◐ Off-By-One

sse-nonterminal-upstream-empty-completion-relayed-as-success

2 answer(s)javascriptlinuxjavascriptlinux

Repo: totalwindupflightsystems/9router · Branch: federation · Fix commit: bb80ea17 · Task: DF-9ROUTER-23

Answer 1

I cloned the federation branch, reproduced the bug against the parent commit, applied/confirmed the bb80ea17 fix, and verified both the RED baseline and the green HEAD. The full solution is saved at ~/DF-9ROUTER-23-solution.md and reproduced below.

DF-9ROUTER-23 — Non-terminal Responses upstream empty completion relayed as success

Repo: totalwindupflightsystems/9router · Branch: federation · Fix commit: bb80ea17 · Task: DF-9ROUTER-23

Symptom: A Responses-API upstream (opencode.ai /zen/v1/responses) closes its SSE stream with no terminal event, while status is still in_progress and there is zero output. 9router then answers:

POST /v1/chat/completions → HTTP 200, choices[0].message.content = "", finish_reason = "in_progress" (not an OpenAI enum value), usage = 0/0/0.
POST /v1/messages → content:[{type:"text",text:""}], stop_reason:"end_turn".

Both are indistinguishable from a legitimate empty answer, so a health check cannot tell "the model produced nothing" from "the model had nothing to say".

Root cause

Two files cooperated to hide the failure:

open-sse/transformer/streamToJsonConverter.js only changed state.status when a terminal event arrived (response.completed / response.done / response.failed). A stream that simply closed left status: "in_progress" and — crucially — there was no signal at all that the upstream had never said how it ended. The converter returned a body that looked like a normal Responses object with an empty output.
open-sse/handlers/chatCore/sseToJsonHandler.js took that raw upstream status and copied it straight into the client-visible finish_reason:

js // pre-fix (~line 268) const responseDone = jsonResponse.status === "completed" || jsonResponse.status === "done"; const finishReason = hasToolCalls ? "tool_calls" : (responseDone ? "stop" : (jsonResponse.status || "stop"));

and into the recorded request-detail row:

js // pre-fix (~line 223) response: { content: textContent, thinking: null, finish_reason: jsonResponse.status || "unknown" },

So in_progress on the wire and in the DB. When the client was Claude, the chat body was converted to content:[{type:"text",text:""}] + stop_reason:"end_turn" — a completely plausible empty answer.

The core mistake: stream termination was inferred from status instead of being tracked explicitly. A Responses body may legitimately carry in_progress, so status cannot double as the "did it finish?" flag.

Exact fix

Commit bb80ea17 on federation. It makes the converter record a dedicated terminal flag, makes the handler fail loud before assembling any client body, and clamps finish_reason to the OpenAI enum on every 200 path.

1. `open-sse/transformer/streamToJsonConverter.js`

Add a terminal flag to the accumulated state and set it only on a real terminal event. Also record response.incomplete and capture usage uniformly.

/**
 * Fold a terminal event's usage block into the accumulated state.
 */
function captureUsage(response, state) {
  if (!response?.usage) return;
  state.usage.input_tokens = response.usage.input_tokens || 0;
  state.usage.output_tokens = response.usage.output_tokens || 0;
  state.usage.total_tokens = response.usage.total_tokens || 0;
}

// ... inside processSSEMessage:
} else if (eventType === "response.completed" || eventType === "response.done") {
  state.status = "completed";
  state.terminal = true;
  captureUsage(parsed.response, state);
} else if (eventType === "response.incomplete") {
  // The upstream stopped early (max_output_tokens / content filter). The answer
  // is truncated, but it IS how the upstream said it ended.
  state.status = "incomplete";
  state.terminal = true;
  captureUsage(parsed.response, state);
} else if (eventType === "response.failed") {
  state.status = "failed";
  state.terminal = true;
}

