes = ?all_routes,
                     using = %all_routes.first().unwrap_or(&"none"),
@@ -216,7 +278,7 @@ impl OrchestratorService {
                 );
             }
 
-            if let Some((route_name, _)) = routes.first() {
+            if let Some((route_name, _)) = selection.routes.first() {
                 let top_pref = inline_top_map
                     .as_ref()
                     .and_then(|m| m.get(route_name))
@@ -227,7 +289,16 @@ impl OrchestratorService {
                         Some(svc) => svc.rank_models(&pref.models, &pref.selection_policy).await,
                         None => pref.models.clone(),
                     };
-                    Some((route_name.clone(), ranked))
+                    let activated_skills = resolve_activated_skills(
+                        &self.skills_catalog,
+                        pref.skills.as_deref().unwrap_or(&[]),
+                        &selection.skills,
+                    );
+                    Some(RouteDecision {
+                        route_name: route_name.clone(),
+                        models: ranked,
+                        activated_skills,
+                    })
                 } else {
                     None
                 }
@@ -239,7 +310,7 @@ impl OrchestratorService {
         };
 
         info!(
-            selected_model = ?result,
+            selected_route = ?result.as_ref().map(|r| (&r.route_name, r.models.first(), r.activated_skills.iter().map(|s| s.name.as_str()).collect::<Vec<_>>())),
             "plano-orchestrator determined route"
         );
 
@@ -253,7 +324,7 @@ impl OrchestratorService {
         messages: &[Message],
         usage_preferences: Option<Vec<AgentUsagePreference>>,
         request_id: Option<String>,
-    ) -> Result<Option<Vec<(String, String)>>> {
+    ) -> Result<Option<OrchestratorSelection>> {
         if messages.is_empty() {
             return Ok(None);
         }
@@ -328,6 +399,61 @@ impl OrchestratorService {
     }
 }
 
+/// Build the orchestrator-visible skills catalog (deduplicated by name) from
+/// the union of every skill name referenced under
+/// `routing_preferences[].skills`. Skills that are not referenced by any
+/// route are excluded — they would just clutter the prompt with no way for
+/// the orchestrator to attach them to a route.
+fn build_skills_catalog_for_routes(
+    catalog: &[SkillRef],
+    routes: &HashMap<String, TopLevelRoutingPreference>,
+) -> Vec<SkillRef> {
+    let mut referenced: std::collections::HashSet<&str> = std::collections::HashSet::new();
+    for route in routes.values() {
+        if let Some(names) = route.skills.as_ref() {
+            for name in names {
+                referenced.insert(name.as_str());
+            }
+        }
+    }
+
+    let mut out: Vec<SkillRef> = Vec::new();
+    let mut seen: std::collections::HashSet<String> = std::collections::HashSet::new();
+    for skill in catalog {
+        if referenced.contains(skill.name.as_str()) && seen.insert(skill.name.clone()) {
+            out.push(skill.clone());
+        }
+    }
+    out
+}
+
+/// Filter the orchestrator-selected skill names down to the SKILL.md bodies
+/// allowed for the chosen route, preserving the order the orchestrator
+/// returned. Unknown names (either not in the catalog or not allowed by the
+/// route) are silently dropped; the orchestrator can hallucinate.
+fn resolve_activated_skills(
+    catalog: &[SkillRef],
+    route_allowlist: &[String],
+    selected: &[String],
+) -> Vec<SkillRef> {
+    let allowed: std::collections::HashSet<&str> =
+        route_allowlist.iter().map(String::as_str).collect();
+    let mut out: Vec<SkillRef> = Vec::with_capacity(selected.len());
+    let mut taken: std::collections::HashSet<&str> = std::collections::HashSet::new();
+    for name in selected {
+        if !allowed.contains(name.as_str()) {
+            continue;
+        }
+        if !taken.insert(name.as_str()) {
+            continue;
+        }
+        if let Some(skill) = catalog.iter().find(|s| &s.name == name) {
+            out.push(skill.clone());
+        }
+    }
+    out
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

crates/brightstaff/src/router/orchestrator_model.rs

@@ -10,20 +10,37 @@ pub enum OrchestratorModelError {
 
 pub type Result<T> = std::result::Result<T, OrchestratorModelError>;
 
+/// The result of running Plano-Orchestrator over a conversation: zero or more
+/// selected routes (each mapped to its upstream model name) plus zero or more
+/// selected Agent Skills. Skills are filtered down by the consumer to the
+/// catalog defined under `routing_preferences[].skills` for the chosen route.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct OrchestratorSelection {
+    pub routes: Vec<(String, String)>,
+    pub skills: Vec<String>,
+}
+
+impl OrchestratorSelection {
+    pub fn is_empty(&self) -> bool {
+        self.routes.is_empty() && self.skills.is_empty()
+    }
+}
+
 /// OrchestratorModel trait for handling orchestration requests.
-/// Returns multiple routes as the model output format is:
+/// Returns multiple routes and skills as the model output format is:
-/// {"route": ["route_name_1", "route_name_2", ...]}
+/// {"route": ["route_name_1", ...], "skills": ["skill_name_1", ...]}
 pub trait OrchestratorModel: Send + Sync {
     fn generate_request(
         &self,
         messages: &[Message],
         usage_preferences: &Option<Vec<AgentUsagePreference>>,
     ) -> ChatCompletionsRequest;
-    /// Returns a vector of (route_name, model_name) tuples for all matched routes.
+    /// Parses the orchestrator's raw model output into selected routes (each
+    /// mapped to a model) and selected skill names.
     fn parse_response(
         &self,
         content: &str,
         usage_preferences: &Option<Vec<AgentUsagePreference>>,
-    ) -> Result<Option<Vec<(String, String)>>>;
+    ) -> Result<Option<OrchestratorSelection>>;
     fn get_model_name(&self) -> String;
 }

crates/brightstaff/src/router/orchestrator_model_v1.rs

@@ -1,12 +1,12 @@
 use std::collections::HashMap;
 
