Repository navigation
Description
Summary
Several 3.2.x storyboards call a tool the spec lets a seller leave out, with no requires_tool on the step and no storyboard gate that implies the tool. The SDK runs a storyboard when the agent advertises any one of its required_tools, then skips each step whose tool isn't advertised as missing_tool. The hosted grader counts that skip as a failed step unless the pinned step's requires_tool names the missing tool, and it counts the stateful steps that cascade from it as dependency_failed. A seller that leaves out an optional surface therefore gets a failing or partial storyboard for it.
When #6753 made wholesale discovery optional, its grading audit found nothing that graded a seller on a wholesale path it hadn't opted into, relying on capability gates and requires_tool. Since then, #6619 (sales-dooh) and #7444 (product_coverage_filters) added list_products steps with neither, and the hosted classification from #7454 counts an ungated missing_tool skip as a failure. change_rights_state_projection (#6794) and revenue_share_pricing (#6095) have the same shape for control_media_buy and sync_event_sources.
Three of these fail on our own seller in a local run that we re-graded with the registry's policy: product_coverage_filters, revenue_share_pricing and change_rights_state_projection. Restricted to ten tools the spec makes optional, a scan of the 3.2.4 bundle finds 20 such steps in 9 storyboards. The nine files are identical in 3.2.5 (released 2026-10-16), in static/compliance/source on main, and in the 3.3.0-beta.0 snapshot that the release-bot draft #7897 prepares. #7586 reported the same pattern for the provenance storyboards, and #7599 / #7646 fixed it there.
Our stake: our sales agent (https://sell.nofluffadvisory.com/mcp) is graded on 3.1.24 and stays on 3.1 until 3.3. Its 3.2 build is behind a flag that is off, so nothing here changes our current card. It matters for any seller that moves to 3.2 without the wholesale feed, control_media_buy or conversion tracking. Eight of the nine storyboards declare requires: [controller], so in practice that means sellers that expose comply_test_controller. No hosted card shows it yet. Of the 35 agents in the public registry, 32 answered on 2026-10-16 at 19:54Z, two of them opted out of grading. The only one graded on 3.2.x is our signals agent (adcp.signal-stack.io, 3.2.4), which doesn't run these storyboards.
Evidence
Links are pinned to adcp main 68c90b6 and to @adcp/sdk 15.9.2 (90a40c3), which the registry has pinned since #8231.
Runner.
required_toolsis the "Lenient any-of" gate (storyboard-schema.yaml L54-L56).hasAnyRequiredTool("Storyboard applicability is any-of; it does not authorize individual steps") runs the storyboard if the agent advertises any listed tool (runner.ts L3353-L3371).- A step whose task the agent doesn't advertise returns
skipped: true, skip_reason: "missing_tool"(L6127-L6139). A step whoserequires_toolisn't advertised returns the same reason (L6041-L6046). - A missing-tool skip on a stateful step trips the phase cascade, unless it is the phase's only stateful step or a same-phase peer declares
provides_state_forfor it (L4190-L4241).
Grader.
classifyComplianceStepreturnsfailedformissing_tool, except on the synthetic/missing_toolgate (step-disposition.ts L37-L43).- For a
missing_toolorprerequisite_failedskip,classifyRunStepreturnsnot_applicableonly when that step's own pinnedrequires_toolnames a tool the agent doesn't advertise (compliance-testing.ts L817-L841). A dependent step that calls an advertised tool staysdependency_failed. - This is deliberate (fix(compliance): classify hosted skips and expose redacted run provenance #7454) and tested. compliance-publication.test.ts L70-L77 keeps production
missing_toolfailing. derive-storyboard-statuses.test.ts L535-L572 keeps the steps after a pinned optional-tool skip as dependency failures. - The schema build also treats
requires_toolas the marker of optional surface (build-schemas.cjs L1097-L1103).
Spec.
- 3.2 sellers advertise "any supported subset" of the compact tools in
lifecycle_tools,list_productsandcontrol_media_buyincluded. The same paragraph says a seller that omitslist_products"is fully conformant", and that wholesale-path conformance scenarios "are capability-gated and grade not applicable" (get_adcp_capabilities.mdx L623). - list_products.mdx L10: wholesale support is optional, and "no conformance scenario requires the wholesale path" (Media-buy: declare seller discovery posture — a wholesale feed is not a capability signal #6729 / docs: wholesale discovery is optional and never a capability signal #6753).
sync_event_sourcesandlog_eventare required whenmedia_buy.conversion_trackingis present, andsync_audienceswhenmedia_buy.audience_targetingis present (L1751-L1752).
Observed on our seller
The run.
- Local
adcp storyboard runon 2026-10-14 at 12:50Z,@adcp/sdk15.8.1, bundle 3.2.3, against a local instance of our sales agent with 3.2 serving on. - The agent advertises 15 tools. None of them is
list_products,control_media_buyorsync_event_sources. - The CLI reports
missing_tool=6. Those are the six steps in the table.
How we graded it.