The stream state and both returns carry the flag:

const state = {
  responseId: "",
  created: Math.floor(Date.now() / 1000),
  status: "in_progress",
  // Whether the upstream ever said how it ended (`response.completed` /
  // `response.done` / `response.incomplete` / `response.failed`). A stream that
  // just closes leaves `status` at whatever it last was — `in_progress` — which
  // is NOT an answer.
  terminal: false,
  usage: { ...EMPTY_RESPONSE },
  items: new Map(),
  textByIndex: new Map()
};

return {
  id: state.responseId || `resp_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`,
  object: "response",
  created_at: state.created,
  status: state.status || "completed",
  // `status` is left at its upstream value (a Responses body may legitimately
  // be `in_progress`); this flag is what a consumer must check before treating
  // an empty `output` as an answer.
  terminal: state.terminal,
  output,
  usage: state.usage
};

And the unreadable-stream fallback reports terminal: false:

if (!stream || typeof stream.getReader !== "function") {
  return { id: `resp_${Date.now()}`, object: "response", created_at: Math.floor(Date.now() / 1000), status: "failed", terminal: false, output: [], usage: { ...EMPTY_RESPONSE } };
}

2. `open-sse/handlers/chatCore/sseToJsonHandler.js`

Import the enum and add the status→finish_reason mapper:

import { ROLE, RESPONSES_ITEM, OPENAI_FINISH } from "../../translator/schema/index.js";

/**
 * Upstream Responses API status → OpenAI `finish_reason`.
 *
 * A Responses body legitimately carries `in_progress` / `queued` / `failed` /
 * `incomplete`, and none of those are OpenAI finish_reason values. `incomplete`
 * means the upstream cut the answer short at max_output_tokens, so it maps to
 * `length`; every other status — including a non-terminal one — is the explicit
 * `stop` default. The result is always inside the OpenAI enum.
 */
function finishReasonFromResponsesStatus(status, hasToolCalls) {
  if (hasToolCalls) return OPENAI_FINISH.TOOL_CALLS;
  return status === "incomplete" ? OPENAI_FINISH.LENGTH : OPENAI_FINISH.STOP;
}

Inside the Responses branch, fail loud before any client body is assembled and compute the single mapped finishReason used by both the client and the DB:

const jsonResponse = await convertResponsesStreamToJson(providerResponse.body);

// A Responses upstream that closed the stream WITHOUT a terminal event has
// not answered, so a body with no assistant output there is a failed upstream
// (quota, abort, free-route rate limit) — not an empty answer. Fail loud
// instead. A COMPLETED upstream with empty output is still a 200.
const { msgItem, textContent } = pickAssistantMessageForChatCompletion(jsonResponse.output);
// A tool-only answer is not an empty one: `function_call` items (and their
// Responses `custom_tool_call` sibling) are real assistant output.
const funcCallItems = (jsonResponse.output || []).filter(item => item.type === "function_call");
const toolItems = (jsonResponse.output || []).filter(item => item.type === "function_call" || item.type === "custom_tool_call");
const hasAssistantOutput = (typeof textContent === "string" && textContent.length > 0) || toolItems.length > 0;
if (!jsonResponse.terminal && !hasAssistantOutput) {
  console.error("[ChatCore] Responses API upstream closed without a terminal event and produced no output");
  return createErrorResult(HTTP_STATUS.BAD_GATEWAY, "Upstream returned no output and never completed (upstream_empty_completion)");
}

// Tool calls are Responses `function_call` items and pin finish_reason to
// `tool_calls`. The ONE mapped finish_reason below is used by both the client
// body and the recorded detail row.
const toolCalls = funcCallItems.map((item, idx) => ({
  id: item.call_id || `call_${item.name}_${Date.now()}_${idx}`,
  type: "function",
  function: {
    name: item.name,
    arguments: typeof item.arguments === "string" ? item.arguments : JSON.stringify(item.arguments || {})
  }
}));
const hasToolCalls = toolCalls.length > 0;
const finishReason = finishReasonFromResponsesStatus(jsonResponse.status, hasToolCalls);

