[codex] Fix available-actions active-buy storyboard

[bokelley]

Summary

Fixes the available-actions storyboard so it first drives the buy active with start_time: "asap", a synced creative, and an assigned creative before running pause-shaped negative probes. Also fixes dev compliance latest metadata to use semver and adds a temporary ledgered SDK request-builder shim so local/CI storyboard runs preserve asap until the upstream SDK release lands.

Verification

Precommit passed unit, lint, and typecheck stages; pre-push passed version sync, current-source storyboard matrix, 3.0 compatibility storyboard matrix, and skipped docs checks because no docs changed.

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
adcp	🟢 Ready	View Preview	Jun 6, 2026, 10:57 AM

💡 Tip: Enable Workflows to automatically generate PRs for you.

[aao-release-bot]

Approving. The fix is the right shape: drive the buy to a real terminal state before probing its action surface, instead of asserting available_actions[] semantics against a buy still stuck in pending_creatives.

The core change is sound. The negative probe at direct_decrease_wrong_status was previously testing the wrong prerequisite — it fired wrong_status only because the buy never left pending_creatives, which is creative-review noise, not action-discovery enforcement. Inverting decrease_budget to allowed_statuses: ["paused"] and driving the buy active via start_time: "asap" + an assigned creative makes the probe test what the scenario name claims: a product-supported action gated out by the buy's current status.

Things I checked

• Lifecycle wiring is complete. New sync_creatives step plumbs products[0].format_ids[0] into $context.available_actions_format_id, then create_media_buy assigns it via creative_assignments alongside start_time: "asap". ad-tech-protocol-expert confirmed asap is a valid start-timing.json sentinel and that asap + assigned creative under auto_approve lands active — matching the sibling pending_creatives_to_start.yaml pattern.
• Capability gate is honest. requires_capability: media_buy.creative_approval_mode == auto_approve is a real path (get-adcp-capabilities-response.json:266, enum ["auto_approve","require_human"]) lifted verbatim from pending_creatives_to_start.yaml. Manual-review sellers grade not_applicable rather than false-failing on a prerequisite they can't satisfy. Scenario isn't in any required-clean floor, so not_applicable drops nothing.
• Stale prose got fixed. The direct_decrease_wrong_status narrative now reads "The created buy is active" instead of "still pending_creatives." Self-consistent with the inverted assertion — no contradiction left on the wire.
• build-compliance.cjs version is in scope. Both buildTo(latestDir, version, ...) sites sit inside main() after const version = getVersion(); stamping latest/ with real semver instead of the literal 'latest'. No consumer keys off the literal string value.
• Shim is ledgered and fail-loud. patch-sdk-storyboard-request-builder.cjs is idempotent (sampleStart === 'asap' early-exit), backs up before patching, and hard-exits 1 if the expected window is gone. Ledger entry, lint-shim assertion, and restore loop over both backups are all updated in lockstep. field_value_or_absent on legacy status correctly handles the 3.1 status-rename deprecation window (#4908).

Follow-ups (non-blocking — file as issues)

• Request-builder patch now runs unconditionally (run-storyboards-matrix.sh:175). The pre-existing SDK reach-ins were OVERLAY=1-gated; this new patch is not, so the asap-preservation behavior also applies to --latest-3.0 released-bundle compat runs and to other asap scenarios (proposal_finalize_asap_timing.yaml, etc.). The trap/restore is symmetric so nothing leaks — this is blast-radius, not corruption. PR body says the 3.0 matrix passed pre-push; if the universal application is deliberate, state that in the ledger entry. Otherwise gate it behind OVERLAY=1 to match the other reach-ins.

Minor nits (non-blocking)

1. Backup precedes the match guard. patch-sdk-storyboard-request-builder.cjs:23-25 writes .adcp-overlay-backup before checking that legacyWindow exists. A future SDK lacking both the legacy window and the asap marker leaves a spurious (identical) backup on the exit-1 path. run-storyboards-matrix.sh's EXIT trap cleans it, but overlay-compliance-cache.sh (set -euo pipefail) aborts first. Move the backup after both early-exit guards.

Two clean expert verdicts: code-reviewer — no blockers, one Medium noted above; ad-tech-protocol-expert — sound-with-caveats, and every caveat it raised is already closed by the diff it couldn't fetch.

LGTM. Follow-ups noted below.

[aao-release-bot]

Approving. Drives the buy to active before the negative probes so the storyboard tests available_actions enforcement, not a pending-creative lifecycle artifact — the prerequisite and the thing under test are finally separated.

Things I checked

• The status flip is internally consistent. decrease_budget allowed_statuses goes ["active"]→["paused"] in the fixture (yaml L88) and the get_products assertion (L170), and the negative-probe expected prose now reads "only for paused buys. The created buy is active" (L529-531). The probe fires decrease_budget against an active buy and correctly expects ACTION_NOT_ALLOWED. No stale assertion stranded — decrease_budget is never asserted into buy-level available_actions.
• The active-buy path is wired end to end. requires_capability: media_buy.creative_approval_mode == auto_approve + accepts_creatives + the new sync_creatives step + context_outputs capturing available_actions_format_id + $context.available_actions_format_id in the sync request all line up. Mirrors the merged pending_creatives_to_start.yaml gate pattern. Manual-review sellers grade not_applicable rather than false-failing on the lifecycle prerequisite.
• media_buy_status: active + field_value_or_absent on legacy status is the sanctioned dual-field handling for the 3.1 deprecation window. media_buy_status is the correct 3.1 wire field, active is a valid enum (enums/media-buy-status.json), start_time: "asap" is spec-valid (core/start-timing.json, create_media_buy.mdx L165).
• build-compliance.cjs:765,782 ('latest'→version) flows to generateIndex and stamps latest/index.json. No consumer compares those fields to the literal "latest"; the ^3\.0\.[0-9]+$ reader in run-storyboards-matrix.sh is unaffected. Aligns dev latest/ with the release build, which already stamped real semver.
• run-storyboards-matrix.sh trap moved out of the OVERLAY==1 block. restore_sdk_generated_schema is idempotent (acts only if the backup dir exists). Always-on EXIT trap now also cleans up the --latest-3.0/--compliance-dir paths the old code leaked. Safe and an improvement.
• Changeset present, empty frontmatter = no version bump — correct for a compliance-scenario + build-infra change with no published-package wire impact.

Follow-ups (non-blocking — file as issues)

• active-on-create is a hard pin against a spec MAY. specification.mdx L143 lets auto_approve sellers return active, pending_creatives, or pending_start from create_media_buy (seller's choice based on platform setup time). The asap + assigned-creative + auto_approve combination makes active a reasonable deterministic reading, but a conformant seller with non-trivial setup time could legitimately return pending_start and false-fail. Consider accepting active-or-pending_start and driving a flight-start transition before the negative probe.
• confirmed_at is not pinned. create-media-buy-response.json makes active + confirmed_at: null a oneOf violation. The scenario asserts active but never asserts confirmed_at != null, so a provisional buy could pass the active check while silently violating the response oneOf. Worth a field_present on confirmed_at alongside the active assertion.

The third drift-cleanup in the lineage that ends with the SDK shim ledger swapping "SDK package" for "protocol bundle" in a single sentence — notable that the prose churn ships in the same PR as the fix, but it's accurate.

LGTM. Follow-ups noted below.