Rosetta Add Framework

Rosetta — Add Framework Orchestrator

End-to-end workflow for bringing a new agent framework into the repo. Mirrors the existing structure: every new framework gets a no-observability/<framework>-py/ (or -ts/) directory, a phoenix/<framework>-py/ directory, and an ax/<framework>-py/ directory, each running an identical Wonder Toys shopping agent.

Inputs

• <framework> — framework name in the form used by directory naming, e.g. agno, google-adk, smolagents, crewai. Use the slug from Arize's docs (https://arize.com/docs/ax/integrations/python-agent-frameworks/<slug>/<slug>-tracing).
• <language> — py for Python, js/ts for TypeScript, java for Java. Affects which existing tier is used as the clone source. Note: Java tiers don't fit the current Node/Python repo shape — see the Java section under TODO before attempting.
• --start-from <phase> (optional) — resume from discover | research | build-noobs | build-phoenix | build-ax | playwright | docs. Useful when a phase fails mid-flow.

TODO — frameworks not yet implemented

Snapshot (last refreshed 2026-06-12 from https://arize.com/docs/llms.txt). Each phase-1 discover run re-fetches and shows the diff.

Python (Tier A — clear agent frameworks)

• Agno
• AutoGen
• AWS Strands
• Bedrock AgentCore — new Arize page under aws-strands docs path; investigate whether it's a distinct framework from AWS Strands or a Bedrock-specific variant
• CrewAI
• DSPy
• Google ADK
• Haystack
• LlamaIndex Workflows
• OpenAI Agents SDK
• [~] Pipecat — skipped: not-chat-shaped (real-time voice/multimodal framework; audio I/O pipeline only, no text-only chat mode)
• Pydantic AI
• Semantic Kernel
• Smolagents (Hugging Face)
• BeeAI

TypeScript (Tier A)

• BeeAI
• Claude Agent SDK — synthesizes LLM spans in the chat route (instrumentor emits AGENT+TOOL only); a CHAIN parent wraps the run so AGENT+LLM share one trace
• OpenAI Agents SDK

Java (Tier A)

⚠️ Adding a Java tier is a significantly larger lift than adding another Python or TypeScript framework. The repo's current shape assumes a Node/Python stack — every existing tier is either a Next.js monolith or a FastAPI Python backend + Next.js frontend. A Java tier needs:

• A JVM build setup (Maven or Gradle) added to the repo
• A new backend serving the same /chat SSE protocol from Java (e.g. Spring Boot)
• Either a separate Java→Next.js frontend bridge, or a Java-rendered UI
• Java-specific tracing setup matching the Arize Java docs
• Adjustments to scripts/start.sh patterns since the existing one is uv/npm-only

Decide whether the project wants this dimension before starting. If yes, the per-framework work is then:

• LangChain4j
• Spring AI
• Arconia

Borderline — decide before building

• Guardrails AI (output validation)
• Instructor (structured output)
• MCP / Model Context Protocol (wire protocol)
• Portkey (LLM gateway)
• Together AI (LLM provider)
• OpenInference Annotation Tracing (Java) — @Chain/@LLM/@Tool/@Agent decorators applied via ByteBuddy for hand-built agents; not a framework

Phases

Execute in order. Each phase has a dedicated skill — read it and follow.

Viability gate (phase 2)

Before any code is written, verify three things. If any check fails, abort with a structured skip-with-reason report (see "Skip-with-reason vocabulary" below) and do not create any tier directories.

Check 1 — Anthropic Claude support

The repo standardises on Anthropic Claude. The framework must support it (directly via an Anthropic provider, via LiteLLM, via Bedrock, or similar).

• Web-fetch the framework's "models" / "providers" docs page
• Look for: an Anthropic provider class, a model string like anthropic/..., a LiteLlm / LiteLLM wrapper, or a Bedrock Anthropic integration
• If none found → skip with reason no-anthropic-support

Check 2 — Streaming text API

The Wonder Toys agent streams SSE deltas to the UI. The framework must expose a token-level streaming API for text responses (not just final aggregates).