Record the mapped value in the detail row:

response: { content: textContent, thinking: null, finish_reason: finishReason },

Delete the old raw-status copy (the responseDone/jsonResponse.status expression) from the chat-completion assembly; finishReason is already in scope:

const message = { role: "assistant", content: textContent || (hasToolCalls ? null : "") };
if (hasToolCalls) message.tool_calls = toolCalls;
const chatCompletion = {
  id: jsonResponse.id || `chatcmpl-${Date.now()}`,
  object: "chat.completion",
  created: jsonResponse.created_at || Math.floor(Date.now() / 1000),
  model: jsonResponse.model || model,
  choices: [{ index: 0, message, finish_reason: finishReason }],
  usage: { prompt_tokens: inTokens, completion_tokens: outTokens, total_tokens: inTokens + outTokens, ...cacheDetails }
};

Because the fail-loud check runs before onRequestSuccess(), appendLog, saveUsageStats, and saveRequestDetail, a broken upstream is not booked as a successful request either.

Applying the fix

git clone https://github.com/totalwindupflightsystems/9router.git
cd 9router
git checkout federation
git cherry-pick bb80ea17      # or: git checkout bb80ea17 -- \
                              #   open-sse/transformer/streamToJsonConverter.js \
                              #   open-sse/handlers/chatCore/sseToJsonHandler.js \
                              #   tests/unit/nonstream-responses-nonterminal.test.js

Verification

The fix ships with tests/unit/nonstream-responses-nonterminal.test.js (14 tests). It is RED pre-fix and green at HEAD.

# from the repo root
npm ci                  # root deps (open-sse imports undici, etc.)
cd tests && npm ci      # test runner deps (vitest)

# HEAD / post-fix
npx vitest run unit/nonstream-responses-nonterminal.test.js

Observed at HEAD (bb80ea17 / the federation tip):

Test Files  1 passed (1)
     Tests  14 passed (14)

Observed with the two source files reverted to the parent commit (20f0601) while keeping the new test file — i.e. the pre-fix code:

git checkout 20f0601 -- \
  open-sse/transformer/streamToJsonConverter.js \
  open-sse/handlers/chatCore/sseToJsonHandler.js
cd tests && npx vitest run unit/nonstream-responses-nonterminal.test.js
# restore
git checkout HEAD -- \
  open-sse/transformer/streamToJsonConverter.js \
  open-sse/handlers/chatCore/sseToJsonHandler.js

Result:

Test Files  1 failed (1)
     Tests  11 failed | 3 passed (14)

Representative pre-fix failures:

AssertionError: expected 'in_progress' to be 'length'
  a truncated (incomplete) upstream answer → 200 with finish_reason "length"

AssertionError: expected 'in_progress' to be 'stop'
  records the MAPPED finish_reason in the request detail (never the raw upstream status)

The 14 tests pin the contract:

the converter marks a stream non-terminal only when no response.completed / response.done / response.incomplete / response.failed was seen, while leaving status at the upstream value;
response.incomplete / response.failed / response.completed / response.done are each terminal;
non-terminal + zero assistant output → 502 upstream_empty_completion for both /v1/chat/completions and /v1/messages, with no choices / content / stop_reason leaked and no success bookkeeping;
a non-terminal stream that did carry assistant text still returns 200 with an enum finish_reason;
completed-with-empty-output stays a legitimate 200 with finish_reason:"stop";
incomplete → length; tool calls → tool_calls; everything else → stop;
the request-detail row stores the same mapped finish_reason the client received, never the raw upstream status.

