Spawns a real forgejo-mcp --transport stdio subprocess, runs the
canonical MCP handshake (initialize → notifications/initialized →
tools/list → tools/call get_forgejo_mcp_server_version) through the
bridge, and verifies each step. Validates that the opaque-pipe
assumption holds end-to-end: every JSON-RPC message round-trips
correctly with no hand-rolled framing surprises.
Binary discovery (skipped if none found):
1. $FORGEJO_MCP_BIN
2. ../../../forgejo-mcp/forgejo-mcp (sibling-repo built binary)
3. go build of ../../../forgejo-mcp into a temp dir
Fake Forgejo (httptest.Server) covers the two probes the SDK and
forgejo-mcp's testConnection do at startup:
- GET /api/v1/version → returns 11.0.0 (>=1.11.0 satisfies SDK gate)
- GET /api/v1/user → returns a minimal authenticated user blob
Skipped under -short. Catch-all 404 handler logs unexpected probes so
new SDK or forgejo-mcp behaviour surfaces clearly in test output.
Closes forgejo-mcp-broker-xot. Phase 4 complete.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds internal/bridge: connects HTTP-side MCP clients to a stdio-side
child via JSON-RPC framing. Decoupled from internal/supervisor — takes
io.Writer + *bufio.Reader + done channel directly so it tests cleanly
with io.Pipe pairs and could later wrap something other than a child
process.
Routing model: one reader goroutine consumes child stdout line-by-line.
Each line is parsed only enough to extract the JSON-RPC `id` field
(string/number/null kept as raw JSON, so `1` and `"1"` don't collide).
HTTP requests register a per-id waiter channel before forwarding their
body to the child; the reader delivers the response to whichever waiter
matches. Concurrent in-flight requests are safe; a duplicate id while
the first is still pending returns 409.
HandleSSE response shapes:
- request with id + child reply → 200 text/event-stream, one
`event: message` SSE event carrying the JSON-RPC response
- request without id (notification) → 204 No Content (no waiter
needed; MCP notifications are fire-and-forget)
- empty body → 400
- duplicate in-flight id → 409
- send-to-child fails → 502
- client disconnect mid-wait → bridge cleans up its waiter; child
keeps running, other in-flight requests unaffected
- child exits before reply → SSE `error` event with reason=child_exited
Tests cover all of the above plus stale unsolicited replies, malformed
lines from the child, and reader robustness across both. 90.0%
coverage. The remaining gap is splitLines' empty-data branch (only
reachable if the child sends a literal `\n` line).
Closes forgejo-mcp-broker-am1.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