-use common::configuration::{AgentUsagePreference, OrchestrationPreference};
+use common::configuration::{AgentUsagePreference, OrchestrationPreference, SkillRef};
 use hermesllm::apis::openai::{ChatCompletionsRequest, Message, MessageContent, Role};
 use hermesllm::transforms::lib::ExtractText;
 use serde::{ser::Serialize as SerializeTrait, Deserialize, Serialize};
 use tracing::{debug, warn};
 
-use super::orchestrator_model::{OrchestratorModel, OrchestratorModelError};
+use super::orchestrator_model::{OrchestratorModel, OrchestratorModelError, OrchestratorSelection};
 
 pub const MAX_TOKEN_LEN: usize = 8192; // Default max token length for the orchestration model
 
@@ -138,10 +138,47 @@ Return your answer strictly in JSON as follows:
 If no routes are needed, return an empty list for `route`.
 "#;
 
+/// System prompt used when one or more Agent Skills are attached to candidate
+/// routes. Adds a `<skills>` block alongside `<routes>` and asks the model to
+/// also pick zero or more skills that should be loaded into the downstream
+/// LLM's system prompt.
+pub const ARCH_ORCHESTRATOR_V1_SYSTEM_PROMPT_WITH_SKILLS: &str = r#"
+You are a helpful assistant that selects the most suitable routes and Agent Skills based on user intent.
+You are provided with a list of available routes enclosed within <routes></routes> XML tags:
+<routes>
+{routes}
+</routes>
+
+You are provided with a list of available Agent Skills enclosed within <skills></skills> XML tags:
+<skills>
+{skills}
+</skills>
+
+You are also given the conversation context enclosed within <conversation></conversation> XML tags:
+<conversation>
+{conversation}
+</conversation>
+
+## Instructions
+1. Analyze the latest user intent from the conversation.
+2. Compare it against the available routes to find which routes can help fulfill the request.
+3. Independently compare it against the available skills; pick the skills whose descriptions match what the user is trying to do. Skills can be combined with any route. Activating a skill loads detailed instructions into the next response's system prompt.
+4. Respond only with exact names from <routes> and <skills>.
+5. If no routes or skills can help, return empty lists.
+
+## Response Format
+Return your answer strictly in JSON as follows:
+{{"route": ["route_name_1", "..."], "skills": ["skill_name_1", "..."]}}
+Use empty lists for `route` and/or `skills` when nothing applies.
+"#;
+
 pub type Result<T> = std::result::Result<T, OrchestratorModelError>;
 pub struct OrchestratorModelV1 {
     agent_orchestration_json_str: String,
     agent_orchestration_to_model_map: HashMap<String, String>,
+    /// Pre-rendered `<skills>` block (one JSON entry per skill, name +
+    /// description). Empty when no skills are attached to any route.
+    skills_catalog_json_str: String,
     orchestration_model: String,
     max_token_length: usize,
 }
@@ -151,10 +188,27 @@ impl OrchestratorModelV1 {
         agent_orchestrations: HashMap<String, Vec<OrchestrationPreference>>,
         orchestration_model: String,
         max_token_length: usize,
+    ) -> Self {
+        Self::with_skills(
+            agent_orchestrations,
+            Vec::new(),
+            orchestration_model,
+            max_token_length,
+        )
+    }
+
+    /// Like `new`, but additionally seeds the orchestrator with an Agent
+    /// Skills catalog. When `skills_catalog` is empty the orchestrator uses
+    /// the routes-only system prompt; otherwise it asks the model to also
+    /// pick zero or more skills from the catalog.
+    pub fn with_skills(
+        agent_orchestrations: HashMap<String, Vec<OrchestrationPreference>>,
+        skills_catalog: Vec<SkillRef>,
+        orchestration_model: String,
+        max_token_length: usize,
     ) -> Self {
         let agent_orchestration_values: Vec<OrchestrationPreference> =
             agent_orchestrations.values().flatten().cloned().collect();
-        // Format routes: each route as JSON on its own line with standard spacing
         let agent_orchestration_json_str = agent_orchestration_values
             .iter()
             .map(to_spaced_json)
@@ -165,19 +219,48 @@ impl OrchestratorModelV1 {
             .flat_map(|(model, prefs)| prefs.iter().map(|pref| (pref.name.clone(), model.clone())))
             .collect();
 
+        let skills_catalog_json_str = render_skills_catalog(&skills_catalog);
+
         OrchestratorModelV1 {
             orchestration_model,
             max_token_length,
             agent_orchestration_json_str,
             agent_orchestration_to_model_map,
+            skills_catalog_json_str,
         }
     }
 }
 