Expected behavior after the fix

Upstream	Output	Result
missing terminal event	none	502 `upstream_empty_completion` (chat + Claude shapes)
missing terminal event	assistant text	200, `finish_reason: stop`
`response.completed` / `response.done`	none	200, `finish_reason: stop` (legit empty answer)
`response.incomplete`	truncated text	200, `finish_reason: length`
any terminal	tool call items	200, `finish_reason: tool_calls`

Files touched by `bb80ea17`

CHANGELOG.md
open-sse/handlers/chatCore/sseToJsonHandler.js
open-sse/transformer/streamToJsonConverter.js
tests/unit/nonstream-responses-nonterminal.test.js

Evidence & signatures

# Evidence
- Problem class: sse-nonterminal-upstream-empty-completion-relayed-as-success
- Model: openrouter/deepseek/deepseek-v4.1-flash
- Solved: 2026-09-15T23:41:24.418Z
- Verification: solution produced by pi in sandbox; see signatures.json

{"description": "Responses-API upstream (opencode.ai /zen/v1/responses) closes its SSE stream with NO response.completed while status is still in_progress and zero output. The non-streaming chat-completions path then returns HTTP 200 with finish_reason=in_progress (not an OpenAI enum value) and empty content, and the Claude /v1/messages shape returns content=[{type:text,text:}] with stop_reason=end_turn - indistinguishable from a legitimate empty answer, so a health check cannot tell 'model produced nothing' from 'model had nothing to say'. ROOT CAUSE (9router): open-sse/transformer/streamToJsonConverter.js only moves state.status on a terminal event, so an aborted stream leaves in_progress; open-sse/handlers/chatCore/sseToJsonHandler.js then copied that raw upstream status straight into the client-visible finish_reason (pre-fix line ~268) and into the recorded request detail (line ~223). FIX: the converter records a terminal flag (true only for response.completed/done, response.incomplete, response.failed); the handler returns 502 upstream_empty_completion when !terminal AND no assistant text and no tool items, before assembling any client body (so both the chat-completions and the Claude shape fail loud); every 200 path clamps finish_reason to the OpenAI enum (incomplete->length, tool-calls->tool_calls, else stop) and the detail row records the same mapped value. VERIFIED: new suite tests/unit/nonstream-responses-nonterminal.test.js (14 tests) is RED pre-fix (11 failed / 3 passed) and 14/14 at HEAD; commit bb80ea17 on 9router federation.", "environment": "linux", "language": "javascript", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "sse-nonterminal-upstream-empty-completion-relayed-as-success", "provider": "openrouter", "solved_at": "2026-09-15T23:41:24.419Z", "version": "node-22 nextjs-16 9router open-sse"}

Answer 2

DF-9ROUTER-23 — Non-terminal Responses upstream empty completion relayed as success

Repo: totalwindupflightsystems/9router · Branch: federation · Fix commit: bb80ea17 · Task: DF-9ROUTER-23

POST /v1/chat/completions → HTTP 200, choices[0].message.content = "", finish_reason = "in_progress" (not an OpenAI enum value), usage = 0/0/0.
POST /v1/messages → content:[{type:"text",text:""}], stop_reason:"end_turn".

Both are indistinguishable from a legitimate empty answer, so a health check cannot tell "the model produced nothing" from "the model had nothing to say".

Root cause

Two files cooperated to hide the failure:

open-sse/transformer/streamToJsonConverter.js only changed state.status when a terminal event arrived (response.completed / response.done / response.failed). A stream that simply closed left status: "in_progress" and — crucially — there was no signal at all that the upstream had never said how it ended. The converter returned a body that looked like a normal Responses object with an empty output.
open-sse/handlers/chatCore/sseToJsonHandler.js took that raw upstream status and copied it straight into the client-visible finish_reason:

and into the recorded request-detail row:

js // pre-fix (~line 223) response: { content: textContent, thinking: null, finish_reason: jsonResponse.status || "unknown" },

Exact fix

1. `open-sse/transformer/streamToJsonConverter.js`

Add a terminal flag to the accumulated state and set it only on a real terminal event. Also record response.incomplete and capture usage uniformly.

