Reported issues for zendesk-mcp-server
Pod holds 19 of 58 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 zendesk-mcp-server.
Most discussed
find_translation_gaps: try the translations sideload to drop the per-node fan-out
Context
find_translation_gaps (added in #225, closes #224) audits the category → section tree for a target locale and reports every node with no translation or an unpublished draft. It gets the draft state by calling GET /{sections|categories}/{id}/translations once per node, in waves of 5, bounded by ZENDESK_TRANSLATION_GAP_SCAN_MAX_NODES (default 60). On a 58-node tenant that is ~61 requests for one audit, and past the cap the report has to admit it left nodes unchecked.
That d
Read the thread · 2026-08-11 · closed · 4 comments
Exit on stdin close in stdio mode
Running in stdio mode behind npx, the server never receives SIGTERM. The npm exec and sh -c links in the chain do not relay signals, so the only sign that the client is gone is EOF on stdin.
Today the process still exits, but only because nothing keeps the event loop alive once stdin closes. That is circumstantial, not by design: adding an HTTP keep-alive agent, a refresh timer or a persistent cache would turn it into a process that outlives every session.
Suggestion: listen for stdin
Read the thread · 2026-08-20 · closed · 3 comments
Can't pull a specific Help Center article as an MCP resource (for @-mention / context)
What
Help Center articles can only be reached through tools (search_articles, get_article, section tools). There's no MCP resource to reference a specific article, so clients that let users attach/@-mention resources can't pin an article as context.
Why / the problem
The server already ships one pull-only resource (zendesk-hc://topology), which proves the pattern is valuable. Extending it to articles would let a user surface search results, then **pin the relevant article as a
Read the thread · 2026-07-04 · closed · 3 comments
Ticket field definitions are not readable — custom-field writes are blind
What
There is no way to read ticket field definitions — the list of ticket fields (system + custom), their ids, types, and the valid option values for dropdown/multiselect fields.
Why / the problem
create_ticket and update_ticket already accept custom_fields as { id, value } pairs, but nothing lets the model discover those ids or the values a field will accept. Today the descriptions punt with "field ids come from your Zendesk admin settings" — i.e. the assistant writes
Read the thread · 2026-07-04 · closed · 3 comments
Support markdown-inline images in comment bodies for get_ticket_attachments
Context
Follow-up from #13 (merged). During manual testing, @florentleveque identified an edge case not covered by get_ticket_attachments: images referenced inline in the comment body as markdown  are not enrolled in the comment's attachments array returned by the Zendesk API, so they are invisible to the new tool.
Current behavior
get_ticket_attachments collects images via comment.attachments[]. Images that authors embed directly in the message body (drag-and-d
Read the thread · 2026-05-18 · closed · 3 comments
No section or category translation tools: articles get published in a target locale under sections that only exist in the source locale
Problem
Translations are only reachable at the article level. list_article_translations takes an article_id and nothing else, and create_article_translation / update_article_translation are article-scoped too. Sections and categories have no translation tool at all.
Reading is partly possible as a side effect: list_sections and list_categories accept a locale, and the filtered listing only returns the ones that have a translation in that locale. That is enough to spot a gap,
Read the thread · 2026-08-11 · closed · 2 comments
docs: document the SSE keep-alive and proxy-buffering behaviour of the HTTP transport
Summary
Since @modelcontextprotocol/sdk 1.30.0 the Streamable HTTP transport keeps its SSE streams alive on its own: it writes periodic comment frames and asks intermediaries not to buffer the stream. Our node:http transport is a thin wrapper around the SDK's, so we inherit all of it — and docs/http-deployment.md says nothing about it, even though the doc's whole audience is people running the server behind a reverse proxy.
This is documentation only; no behaviour change is proposed.