+/// JSON shape suitable for the `<skills>` block in the orchestrator prompt:
+/// `{"name": "...", "description": "..."}`. Only metadata that helps the
+/// orchestrator pick a skill is included; the full SKILL.md body is injected
+/// separately, after a skill has been selected.
+#[derive(Debug, Clone, Serialize)]
+struct SkillCatalogEntry<'a> {
+    name: &'a str,
+    description: &'a str,
+}
+
+fn render_skills_catalog(skills: &[SkillRef]) -> String {
+    skills
+        .iter()
+        .map(|s| SkillCatalogEntry {
+            name: &s.name,
+            description: s.catalog_description(),
+        })
+        .map(|entry| to_spaced_json(&entry))
+        .collect::<Vec<String>>()
+        .join("\n")
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 struct AgentOrchestratorResponse {
-    /// The route field now expects an array of route names: ["route_name_1", "route_name_2", ...]
+    /// The route field expects an array of route names: ["route_name_1", "route_name_2", ...].
     pub route: Option<Vec<String>>,
+    /// Optional array of Agent Skill names the orchestrator chose to activate.
+    /// Absent or empty when no skills should be loaded for this turn.
+    #[serde(default)]
+    pub skills: Option<Vec<String>>,
 }
 
 const TOKEN_LENGTH_DIVISOR: usize = 4; // Approximate token length divisor for UTF-8 characters
@@ -209,7 +292,13 @@ impl OrchestratorModel for OrchestratorModelV1 {
         // Ensure the conversation does not exceed the configured token budget.
         // We use `len() / TOKEN_LENGTH_DIVISOR` as a cheap token estimate to
         // avoid running a real tokenizer on the hot path.
-        let mut token_count = ARCH_ORCHESTRATOR_V1_SYSTEM_PROMPT.len() / TOKEN_LENGTH_DIVISOR;
+        let template_len = if self.skills_catalog_json_str.is_empty() {
+            ARCH_ORCHESTRATOR_V1_SYSTEM_PROMPT.len()
+        } else {
+            ARCH_ORCHESTRATOR_V1_SYSTEM_PROMPT_WITH_SKILLS.len()
+                + self.skills_catalog_json_str.len()
+        };
+        let mut token_count = template_len / TOKEN_LENGTH_DIVISOR;
         let mut selected_messages_list_reversed: Vec<Message> = vec![];
         for (selected_messsage_count, message) in messages_vec.iter().rev().enumerate() {
             let message_text = message.content.extract_text();
@@ -289,14 +378,16 @@ impl OrchestratorModel for OrchestratorModelV1 {
         // Generate the orchestrator request message based on the usage preferences.
         // If preferences are passed in request then we use them;
         // Otherwise, we use the default orchestration modelpreferences.
-        let orchestrator_message =
+        let routes_block = match convert_to_orchestrator_preferences(usage_preferences_from_request)
-            match convert_to_orchestrator_preferences(usage_preferences_from_request) {
+        {
-                Some(prefs) => generate_orchestrator_message(&prefs, &selected_conversation_list),
+            Some(prefs) => prefs,
-                None => generate_orchestrator_message(
+            None => self.agent_orchestration_json_str.clone(),
-                    &self.agent_orchestration_json_str,
-                    &selected_conversation_list,
-                ),
         };
+        let orchestrator_message = generate_orchestrator_message(
+            &routes_block,
+            &self.skills_catalog_json_str,
+            &selected_conversation_list,
+        );
 
         ChatCompletionsRequest {
             model: self.orchestration_model.clone(),
@@ -316,7 +407,7 @@ impl OrchestratorModel for OrchestratorModelV1 {
         &self,
         content: &str,
         usage_preferences: &Option<Vec<AgentUsagePreference>>,
-    ) -> Result<Option<Vec<(String, String)>>> {
+    ) -> Result<Option<OrchestratorSelection>> {
         if content.is_empty() {
             return Ok(None);
         }
@@ -326,20 +417,21 @@ impl OrchestratorModel for OrchestratorModelV1 {
 
         let selected_routes = orchestrator_response.route.unwrap_or_default();
 
-        // Filter out empty routes
         let valid_routes: Vec<String> = selected_routes
             .into_iter()
             .filter(|route| !route.is_empty())
             .collect();
 
-        if valid_routes.is_empty() {
+        let selected_skills: Vec<String> = orchestrator_response
-            return Ok(None);
+            .skills
-        }
+            .unwrap_or_default()
+            .into_iter()
+            .filter(|s| !s.is_empty())
+            .collect();
 
-        let mut result: Vec<(String, String)> = Vec::new();
+        let mut routes: Vec<(String, String)> = Vec::new();
 
         if let Some(usage_preferences) = usage_preferences {
-            // If usage preferences are defined, we need to find the model that matches each selected route
             for selected_route in valid_routes {
                 let model_name: Option<String> = usage_preferences
                     .iter()
@@ -351,7 +443,7 @@ impl OrchestratorModel for OrchestratorModelV1 {
                     .map(|pref| pref.model.clone());
 
                 if let Some(model_name) = model_name {
-                    result.push((selected_route, model_name));
+                    routes.push((selected_route, model_name));
                 } else {
                     warn!(
                         route = %selected_route,
@@ -361,14 +453,13 @@ impl OrchestratorModel for OrchestratorModelV1 {
                 }
             }
         } else {
-            // If no usage preferences are passed in request then use the default orchestration model preferences
             for selected_route in valid_routes {
                 if let Some(model) = self
                     .agent_orchestration_to_model_map
                     .get(&selected_route)
                     .cloned()
                 {
-                    result.push((selected_route, model));
+                    routes.push((selected_route, model));
                 } else {
                     warn!(
                         route = %selected_route,
@@ -379,11 +470,14 @@ impl OrchestratorModel for OrchestratorModelV1 {
             }
         }
 