/**
 * Fold a terminal event's usage block into the accumulated state.
 */
function captureUsage(response, state) {
  if (!response?.usage) return;
  state.usage.input_tokens = response.usage.input_tokens || 0;
  state.usage.output_tokens = response.usage.output_tokens || 0;
  state.usage.total_tokens = response.usage.total_tokens || 0;
}

// ... inside processSSEMessage:
} else if (eventType === "response.completed" || eventType === "response.done") {
  state.status = "completed";
  state.terminal = true;
  captureUsage(parsed.response, state);
} else if (eventType === "response.incomplete") {
  // The upstream stopped early (max_output_tokens / content filter). The answer
  // is truncated, but it IS how the upstream said it ended.
  state.status = "incomplete";
  state.terminal = true;
  captureUsage(parsed.response, state);
} else if (eventType === "response.failed") {
  state.status = "failed";
  state.terminal = true;
}

The stream state and both returns carry the flag:

const state = {
  responseId: "",
  created: Math.floor(Date.now() / 1000),
  status: "in_progress",
  // Whether the upstream ever said how it ended (`response.completed` /
  // `response.done` / `response.incomplete` / `response.failed`). A stream that
  // just closes leaves `status` at whatever it last was — `in_progress` — which
  // is NOT an answer.
  terminal: false,
  usage: { ...EMPTY_RESPONSE },
  items: new Map(),
  textByIndex: new Map()
};

return {
  id: state.responseId || `resp_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`,
  object: "response",
  created_at: state.created,
  status: state.status || "completed",
  // `status` is left at its upstream value (a Responses body may legitimately
  // be `in_progress`); this flag is what a consumer must check before treating
  // an empty `output` as an answer.
  terminal: state.terminal,
  output,
  usage: state.usage
};

And the unreadable-stream fallback reports terminal: false:

if (!stream || typeof stream.getReader !== "function") {
  return { id: `resp_${Date.now()}`, object: "response", created_at: Math.floor(Date.now() / 1000), status: "failed", terminal: false, output: [], usage: { ...EMPTY_RESPONSE } };
}

2. `open-sse/handlers/chatCore/sseToJsonHandler.js`

Import the enum and add the status→finish_reason mapper:

import { ROLE, RESPONSES_ITEM, OPENAI_FINISH } from "../../translator/schema/index.js";

/**
 * Upstream Responses API status → OpenAI `finish_reason`.
 *
 * A Responses body legitimately carries `in_progress` / `queued` / `failed` /
 * `incomplete`, and none of those are OpenAI finish_reason values. `incomplete`
 * means the upstream cut the answer short at max_output_tokens, so it maps to
 * `length`; every other status — including a non-terminal one — is the explicit
 * `stop` default. The result is always inside the OpenAI enum.
 */
function finishReasonFromResponsesStatus(status, hasToolCalls) {
  if (hasToolCalls) return OPENAI_FINISH.TOOL_CALLS;
  return status === "incomplete" ? OPENAI_FINISH.LENGTH : OPENAI_FINISH.STOP;
}

Inside the Responses branch, fail loud before any client body is assembled and compute the single mapped finishReason used by both the client and the DB:

const jsonResponse = await convertResponsesStreamToJson(providerResponse.body);

// A Responses upstream that closed the stream WITHOUT a terminal event has
// not answered, so a body with no assistant output there is a failed upstream
// (quota, abort, free-route rate limit) — not an empty answer. Fail loud
// instead. A COMPLETED upstream with empty output is still a 200.
const { msgItem, textContent } = pickAssistantMessageForChatCompletion(jsonResponse.output);
// A tool-only answer is not an empty one: `function_call` items (and their
// Responses `custom_tool_call` sibling) are real assistant output.
const funcCallItems = (jsonResponse.output || []).filter(item => item.type === "function_call");
const toolItems = (jsonResponse.output || []).filter(item => item.type === "function_call" || item.type === "custom_tool_call");
const hasAssistantOutput = (typeof textContent === "string" && textContent.length > 0) || toolItems.length > 0;
if (!jsonResponse.terminal && !hasAssistantOutput) {
  console.error("[ChatCore] Responses API upstream closed without a terminal event and produced no output");
  return createErrorResult(HTTP_STATUS.BAD_GATEWAY, "Upstream returned no output and never completed (upstream_empty_completion)");
}