- We re-graded the run with a reimplementation of
classifyRunStep/deriveStoryboardStatusesfrom the registry source above, using the 3.2.4 bundle as the pinned metadata. - We then graded it again with
requires_tooladded to those six steps in a scratch copy of the bundle.
Why a 3.2.3 run stands in for 3.2.4 / 15.9.2.
product_coverage_filtersandrevenue_share_pricingdiffer between 3.2.3 and 3.2.4 only byrequires: [controller], and we advertise the controller.change_rights_state_projectionis identical.runner.js,agent-routing.js,compliance.jsandcomply.jsare byte-identical in 15.8.1 and 15.9.2.
| Storyboard | Steps without requires_tool |
Hosted policy, as published | Hosted policy, requires_tool added to those steps |
|---|---|---|---|
product_coverage_filters |
compact_coverage/compact_country_and_metro (list_products) |
failing 0/2 | failing 0/1: the list_products step is N/A |
revenue_share_pricing |
setup/sync_event_sources |
partial 1/7 (1 failed, 5 dependency_failed) | partial 1/6: the step is N/A, the 5 later steps stay dependency_failed |
change_rights_state_projection |
4 control_media_buy steps: L196, L226, L312, L348 |
failing 0/6 | failing 0/2: the 4 steps are N/A, both get_media_buys reads still fail |
product_coverage_filters. The compact step is non-stateful and comes last, so requires_tool: list_products on the step is enough. The failure that remains, legacy_country_and_metro, is our bug: we don't apply coverage filters on get_products.
revenue_share_pricing. sync_event_sources is stateful and shares its phase with sync_accounts. The runner therefore cascades prerequisite_failed into all five later steps, and the grader keeps those as dependency failures. A step-level requires_tool doesn't fix this storyboard. The documented revenue-share settlement flow also starts with the buyer registering the event source through sync_event_sources (pricing-models.mdx L414-L416).
change_rights_state_projection. This storyboard has no gate besides requires: [controller], and its required_tools include get_media_buys, so every seller with that tool and the test controller runs it.
- The first read already asserts an
available_actionsentry withtask: control_media_buy(L156-L163), so the whole storyboard assumes the tool. - In our run, the paused-state read ran after the pause step was skipped. It failed with
Expected "paused", got "active". - If the active read passes, the runner source says the skipped pause instead cascades
prerequisite_failedinto the paused read. We read that in the source; we didn't run it. - Our agent also fails the active read because we don't implement accepted-proposal change rights. That part is our gap.
Scan of the 3.2.4 bundle
Method.
- Load all 213 storyboards in
dist/compliance/3.2.4with the SDK'slistAllComplianceStoryboards. - Keep steps whose
taskis a tool listed indist/schemas/3.2.4/index.json, that have norequires_tool, and whose storyboard can run without that tool (itsrequired_toolsis empty or names another tool). - Drop
comply_test_controllersteps inrequires: [controller]storyboards.
Result. That leaves 1,077 steps. We triaged only the steps whose task the spec text above makes optional: list_products, request_proposals, refine_proposals, decline_proposals, buy_products, accept_proposal, control_media_buy, sync_event_sources, log_event and sync_audiences. That is 84 steps in 26 storyboards.
64 of the 84 sit behind a gate that implies the tool:
- a
media_buy.conversion_trackingormedia_buy.audience_targetinggate; - an opt-in compact-lifecycle probe scenario in
compliance_testing.scenarios; proposal_refinement, which is valid only withrefine_proposals(L645);- a specialism defined around the tool.
The other 20 fall into two groups.
No gate related to the tool:
| Storyboard | Step | Task | Stateful | Cascade |
|---|---|---|---|---|
media_buy_seller/product_coverage_filters |
compact_coverage/compact_country_and_metro |
list_products |
no | none |
media_buy_seller/revenue_share_pricing |
setup/sync_event_sources |
sync_event_sources |
yes | all 5 later steps |
media_buy_seller/change_rights_state_projection |
the 4 steps above | control_media_buy |
yes | get_paused_buy and the later control steps |
sales_dooh (stable specialism; runs only for sellers that claim sales-dooh) |
product_discovery/get_dooh_products |
list_products |
yes | none by cascade (only stateful step in its phase), but later steps read $context.dooh_product_id from it (L379, L444, L549) |
Gated on a capability that doesn't imply the tool. These only catch a seller that declares the capability without the tool, and some may be intended:
| Storyboard | Gate | Steps |
|---|---|---|
frequency_cap_capability_updates |
media_buy.frequency_capping present ("retains its legacy meaning", L682) |
control_media_buy L211, L236, L262, L287 |
media_buy_frequency_cap_updates |
media_buy.aggregate_frequency_capping present |
control_media_buy L271 (cascades into clear_root_cap_and_add_package), L413 |
availability_windows |
media_buy.availability_horizon equals true; required_tools: [list_products, buy_products] |
list_products L143, L202, L309; buy_products L360, L404. Only affects a seller with one of the two tools. #6753's audit accepted this gate as the opt-in. |
typed_proposal_negotiation |
lifecycle_tools contains request_proposals, plus proposal_refinement |
accept_proposal L611 |
refine_frequency_cap_negotiation |
proposal_refinement (implies refine_proposals only), plus aggregate capping |
request_proposals L207 |
Not triaged: the deprecated facades. In 3.2, get_products, create_media_buy and update_media_buy are compatibility facades that a server advertises when it supports callers that use them (schemas.mdx L289-L292), and get_products is itself in the L623 subset. 351 of the 1,077 candidate steps call one of the three, so a compact-only seller may hit the same grading there. #8201 raises this for sales-dooh. We haven't triaged it.