-        if result.is_empty() {
+        if routes.is_empty() && selected_skills.is_empty() {
             return Ok(None);
         }
 
-        Ok(Some(result))
+        Ok(Some(OrchestratorSelection {
+            routes,
+            skills: selected_skills,
+        }))
     }
 
     fn get_model_name(&self) -> String {
@@ -391,7 +485,11 @@ impl OrchestratorModel for OrchestratorModelV1 {
     }
 }
 
-fn generate_orchestrator_message(prefs: &str, selected_conversation_list: &Vec<Message>) -> String {
+fn generate_orchestrator_message(
+    prefs: &str,
+    skills_catalog: &str,
+    selected_conversation_list: &Vec<Message>,
+) -> String {
     // Format conversation with 4-space indentation (equivalent to Python's json.dumps(obj, indent=4))
     let formatter = serde_json::ser::PrettyFormatter::with_indent(b"    ");
     let mut conversation_buf = Vec::new();
@@ -399,9 +497,19 @@ fn generate_orchestrator_message(prefs: &str, selected_conversation_list: &Vec<M
     SerializeTrait::serialize(&selected_conversation_list, &mut serializer).unwrap();
     let conversation_json = String::from_utf8(conversation_buf).unwrap_or_default();
 
+    let template = if skills_catalog.is_empty() {
         ARCH_ORCHESTRATOR_V1_SYSTEM_PROMPT
+    } else {
+        ARCH_ORCHESTRATOR_V1_SYSTEM_PROMPT_WITH_SKILLS
+    };
+
+    let mut out = template
         .replace("{routes}", prefs)
-        .replace("{conversation}", &conversation_json)
+        .replace("{conversation}", &conversation_json);
+    if !skills_catalog.is_empty() {
+        out = out.replace("{skills}", skills_catalog);
+    }
+    out
 }
 
 fn convert_to_orchestrator_preferences(
@@ -1349,23 +1457,30 @@ If no routes are needed, return an empty list for `route`.
         let orchestrator =
             OrchestratorModelV1::new(agent_orchestrations, "test-model".to_string(), 2000);
 
+        fn routes(pairs: &[(&str, &str)]) -> OrchestratorSelection {
+            OrchestratorSelection {
+                routes: pairs
+                    .iter()
+                    .map(|(r, m)| (r.to_string(), m.to_string()))
+                    .collect(),
+                skills: Vec::new(),
+            }
+        }
+
         // Case 1: Valid JSON with single route in array
         let input = r#"{"route": ["Image generation"]}"#;
         let result = orchestrator.parse_response(input, &None).unwrap();
-        assert_eq!(
+        assert_eq!(result, Some(routes(&[("Image generation", "gpt-4o")])));
-            result,
-            Some(vec![("Image generation".to_string(), "gpt-4o".to_string())])
-        );
 
         // Case 2: Valid JSON with multiple routes in array
         let input = r#"{"route": ["Image generation", "Code generation"]}"#;
         let result = orchestrator.parse_response(input, &None).unwrap();
         assert_eq!(
             result,
-            Some(vec![
+            Some(routes(&[
-                ("Image generation".to_string(), "gpt-4o".to_string()),
+                ("Image generation", "gpt-4o"),
-                ("Code generation".to_string(), "gpt-4o".to_string())
+                ("Code generation", "gpt-4o")
-            ])
+            ]))
         );
 
         // Case 3: Valid JSON with empty array
@@ -1396,14 +1511,65 @@ If no routes are needed, return an empty list for `route`.
         // Case 7: Single quotes and \n in JSON
         let input = "{'route': ['Image generation']}\\n";
         let result = orchestrator.parse_response(input, &None).unwrap();
-        assert_eq!(
+        assert_eq!(result, Some(routes(&[("Image generation", "gpt-4o")])));
-            result,
-            Some(vec![("Image generation".to_string(), "gpt-4o".to_string())])
-        );
 
         // Case 8: Array with unknown route (not in orchestrations map)
         let input = r#"{"route": ["Unknown route"]}"#;
         let result = orchestrator.parse_response(input, &None).unwrap();
         assert_eq!(result, None);
+
+        // Case 9: Routes plus selected skills are propagated through.
+        let input = r#"{"route": ["Image generation"], "skills": ["pdf-processing"]}"#;
+        let result = orchestrator.parse_response(input, &None).unwrap().unwrap();
+        assert_eq!(result.routes.len(), 1);
+        assert_eq!(result.skills, vec!["pdf-processing".to_string()]);
+
+        // Case 10: Skills-only selection (no routes) still surfaces as Some.
+        let input = r#"{"route": [], "skills": ["pdf-processing"]}"#;
+        let result = orchestrator.parse_response(input, &None).unwrap().unwrap();
+        assert!(result.routes.is_empty());
+        assert_eq!(result.skills, vec!["pdf-processing".to_string()]);
+    }
+
+    #[test]
+    fn test_system_prompt_with_skills_block() {
+        let orchestrator = OrchestratorModelV1::with_skills(
+            HashMap::from([(
+                "gpt-4o".to_string(),
+                vec![OrchestrationPreference {
+                    name: "Image generation".to_string(),
+                    description: "generating image".to_string(),
+                }],
+            )]),
+            vec![SkillRef {
+                name: "pdf-processing".to_string(),
+                description: "Extract structured data from PDFs.".to_string(),
+                path: None,
+                base_dir: None,
+                body: None,
+                scope: None,
+                compatibility: None,
+                license: None,
+                metadata: None,
+                allowed_tools: None,
+            }],
+            "test-model".to_string(),
+            usize::MAX,
+        );
+
+        let conversation: Vec<Message> = serde_json::from_str(
+            r#"[{"role": "user", "content": "extract the invoice totals from this pdf"}]"#,
+        )
+        .unwrap();
+        let req = orchestrator.generate_request(&conversation, &None);
+        let prompt = req.messages[0].content.extract_text();
+
+        assert!(prompt.contains("<skills>"));
+        assert!(prompt.contains("</skills>"));
+        assert!(prompt.contains(r#""name": "pdf-processing""#));
+        assert!(prompt.contains("Extract structured data from PDFs."));
+        // Response format documentation must mention the `skills` array
+        // so the orchestrator emits it.
+        assert!(prompt.contains(r#""skills""#));
     }
 }

