Coverage

This matrix comes from the mapping registry, the same code that does the translation, so it cannot drift from it. When a manifest uses a section that is not fully supported, /v1/translate and /reconcile return a gap for it with the reason below.

Status	Means
supported	Translates with the same behaviour
emulated	Reaches the same goal by a different mechanism
partial	Translates, with a named difference
unsupported	Dropped, and reported

Agent

Identity, prompt and model

Section	Status	What happens
identity	supported	The Agent Engine: display name, description, labels, the ADK config
identity.other	unsupported	Other identity keys (an avatar, say) have no place on an Agent Engine
prompts	partial	The system prompt becomes the instruction. A prompt with braces goes out literally: Astromesh renders it as Jinja2, the native agent does not render it. Named templates are dropped
model	partial	A Gemini model through litellm (vertex_ai/gemini-…, gemini/gemini-…) translates. Any other model is reported: Vertex would not serve it
model.fallback	unsupported	The native agent calls only the first candidate
model.routing	unsupported	Routing strategies and per-role models are dropped: one model answers every role
model.parameters	partial	Sampling settings go to generate_content_config; any other parameter is named

Orchestration

Section	Status	What happens
orchestration	partial	react translates. pipeline is emulated with a SequentialAgent. plan_and_execute runs as a plain tool-calling loop (ADK configs have no planner). parallel_fan_out, supervisor, swarm and glyph are unsupported
orchestration.limits	unsupported	max_iterations, timeout, narrate, max_repairs: an ADK config has no iteration cap or timeout (ADK’s max_llm_calls, 500 by default, is set by the caller)

Memory and knowledge

Section	Status	What happens
memory.conversational	partial	Agent Engine Sessions. They send the whole history every turn, so max_turns-style tuning is not applied and token use grows with the conversation
memory.semantic	partial	The engine’s Memory Bank. After each turn an LLM extracts facts from the session, instead of storing embeddings of raw turns. similarity_threshold and max_results are not applied
memory.episodic	unsupported	No Agent Engine homolog; the native agent has no event log
memory.backend	partial	Declared backends (redis, postgres…) are replaced by Sessions and Memory Bank
knowledge	emulated	RAG Engine retrieval over the pipeline’s corpus. Astromesh injects context every turn; the native agent retrieves only when the model calls rag_query. Without knowledge.top_k, Vertex picks the count

Tools

Section	Status	What happens
tools.client	supported	A long-running tool the caller answers
tools.builtin.grounded	emulated	web_search becomes Google Search grounding: a server-side capability the loop does not see as a tool call, billed per query. Unsupported on non-Gemini models
tools.builtin.unmapped	unsupported	Other builtins run Astromesh Python with no GCP counterpart
tools.agent	unsupported	Calling another agent would need that agent inlined as a sub-agent, and each mapper sees one manifest at a time
tools.catalog	unsupported	integration, api and mcp tools resolve credentials per run inside Astromesh, and a writes: true one proposes instead of executing. An ADK toolset would execute it, turning a proposal into an action
tools.unloadable	unsupported	Types the core’s loader skips (internal, webhook, rag, mcp_*…); a tool without type counts as internal
tools.rate_limit	unsupported	The native agent has no per-tool limit
tools.other	unsupported	Any unrecognized type

Guardrails

Section	Status	What happens
guardrails.pii_detection	partial	A Model Armor template with Sensitive Data Protection, on the side the rule is declared
guardrails.pii_behavior	partial	Model Armor blocks a message with PII; Astromesh redacts it and lets it through. ADK applies Model Armor only to Gemini models
guardrails.content_filter	unsupported	Regex replacement with [FILTERED]; Model Armor classifies by category and has no pattern replacement
guardrails.topic_filter	unsupported	The GCP homolog, a semantic governance policy, requires registering the agent in Agent Registry, which Prisma does not do yet
guardrails.max_length	unsupported	No configurable GCP homolog
guardrails.cost_limit	unsupported	max_output_tokens is the closest knob, but it caps every model call, tool calls included, so it is not set
guardrails.other	unsupported	Any unrecognized guardrail type

Observability and the rest

Section	Status	What happens
observability.tracing	supported	Agent Engine telemetry, turned on with a deploy-time flag
observability.metrics	unsupported	Agent Engine exposes request count, latency and CPU/memory; token and cost metrics have no built-in
observability.otlp_export	unsupported	Exporting to an arbitrary OTLP endpoint is not a supported GCP capability
observability.other	unsupported	Unrecognized keys
spec.other	unsupported	Any spec section no row above maps (for example chain, output_schema, prefetch)

RAGPipeline

Section	Status	What happens
corpus	supported	A RAG Engine corpus, <name>-corpus
embeddings	partial	Only Vertex-managed embedding models are accepted; another provider is reported
reranking	partial	RAG Engine’s ranking differs from Astromesh reranking
vector_store	partial	Replaced by RAG Engine’s managed store

Note

The contract test copies the core’s agent.schema.json (from core 0.61.0) and fails when a key in it reaches neither the native agent nor a named gap. Sections the core added later land in spec.other until Prisma decides on them.