Konnect platform router
Goal
Turn broad Konnect requests into one primary downstream skill with a clear next
step and one concrete thing that skill should prove next.
Use this skill as a classifier and handoff layer. Do not turn it into a
product guide or a second copy of specialist workflows.
Shared Operating Defaults
- Prefer the shared
kong-konnect MCP server when the request depends on live
Konnect state.
- If
kong-konnect MCP is unavailable, say so early and continue with
kongctl-query, repository artifacts, logs, or user-provided state.
- Preserve the repository's existing declarative toolchain. Do not convert the
user between
decK, kongctl, and Terraform unless they ask.
- Prefer current live state and current product docs over stale assumptions
when behavior, product coverage, or limits may have changed.
- When a downstream CLI appears blocked by the agent environment, let the
specialist tool skill handle sandbox retry guidance instead of classifying it
as a product failure here.
These defaults are intentionally short. Specialist skills may restate the same
guidance because they can trigger directly without this router.
Classification Order
1. Decide whether routing is needed at all
Bypass this router when the request already has one obvious owner.
Hand off directly when the user already named:
- a specialist workflow such as gateway triage, API publication, Event Gateway,
or AI Gateway
- a declarative toolchain such as
kongctl, Terraform, or decK
- a repo-authoring task that belongs to
kong-skill-authoring rather than a
Konnect product workflow
Use this router only when the request is broad, cross-surface, or still
ambiguous after the first read.
2. Decide whether this is diagnosis, inspection, or implementation
Classify the user's immediate need first:
- diagnosing current Konnect behavior
- asking for exact read-only inspection
- planning or authoring declarative changes
- asking which domain owns the problem
If the request is already a direct declarative-tool request, hand off to the
tool skill after checking whether a domain skill should diagnose first.
3. Route broad Konnect domain requests
Use one primary destination:
konnect-access-scope
- auth failures, missing permissions, wrong region, wrong org, role or team
issues, resources that likely exist but cannot be seen or edited
konnect-control-plane-bootstrap
- creating a control plane, first-run topology, hosted versus self-hosted
data plane choices, early ownership boundaries
konnect-gateway-triage
- disconnected data planes, config rollout problems, wrong control plane,
live gateway drift, traffic failures that may be control-plane or
data-plane related
konnect-observability-triage
- missing analytics, Explorer, Debugger, dataset visibility, observability
gaps after traffic already exists
konnect-api-catalog
- managed APIs, versions, specs, packages, catalog readiness, governance
state before publication
konnect-api-publish
- API publication from managed API through Catalog and Dev Portal,
visibility gaps, published-but-not-discoverable problems
konnect-app-auth
- developer app registration, credentials, approval flows, auth strategies,
published APIs that developers still cannot use
konnect-ai-gateway
- provider routing, AI Proxy setup, guardrails, prompt controls, token or
latency issues, governance around LLM traffic
konnect-event-gateway
- listeners, clusters, auth setup, hostname mapping, routing or policy
issues in Event Gateway
4. Route direct read-only inspection
Use kongctl-query when the user mainly needs:
- exact read-only command syntax
- list/get verification
- formatted CLI output
- auth verification without declarative changes
5. Route declarative implementation
After domain classification, preserve the toolchain already present in the repo:
kongctl-declarative
kongctl YAML repos, plan/apply workflows, Konnect APIOps, adoption into
kongctl
terraform-konnect
- HCL-managed Konnect resources, Terraform plan/apply/import workflows
deck-gateway
- Gateway entity config in
decK state files, diff/sync/dump, Gateway-side
config generation
terraform-kong-gateway
- self-managed Gateway entities managed through the official Terraform
provider
If the user asks for config changes but the failure domain is still ambiguous,
route to the domain skill first and let it hand off to the tool skill once the
missing link is known.
Routing Rules
- Do not send Portal visibility problems straight to
konnect-app-auth if the
API may not be published yet. Publication comes before consumer auth.
- Do not send managed-API governance or package questions straight to
konnect-api-publish when the issue is upstream catalog readiness.
- Do not send disconnected data planes to declarative tool skills before
gateway health is understood.
- Do not treat missing analytics as a gateway rollout problem until control
plane, data plane, and traffic presence are reasonably confirmed.
- Do not assume access problems are product-state problems when the caller may
simply be in the wrong org, region, or team scope.
- Do not keep repo skill-review or skill-authoring work in this router just
because the word "Konnect" appears in the skill package path.
