| name | upgrade-pipecat |
| description | Upgrade the Nemotron Voice Agent to a new Pipecat (pipecat-ai) version. Reads release notes for every release in range, diffs old vs new, discovers every example pipeline and Pipecat call site, implements changes, then runs multi-agent gap analysis until clean. Generic across Pipecat versions. |
| version | 2.0.0 |
| metadata | {"author":"NVIDIA Voice Agent Team <nemotron-voice-agent@nvidia.com>","tags":["upgrade","migration","pipecat","voice-agent","nemotron"]} |
Pipecat Version Upgrade — Nemotron Voice Agent
Autonomously migrate this repo to a new pipecat-ai version. Repo root is the working directory.
Invocation
/upgrade-pipecat new=<target> [old=<source>]
new (required): target pipecat-ai version. PyPI version (1.5.0), local pipecat checkout path
(full diff), git tag (v1.5.0), or docs URL (https://docs.pipecat.ai/, least complete).
old (optional): current version. A version string enables CHANGELOG/git diff. Omit to auto-detect from
pyproject.toml (pipecat-ai[...]==X.Y.Z) + uv.lock.
The only input is the pipecat-ai version. Latest version and release notes come from the canonical
https://github.com/pipecat-ai/pipecat/releases. The skill scans BOTH dependency surfaces for every Pipecat
package and reads each one's own release notes:
- Server (Python) —
pyproject.toml/uv.lock: every pipecat* dependency (e.g. pipecat-ai-subagents,
pipecat-ai-flows, …).
- Client (npm) —
client/package.json: every @pipecat-ai/* dependency (e.g. client-js, client-react,
*-transport, …).
Extras and every dependency change are derived from these notes + the lockfiles — nothing about specific
packages is hardcoded here.
Phases (run in order)
- Explore & Discover — workflows/01-explore-discover.md. Read release
notes for every release in range, analyze against the repo, confirm with source diff + per-example scan.
- Plan & Implement — workflows/02-plan-implement.md.
- Gap Analysis Loop — workflows/03-gap-analysis-loop.md.
- Deploy & Validate — workflows/04-deploy-validate.md.
Pipecat docs MCP — use for any doubt
For ANY Pipecat uncertainty (signature, moved module, intended usage, migration path), query the
pipecat-docs MCP tool search_daily_knowledge_sources (backed by https://daily-docs.mcp.kapa.ai) instead of
guessing. Pass one complete sentence as query; cite the returned source_url in the change log. For exact
signatures, trust the actual installed/new source; use the MCP for intent and migration guidance.
Principles
- Release-notes-driven (hard gate): read the
pipecat-ai release notes + CHANGELOG for EVERY release in
range — and each pipecat* subpackage's own notes — before doing anything else. Do not start the diff, the
plan, or any edit until this is done and recorded. Most missed migrations come from skipping this. The source
diff only confirms and completes the notes.
- Dependencies follow the notes, not assumptions: discover the repo's current Pipecat-related dependencies
from
pyproject.toml/uv.lock, then let the pipecat-ai notes dictate what happens to each — bumped,
renamed, newly required, or folded into core (dependency removed + imports migrated). Never hardcode or assume
a companion package stays separate, stays present, or co-versions.
- Discovery-first: never assume module paths, frame names, service constructors, or processor APIs — scan
the installed package and new source. Pipecat reorganizes its module tree between versions.
- Generic: works for any transition; discover changes, hardcode nothing.
- Examples are the unit of work: 5 examples (
generic, multilingual, omni_assistant,
omni_assistant_subagents, frontend_backend_agent), each with its own pipeline.py. One agent per example + one
cross-cutting agent for src/examples/shared/ and src/server.py.
- Server + client move together (RTVI contract): the RTVI wire protocol couples the Python server to the
@pipecat-ai/* client packages, so they must be upgraded in lockstep. Bump client/package.json to versions
compatible with the target pipecat-ai, migrate client/src/ RTVI usage (renamed events/messages), and gate
on npm lint+build. Discover the client packages — don't assume which exist.
- Iterative convergence: gap analysis loops until a pass finds zero gaps.
- Validation gates:
uv sync, import smoke tests, ruff, pytest, Compose deploy. Human input only for
plan confirmation and review.