Repository navigation
Description
Background
adcp#4907 proposes bidirectional storyboard interop CI: before @adcp/sdk releases, the candidate SDK should run storyboards against Python's reference seller.
For Phase 1, the intended target is examples/seller_agent.py, because this repo already uses that file as its own storyboard reference seller. Rather than copy the Python CI shell steps into adcp-client, this repo should own a small reusable harness that both Python CI and cross-repo release-gating can call.
This is separate from making the job blocking (adcp-client-python#814). That issue owns enforcement. This issue owns centralizing the recipe.
Proposed shape
Add a script such as:
scripts/ci/run_storyboard_reference_seller.sh
The script should:
- Assume it is running from an
adcp-client-pythoncheckout. - Install or use an
adcpstoryboard runner based on an explicit input. - Install Python dependencies needed by
examples/seller_agent.py. - Boot
examples/seller_agent.pyon a configurable port. - Run
adcp storyboard run http://127.0.0.1:$PORT/mcp media_buy_seller --json --allow-http. - Write a JSON result file to a configurable path.
- Fail non-zero when
overall_status != passing. - Fail non-zero when
controller_detectedis false, unless the repo deliberately changes that invariant.
A minimal interface could be:
ADCP_SDK_VERSION=7.10.2 \
ADCP_PORT=3001 \
STORYBOARD_RESULT_PATH=storyboard-result.json \
./scripts/ci/run_storyboard_reference_seller.shFor adcp-client release-gating, the script should also support a candidate SDK tarball or preinstalled adcp binary, for example:
ADCP_SDK_TARBALL=/path/to/adcp-sdk.tgz \
ADCP_PORT=3001 \
./scripts/ci/run_storyboard_reference_seller.shExact environment variable names are flexible; the important contract is that the caller can choose between a pinned published SDK and a candidate SDK build.
Scope
- Extract the repeated
examples/seller_agent.pystoryboard setup/run/assert logic from.github/workflows/ci.ymlinto the script. - Update Python CI to call the script instead of duplicating the shell inline.
- Keep the script local-run friendly: no GitHub-only assumptions except optional artifact paths.
- Do not add the
adcp-clientcross-repo workflow here; that remainsadcp-client#1916.
Acceptance
- Python CI uses the shared script for the
examples/seller_agent.pystoryboard job. - The script supports a pinned published
@adcp/sdkversion. - The script supports a candidate/local SDK runner path, tarball, or preinstalled binary suitable for
adcp-client#1916. - The script exits non-zero on storyboard failure.
- The script documents its required inputs in comments or a short README section.
-
adcp-client#1916can call this script from a taggedadcp-client-pythoncheckout without copying the CI recipe.
References
- Primary RFC: RFC: bidirectional storyboard interop CI — @adcp/sdk releases gated on reference-impl conformance adcp#4907
- TS-side release-gating workflow: feat(ci): wire bidirectional storyboard interop CI per adcp#4907 Phase 1 adcp-client#1916
- Make the Python storyboard job blocking: ci: make examples/seller_agent.py storyboard job blocking #814
Activity
bokelley commentedon May 22, 2026
Triage
Classification: Feature request (CI infrastructure)
Bucket(s): ci, examples (no matching repo labels found — gap noted)
Status: drafting-pr
What the experts said:
ADCP_SDK_VERSION/ADCP_SDK_TARBALLexclusivity and stdout fallback approvedMy take: Well-specified issue with clear acceptance criteria. The new script improves on the inline CI steps by adding an isolated venv, guaranteed agent cleanup via
trap, and input validation for the tarball mode — making it safe to adopt fromadcp-client#1916.Drafting PR:
scripts/ci/run_storyboard_reference_seller.sh(new) + CI job slimmed to a single step calling the script. → #820Triaged by Claude Code. Session: https://claude.ai/code/session_01BdEqT6X8riNXLLCZUxUTMx
Generated by Claude Code