decimalai 0.16.0 — native mode

Added: native mode — a framework's own skill loader, seen by DecimalAI

Eight frameworks load the skills decimalai skills add installed with a loader of
their own. Each now has a one-line helper that points that loader at
.agents/skills, and DecimalAI records every skill the loader hands the model as an
activation carrying the exact skill id and version id of the bytes the model
received. Each helper is in its framework's module (import decimalai.adk first):

framework	helper
Claude Agent SDK	ClaudeAgentOptions(plugins=[decimalai.claude_agent_sdk.skills_plugin()])
Google ADK	SkillToolset(skills=decimalai.adk.skills(), ...)
CrewAI	Agent(skills=decimalai.crewai.skills(), tools=decimalai.crewai.skill_tools())
OpenAI Agents (sandbox)	SandboxAgent(capabilities=[Shell(), decimalai.openai_agents.sandbox_skills()], ...)
Deep Agents	create_deep_agent(model, **decimalai.deepagents.skills())
Microsoft Agent Framework	Agent(..., context_providers=[decimalai.agent_framework.skills()])
Pydantic AI harness	Agent(model, capabilities=[LocalWorkspace(project), decimalai.pydantic_ai_harness.skills(project), FileSystem(), Shell()])
Strands	Agent(plugins=[decimalai.strands.skills()])

• The version is the one the model received, not the one on disk. ADK,
  Microsoft Agent Framework and Strands keep serving a skill they already loaded after
  decimalai skills update wrote a new version; the activation then names the old
  version, because that is what the model read. The version comes from the delivered
  copy's stamp, else the lock by the body's hash, else any version this process has
  seen. A body that is none of them is reported with its skill and hash and no
  version, and the platform resolves it if it can.
• A skill loaded twice in one run is one activation.
• The router stays out of an agent whose own loader serves its skills (ADK with a
  SkillToolset, CrewAI with skills=, an OpenAI SandboxAgent with Skills, Deep
  Agents, a Pydantic AI agent with the harness's Skills): no menu, no
  load_skill of DecimalAI's, so the model sees one catalog. Before, the OpenAI
  adapter renamed its own tool to decimalai_load_skill beside the sandbox's, and
  CrewAI warned that two catalogs reached the model.
• Gap-fillers. decimalai.crewai.skill_tools() gives CrewAI the resource reader and
  script runner it lacks (confined to installed skills; scripts=False leaves the
  runner out). decimalai.agent_framework.skills() adds the script runner Microsoft
  Agent Framework does not ship, and runs skill loads and reads without asking for
  approval (scripts still ask unless auto_approve_scripts=True).
  sandbox_skills() answers an unknown skill name instead of aborting the run, and
  warns when the sandbox would list no skills (it skips symlinked folders silently).
• Microsoft Agent Framework and Strands need no adapter: with decimalai.init(otel=True)
  their own OpenTelemetry spans become the trace, and the load is recorded on it.
• The Pydantic AI harness is graded at 0.57.0 (the conformance-tests floor), which keeps a
  user-only skill (disable-model-invocation: true) out of the model's catalog, as only
  Claude Code did before.

Deprecated (unchanged, and kept one release longer)

The 0.14 disk verbs (decimalai skills export <ref>, decimalai skills install <ref>,
SkillRouter.install(), install_skill(), pull_missing(), export_to_disk() and the
disk_export writers) and the version 1 lock fallback still work in 0.16.0, with the same
warning. The 0.15.0 notes said they would stop working in 0.16; they now stop working in
0.17, and the warnings say so.

Fixed

• decimalai.flush() now also exports the spans the init(otel=True) pipeline still
  holds. It drained only CrewAI's, so a short script could exit before a run traced
  through OpenTelemetry became a trace.
• The OpenTelemetry exporter joins each tool span's result onto the model's tool call
  (by gen_ai.tool.call.id), so a trace shows what a tool returned even when its
  input, cut at 200 characters, does not.