• Look for: an async for chunk in agent.run_stream(...) shape, an event-bus with text-delta events, a RunConfig(streaming_mode=SSE)-style flag, or callbacks on partial tokens
• If the framework only returns final aggregated responses → skip with reason no-streaming-api
• If it streams only audio/video (e.g. Pipecat) → skip with reason not-chat-shaped

Check 3 — OpenInference instrumentation

The framework must have a published openinference-instrumentation-<framework> (or equivalent) package, or be auto-instrumented by a more general OpenInference instrumentor.

# Quick existence check via PyPI
curl -sf "https://pypi.org/pypi/openinference-instrumentation-${FRAMEWORK}/json" > /dev/null \
  && echo "instrumentation: yes" || echo "instrumentation: NO"

• If the package doesn't exist AND no general instrumentor covers it → skip with reason no-instrumentation

Check 4 — Framework shape fits "chat-to-purchase"

Some frameworks aren't conversational by design — they assume batch task execution (CrewAI was on the edge of this; we adapted), declarative pipelines (DSPy), voice/video streams (Pipecat), or batch inference. Apply judgment:

• If the framework has no conversational primitive (session/thread/message history) and no easy way to fake one by threading history into a task description → skip with reason not-chat-shaped
• Borderline cases: build it but document the awkwardness inline (CrewAI's per-turn fresh Crew is an acceptable adaptation)

Skip-with-reason vocabulary

When a framework can't be built, write a structured marker so the TODO list and PR history stay honest. Apply each in the orchestrator's TODO and in MEMORY.md:

Record skips in .claude/skills/rosetta-add-framework/SKILL.md (this file) under the TODO section by changing [ ] to [~] and appending the reason — e.g. - [~] Pipecat — skipped: not-chat-shaped (voice/video only). The discover phase preserves these markers across runs.

Research checklist (phase 2)

Before building, gather these answers. Save anything non-obvious to project memory so the next framework's research is faster.

Framework API

1. Agent construction — class name, constructor signature, how to pass system prompt and Anthropic Claude
2. Tool registration — decorator or constructor list? Does it accept Annotated[..., Field(description=...)] for params?
3. Streaming API — what method/event stream emits text deltas? How is a tool-call event distinguished from a text-delta event?
4. Conversation memory — does it have a session/thread primitive, or does the caller manage history? How is history passed across turns?
5. Anthropic Claude — model string format. Does the framework require a specific Sonnet version (some require strict-tools support, only available on 4.5+)?
6. Installed version — pin to a tested version in requirements.txt

Arize tracing

7. Phoenix pattern — fetch https://arize.com/docs/phoenix/integrations/python/<framework>/<framework>-tracing or equivalent. What instrumentor + span processor? Does register() route to the right project, or is the manual TracerProvider(resource=Resource.create({PROJECT_NAME: ...})) pattern required?
8. AX pattern — fetch https://arize.com/docs/ax/integrations/python-agent-frameworks/<framework>/<framework>-tracing. Same questions.
9. Session propagation — does the framework's OpenInference instrumentation auto-emit session.id, or is using_session(user_id) wrap required? Grep the installed openinference-instrumentation-<framework> package for session.id references.

Check memory first

Before web-fetching, check project memory for framework_<name>_* keys — earlier work on similar frameworks may have already documented these answers.

Session.id audit (phase 9)

After all 3 tiers are built and tested:

1. Run the AX tier, fire a single chat with x-user-id: session-audit-$(date +%s)
2. Wait 30s for ingestion
3. Query: ax traces list <project> -s $SPACE -o json --limit 10 and check whether the most recent trace has attributes.session.id set
4. If set — done. Session works automatically via the framework's primitives or its OpenInference instrumentation.
5. If not set — add using_session(user_id) wrap to agent.py following the CrewAI pattern:

   from contextlib import nullcontext
   try:
       from openinference.instrumentation import using_session
   except ImportError:
       def using_session(session_id: str):
           return nullcontext()
   ...
   with using_session(user_id):
       # the agent call that emits spans
       ...
   

   Apply identically to all 3 tiers (the try/except keeps no-observability working). Re-test.

