Context7 Platform -- Up-to-date docs for AI agents
chore(release): version packages (#3021)
github-actions[bot]ef14a58dfedb
diff --git a/.changeset/mcp-api-fetch-timeout.md b/.changeset/mcp-api-fetch-timeout.mddeleted file mode 100644index 68e9c83c..00000000--- a/.changeset/mcp-api-fetch-timeout.md+++ /dev/null@@ -1,5 +0,0 @@-----"@upstash/context7-mcp": patch------Add a 60s `AbortSignal.timeout()` to both Context7 API calls in `lib/api.ts`. Without a signal a stalled backend call rides undici's ~300s default before failing. 60s is generous: these are vector queries with p99.9 ~3.2s, and no request exceeded 30s across a full day of production traffic.diff --git a/.changeset/mcp-drop-self-cancellations.md b/.changeset/mcp-drop-self-cancellations.mddeleted file mode 100644index ecb7c0a1..00000000--- a/.changeset/mcp-drop-self-cancellations.md+++ /dev/null@@ -1,5 +0,0 @@-----"@upstash/context7-mcp": patch------Disable SSE keepalive heartbeats on the HTTP handler (`keepAliveMs: 0`). Every tool is a millisecond vector query, so no legitimate exchange needs a heartbeat — but a hung exchange kept alive by heartbeats can never be reaped by a proxy's stream idle timeout. One such hang is deterministic: a 2025-era JSON-RPC batch carrying a request plus its own `notifications/cancelled` gets no response for the cancelled request (per spec), the SDK transport then never closes the stream, and heartbeats kept it alive until the gateway's 1200s hard cap — the dominant source of leaked upstream connections in the 2026-08-11 mcp.context7.com outage. With heartbeats off, silent hangs go idle and the proxy reaps them at its idle timeout.diff --git a/packages/mcp/CHANGELOG.md b/packages/mcp/CHANGELOG.mdindex 6127d302..85218d76 100644--- a/packages/mcp/CHANGELOG.md+++ b/packages/mcp/CHANGELOG.md@@ -1,5 +1,12 @@# @upstash/context7-mcp+## 4.0.2++### Patch Changes++- 67528f2: Add a 60s `AbortSignal.timeout()` to both Context7 API calls in `lib/api.ts`. Without a signal a stalled backend call rides undici's ~300s default before failing. 60s is generous: these are vector queries with p99.9 ~3.2s, and no request exceeded 30s across a full day of production traffic.+- c68104e: Disable SSE keepalive heartbeats on the HTTP handler (`keepAliveMs: 0`). Every tool is a millisecond vector query, so no legitimate exchange needs a heartbeat — but a hung exchange kept alive by heartbeats can never be reaped by a proxy's stream idle timeout. One such hang is deterministic: a 2025-era JSON-RPC batch carrying a request plus its own `notifications/cancelled` gets no response for the cancelled request (per spec), the SDK transport then never closes the stream, and heartbeats kept it alive until the gateway's 1200s hard cap — the dominant source of leaked upstream connections in the 2026-08-11 mcp.context7.com outage. With heartbeats off, silent hangs go idle and the proxy reaps them at its idle timeout.+## 4.0.1### Patch Changesdiff --git a/packages/mcp/package.json b/packages/mcp/package.jsonindex 1b071f07..cfbb9910 100644--- a/packages/mcp/package.json+++ b/packages/mcp/package.json@@ -1,6 +1,6 @@{"name": "@upstash/context7-mcp",- "version": "4.0.1",+ "version": "4.0.2","mcpName": "io.github.upstash/context7","description": "MCP server for Context7","scripts": {