Read the thread · 2026-08-04 · closed · 2 comments
Logging a field with a circular reference crashes the caller
What happens
Passing a value containing a circular reference to any logger method throws RangeError: Maximum call stack size exceeded out of the call that emitted the log.
const log = createLogger('debug');
const circular: Record<string, unknown> = {};
circular['self'] = circular;
log.info('evt', { circular }); // throws
RangeError: Maximum call stack size exceeded
❯ isSensitive src/utils/logger.ts:54:39
❯ redactValue src/utils/logger.ts:64:18
❯ redactValue src/utils/lo
[Read the thread](https://github.com/fruggr/zendesk-mcp-server/issues/198) · 2026-08-02 · closed · 2 comments
## Most recent
### Migrate to MCP SDK v2, shipped through a dedicated `beta` release channel
## Summary
Execute the migration from `@modelcontextprotocol/sdk@1.x` to the v2 package split
(`@modelcontextprotocol/core` / `server` / `client` / `node`), and — because of what
this change touches — ship it **through a pre-release channel with a real beta period
of several days to weeks**, not straight to `@latest`.
This issue is the *migration*, deliberately separate from #213, which is the
*evaluation* and explicitly leaves the migration out of its own scope. #213's written
recommendation
[Read the thread](https://github.com/fruggr/zendesk-mcp-server/issues/231) · 2026-08-12 · open · 0 comments
### Publish the mutation score to the Stryker dashboard, so the rollout is measurable over time
Sub-task of #208, and **first** — not because it raises a score, but because nothing else in the rollout is currently measurable over time.
## The gap
#197 shipped reporting that keeps nothing. The HTML report goes out as a CI artifact (30 days for the baseline job, 7 for PRs) and then vanishes; the cached `stryker-incremental.json` is current state, overwritten on every push to `main`, not a history. So "where was `src/auth` a month ago" has no answer today, and #208's table is a snapshot mai
[Read the thread](https://github.com/fruggr/zendesk-mcp-server/issues/214) · 2026-08-04 · closed · 1 comment
### Evaluate the MCP SDK v2 migration — Renovate is structurally blind to it
## Summary
The MCP TypeScript SDK has moved to a **v2 published under new package names**. We depend on `@modelcontextprotocol/sdk`, whose latest release is `1.30.0` — the final version of the v1 line. **No automation will ever tell us about v2**, because it is a package split, not a version bump.
This issue is an **evaluation**, not a decision to migrate. Its output is a written recommendation (migrate now / later / never, and in what order), not code.
## Why Renovate will never propose it
[Read the thread](https://github.com/fruggr/zendesk-mcp-server/issues/213) · 2026-08-04 · closed · 1 comment
### Bring src/tools/** into the mutation scope
Sub-task of #208, and the only one that is **blocked** rather than merely unstarted. `src/tools/**` has never been measured because it is excluded from `mutate` in `stryker.config.mjs`.
## Why it is out
Stryker's `StringLiteral` mutator empties every string, and this directory is mostly `.describe()` calls — the per-parameter descriptions the MCP tool surface is built from and that `AGENTS.md` requires for the Glama quality bar. Each one yields a survivor that says nothing about test quality,
[Read the thread](https://github.com/fruggr/zendesk-mcp-server/issues/212) · 2026-08-04 · open · 1 comment
### Raise the mutation score of src/client
Sub-task of #208. `src/client/zendesk-api.ts` is the single HTTP path every tool goes through, so an unasserted branch here is a bug surface for the whole tool set.
## Measured
Full-scope baseline on `main` ([run 14](https://github.com/fruggr/zendesk-mcp-server/actions/runs/30934244169), commit `31df70c`, 2026-08-04): **68.04 % total / 71.74 % covered** — 66 killed, 26 survived, 5 never reached, out of 97 mutants.