crates/brightstaff/src/router/stress_tests.rs

@@ -33,6 +33,7 @@ mod tests {
                     "openai/gpt-4o".to_string(),
                     "openai/gpt-4o-mini".to_string(),
                 ],
+                skills: None,
                 selection_policy: SelectionPolicy {
                     prefer: SelectionPreference::None,
                 },
@@ -44,6 +45,7 @@ mod tests {
                     "anthropic/claude-3-sonnet".to_string(),
                     "openai/gpt-4o-mini".to_string(),
                 ],
+                skills: None,
                 selection_policy: SelectionPolicy {
                     prefer: SelectionPreference::None,
                 },

crates/common/src/configuration.rs

@@ -163,6 +163,12 @@ pub struct TopLevelRoutingPreference {
     pub name: String,
     pub description: String,
     pub models: Vec<String>,
+    /// Agent Skills associated with this route. When Plano-Orchestrator
+    /// selects this route, every skill listed here is also offered to the
+    /// orchestrator in the `<skills>` block; selected skills have their
+    /// SKILL.md body prepended to the upstream system prompt.
+    #[serde(default)]
+    pub skills: Option<Vec<String>>,
     #[serde(default)]
     pub selection_policy: SelectionPolicy,
 }
@@ -224,6 +230,17 @@ pub struct Configuration {
     pub state_storage: Option<StateStorageConfig>,
     pub routing_preferences: Option<Vec<TopLevelRoutingPreference>>,
     pub model_metrics_sources: Option<Vec<MetricsSource>>,
+    /// Agent Skills (https://agentskills.io) installed for this project.
+    ///
+    /// The Plano CLI discovers `.plano/skills/<name>/SKILL.md` files at render
+    /// time and materializes them into this list with `body` already loaded so
+    /// downstream consumers do not need filesystem access. Skills are scoped
+    /// to specific routes via `routing_preferences[].skills`; Plano-Orchestrator
+    /// receives a `<skills>` block alongside `<routes>` for any skills attached
+    /// to candidate routes, and selected skills have their SKILL.md body
+    /// injected into the upstream system prompt.
+    #[serde(default)]
+    pub skills: Option<Vec<SkillRef>>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize, Default)]
@@ -611,6 +628,45 @@ pub struct PromptTarget {
     pub auto_llm_dispatch_on_response: Option<bool>,
 }
 
+/// An Agent Skill (https://agentskills.io) as materialized by the Plano CLI.
+///
+/// At runtime brightstaff and the WASM filters reason over the catalog
+/// (`name` + `description`) and, when a skill is selected, inject the
+/// pre-loaded `body` into the downstream system prompt.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+pub struct SkillRef {
+    pub name: String,
+    pub description: String,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub path: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub base_dir: Option<String>,
+    /// Full SKILL.md markdown body (post-frontmatter). Inlined here at render
+    /// time so the WASM sandbox does not need filesystem access.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub body: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub scope: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub compatibility: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub license: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub metadata: Option<HashMap<String, String>>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub allowed_tools: Option<String>,
+}
+
+impl SkillRef {
+    /// Best-effort short summary suitable for the `<skills>` block sent to
+    /// Plano-Orchestrator: only the public-facing description, never the
+    /// full SKILL.md body. The body is injected separately, after a skill
+    /// has been selected.
+    pub fn catalog_description(&self) -> &str {
+        &self.description
+    }
+}
+
 // convert PromptTarget to ChatCompletionTool
 impl From<&PromptTarget> for ChatCompletionTool {
     fn from(val: &PromptTarget) -> Self {
@@ -807,4 +863,34 @@ disable_signals: false
         let overrides: super::Overrides = serde_yaml::from_str(yaml_missing).unwrap();
         assert_eq!(overrides.disable_signals, None);
     }
+
+    #[test]
+    fn test_top_level_routing_preference_skills_deserialize() {
+        let yaml = r#"
+name: code review
+description: reviewing, analyzing, and suggesting improvements to existing code
+models:
+  - openai/gpt-4o
+skills:
+  - code-review-skill
+"#;
+        let pref: super::TopLevelRoutingPreference = serde_yaml::from_str(yaml).unwrap();
+        assert_eq!(pref.name, "code review");
+        assert_eq!(
+            pref.skills.as_deref(),
+            Some(&["code-review-skill".to_string()][..])
+        );
+    }
+
+    #[test]
+    fn test_top_level_routing_preference_skills_optional() {
+        let yaml = r#"
+name: code generation
+description: generating new code
+models:
+  - openai/gpt-4o
+"#;
+        let pref: super::TopLevelRoutingPreference = serde_yaml::from_str(yaml).unwrap();
+        assert!(pref.skills.is_none());
+    }
 }