Validation Checklist
Before handing off, verify that you can state:
- whether the request is broad enough to need routing at all
- whether the request is diagnosis, read-only inspection, implementation, or
broad triage
- which Konnect domain or tool owns the problem first, and which nearest
adjacent skill you ruled out
- whether live state inspection or repository state is the first source of
truth
- whether an existing repo toolchain should be preserved
- which one downstream skill should own the next step and what fact it should
prove first
Output Style
When responding from this router:
- name the primary downstream skill explicitly
- explain the routing reason in one short paragraph
- mention
kong-konnect MCP only when live state matters
- if routing confidence is low, name the missing fact that would separate the
top two candidates instead of guessing
- avoid giving long step-by-step instructions that belong in the downstream
skill
1---2name: konnect-platform-router3description: Route broad or ambiguous Konnect requests to the first owning specialist skill. Use when multiple Konnect surfaces are involved or the right owner is unclear, not when the user already named a specific workflow, toolchain, or non-Konnect task.4license: MIT5---67# Konnect platform router89## Goal1011Turn broad Konnect requests into one primary downstream skill with a clear next12step and one concrete thing that skill should prove next.1314Use this skill as a classifier and handoff layer. Do not turn it into a15product guide or a second copy of specialist workflows.1617## Shared Operating Defaults1819- Prefer the shared `kong-konnect` MCP server when the request depends on live20 Konnect state.21- If `kong-konnect` MCP is unavailable, say so early and continue with22 `kongctl-query`, repository artifacts, logs, or user-provided state.23- Preserve the repository's existing declarative toolchain. Do not convert the24 user between `decK`, `kongctl`, and Terraform unless they ask.25- Prefer current live state and current product docs over stale assumptions26 when behavior, product coverage, or limits may have changed.27- When a downstream CLI appears blocked by the agent environment, let the28 specialist tool skill handle sandbox retry guidance instead of classifying it29 as a product failure here.3031These defaults are intentionally short. Specialist skills may restate the same32guidance because they can trigger directly without this router.3334## Classification Order3536### 1. Decide whether routing is needed at all3738Bypass this router when the request already has one obvious owner.3940Hand off directly when the user already named:4142- a specialist workflow such as gateway triage, API publication, Event Gateway,43 or AI Gateway44- a declarative toolchain such as `kongctl`, Terraform, or `decK`45- a repo-authoring task that belongs to `kong-skill-authoring` rather than a46 Konnect product workflow4748Use this router only when the request is broad, cross-surface, or still49ambiguous after the first read.5051### 2. Decide whether this is diagnosis, inspection, or implementation5253Classify the user's immediate need first:5455- diagnosing current Konnect behavior56- asking for exact read-only inspection57- planning or authoring declarative changes58- asking which domain owns the problem5960If the request is already a direct declarative-tool request, hand off to the61tool skill after checking whether a domain skill should diagnose first.6263### 3. Route broad Konnect domain requests6465Use one primary destination:6667- `konnect-access-scope`68 - auth failures, missing permissions, wrong region, wrong org, role or team69 issues, resources that likely exist but cannot be seen or edited70- `konnect-control-plane-bootstrap`71 - creating a control plane, first-run topology, hosted versus self-hosted72 data plane choices, early ownership boundaries73- `konnect-gateway-triage`74 - disconnected data planes, config rollout problems, wrong control plane,75 live gateway drift, traffic failures that may be control-plane or76 data-plane related77- `konnect-observability-triage`78 - missing analytics, Explorer, Debugger, dataset visibility, observability79 gaps after traffic already exists80- `konnect-api-catalog`81 - managed APIs, versions, specs, packages, catalog readiness, governance82 state before publication83- `konnect-api-publish`84 - API publication from managed API through Catalog and Dev Portal,85 visibility gaps, published-but-not-discoverable problems86- `konnect-app-auth`87 - developer app registration, credentials, approval flows, auth strategies,88 published APIs that developers still cannot use89- `konnect-ai-gateway`90 - provider routing, AI Proxy setup, guardrails, prompt controls, token or91 latency issues, governance around LLM traffic92- `konnect-event-gateway`93 - listeners, clusters, auth setup, hostname mapping, routing or policy94 issues in Event Gateway9596### 4. Route direct read-only inspection9798Use `kongctl-query` when the user mainly needs:99100- exact read-only command syntax101- list/get verification102- formatted CLI output103- auth verification without declarative changes104105### 5. Route declarative implementation106107After domain classification, preserve the toolchain already present in the repo:108109- `kongctl-declarative`110 - `kongctl` YAML repos, plan/apply workflows, Konnect APIOps, adoption into111 `kongctl`112- `terraform-konnect`113 - HCL-managed Konnect resources, Terraform plan/apply/import workflows114- `deck-gateway`115 - Gateway entity config in `decK` state files, diff/sync/dump, Gateway-side116 config generation117- `terraform-kong-gateway`118 - self-managed Gateway entities managed through the official Terraform119 provider120121If the user asks for config changes but the failure domain is still ambiguous,122route to the domain skill first and let it hand off to the tool skill once the123missing link is known.124125## Routing Rules126127- Do not send Portal visibility problems straight to `konnect-app-auth` if the128 API may not be published yet. Publication comes before consumer auth.129- Do not send managed-API governance or package questions straight to130 `konnect-api-publish` when the issue is upstream catalog readiness.131- Do not send disconnected data planes to declarative tool skills before132 gateway health is understood.133- Do not treat missing analytics as a gateway rollout problem until control134 plane, data plane, and traffic presence are reasonably confirmed.135- Do not assume access problems are product-state problems when the caller may136 simply be in the wrong org, region, or team scope.137- Do not keep repo skill-review or skill-authoring work in this router just138 because the word "Konnect" appears in the skill package path.139140## Validation Checklist141142Before handing off, verify that you can state:143144- whether the request is broad enough to need routing at all145- whether the request is diagnosis, read-only inspection, implementation, or146 broad triage147- which Konnect domain or tool owns the problem first, and which nearest148 adjacent skill you ruled out149- whether live state inspection or repository state is the first source of150 truth151- whether an existing repo toolchain should be preserved152- which one downstream skill should own the next step and what fact it should153 prove first154155## Output Style156157When responding from this router:158159- name the primary downstream skill explicitly160- explain the routing reason in one short paragraph161- mention `kong-konnect` MCP only when live state matters162- if routing confidence is low, name the missing fact that would separate the163 top two candidates instead of guessing164- avoid giving long step-by-step instructions that belong in the downstream165 skill