# Reported issues for Kody

Pod holds 17 of 26 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user.

Back to [Kody](/mcp/kody).

## Most discussed

### PLAN: heykody.app → kody.codes cutover (awaiting explicit GO from Kent)

> [!IMPORTANT]
> ## ⛔ GATE: PLAN ONLY — nothing below executes until Kent comments "go" (or tells the agent directly)
> No DNS changes, no Cloudflare changes, no OAuth console changes, no variable flips, no merges to `main`. Branch prep marked **[preparable]** may be built and CI'd but stays unmerged. Everything marked **[blocking-Kent]** or **[flip]** waits for the gate.

Decision: the public apex moves **heykody.app → kody.codes**. heykody.app keeps working as a redirect indefinitely. This is 

[Read the thread](https://github.com/kentcdodds/kody/issues/1417) · 2026-08-13 · closed · 3 comments

### Future work: untangle MCP from the Remix app; nightly test-suite cleanup automation

Tracking issue for two deferred items from the mono-worker extraction program (ADR 0016).

## 1. Untangle MCP from the Remix app

Today the MCP server and the Remix app live in the same worker (`packages/worker`) and share auth/session/D1 plumbing. ADR 0016 deliberately stopped the extraction at the runtime and jobs lanes because MCP↔app coupling is the tightest seam — but it remains worth investigating.

Motivations:
- Independent deploy/rollback for the MCP surface (the fastest-evolving protoc

[Read the thread](https://github.com/kentcdodds/kody/issues/1399) · 2026-08-12 · closed · outside contributor · 3 comments

### Repro needed: opaque Spotify device ID rejected as sensitive by a safety guard

From platform feedback `aea5c250`: passing an opaque Spotify device ID to
`@kentcdodds/spotify/play-context` triggered a safety rejection ("classified as a
credential / sensitive personal identifier"); retrying without the device ID succeeded.

I could not locate a platform-side classifier that would reject an opaque provider
resource ID — nothing in the secrets or capability-input path
(`packages/worker/src/mcp/secrets/…`) does content classification of argument values.
So this is either packag

[Read the thread](https://github.com/kentcdodds/kody/issues/1404) · 2026-08-12 · closed · outside contributor · 2 comments

### Cleanup: remove MCP agent session one-shot backfill tooling after production marker

## Context

Migration `0086` + the operator sweep (`npm run backfill:mcp-agent-sessions` / `🧭 Backfill MCP agent sessions` workflow) index dormant MCP Agent Durable Objects into `mcp_agent_sessions` and write `deployment_backfill_markers` key `mcp_agent_sessions` version `1`. Account deletion fails closed until that marker exists.

Once production dry-run is clean and `mode=execute` has written the marker, the one-shot operator surface is dead weight. Steady-state indexing already happens via `r

[Read the thread](https://github.com/kentcdodds/kody/issues/1051) · 2026-07-30 · closed · 2 comments

### MCP readiness observation: MCP handshake did not return a valid tool list

**SaSame MCP Factory — automated observation notice** (MCP handshake did not return a valid tool list)

- **MCP name (per official MCP registry):** `io.github.kentcdodds/kody`
- **Endpoint observed:** `https://heykody.dev/mcp`
- **Observed at:** 2026-07-28T13:05:06.938Z (UTC)
- **Observed behavior:** grade D (0/10 criteria), including a failed `tools/list` step (C1 — Protocol handshake conformance).
- **Top gap:** C1 Protocol handshake conformance — initialize result keys: (none, status 401)

**

[Read the thread](https://github.com/kentcdodds/kody/issues/1005) · 2026-07-28 · closed · external user · 2 comments

### feat: add secret_set capability so codemode skills can persist refreshed OAuth tokens

## Problem

Skills that call OAuth-protected APIs (e.g. Spotify) need to refresh access tokens when they expire. Currently:

- Skills can **read** secrets via `{{secret:name}}` placeholders in fetch calls
- Skills can **list** secret metadata via `secret_list`
- Skills **cannot write/update** secrets from codemode

This means a skill can do a token refresh via fetch to the token endpoint, get a new access token back, but has no way to persist it. The next skill invocation has to refresh again, a

[Read the thread](https://github.com/kentcdodds/kody/issues/92) · 2026-03-29 · closed · 2 comments

### MCP server listed on ATX - claim or correct the record

Hi - I run [ATX](https://atxregistry.org), an independent registry that records
evidence about publicly reachable MCP servers. This server is listed:

https://atxregistry.org/s/<HOST>

Nothing is needed from you. Flagging it in case you'd like to correct anything,
or claim the domain to get a verified-operator badge and trigger rescans after
a change - one DNS TXT record, no account.

Collection is unauthenticated `initialize` + `tools/list` only, no tool is ever
invoked, robots.txt is honoured,

[Read the thread](https://github.com/kentcdodds/kody/issues/1756) · 2026-08-25 · closed · external user · 1 comment

### Adopt Agents MCP SDK v2 / agents/mcp/server path

## Context

We upgraded to `agents@0.20.1` and installed required peers `@modelcontextprotocol/client@2.0.0` + `@modelcontextprotocol/server@2.0.0`, while continuing to serve MCP via the legacy **McpAgent** + `@modelcontextprotocol/sdk@1.30.0` path.

`agents@0.20` adds:

- MCP SDK **v2** client/server support
- `agents/mcp/server` entry that can serve stateless MCP without retaining McpAgent / WorkerTransport / SDK v1 modules
- Client discovery via `server/discover` with legacy `initialize` fall

[Read the thread](https://github.com/kentcdodds/kody/issues/1191) · 2026-08-03 · closed · 1 comment

## Most recent

### Home connector observability gaps can hide disconnects and tool-snapshot failures

## Summary

We investigated why scheduled shade automation events were failing with `Tool "home_bond_shade_set_position" not found` around 2026-04-26 02:09-02:16 UTC and 03:30 UTC, while the `kody-home-connector` Sentry project had no new events after 2026-03-25.