crates/common/src/lib.rs

@@ -8,6 +8,7 @@ pub mod path;
 pub mod pii;
 pub mod ratelimit;
 pub mod routing;
+pub mod skills_runtime;
 pub mod stats;
 pub mod tokenizer;
 pub mod traces;

crates/common/src/skills_runtime.rs

@@ -0,0 +1,215 @@
+//! Runtime helpers for handling Agent Skills selected by Plano-Orchestrator.
+//!
+//! These functions live in `common` (rather than `brightstaff` or a WASM
+//! crate) so they can be unit-tested on the native target without dragging
+//! in proxy-wasm host-call symbols or tokio runtime dependencies.
+
+use crate::configuration::{SkillRef, TopLevelRoutingPreference};
+
+/// Filter `skills` down to the subset attached to `route_name` via
+/// `routing_preferences[].skills`. When the selected route has no `skills:`
+/// list, returns an empty vector — skills are scoped to routes, not global.
+///
+/// `routing_preferences` is the source of truth for which skills are even
+/// eligible for the orchestrator to activate on a given route.
+pub fn skills_for_route<'a>(
+    skills: &'a [SkillRef],
+    routing_preferences: &[TopLevelRoutingPreference],
+    route_name: &str,
+) -> Vec<&'a SkillRef> {
+    let Some(route) = routing_preferences.iter().find(|p| p.name == route_name) else {
+        return Vec::new();
+    };
+    let Some(allow) = route.skills.as_ref() else {
+        return Vec::new();
+    };
+    let mut out: Vec<&SkillRef> = Vec::with_capacity(allow.len());
+    for name in allow {
+        if let Some(skill) = skills.iter().find(|s| &s.name == name) {
+            out.push(skill);
+        }
+    }
+    out
+}
+
+/// Resolve a list of orchestrator-selected skill names to their `SkillRef`s.
+/// Unknown names are dropped silently — the orchestrator can hallucinate.
+/// Results are deduplicated by name, preserving the order Plano-Orchestrator
+/// returned.
+pub fn resolve_selected_skills<'a>(
+    skills: &'a [SkillRef],
+    selected_names: &[String],
+) -> Vec<&'a SkillRef> {
+    let mut out: Vec<&SkillRef> = Vec::with_capacity(selected_names.len());
+    for name in selected_names {
+        if out.iter().any(|s| &s.name == name) {
+            continue;
+        }
+        if let Some(skill) = skills.iter().find(|s| &s.name == name) {
+            out.push(skill);
+        }
+    }
+    out
+}
+
+/// Append the bodies of activated skills to a system prompt, wrapped in
+/// `<skill_content name="...">` tags so a future context-management pass can
+/// recognize and recompact them.
+///
+/// Returns `None` only if no base system prompt was supplied and no skills
+/// were activated. When skills are present the wrapper text always appears so
+/// the downstream model receives a clear, well-structured instruction block.
+pub fn augment_system_prompt_with_skills(
+    base_system_prompt: Option<String>,
+    activated_skills: &[&SkillRef],
+) -> Option<String> {
+    if activated_skills.is_empty() {
+        return base_system_prompt;
+    }
+    let mut buf = String::new();
+    if let Some(base) = base_system_prompt {
+        if !base.is_empty() {
+            buf.push_str(&base);
+            buf.push('\n');
+            buf.push('\n');
+        }
+    }
+    buf.push_str(
+        "The following Agent Skills have been activated for this request. \
+         Follow their instructions when relevant; resolve relative paths \
+         against each skill's base directory.\n\n",
+    );
+    for skill in activated_skills {
+        buf.push_str(&format!("<skill_content name=\"{}\"", skill.name));
+        if let Some(base_dir) = skill.base_dir.as_deref() {
+            buf.push_str(&format!(" base_dir=\"{}\"", base_dir));
+        }
+        buf.push_str(">\n");
+        if let Some(body) = skill.body.as_deref() {
+            buf.push_str(body.trim_end());
+            buf.push('\n');
+        } else {
+            buf.push_str(&format!("(skill description) {}\n", skill.description));
+        }
+        buf.push_str("</skill_content>\n\n");
+    }
+    Some(buf.trim_end().to_string())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::configuration::SelectionPolicy;
+
+    fn skill(name: &str, body: &str) -> SkillRef {
+        SkillRef {
+            name: name.to_string(),
+            description: format!("desc for {}", name),
