# Appendix: search of the MCP specification and proposals, 15 November 2026

IsoKron, November 2026. Appendix to "Do popular packaged MCP servers say whether a write happened?".

This appendix records the search behind the dated statement in the article's section "What the underlying APIs already provide". The question: does the MCP specification, or any proposal filed against it, define a field in a tool result for whether a write took effect, or for an unknown outcome?

**Result.** The specification defines no such field. One open draft proposal (SEP-2848) defines both, for tool calls held for out-of-band approval. Two closed proposals for signed audit records define a call-outcome status with no value for unknown.

## 1. Setup

- **Date:** 15 November 2026. Searches were run on that date; GitHub counts change over time.
- **Tools:** `curl` on modelcontextprotocol.io and raw.githubusercontent.com; the GitHub REST search endpoint `search/issues`, called read-only with `gh api -X GET search/issues -f q='<terms> repo:modelcontextprotocol/modelcontextprotocol is:issue'` (and `is:pull-request`); the pull-request pages of #2848 and #3182 in a browser.
- **Repositories:**
  - `modelcontextprotocol/modelcontextprotocol`: specification, schema, SEPs (as issues and pull requests, and as `seps/*.md` files).
  - `modelcontextprotocol/specification`: the specification page links its schema there. Raw files at that path are byte-identical to `modelcontextprotocol/modelcontextprotocol` (`schema/2026-07-28/schema.ts` compared with `cmp`). GitHub search rejects it as a repository (HTTP 422, "cannot be searched … do not exist"), consistent with a renamed repository. It was therefore covered by searching `modelcontextprotocol/modelcontextprotocol`.

## 2. The specification

- **Current version:** **2026-07-28**. Source: https://modelcontextprotocol.io/specification/versioning ("The **current** protocol version is 2026-07-28"); https://modelcontextprotocol.io/specification/latest resolves to it. Earlier versions in the docs index: 2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05.
- **Tools section read in full:** https://modelcontextprotocol.io/specification/2026-07-28/server/tools, plus `CallToolResult` and `ToolAnnotations` in https://github.com/modelcontextprotocol/modelcontextprotocol/blob/main/schema/2026-07-28/schema.ts.

| Item | What the 2026-07-28 text says | Bears on the question? |
|---|---|---|
| `isError` | "Whether the tool call ended in an error. If not set, this is assumed to be false (the call was successful)." Tool-originated errors "SHOULD be reported inside the result object, with `isError` set to true, _not_ as an MCP protocol-level error response. Otherwise, the LLM would not be able to see that an error occurred and self-correct." | Says the call failed, not whether a write took effect. No unknown value. |
| Error handling | Two mechanisms: protocol errors (JSON-RPC: unknown tool, malformed request, server errors) and tool execution errors (API failures, input validation, business logic) in the result with `isError: true`. | No outcome field. |
| `structuredContent` / `outputSchema` | Any JSON value, conforming to `outputSchema` if declared. No reserved members. | A place to put an outcome; defines none. |
| `ToolAnnotations` | `readOnlyHint`, `destructiveHint`, `idempotentHint` ("calling the tool repeatedly with the same arguments will have no additional effect on its environment"), `openWorldHint`. All "hints", per tool, untrusted unless from a trusted server. | Describe the tool, not one call's outcome. |
| `resultType` | New in 2026-07-28: `"complete" \| "input_required" \| string` on results. | Request flow (multi round-trip), not write outcome. |
| Tasks extension | https://modelcontextprotocol.io/extensions/tasks/overview: task status `working`, `input_required`, `completed`, `failed`, `cancelled`. | Task lifecycle, not whether a write applied. No unknown value. |

Full-text check of the whole documentation (`https://modelcontextprotocol.io/llms-full.txt`, 2.38 MB, includes specification, docs and published SEPs), case-insensitive occurrence counts: `unknown outcome` 0, `outcome is unknown` 0, `partial success` 0, `idempot` 10 (the `idempotentHint` definition, the annotations working-group charter, and the Tasks SEP rationale on client-generated task IDs), `at most once` 2 (neither about tool calls), `exactly once` 1 (Skills resources), `isError` 46.

## 3. Issues and pull requests in `modelcontextprotocol/modelcontextprotocol`

Query form: `<terms> repo:modelcontextprotocol/modelcontextprotocol is:issue` and `… is:pull-request`. GitHub matches title, body and comments; quoted terms are phrases. Counts are `total_count`.

