Skip to content

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

TypeScript

docs: explain undeclared bodies and why the gauge skips image parts

Alfredo Lopez826e7ec586bd
diff --git a/codex-llm-proxy/src/server.ts b/codex-llm-proxy/src/server.ts
index 75110a0f8..226acca78 100644
--- a/codex-llm-proxy/src/server.ts
+++ b/codex-llm-proxy/src/server.ts
@@ -135,9 +135,12 @@ function boundedErrorHandler(err: unknown, _req: Request, res: Response, _next:
* the shared byte budget, so the bodies in memory are bounded by bytes rather
* than by the stream gate's request count. A `Transfer-Encoding` body is
* refused with 411 instead of being read unbounded; a non-numeric
- * `Content-Length` is refused with 400. Bodies over the limit go to `parse`
- * without a reservation: on the admin app express.json answers 413 from the
- * header without buffering them, and on the runtime app
+ * `Content-Length` is refused with 400. A request with neither header has no
+ * body under HTTP/1.1 (RFC 9112 §6.3): `parse` reads nothing, and bytes sent
+ * after its headers are parsed as the next request, which Node answers 400.
+ * Bodies over the limit go to `parse` without a reservation: on the admin app
+ * express.json answers 413 from the header without buffering them, and on the
+ * runtime app
* `selectTransportBudget` either answers 413 or admits them through the visual
* gate. A granted body must be read and parsed within `readDeadlineMs` of the
* grant, or it is answered 408 `request_timeout`, its reservation released and
@@ -310,8 +313,9 @@ export function createProxyApps(
// above the ordinary cap can be a visual envelope. Those bodies take the
// 2-wide gate (the image bytes stay resident for the ChatGPT stream). Every
// smaller body, including every valid V1, stays on the ordinary parser and
- // later the 8-wide stream gate. A missing length cannot be upgraded: it is
- // parsed at the ordinary cap, so a chunked caller cannot occupy a visual slot.
+ // later the 8-wide stream gate. A missing length cannot be upgraded: body
+ // admission already refused a chunked body with 411, and a request with
+ // neither header has no body, so only a declared length takes a visual slot.
const selectTransportBudget = (req: GatedRequest, res: Response, next: NextFunction): void => {
const declared = contentLengthBytes(req)
if (declared !== null && declared > config.maxVisualBodyBytes) {
diff --git a/grok-llm-proxy/src/server.ts b/grok-llm-proxy/src/server.ts
index bd7a6ce6c..162541b28 100644
--- a/grok-llm-proxy/src/server.ts
+++ b/grok-llm-proxy/src/server.ts
@@ -115,8 +115,10 @@ function boundedErrorHandler(err: unknown, _req: Request, res: Response, _next:
* the shared byte budget, so the bodies in memory are bounded by bytes rather
* than by the stream gate's request count. A `Transfer-Encoding` body is
* refused with 411 instead of being read unbounded; a non-numeric
- * `Content-Length` is refused with 400. Bodies over the limit go to `parse`
- * without a reservation, so express.json answers 413 from the header without
+ * `Content-Length` is refused with 400. A request with neither header has no
+ * body under HTTP/1.1 (RFC 9112 §6.3): `parse` reads nothing, and bytes sent
+ * after its headers are parsed as the next request, which Node answers 400.
+ * Bodies over the limit go to `parse` without a reservation, so express.json answers 413 from the header without
* buffering them. A granted body must be read and parsed within
* `readDeadlineMs` of the grant, or it is answered 408 `request_timeout`, its
* reservation released and its connection closed. Otherwise the reservation is
diff --git a/mcp-host/package-lock.json b/mcp-host/package-lock.json
index 4fa2d45b5..661c329b3 100644
--- a/mcp-host/package-lock.json
+++ b/mcp-host/package-lock.json
@@ -1,12 +1,12 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.435",
+ "version": "0.1.436",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "clerum-mcp-host",
- "version": "0.1.435",
+ "version": "0.1.436",
"license": "SEE LICENSE IN LICENSE",
"dependencies": {
"@anthropic-ai/sdk": "^0.32.0",
diff --git a/mcp-host/package.json b/mcp-host/package.json
index 3b64c7700..6333755e4 100644
--- a/mcp-host/package.json
+++ b/mcp-host/package.json
@@ -1,6 +1,6 @@
{
"name": "clerum-mcp-host",
- "version": "0.1.435",
+ "version": "0.1.436",
"description": "MCP Host - reads Host CRD and provides LLM access via OpenAI/Claude",
"main": "dist/main.js",
"engines": {
diff --git a/mcp-host/src/core/tokenizer/heuristic.ts b/mcp-host/src/core/tokenizer/heuristic.ts
index 93edeae3f..d5ead051b 100644
--- a/mcp-host/src/core/tokenizer/heuristic.ts
+++ b/mcp-host/src/core/tokenizer/heuristic.ts
@@ -27,6 +27,14 @@
* decides the tier from this heuristic and uses an exact counter only to record
* the delta. It is exact nowhere — it ignores `contentParts` and
* provider-specific framing.
+ *
+ * Ignoring `contentParts` is deliberate. Their text is also in `content`, which
+ * is counted. Their image bytes are not tokens: providers price an image by its
+ * dimensions, and `bytes / 4` would read one 3 MiB image as ~1M tokens, forcing
+ * a truncation that cannot remove it, because the image sits in the protected
+ * tail. Image bytes are bounded by their own budgets instead: the contract
+ * measures `maxRequestBodyBytes` without them and caps them through
+ * `maxVisualRequestBodyBytes` and `VISUAL_LIMITS`.
*/
import type { ChatMessage, ToolDefinition } from '../types'