+            path: Some(format!("/skills/{}/SKILL.md", name)),
+            base_dir: Some(format!("/skills/{}", name)),
+            body: Some(body.to_string()),
+            scope: Some("project".to_string()),
+            compatibility: None,
+            license: None,
+            metadata: None,
+            allowed_tools: None,
+        }
+    }
+
+    fn route(name: &str, skill_names: Option<Vec<&str>>) -> TopLevelRoutingPreference {
+        TopLevelRoutingPreference {
+            name: name.to_string(),
+            description: format!("desc for {}", name),
+            models: vec!["openai/gpt-4o".to_string()],
+            skills: skill_names.map(|v| v.into_iter().map(String::from).collect()),
+            selection_policy: SelectionPolicy::default(),
+        }
+    }
+
+    #[test]
+    fn skills_for_route_returns_attached_skills() {
+        let catalog = vec![
+            skill("pdf-processing", "extract"),
+            skill("code-review", "review"),
+        ];
+        let routes = vec![
+            route("code review", Some(vec!["code-review"])),
+            route("doc work", Some(vec!["pdf-processing"])),
+        ];
+        let resolved = skills_for_route(&catalog, &routes, "code review");
+        assert_eq!(resolved.len(), 1);
+        assert_eq!(resolved[0].name, "code-review");
+    }
+
+    #[test]
+    fn skills_for_route_empty_when_route_has_no_skills_list() {
+        let catalog = vec![skill("pdf-processing", "extract")];
+        let routes = vec![route("code review", None)];
+        let resolved = skills_for_route(&catalog, &routes, "code review");
+        assert!(resolved.is_empty());
+    }
+
+    #[test]
+    fn skills_for_route_empty_when_route_missing() {
+        let catalog = vec![skill("pdf-processing", "extract")];
+        let routes = vec![route("code review", Some(vec!["pdf-processing"]))];
+        let resolved = skills_for_route(&catalog, &routes, "no-such-route");
+        assert!(resolved.is_empty());
+    }
+
+    #[test]
+    fn skills_for_route_drops_unknown_skill_names() {
+        let catalog = vec![skill("pdf-processing", "extract")];
+        let routes = vec![route(
+            "code review",
+            Some(vec!["pdf-processing", "ghost-skill"]),
+        )];
+        let resolved = skills_for_route(&catalog, &routes, "code review");
+        assert_eq!(resolved.len(), 1);
+        assert_eq!(resolved[0].name, "pdf-processing");
+    }
+
+    #[test]
+    fn resolve_selected_skills_drops_unknown_and_dedupes() {
+        let catalog = vec![
+            skill("pdf-processing", "extract"),
+            skill("code-review", "review"),
+        ];
+        let names = vec![
+            "code-review".to_string(),
+            "ghost".to_string(),
+            "code-review".to_string(),
+            "pdf-processing".to_string(),
+        ];
+        let resolved = resolve_selected_skills(&catalog, &names);
+        assert_eq!(resolved.len(), 2);
+        assert_eq!(resolved[0].name, "code-review");
+        assert_eq!(resolved[1].name, "pdf-processing");
+    }
+
+    #[test]
+    fn augment_passthrough_with_no_skills() {
+        let augmented = augment_system_prompt_with_skills(Some("you are helpful".to_string()), &[]);
+        assert_eq!(augmented.as_deref(), Some("you are helpful"));
+    }
+
+    #[test]
+    fn augment_includes_skill_bodies() {
+        let s = skill("pdf-processing", "extract text and tables");
+        let augmented =
+            augment_system_prompt_with_skills(Some("you are helpful".to_string()), &[&s])
+                .expect("augmented");
+        assert!(augmented.starts_with("you are helpful"));
+        assert!(augmented.contains("<skill_content name=\"pdf-processing\""));
+        assert!(augmented.contains("extract text and tables"));
+        assert!(augmented.contains("base_dir=\"/skills/pdf-processing\""));
+    }
+
+    #[test]
+    fn augment_without_base_prompt_still_works() {
+        let s = skill("code-review", "look at diffs");
+        let augmented = augment_system_prompt_with_skills(None, &[&s]).expect("augmented");
+        assert!(augmented.contains("<skill_content name=\"code-review\""));
+        assert!(augmented.contains("look at diffs"));
+    }
+}