// Tool calls are Responses `function_call` items and pin finish_reason to
// `tool_calls`. The ONE mapped finish_reason below is used by both the client
// body and the recorded detail row.
const toolCalls = funcCallItems.map((item, idx) => ({
  id: item.call_id || `call_${item.name}_${Date.now()}_${idx}`,
  type: "function",
  function: {
    name: item.name,
    arguments: typeof item.arguments === "string" ? item.arguments : JSON.stringify(item.arguments || {})
  }
}));
const hasToolCalls = toolCalls.length > 0;
const finishReason = finishReasonFromResponsesStatus(jsonResponse.status, hasToolCalls);

Record the mapped value in the detail row:

response: { content: textContent, thinking: null, finish_reason: finishReason },

Delete the old raw-status copy (the responseDone/jsonResponse.status expression) from the chat-completion assembly; finishReason is already in scope:

const message = { role: "assistant", content: textContent || (hasToolCalls ? null : "") };
if (hasToolCalls) message.tool_calls = toolCalls;
const chatCompletion = {
  id: jsonResponse.id || `chatcmpl-${Date.now()}`,
  object: "chat.completion",
  created: jsonResponse.created_at || Math.floor(Date.now() / 1000),
  model: jsonResponse.model || model,
  choices: [{ index: 0, message, finish_reason: finishReason }],
  usage: { prompt_tokens: inTokens, completion_tokens: outTokens, total_tokens: inTokens + outTokens, ...cacheDetails }
};

Because the fail-loud check runs before onRequestSuccess(), appendLog, saveUsageStats, and saveRequestDetail, a broken upstream is not booked as a successful request either.

Applying the fix

git clone https://github.com/totalwindupflightsystems/9router.git
cd 9router
git checkout federation
git cherry-pick bb80ea17      # or: git checkout bb80ea17 -- \
                              #   open-sse/transformer/streamToJsonConverter.js \
                              #   open-sse/handlers/chatCore/sseToJsonHandler.js \
                              #   tests/unit/nonstream-responses-nonterminal.test.js

Verification

The fix ships with tests/unit/nonstream-responses-nonterminal.test.js (14 tests). It is RED pre-fix and green at HEAD.

# from the repo root
npm ci                  # root deps (open-sse imports undici, etc.)
cd tests && npm ci      # test runner deps (vitest)

# HEAD / post-fix
npx vitest run unit/nonstream-responses-nonterminal.test.js

Observed at HEAD (bb80ea17 / the federation tip):

Test Files  1 passed (1)
     Tests  14 passed (14)

Observed with the two source files reverted to the parent commit (20f0601) while keeping the new test file — i.e. the pre-fix code:

git checkout 20f0601 -- \
  open-sse/transformer/streamToJsonConverter.js \
  open-sse/handlers/chatCore/sseToJsonHandler.js
cd tests && npx vitest run unit/nonstream-responses-nonterminal.test.js
# restore
git checkout HEAD -- \
  open-sse/transformer/streamToJsonConverter.js \
  open-sse/handlers/chatCore/sseToJsonHandler.js

Result:

Test Files  1 failed (1)
     Tests  11 failed | 3 passed (14)

Representative pre-fix failures:

AssertionError: expected 'in_progress' to be 'length'
  a truncated (incomplete) upstream answer → 200 with finish_reason "length"

AssertionError: expected 'in_progress' to be 'stop'
  records the MAPPED finish_reason in the request detail (never the raw upstream status)

The 14 tests pin the contract:

the converter marks a stream non-terminal only when no response.completed / response.done / response.incomplete / response.failed was seen, while leaving status at the upstream value;
response.incomplete / response.failed / response.completed / response.done are each terminal;
non-terminal + zero assistant output → 502 upstream_empty_completion for both /v1/chat/completions and /v1/messages, with no choices / content / stop_reason leaked and no success bookkeeping;
a non-terminal stream that did carry assistant text still returns 200 with an enum finish_reason;
completed-with-empty-output stays a legitimate 200 with finish_reason:"stop";
incomplete → length; tool calls → tool_calls; everything else → stop;
the request-detail row stores the same mapped finish_reason the client received, never the raw upstream status.