### Main term list

| Terms | Issues | PRs |
|---|---|---|
| idempotency | 14 | 13 |
| idempotent | 10 | 12 |
| idempotentHint | 11 | 9 |
| idempotency key | 13 | 11 |
| "idempotency key" | 7 | 5 |
| "unknown outcome" | 1 | 1 |
| "outcome unknown" | 1 | 1 |
| "partial success" | 2 | 2 |
| "partial failure" | 3 | 2 |
| "write result" | 7 | 3 |
| "did it apply" | 0 | 0 |
| retry | 71 | 65 |
| retryable | 3 | 0 |
| outcome | 39 | 45 |
| conflict | 47 | 249 |
| "structuredContent" error | 17 | 18 |
| "isError" status | 20 | 19 |
| isError | 28 | 27 |
| "error code" tool result | 31 | 30 |
| "tool result" status | 89 | 66 |
| side effect | 37 | 36 |
| "at most once" | 2 | 1 |
| "exactly once" | 2 | 5 |
| SEP idempotency | 14 | 10 |
| SEP retry | 53 | 47 |
| SEP isError | 19 | 18 |
| SEP structuredContent | 17 | 13 |
| "tool error" structured | 25 | 18 |

### Follow-up terms (added after reading the first hits)

| Terms | Issues | PRs |
|---|---|---|
| "outcome-unknown" | 1 | 1 |
| indeterminate | 2 | 1 |
| ambiguous outcome | 14 | 19 |
| disposition | 4 | 3 |
| "not executed" | 6 | 6 |
| "may have applied" | 0 | 0 |
| "side effect occurred" | 0 | 0 |
| duplicate execution | 16 | 31 |
| dedup | 1 | 3 |
| replay | 34 | 43 |
| "lost response" | 2 | 3 |
| receipt | 19 | 24 |
| "write-result" | 7 | 3 |
| "write-safety" | 1 | 0 |
| isokron | 0 | 0 |

**Screening.** Up to 100 items per query and type were listed (only "conflict" PRs, 249, exceeded that). That gave 477 distinct items from the main list and 45 more from the first twelve follow-up terms; the last three ("write-result", "write-safety", isokron) were listed separately (11 items). Every title was screened; 25 bodies were read in full, and the full SEP text was read for SEP-2848, the Tool Outcome Attestation extension and the `ToolFailure` schema. No issue or PR mentions IsoKron or proposes `write-result/0.1`; the "write-result" and "write-safety" hits are unrelated uses of the words.

## 4. Relevant items

### Defines such a field

- **SEP-2848: Asynchronous Approval for Tool Calls** (PR [#2848](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2848), **open**, draft, Extensions Track, created 3 June 2026, "Sponsor: None (seeking sponsor)"). Extension `io.modelcontextprotocol/tool-approval`: a server that denies a call pending out-of-band approval returns a task, and on approval executes the tool. SEP text (`seps/2848-asynchronous-approval-for-tool-calls.md` at the PR head), section "Execution disposition": `io.modelcontextprotocol/tool-approval-disposition` "is this extension's machine-readable outcome. It is a string and MUST be present in the `_meta` of the `CallToolResult` carried in a completed task's `result` field, on every completed approval task". Values: `approved-executed` ("authorized and executed once"), `denied-not-executed` ("did not run … so no side effect occurred"), `execution-error` ("ran but failed; a side effect may have occurred"), `outcome-unknown` ("execution was claimed but the server cannot determine whether invocation or the side effect completed"). "A client that does not recognize a disposition value MUST treat the operation as possibly executed and MUST NOT retry it automatically." **Scope:** only approval-gated calls under this extension; not ordinary `tools/call` results. The SEP's open questions note the execution-outcome values "could move into SEP-2643 or Tasks".

### Define a call-outcome status, without an unknown value (closed)

- **SEP: Server-Side Signed Execution Record for MCP Tool Calls** (PR [#2828](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2828), closed unmerged 18 July 2026). A signed "outcome record, emitted after execution: the `executed` / `refused` / `errored` status and a commitment over the result". An audit record for governance, not a result field for the caller.
- **Tool Outcome Attestation extension, `dev.agentstatus/toa`** (issue [#3350](https://github.com/modelcontextprotocol/modelcontextprotocol/issues/3350), closed 22 September 2026; SEP text hosted outside the repository at https://github.com/dulrajnr/mcp-sep-toa). A signed attestation bound in `result._meta`, with optional `disposition`: `delivered`, `failed`, `refused`, `unavailable` ("server/tool was unreachable or could not be invoked"). No value for unknown.