None of the 117 open PRs here changes these nine files, step-disposition.ts or classifyRunStep. #7897 only copies the nine files unchanged into the 3.3.0-beta.0 snapshot.
Ask
- Gate these steps so that a seller without an optional tool grades not applicable, on main and on 3.2.x (as fix(compliance): backport provenance requires_tool gates to 3.1.x #7646 did for 3.1.x).
- For non-stateful steps that nothing depends on, add
requires_tool: <tool>.product_coverage_filters/compact_country_and_metrois the clear case, and it is what L623 already promises for the wholesale path. - Where later steps need the optional step's state,
requires_toolon that one step isn't enough under the current grader, because the dependents stay dependency failures. Gate the storyboard instead.change_rights_state_projection:requires_capability: {path: media_buy.lifecycle_tools, contains: control_media_buy}.typed_proposal_negotiationalready gates onmedia_buy.lifecycle_toolscontainingrequest_proposals, with the same rationale in its comment (L15-L21).revenue_share_pricing:requires_capability: {path: media_buy.conversion_tracking, present: true}. That is the existing declaration that makessync_event_sourcesmandatory.sales_dooh:requires_toolonget_dooh_productsalone would leave the later steps without a product ID. Either give the specialism a discovery path that works withoutlist_products, or say thatsales-doohrequireslist_products, which list_products.mdx L10 currently rules out. compliance: opt the 3.2 compact proposal storyboards andsales-doohintofixture_resolution(discover), and allowlist_productsdiscovery at 3.2 #8201 asks the related surface question for this specialism.
- The second table is your call, storyboard by storyboard.
- For non-stateful steps that nothing depends on, add
- If you'd rather fix this in the grader, grade
missing_toolas not applicable when the seller's own declarations make the tool optional, and carry that to the steps that cascade from it. Examples: a compact tool missing frommedia_buy.lifecycle_tools, orsync_event_sources/log_eventwithoutmedia_buy.conversion_tracking. That narrows a rule fix(compliance): classify hosted skips and expose redacted run provenance #7454 made deliberately and tests, so the storyboard change looks smaller to us. It could also cover the facade case above without editing each storyboard. - A build check that flags a step calling one of these tools without
requires_toolor a related gate would catch the next one. docs: wholesale discovery is optional and never a capability signal #6753's audit was a one-off, and feat(compliance): add DOOH sales profile #6619 and fix(media-buy): preserve legacy discovery coverage and request semantics #7444 landed after it.
Out of scope: whether a seller that has the tool but doesn't sell the feature (revenue share, change rights) should run these storyboards. That's a separate question. For revenue share, media_buy.supported_pricing_models already exists in 3.2.4 and could be combined with the gate above under requires_all_capabilities.
What isn't verified:
- No hosted run. Our 3.2 build isn't serving publicly, and no sales agent that answered the registry has a 3.2 card.
- The grading above is our reimplementation of the registry code. It matches steps by title and task, because CLI JSON has no
step_id, and it skips the observations pass. - The run used 15.8.1 and bundle 3.2.3. That it carries over to 15.9.2 and 3.2.4 rests on the byte-identical runner files and the bundle diff above.
- We haven't run the six storyboards outside our run against any agent. Their behavior comes from the storyboard text and the runner source, and so do the cascade counts.
- We didn't emulate the storyboard-level gates in the Ask. We expect them to grade
not_applicable, since the runner reports an unmet storyboard gate ascapability_unsupportedand the grader maps that tonot_applicable(step-disposition.ts L16-L22). - We didn't check whether the tools in the other 993 candidate steps are optional, beyond the facade note above.
Related: #7586, #7599, #7646 (same fix for the provenance steps); #6729, #6753 (wholesale is optional; the grading audit); #6619, #7444, #6794, #6095 (where these storyboards came from); #7454 (hosted skip classification); #8201 (sales-dooh discovery surface); #7723, #7728, #6907, #7824 (other applicability leaks); #8231; #7897; adcp-client#2890, adcp-client#2401, adcp-client#2652.
Activity
EvgenyAndroid commentedon Oct 16, 2026
Related 3.2 applicability findings from the same sweep, filed separately because each needs a different fix: #8251 (prerelease
3.2-beta.6pins), #8252 (refine phase not gated onbuying_modes), #8254 (optional features with no capability gate). Also #8199, the provenance truth-of-claim verifier, with fix PR #8218 open.