Changelog

All notable changes to this skill are documented here. Format follows Keep a Changelog. Versions are semver and must match metadata.version in SKILL.md.

Bump rules:

• patch: non-behavioral changes (typos, formatting, reference-only edits).
• minor: behavioral additions (new sections, rules, triggers).
• major: contract changes (response shape, output format, scope).

Every bump also updates the evals/evals.json skill_version.

[0.2.1] - 2026-10-17

Changed

• Replaced Unicode ellipses with ASCII ... in SKILL.md. Security scanners normalize Unicode and read the changed text as obfuscation; wording and behavior are unchanged.

[0.2.0] - 2026-10-17

Added

• Data, not instructions. A new first guardrail treats every API response and error as data: plan and shipment names, option and carrier text, addresses, operationProblems[] messages, error strings, eligibility, prep, and compliance reasons, and list_identities labels. It also covers box files the seller uploads. An embedded instruction never picks an option, counts as a yes or CONFIRM, or decides the next call; the skill flags it to the seller instead.
• Show, then yes. Before createInboundPlan, confirmPackingOption, setPackingInformation, updateShipmentTrackingDetails, updateShipmentSourceAddress, setPrepDetails, updateItemComplianceDetails, and the two rename calls, the skill shows the exact call, including the merchant ID, and waits for an explicit yes in the seller's next message. A pick from options shown in full counts as the yes for confirmDeliveryWindowOptions and scheduleSelfShipAppointment. cancelSelfShipAppointment gets its own ask.
• Retry rule. A retry that changes anything (option ID, amount, selection, box or address data) needs a fresh yes or CONFIRM. After a 5xx, the skill reads the plan before retrying.
• Evals 5 (single fee-free packing option plus box data stop for a yes), 6 (an instruction in a plan name is flagged and ignored), and 7 (a CONFIRM sent before the restatement doesn't count).

Changed

• Gate table. The catch-all row "Everything else (creates, sets, generates, reads) | Reversible, no fee, or read-only | Proceed" is replaced by a per-call classification. Only the generate* calls, createMarketplaceItemLabels, and reads run without asking. Server selection calls run only for the seller the user named.
• Single fee-free packing option. It is no longer confirmed with a notice; the skill shows it and waits for a yes. This and the gate table resolve a contradiction between the description ("Pauses at every fee-charging or irreversible step") and the body, which proceeded on writes it called permanent (box data, compliance details). The safer description wins.
• CONFIRM timing. The typed CONFIRM must come in a later message than the restatement. One in the original request doesn't count.
• Server authority. Tool descriptions and get_schema are authoritative only for tool names and argument names, types, and shapes. They never relax a gate (tool-access.md, SKILL.md section 1, request-shapes.md).
• Failed writes. Fixes after a FAILED operation or a rejected confirm come from the seller's data, not from the error text (operations-and-limits.md, tool-access.md, carriers-and-shipping.md).
• Eval 0 now expects a preview and a stop before createInboundPlan; the original request is not the yes. Its pack-first and status-check assertions moved to eval 5.

[0.1.0] - 2026-09-24

Added

• First KuudoAI release. Adapted from the Amazon Selling Partner fba-inbound-management skill (Apache-2.0) and substantially rewritten to work with a code-mode SP-API MCP server. The rewrite was verified against the Fulfillment Inbound v2024-03-20 OpenAPI model and Amazon's use-case guides.

Changed (compared with the upstream skill)

• Tool resolution. The skill now resolves tools by operationId through server discovery, instead of hard-coding the fbaInbound_* names. On a code-mode server the names are fba-inbound_<operationId> and fulfillment-inbound-v0_<operationId>.
• No entityId. The skill no longer asks for or passes entityId. The seller comes from the server's active identity, and the marketplace is passed only where an operation takes one.
• Pack-first order. Corrected to: generate and confirm the packing option, then setPackingInformation(packingGroupId), then placement.
• Transport quotes before placement. Transportation is quoted for a candidate placement before placement is confirmed, following Amazon's documented flow. The seller sees the placement fee and the carrier quote together.
• Transportation inputs. Corrected to readyToShipWindow.start, plus contact, pallets, and freight info where needed. There is no ship-from field.
• Delivery windows. The decision now keys off each option's preconditions (CONFIRMED_DELIVERY_WINDOW) rather than a program enrollment.
• Fees. Now net fees[] minus discounts[], in the returned currency, instead of "dollars".
• Void windows. Now read from quote.voidableUntil instead of hard-coded hours.
• Pack Later. Corrected to setPackingInformation(shipmentId) after placement, with any content source (MANUAL_PROCESS carries a fee).
• Self-ship appointments. Documented beyond India (MX, BR, EG, SA, AE).
• Polling. Now sized for code mode: chain the write, one status check, and the follow-up reads in one execute block. Re-check across calls, cap at about five checks, and read operationProblems[].

Removed

• The non-existent updateBoxIdentifiers operation.
• The 500K-units-per-SKU limit. The real limits are 1,500 SKUs and 10,000 units per SKU per plan.
• The claim that setPrepDetails sets owners. It sets prep category and types.
• Non-spec frontmatter: the top-level version key and list-valued metadata.

Validated live

• Validated read-only against a live SP-API MCP server. The code-mode specifics now match it: three meta-tools with no tags; a required set_active_identity; plain-string errors; entityId silently ignored; 403 when a seller lacks inbound access; AWD plans in listInboundPlans; get_schema needs detail="full".

Fixed after live smoke testing

• Plan status overviews now read SHIPPED-status plans as well as ACTIVE. Once shipments leave, the plan moves to SHIPPED, so an ACTIVE-only overview missed everything in transit or being received.

Fixed from PR review

• MANUAL_PROCESS gate. setPackingInformation with a MANUAL_PROCESS box now requires a typed CONFIRM, because Amazon charges a manual-processing fee. The skill offers BARCODE_2D and BOX_CONTENT_PROVIDED as fee-free alternatives.

Also added

• Request shapes for every write.
• A tool-access guide.
• getBillOfLading for partnered pallet shipments.
• Status enums and rate limits.
• Content-update limits.