feat(mason): support Unity Catalog agent skills

[junchoi-db]

What

Adds declarative Unity Catalog Skills support to Mason's LangGraph template:

• mason skills list --schema catalog.schema discovers UC skills through the workspace Skills endpoint.
• mason skills add uc catalog.schema.skill --source . writes an exact, idempotent [[skills]] binding to agent.toml.
• Graph construction injects skill name/description metadata only. load_skill fetches instructions lazily, and read_skill_file fetches referenced files lazily through get_skill_files.
• The same binding is used in local development and Databricks Apps; skill bodies and references are never copied into the project or deployment.

[[skills]]
id = "review"
source = { kind = "uc", name = "catalog.schema.skill" }

Why

Custom agents need a declarative skill contract with progressive disclosure while keeping skill content governed in Unity Catalog and fresh at invocation time.

Scope and limits

• Unity Catalog skills only. Local, custom, and path-based skill sources are rejected.
• LangGraph template only; nonempty OpenAI skill declarations fail before dev/deploy side effects.
• Exact three-part UC names only; versions, automatic discovery, and local file copying are not supported.
• Prompt injection is metadata-only. Instruction and reference payloads are fetched remotely and bounded to 1 MiB of UTF-8 text.
• A deployed App principal needs USE_CATALOG, USE_SCHEMA, and READ_VOLUME on each bound
  skill. The legacy READ_SKILL privilege does not authorize metadata or bundle reads.

CLI customer journey

MASON=/path/to/databricks-ai-bridge/integrations/mason/.venv/bin/mason
PROJECT=/tmp/mason-uc-skill-cuj

"$MASON" --profile <workspace-profile> init --framework langgraph \
  --profile <workspace-profile> --repo /path/to/databricks-ai-bridge \
  --ref feat/mason-agent-skills "$PROJECT"

"$MASON" --profile <workspace-profile> skills list \
  --schema catalog.schema
"$MASON" --profile <workspace-profile> skills add uc \
  catalog.schema.skill --source "$PROJECT"

"$MASON" --profile <workspace-profile> dev --source "$PROJECT"
"$MASON" --profile <workspace-profile> deploy <unique-app-name> \
  --source "$PROJECT"

Verification

• Isolated Mason base-install suite: 223 passed on the rebased head.
• Ruff lint/format and ty check: passed.
• Live CLI/local-agent CUJ on product SHA d821203: list, idempotent add, exact manifest binding,
  no local file copy, lazy body loading, and lazy referenced-file loading passed. After the final
  conflict-free rebase, git range-diff maps that product patch identically to 9163537.
• Clean deployed-App CUJ: body prompt called load_skill({"skill_id":"uc-cuj"}); reference prompt
  called load_skill then
  read_skill_file({"skill_id":"uc-cuj","path":"references/proof.txt"}). Both completed with
  HTTP 200 and exact proof markers after granting the App principal READ_VOLUME.
• The generated agent/agent.py was byte-identical to the template before the final deployment;
  the deployment source contained no copied skill body or reference file.
• Temporary App, workspace deployment source, and UC skill were deleted and verified absent.
• Internal CLI/CUJ HTML report SHA-256:
  b944c5b7c340e6964d95d799185a9786fcf6c89165d1843ea1bedf68f2dba891 (byte-identical publish
  round trip). The reproducible commands and decisive proof are included above without exposing
  private workspace/report URLs in this public PR.

Design reference

• Anthropic managed-agent skills

This pull request and its description were written by Isaac.

[junchoi-db]

what is this limit for