### Near misses

- **Issue [#3394](https://github.com/modelcontextprotocol/modelcontextprotocol/issues/3394)**, "Lost-response retry on a mutating `tools/call` re-executes external effects" (**open**, 27 September 2026). Reproduces a duplicate effect after a lost response and asks whether the specification should "explicitly describe the call outcome as ambiguous and warn clients that retrying can re-execute external effects". Proposes non-normative guidance text; "No normative change is proposed". No field.
- **SEP-3182: Request Idempotency** (PR [#3182](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/3182), closed unmerged 23 August 2026). An optional `idempotencyKey` on `tools/call` and a `tools.idempotency` capability; the server returns the original result for a repeated key. Request-side only; the PR page shows no result field marking a replay or an unknown outcome.
- **SEP-3313: Structured Tool-Failure Classification (`ToolFailure`)** (PR [#3313](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/3313), **open**, draft; schema prototype PR [#3312](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/3312) closed 9 November 2026). In `_meta` under `io.modelcontextprotocol/toolFailure` when `isError` is true: `class` (`capability`, `reliability`, `governance`), `code`, `retryable`, `retryNotBefore`, `correlationId`, `agentGuidance`. A repeat signal on failures; nothing on whether the write applied, and no unknown value.
- **Issue [#3003](https://github.com/modelcontextprotocol/modelcontextprotocol/issues/3003)** (open): asks that `structuredContent` be defined for the error path. No outcome field proposed.
- **SEP-2643: Structured Authorization Denials** (PR [#2643](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2643), open): classifies denials. SEP-2848 composes it; it does not itself carry an execution outcome.
- **SEP-2145: Standardize `tools/call` failure reporting** (PR [#2145](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2145), open): which failures are protocol errors and which are tool execution errors. No outcome field.
- **Per-tool behaviour hints, all closed:** SEP-3172 recovery metadata (compensating operation) in `ToolAnnotations` ([#3172](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/3172)); advisory `effect`/`idempotent` hints ([#2745](https://github.com/modelcontextprotocol/modelcontextprotocol/issues/2745)); side-effect classification ([#3025](https://github.com/modelcontextprotocol/modelcontextprotocol/issues/3025)); SEP-2061 action security metadata ([#2061](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2061), moved to an experimental extension). These describe tools, not one call's outcome.
- **Tasks** (SEP-1686 [#1686](https://github.com/modelcontextprotocol/modelcontextprotocol/issues/1686), SEP-2663 [#2663](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2663)): task status values only. SEP-3182 states that SEP-1686 "explicitly deferred a general idempotency mechanism to 'a dedicated proposal'".

## 5. Conclusion and the statement in the article

- **Specification (2026-07-28):** no field in a tool result for whether a write took effect or for an unknown outcome. `isError` says the call failed; annotations describe the tool.
- **Proposals:** one open draft, SEP-2848, defines both (`approved-executed`, `denied-not-executed`, `execution-error`, `outcome-unknown`), scoped to approval-gated calls. Two closed audit-record proposals define a status without an unknown value. The closest general proposals (SEP-3182, SEP-3313, issue #3394) give a request key, a retry flag on failures, and a request for guidance text.

The statement in the article:

> On 15 November 2026 we read the MCP specification (version 2026-07-28) and searched the issues and pull requests in `modelcontextprotocol/modelcontextprotocol`, where proposals (SEPs) are filed, for 43 terms, among them idempotency, "unknown outcome", "partial success", retry, outcome, conflict, `isError` and `structuredContent`; the queries, counts and items read are in the published search log. The specification has no field in a tool result for whether a write took effect or for an unknown outcome. One open draft proposal defines both for a narrow case: SEP-2848, an extension for tool calls held for out-of-band approval, requires a disposition in the result's `_meta` of `approved-executed`, `denied-not-executed`, `execution-error` or `outcome-unknown` ([#2848](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2848)). Two closed proposals for signed audit records define a call-outcome status without an unknown value ([#2828](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2828), [#3350](https://github.com/modelcontextprotocol/modelcontextprotocol/issues/3350)); for ordinary tool calls, an open issue asks whether the specification should call a lost response's outcome ambiguous ([#3394](https://github.com/modelcontextprotocol/modelcontextprotocol/issues/3394)), and an open draft adds a `retryable` flag to failed results ([SEP-3313](https://github.com/modelcontextprotocol/modelcontextprotocol/pull/3313)).

Term count: 28 main + 15 follow-up = 43.