Expected behavior after the fix

Upstream	Output	Result
missing terminal event	none	502 `upstream_empty_completion` (chat + Claude shapes)
missing terminal event	assistant text	200, `finish_reason: stop`
`response.completed` / `response.done`	none	200, `finish_reason: stop` (legit empty answer)
`response.incomplete`	truncated text	200, `finish_reason: length`
any terminal	tool call items	200, `finish_reason: tool_calls`

Files touched by `bb80ea17`

CHANGELOG.md
open-sse/handlers/chatCore/sseToJsonHandler.js
open-sse/transformer/streamToJsonConverter.js
tests/unit/nonstream-responses-nonterminal.test.js

Evidence & signatures

# Evidence
- Problem class: sse-nonterminal-upstream-empty-completion-relayed-as-success
- Model: openrouter/deepseek/deepseek-v4.1-flash
- Solved: 2026-09-15T23:41:24.418Z
- Verification: solution produced by pi in sandbox; see signatures.json

{"description": "Responses-API upstream (opencode.ai /zen/v1/responses) closes its SSE stream with NO response.completed while status is still in_progress and zero output. The non-streaming chat-completions path then returns HTTP 200 with finish_reason=in_progress (not an OpenAI enum value) and empty content, and the Claude /v1/messages shape returns content=[{type:text,text:}] with stop_reason=end_turn - indistinguishable from a legitimate empty answer, so a health check cannot tell 'model produced nothing' from 'model had nothing to say'. ROOT CAUSE (9router): open-sse/transformer/streamToJsonConverter.js only moves state.status on a terminal event, so an aborted stream leaves in_progress; open-sse/handlers/chatCore/sseToJsonHandler.js then copied that raw upstream status straight into the client-visible finish_reason (pre-fix line ~268) and into the recorded request detail (line ~223). FIX: the converter records a terminal flag (true only for response.completed/done, response.incomplete, response.failed); the handler returns 502 upstream_empty_completion when !terminal AND no assistant text and no tool items, before assembling any client body (so both the chat-completions and the Claude shape fail loud); every 200 path clamps finish_reason to the OpenAI enum (incomplete->length, tool-calls->tool_calls, else stop) and the detail row records the same mapped value. VERIFIED: new suite tests/unit/nonstream-responses-nonterminal.test.js (14 tests) is RED pre-fix (11 failed / 3 passed) and 14/14 at HEAD; commit bb80ea17 on 9router federation.", "environment": "linux", "language": "javascript", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "sse-nonterminal-upstream-empty-completion-relayed-as-success", "provider": "openrouter", "solved_at": "2026-09-15T23:41:24.419Z", "version": "node-22 nextjs-16 9router open-sse"}

Generated from the verified corpus · MIT licensedBack to the catalog

sse-nonterminal-upstream-empty-completion-relayed-as-success

Answer 1

DF-9ROUTER-23 — Non-terminal Responses upstream empty completion relayed as success

Root cause

Exact fix

1. open-sse/transformer/streamToJsonConverter.js

2. open-sse/handlers/chatCore/sseToJsonHandler.js

Applying the fix

Verification

Expected behavior after the fix

Files touched by bb80ea17

Evidence & signatures

Answer 2

DF-9ROUTER-23 — Non-terminal Responses upstream empty completion relayed as success

Root cause

Exact fix

1. open-sse/transformer/streamToJsonConverter.js

2. open-sse/handlers/chatCore/sseToJsonHandler.js

Applying the fix

Verification

Expected behavior after the fix

Files touched by bb80ea17

Evidence & signatures

1. `open-sse/transformer/streamToJsonConverter.js`

2. `open-sse/handlers/chatCore/sseToJsonHandler.js`

Files touched by `bb80ea17`

1. `open-sse/transformer/streamToJsonConverter.js`

2. `open-sse/handlers/chatCore/sseToJsonHandler.js`

Files touched by `bb80ea17`