Escaped, by mutator:
| Mutator | Count |
| --- | ---: |
| ConditionalExpressio
[Read the thread](https://github.com/fruggr/zendesk-mcp-server/issues/210) · 2026-08-04 · closed · 1 comment
### Raise the mutation score of src/auth
Sub-task of #208. **The worst-scoring area, and the one where it matters most:** this is the OAuth 2.1 / PKCE code that holds the only credential path the server has.
## Measured
Full-scope baseline on `main` ([run 14](https://github.com/fruggr/zendesk-mcp-server/actions/runs/30934244169), commit `31df70c`, 2026-08-04): `src/auth` scores **58.68 % total / 62.12 % covered** — 223 killed, 136 survived, 21 never reached.
**157 escaped mutants, more than the rest of the scope put together** (the
[Read the thread](https://github.com/fruggr/zendesk-mcp-server/issues/209) · 2026-08-04 · open · 1 comment
### Proxy dispatch leaks raw ZodError JSON: type errors are unreadable in the default namespace mode
## Summary
On the proxy dispatch path, `createStrictParamsParser` handles the unknown-key case with a crafted message but falls through to `throw result.error` (`src/utils/validation.ts:38`) for every other validation failure. That raw `ZodError` reaches the client as its `.message`, which for Zod is a **JSON dump of the issue array**.
This affects `namespace` (the **default**) and `single` modes. In `all` mode the SDK owns the parse, and SDK 1.30.0 just improved that formatting — so the two m
[Read the thread](https://github.com/fruggr/zendesk-mcp-server/issues/206) · 2026-08-04 · open · 0 comments
### Cap attachment payloads: an oversized JSON-RPC message now kills the whole stdio session
## Summary
`@modelcontextprotocol/sdk` 1.30.0 introduces a **10 MB cap on the stdio `ReadBuffer`** (`STDIO_DEFAULT_MAX_BUFFER_SIZE = 10485760`). A message exceeding it does not merely fail the call: the transport emits an error and **closes**, tearing down the MCP session.
Our tool surface can produce such messages in **both directions**, because neither the inbound attachment upload nor the outbound embedded-image response has a total-size budget.
## Evidence
Measured on `StdioServerTranspo
[Read the thread](https://github.com/fruggr/zendesk-mcp-server/issues/205) · 2026-08-04 · open · 0 comments
### pnpm check does not lint the root config files
## What
`pnpm check` and the lefthook pre-commit hook are both scoped to `src/ tests/ scripts/`:
```json
"check": "node scripts/biome.mjs check --error-on-warnings src/ tests/ scripts/",
So no file at the repo root is ever linted by the repo's own gate — including vitest.config.ts, tsdown.config.ts, biome.json itself, and server.json.
How it surfaced
In #197, stryker.config.mjs was added at the root and tripped style/noDefaultExport — which native Stryker ESM configs
Read the thread · 2026-08-03 · closed · 1 comment
Harden the shared Zendesk HTTP client: retry transient failures + contextual network errors
Problem
Every tool goes through executeRequest in src/client/zendesk-api.ts. Two rough edges surfaced during the #170 functional validation (run on a slow Android/Termux link with several requests in flight):
- Opaque network errors. When
fetchitself throws (connection reset, DNS, timeout — theawait fetch(...)path, before any HTTP status), the caller (and the LLM / MCP client) sees a barefetch failedwith nothing naming which operation / article / page failed. HTTP *respo
Read the thread · 2026-07-30 · closed · 2 comments
CLI parser silently ignores value-taking flags with a missing or empty value
Problem
Every value-taking flag branch in parseCliArgs (src/config.ts) is guarded by arg === '--flag' && next. When the value is missing or empty, the guard is falsy and the flag is silently dropped instead of failing at startup. This affects all value-taking flags equally: --mode, --namespace, --tool, --log-level, --transport, --host, --port, --public-url, --cors-origin, --callback-port and (once #173 lands) --hc-resource-scheme.
Two concrete failure shapes:
Read the thread · 2026-07-20 · closed · 1 comment
The remaining reports are on the project's issue tracker.