A source review found multiple observability gaps that can make home-connector websocket disconnects or capability-registration problems effectively silent.

## Findings

### 1) Containerized production can bypass home-connector Sentry

[Read the thread](https://github.com/kentcdodds/kody/issues/268) · 2026-04-26 · closed · outside contributor · 0 comments

### Follow up package-first terminology in talk deck and examples

The package-first PR is merged, but the talk deck still teaches the pre-package-first model in several places.

Follow-up cleanup:
- `docs/talks/kody-mcp-runtime/slides.md` still references `saved skills`, `saved apps`, and the old async-function framing for `execute`.
- The deck should be updated to match the merged package-first architecture: packages as the top-level saved primitive, package-owned app/jobs/exports, and module-oriented execute semantics.

Examples from the merged tree:
- `docs

[Read the thread](https://github.com/kentcdodds/kody/issues/225) · 2026-04-20 · closed · 0 comments

### Fresh repo-backed skill creation can fail with 401 Unauthorized while existing migrated skills run

## Summary

Existing migrated repo-backed skills run successfully, but creating a brand-new repo-backed skill in production can still fail during the save path with `401 Unauthorized`.

## Production observations

Using production Kody MCP after the repo-backed source migration deploy:

1. Existing migrated skills run successfully. Example: `weather-forecast` runs and returns a valid result.
2. Creating a fresh test skill with a new name via `meta_save_skill` failed on the create/save step with:

[Read the thread](https://github.com/kentcdodds/kody/issues/197) · 2026-04-18 · closed · 1 comment

### Remove legacy D1 source storage after repo-backed backfill completes

## Summary

This issue tracks the **final cleanup PR** that should land **after** the repo-backed source backfill is completed and verified in production.

The migration foundation landed in #181, and the production backfill path is being added in #187. Once production data is migrated and verified, we should remove the legacy D1 source-storage and fallback code so repo-backed sources become the only canonical source of truth.

## Goal

After this issue is complete:

- saved apps, saved skills, 

[Read the thread](https://github.com/kentcdodds/kody/issues/188) · 2026-04-17 · closed · 0 comments

### feat: first-class scheduler — run codemode code on cron or datetime

## Overview

Add a `SchedulerDO` (Durable Object) to Kody that allows codemode skills/code to be scheduled for one-shot or recurring (cron) execution, with full lifecycle management via codemode capabilities.

## Motivation

Today `codemode.execute` is purely synchronous and on-demand. There's no way to say "run this skill every morning at 7am" or "run this code once on Friday at 3pm." This feature makes that a first-class primitive.

## Architecture

### Durable Object: `SchedulerDO`

A single 

[Read the thread](https://github.com/kentcdodds/kody/issues/156) · 2026-04-12 · closed · 0 comments

### docs: add raw-content-blocks guide and reference it from execute tool description

## What

The `__mcpContent` sentinel was added in #148 but is completely undocumented. Agents have no way to know it exists.

## Changes needed

### 1. New file: `docs/use/raw-content-blocks.md`

Create a concise guide explaining the feature. Something like:

```md
# Raw MCP content blocks

By default, `execute` serialises its return value as a single `text` content
block. When you need to return a non-text block — most commonly an `image` for
screenshots or charts — return an object with a `__m

[Read the thread](https://github.com/kentcdodds/kody/issues/150) · 2026-04-11 · closed · 0 comments

### execute tool: support returning raw MCP content blocks via __mcpContent sentinel

## Problem

The `execute` tool always serializes its return value as a single `text` content block via `formatExecutionOutput`. This means execute sandbox code cannot return image (or other non-text) MCP content blocks — base64 image data just ends up embedded in a JSON string, not as a real `image` content block that Claude can visually process.

## Proposed Solution

Add a sentinel shape that execute sandbox code can return to opt into raw MCP content blocks:

```ts
return {
  __mcpContent: [

[Read the thread](https://github.com/kentcdodds/kody/issues/148) · 2026-04-11 · closed · 0 comments

### Plan: Progressive disclosure of contextual facts for MCP agents

## Goal

**Create a plan (NOT an implementation) for progressive disclosure of contextual facts to agents interacting with Kody's MCP server.**

The owner (Kent) will review and discuss the plan with the agent before any code is written.

## Problem

Kody's MCP tool descriptions currently front-load a large amount of information that agents may never need. This wastes context window space and makes the initial tool descriptions harder to parse. Agents need to absorb everything up front — OAuth p

[Read the thread](https://github.com/kentcdodds/kody/issues/127) · 2026-04-01 · closed · 0 comments

### createAuthenticatedFetch does not persist rotated refresh tokens

## Bug

`refreshAccessTokenWithConnector` in `packages/worker/src/mcp/execute-modules/codemode-utils.ts` does not persist rotated refresh tokens from the token endpoint response.

## Problem

Many OAuth providers (notably Spotify PKCE) rotate the refresh token on every use — the response includes a new `refresh_token` and the old one is immediately revoked. The current implementation only reads `payload.access_token` and returns it, discarding `payload.refresh_token`. This means:

1. First refre

[Read the thread](https://github.com/kentcdodds/kody/issues/124) · 2026-03-31 · closed · 1 comment

The remaining reports are on [the project's issue tracker](https://github.com/kentcdodds/kody/issues).