Failure handling

• Discover fails (Arize docs unreachable, framework not in supported list) → abort, surface the diff
• Viability gate fails → record the skip-with-reason in this file's TODO (change [ ] to [~] with reason), commit only that change, don't create any tier directories
• Research surfaces a blocker not caught by the gate → same handling as viability-gate failure; tighten the gate's checks so the next framework catches it earlier
• Build/test phase fails → fix the agent.py / requirements.txt, re-run just that phase via --start-from
• Session.id audit fails after fix → document inline in agent.py comments and in commit message; mark the framework as having a session-tagging caveat in the PR description

Failure isolation (rollback) — for long autonomous runs

A long-running orchestrator pass should leave the repo in a clean, committable state regardless of whether any single framework succeeded. If a framework fails mid-build or mid-test:

1. Kill backend processes on ports 8001 (FastAPI), 3000 (Next.js), 6006 (Phoenix). Leave ChromaDB on 8000 alone — it's shared and slow to re-init.
2. Drop any half-built tier directories for this framework. The cleanest check: a tier counts as "complete" only after its corresponding test phase passes. Anything else is removable. Specifically: rm -rf <tier>/<FRAMEWORK_DIR> for any tier whose tier-test skill didn't return a PASS for this run.
3. Remove the framework's .env.local files from any directories that survived the cleanup.
4. Reset shared venv if the failure was a dep-conflict (see Regression check below). Reinstall the previous framework's requirements.txt to restore.
5. Branch hygiene — if the orchestrator created feat/add-<framework> and committed anything, hard-reset the branch back to origin/main or delete the branch outright. Do not push a partial branch.
6. Record the failure under the framework's TODO entry with [!] and a short reason (e.g. - [!] Smolagents — failed: streaming events never carry text deltas). Different marker from [~] so a human can see "tried but broke" vs "skipped on principle".

After rollback, the orchestrator may proceed to the next framework in the TODO (autonomous-run mode) or stop and surface the failure (default).

Regression check (after each Build phase)

After installing a new framework's requirements.txt into the shared .venv, verify that no previously-implemented Python tier's agent.py has broken imports. Dependency conflicts have surfaced before (e.g. installing CrewAI pinned opentelemetry-semconv==0.55b1 which then broke Pydantic AI's instrumentation which needs >=0.62b1).

for tier in no-observability phoenix ax; do
  for fwk_dir in $tier/*-py/; do
    [ "$fwk_dir" = "$TIER/$FRAMEWORK_DIR/" ] && continue   # skip the tier we just built
    fwk_name=$(basename "$fwk_dir")
    /Users/jimbobbennett/github/project-rosetta-stone/.venv/bin/python -c "
import sys; sys.path.insert(0, '$fwk_dir')
try:
    import backend.agent
    print('  OK: $fwk_name')
except Exception as e:
    print(f'  BROKEN: $fwk_name — {type(e).__name__}: {e}')
    sys.exit(1)
" || exit 1
  done
done

If any existing tier breaks, the new framework's install introduced a dep conflict. Mark the new framework as [!] with reason dep-conflict, run the rollback procedure above, and surface the conflict.

(Per-framework venv isolation would prevent this entirely — that's a future refactor. For now, the regression check is the cheap defence.)

State to thread through phases

The orchestrator should maintain these throughout:

• FRAMEWORK — the slug (google-adk)
• FRAMEWORK_DIR — derived (google-adk-py for Python, google-adk for TS — match existing naming)
• LANGUAGE — py or ts
• CLONE_SOURCE — closest existing tier to copy from. For Python: prefer pydantic-ai-py (simplest agent.py); for TS: prefer vercel-ai-sdk. The build skill takes this as input.
• MODEL_OVERRIDE — optional. If the framework can't use the project default claude-sonnet-4-6 (e.g. CrewAI needed Sonnet 4.5), set this here and propagate.

Print these at the start of the run so a human can resume.

End-of-run output

Framework added: <FRAMEWORK>
  Tiers: no-observability, phoenix, ax — all green
  Traces verified: phoenix ✓, ax ✓
  Sessions tagged: ✓ (auto | via using_session wrap)
  Playwright smoke: ✓
  Docs updated: README.md, TODO snapshot in this skill
  PR: <URL>