crates/prompt_gateway/src/stream_context.rs

@@ -209,20 +209,16 @@ impl StreamContext {
             } else {
                 info!("no default prompt target found, forwarding request to upstream llm");
                 let mut messages = Vec::new();
-                // add system prompt
+                if let Some(system_prompt) = self.system_prompt.as_ref().clone() {
-                match self.system_prompt.as_ref() {
-                    None => {}
-                    Some(system_prompt) => {
                     let system_prompt_message = Message {
                         role: SYSTEM_ROLE.to_string(),
-                            content: Some(ContentType::Text(system_prompt.clone())),
+                        content: Some(ContentType::Text(system_prompt)),
                         model: None,
                         tool_calls: None,
                         tool_call_id: None,
                     };
                     messages.push(system_prompt_message);
                 }
-                }
 
                 messages.append(
                     &mut self

docs/source/index.rst

@@ -63,4 +63,5 @@ Built by contributors to the widely adopted `Envoy Proxy <https://www.envoyproxy
       resources/deployment
       resources/configuration_reference
       resources/cli_reference
+      resources/skills
       resources/llms_txt

docs/source/resources/skills.rst

@@ -0,0 +1,177 @@
+.. _agent_skills:
+
+Agent Skills
+============
+
+Plano can load `Agent Skills <https://agentskills.io>`_ — lightweight,
+markdown-defined capabilities — and let Plano-Orchestrator decide *per request*
+which skills to attach to the downstream LLM call. Skills attach to entries in
+``routing_preferences``: when the orchestrator picks a route, it also picks
+zero or more skills from that route's allow-list, and brightstaff injects
+each selected ``SKILL.md`` body into the upstream system prompt before
+forwarding the request.
+
+Why use this?
+-------------
+
+- **Modular instructions.** Ship a skill (markdown + scripts + assets) rather
+  than baking 500-token instructions into every system prompt.
+- **Progressive disclosure.** Skill names and one-line descriptions are
+  always visible to the orchestrator; full instructions load only when a
+  skill is activated.
+- **Per-route scoping.** A ``skills:`` list on a ``routing_preferences``
+  entry constrains which skills can be activated for that route.
+
+Install a skill
+---------------
+
+.. code-block:: bash
+
+   # via the upstream Agent Skills CLI (recommended for multi-skill repos)
+   planoai skills add openai/skills
+
+   # planoai falls back to a direct git clone if `npx` is unavailable; this
+   # path expects a single-skill repo with a SKILL.md at the root.
+   planoai skills add owner/code-review
+
+Where do skills end up?
+~~~~~~~~~~~~~~~~~~~~~~~
+
+Plano looks for skills across three scopes (highest precedence first):
+
+============== ============================== =========================== =====================================
+Scope          Location                       Trust required              Typical installer
+============== ============================== =========================== =====================================
+``project``    ``.plano/skills/<name>/``      Yes — ``planoai skills      ``planoai skills add`` (git fallback)
+                                              trust``
+``user``       ``~/.plano/skills/<name>/``    No (auto-trusted)           manual
+``agents``     ``~/.agents/skills/<name>/``   No (auto-trusted)           ``npx skills add`` / upstream CLI
+============== ============================== =========================== =====================================
+
+The ``agents`` scope is the universal Agent Skills install location used by
+``npx skills add`` (see https://github.com/vercel-labs/add-skill). Because
+``npx skills add`` doesn't know about Plano, it never writes into
+``.plano/skills/``; instead it drops the skill under ``~/.agents/skills/<name>``
+and symlinks it into every recognised agent (Claude Code, Cursor, …). Plano
+treats that directory as an auto-trusted user-tier scope, so anything
+installed there is picked up automatically — no ``planoai skills trust``
+needed.
+
+A ``.plano/skills/.skills.json`` manifest is maintained only for installs
+that land in project scope (the git fallback). The ``agents`` scope owns
+its own bookkeeping in ``~/.agents/``.
+
+Trust the project
+-----------------
+
+Project-level skills are loaded only after you mark the project trusted. This
+matches the recommendation in the `Adding skills support guide
+<https://agentskills.io/client-implementation/adding-skills-support.md>`_:
+
+.. code-block:: bash
+
+   planoai skills trust
+
+   # revoke trust later if needed
+   planoai skills trust --revoke
+
+Skills under ``~/.plano/skills/`` and ``~/.agents/skills/`` are always
+trusted and ignore this setting.
+
+Discover and remove
+-------------------
+
+.. code-block:: bash
+
+   planoai skills list
+   planoai skills remove pdf-processing
+
+Configure routing
+-----------------
+
+Reference installed skills from your ``config.yaml`` in two places:
+
+1. The top-level ``skills:`` catalog (optional — omit to auto-include every
+   discovered skill).
+2. Each ``routing_preferences`` entry that should make a skill eligible for
+   activation. The orchestrator's ``<skills>`` block is built from the union
+   of every ``routing_preferences[].skills`` list; skills not referenced by
+   any route are silently dropped.
+
+.. code-block:: yaml
+
+   skills:
+     - pdf-processing
+     - code-review
+
+   routing_preferences:
+     - name: code review
+       description: |
+         Reviewing pull requests, analyzing diffs, and suggesting improvements
+         to existing code.
+       models:
+         - anthropic/claude-sonnet-4-5
+         - openai/gpt-4.1-2025-04-14
+       skills:
+         - code-review
+
+     - name: document understanding
+       description: |
+         Summarizing PDFs and other long-form documents, extracting structured
+         data such as tables, line items, or signatures.
+       models:
+         - anthropic/claude-sonnet-4-5
+       skills:
+         - pdf-processing
+       selection_policy:
+         prefer: cheapest
+
+When ``planoai up`` runs, the CLI walks ``.plano/skills/`` and
+``~/.plano/skills/``, parses each ``SKILL.md``, and inlines the markdown body
+into the rendered Plano config so the brightstaff orchestrator can attach it
+to the request without any filesystem access.
+
+How routing works
+-----------------
+
+At request time:
+
+1. The brightstaff routing service builds an ``<skills>`` block in the
+   Plano-Orchestrator prompt — alongside the existing ``<routes>`` block —
+   listing every skill referenced by ``routing_preferences[].skills`` with
+   its name and short description.
+2. The orchestrator replies with JSON of the form
+   ``{"route": ["..."], "skills": ["..."]}``.
+3. brightstaff resolves each selected skill name against the chosen route's
+   ``skills:`` allow-list. Names that aren't allowed for that route (or are
+   not in the catalog) are dropped.
+4. The activated ``SkillRef`` bodies are prepended to the upstream request's
+   system prompt — wrapped in
+   ``<skill_content name="..." base_dir="...">…</skill_content>`` tags — and
+   the request is forwarded to the chosen model.
+
+If the orchestrator picks only skills and no route, the request falls back
+to the originally-requested model (or the default) and the skill bodies are
+injected the same way.
+
+Bootstrap from the template
+---------------------------
+
+A ready-made template wires the moving pieces together:
+
+.. code-block:: bash
+
+   planoai init --template skills_routing
+
+Out of scope
+------------
+
+- Hot-reload of ``.plano/skills/`` while Plano is running — re-run
+  ``planoai up`` to pick up new skills.
+- Server-side execution of bundled ``scripts/`` from skills. The upstream
+  client runs scripts as part of the progressive-disclosure model from the
+  `specification <https://agentskills.io/specification.md>`_.
+- Subagent delegation per skill. See the
+  `client-implementation guide
+  <https://agentskills.io/client-implementation/adding-skills-support.md>`_
+  for the advanced pattern.