You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
bug(capabilities): get_adcp_capabilities advertises supported_versions the dispatcher cannot serve #1280
get_adcp_capabilities publishes whatever adcp.supported_versions the seller declares. It never checks that the dispatcher will accept those versions or that the SDK bundles schemas to validate them. A seller can therefore advertise a version that every request is rejected at, and nothing flags it at boot.
Depending on how the seller widened version detection, a buyer that selects the advertised version gets one of two errors:
VERSION_UNSUPPORTED from detect_wire_version, when the version isn't in the accepted wire set; or
VALIDATION_ERROR at / with keyword schema_unavailable, from _missing_explicit_schema_outcome (adcp/validation/schema_validator.py:132-175), when detection accepts it but no bundle exists.
The second is a misleading code. The request body is fine; the server can't speak the version it negotiated.
This happened in production. Embedded Sales Agent advertised 3.1-beta.5, a compatibility alias added while #883 was open. When adcp 7.0.0 started failing closed on explicit versions without a bundle, every 3.1-beta.5 request began failing with schema_unavailable. The capability kept advertising the version for months, and no boot check or log line pointed at the mismatch.
Ask
At boot (validate_capabilities_response_shape, or serve()), reject or warn when an advertised supported_versions entry isn't in the accepted wire set, or has no bundled validator.
When detection accepts an explicit version but no bundle exists, return VERSION_UNSUPPORTED (with supported_versions in details), not VALIDATION_ERROR at /.
Classification: Bug (two related sub-bugs in the validation/capabilities boundary) Bucket(s): validation, handlers (no matching repo labels found; noted for maintainer) Status: ready-for-human Milestone: omitted (no target version specified)
What the experts said:
ad-tech-protocol-expert:VERSION_UNSUPPORTED is the correct wire code for sub-bug 2 — compliance storyboards (exact_nonempty_reject_rc0) assert exactly that for the "server accepted a version it cannot validate" case. Technically breaking but a conformance correction. For sub-bug 1, recommends raising (not warning) at boot when supported_versions is explicitly supplied, scoped only to caller-supplied lists.
code-reviewer: Sub-bug 2 has a blocker-level semantic concern — VERSION_UNSUPPORTED would be actively misleading here, because the dispatcher just accepted the version; the buyer would downgrade unnecessarily. INTERNAL_ERROR is the correct code (server misconfiguration, not a bad request). For sub-bug 1, the warning belongs in capabilities_response() directly (not just in _validate_response_dict), fires only when supported_versions is not None, and must compare against SUPPORTED_WIRE_VERSIONS (not get_supported_adcp_versions()) to avoid false positives on legacy versions like "2.5".
My take: The two sub-bugs need separate decisions. Sub-bug 1's warning-only path is non-breaking and both experts agree on the capabilities_response() location and the SUPPORTED_WIRE_VERSIONS comparison set — that piece could land independently. Sub-bug 2 requires a human call: the protocol expert says VERSION_UNSUPPORTED, the code reviewer says INTERNAL_ERROR, and they make different arguments from spec conformance vs. buyer UX. Either choice is a breaking wire change.
Sub-bug 2 — which code? Protocol expert cites storyboard precedent for VERSION_UNSUPPORTED. Code reviewer argues INTERNAL_ERROR is semantically correct (buyer can't fix a missing schema bundle by downgrading). Which is the intended wire contract?
Sub-bug 1 — warn or raise at boot? Raising at boot on explicitly-supplied bad versions (protocol expert's preference) is breaking for adopters with stale configs. Warning-only (code reviewer) is non-breaking but softer. What's the right tradeoff for adopters who already shipped supported_versions=["3.1-beta.5"]?
Before shipping sub-bug 2's fix, exact_nonempty_reject_rc0 and any partner storyboards asserting VALIDATION_ERROR for schema_unavailable would need updating — is there a storyboard audit step in scope here?
Implementation scope (for when the decisions above are made):
Sub-bug 1 (non-breaking warning path):src/adcp/server/responses.py lines 257–350 — add if supported_versions is not None: guard before the merge block; iterate entries against SUPPORTED_WIRE_VERSIONS from adcp.validation.envelope; emit warnings.warn(..., stacklevel=2) for any miss. No import circularity risk.
Sub-bug 2 (breaking, after code decision):src/adcp/validation/schema_validator.py lines 160–174 (_missing_explicit_schema_outcome) — change returned keyword/code; src/adcp/validation/schema_errors.py line 74 (build_adcp_validation_error_payload) — add a parallel builder or keyword-based routing; src/adcp/server/mcp_tools.py lines 3307–3320 and 3607–3614 — add dispatcher routing on the new keyword.
Activity
bokelley commentedon Oct 2, 2026
Triage
Classification: Bug (two related sub-bugs in the validation/capabilities boundary)
Bucket(s): validation, handlers (no matching repo labels found; noted for maintainer)
Status: ready-for-human
Milestone: omitted (no target version specified)
What the experts said:
VERSION_UNSUPPORTEDis the correct wire code for sub-bug 2 — compliance storyboards (exact_nonempty_reject_rc0) assert exactly that for the "server accepted a version it cannot validate" case. Technically breaking but a conformance correction. For sub-bug 1, recommends raising (not warning) at boot whensupported_versionsis explicitly supplied, scoped only to caller-supplied lists.VERSION_UNSUPPORTEDwould be actively misleading here, because the dispatcher just accepted the version; the buyer would downgrade unnecessarily.INTERNAL_ERRORis the correct code (server misconfiguration, not a bad request). For sub-bug 1, the warning belongs incapabilities_response()directly (not just in_validate_response_dict), fires only whensupported_versions is not None, and must compare againstSUPPORTED_WIRE_VERSIONS(notget_supported_adcp_versions()) to avoid false positives on legacy versions like"2.5".My take: The two sub-bugs need separate decisions. Sub-bug 1's warning-only path is non-breaking and both experts agree on the
capabilities_response()location and theSUPPORTED_WIRE_VERSIONScomparison set — that piece could land independently. Sub-bug 2 requires a human call: the protocol expert saysVERSION_UNSUPPORTED, the code reviewer saysINTERNAL_ERROR, and they make different arguments from spec conformance vs. buyer UX. Either choice is a breaking wire change.Open questions for @bokelley:
VERSION_UNSUPPORTED. Code reviewer arguesINTERNAL_ERRORis semantically correct (buyer can't fix a missing schema bundle by downgrading). Which is the intended wire contract?supported_versions=["3.1-beta.5"]?exact_nonempty_reject_rc0and any partner storyboards assertingVALIDATION_ERRORforschema_unavailablewould need updating — is there a storyboard audit step in scope here?Implementation scope (for when the decisions above are made):
src/adcp/server/responses.pylines 257–350 — addif supported_versions is not None:guard before the merge block; iterate entries againstSUPPORTED_WIRE_VERSIONSfromadcp.validation.envelope; emitwarnings.warn(..., stacklevel=2)for any miss. No import circularity risk.src/adcp/validation/schema_validator.pylines 160–174 (_missing_explicit_schema_outcome) — change returned keyword/code;src/adcp/validation/schema_errors.pyline 74 (build_adcp_validation_error_payload) — add a parallel builder or keyword-based routing;src/adcp/server/mcp_tools.pylines 3307–3320 and 3607–3614 — add dispatcher routing on the new keyword.Triaged by Claude Code. Session: https://claude.ai/code/session_01K6FZpjMxXcq542XLxCdaEz
Generated by Claude Code