iframe-proxy | Sunbelt Computer Software

Repository navigation

History

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

#!/usr/bin/env bash

# gen-corpus-v2.sh — Spec 083 one-time generator for the schema-bearing frozen

# corpus (specs/083-discovery-profiler/datasets/corpus_v2.tools.json).

# What it does (research D4):

# 1. builds mcpproxy from this worktree,

# 2. boots it over the committed Spec-065 snapshot config (7 no-auth

# reference servers) on 127.0.0.1:8093 with a throwaway --data-dir

# (never touches ~/.mcpproxy),

# 3. waits until GET /api/v1/tools reports the expected 45 tools,

# 4. kills the proxy, then exports the full tool definitions WITH input

# schemas from the Bleve index (scripts/gen-corpus-v2-dump), canonicalizes

# (tools sorted by tool_id; all object keys sorted via jq -S) and writes

# the corpus file.

# Why the export reads the index instead of GET /api/v1/tools: that endpoint

# serves schemas from the supervisor StateView, which currently stores a stub

# ({"type":"object","properties":{}} — "TODO: Parse ParamsJSON" in

# internal/runtime/supervisor/supervisor.go), so every schema it returns is

# empty. The Bleve index holds the authoritative ParamsJSON the production

# retrieval funnel ingests — exactly what arm comparison must measure.

# The output is committed once and then immutable (Spec 065 CN-002 precedent:

# a refresh is corpus_v3, never an edit). Maintainers only.

# Config via env:

# PORT listen port (default: 8093 — NOT 8092/8080, avoids

# the docker-compose bench substrate and

# a tray-owned local proxy)

# EXPECTED_TOOLS expected tool count (default: 45, per Spec 065 datasets README)

# WAIT_SECS per-attempt wait for all servers to connect (default: 300)

# OUT output path (default: specs/083-discovery-profiler/datasets/corpus_v2.tools.json)

set -euo pipefail

REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"

cd "$REPO_ROOT"

PORT="${PORT:-8093}"

EXPECTED_TOOLS="${EXPECTED_TOOLS:-45}"

WAIT_SECS="${WAIT_SECS:-300}"

OUT="${OUT:-specs/083-discovery-profiler/datasets/corpus_v2.tools.json}"

CONFIG="specs/065-evaluation-foundation/datasets/snapshot-servers.config.json"

export MCPPROXY_API_KEY="eval-corpus-snapshot"

BASE_URL="http://127.0.0.1:${PORT}"

command -v jq >/dev/null || { echo "error: jq is required" >&2; exit 4; }

[ -f "$CONFIG" ] || { echo "error: snapshot config not found: $CONFIG" >&2; exit 4; }

WORK="$(mktemp -d)"

PROXY_PID=""

cleanup() {

if [ -n "$PROXY_PID" ] && kill -0 "$PROXY_PID" 2>/dev/null; then

kill "$PROXY_PID" 2>/dev/null || true

wait "$PROXY_PID" 2>/dev/null || true

rm -rf "$WORK"

}

trap cleanup EXIT

log() { printf '\n\033[1;34m==>\033[0m %s\n' "$*"; }

log "Building mcpproxy from the worktree"

go build -o "$WORK/mcpproxy" ./cmd/mcpproxy

# Serve from a COPY of the committed config: mcpproxy persists runtime config

# changes back to the file it was started with, and the Spec-065 snapshot

# config is a frozen committed fixture that must never be rewritten.

cp "$CONFIG" "$WORK/config.json"

CONFIG="$WORK/config.json"

# The snapshot config's sqlite server persists to this fixed path; the server

# creates the db file but not the directory.

mkdir -p /tmp/mcpproxy-corpus-snapshot

tool_count() {

curl -fsS -m 10 -H "X-API-Key: $MCPPROXY_API_KEY" "$BASE_URL/api/v1/tools" 2>/dev/null \

| jq -r '.data.tools | length' 2>/dev/null || echo 0

}

start_proxy() {

local data_dir="$1"

mkdir -p "$data_dir"

"$WORK/mcpproxy" serve \

--config "$CONFIG" \

--listen "127.0.0.1:${PORT}" \

--data-dir "$data_dir" \

>"$WORK/proxy.log" 2>&1 &

PROXY_PID=$!

}

stop_proxy() {

if [ -n "$PROXY_PID" ] && kill -0 "$PROXY_PID" 2>/dev/null; then

kill "$PROXY_PID" 2>/dev/null || true

wait "$PROXY_PID" 2>/dev/null || true

PROXY_PID=""

}

# wait_for_tools polls until EXPECTED_TOOLS is reached (echoes the count and

# returns 0) or WAIT_SECS elapses (echoes the last stable count, returns 1).

wait_for_tools() {

local waited=0 count=0

while [ "$waited" -lt "$WAIT_SECS" ]; do

count="$(tool_count)"

if [ "$count" -ge "$EXPECTED_TOOLS" ]; then

echo "$count"

return 0

sleep 5

waited=$((waited + 5))

done

echo "$count"

return 1

}

# Attempt 1 (+ one full retry for transient npx/uvx download failures).

ATTEMPT=1

COUNT=0

DATA_DIR=""

while :; do

log "Booting snapshot proxy on ${BASE_URL} (attempt ${ATTEMPT}, waiting up to ${WAIT_SECS}s for ${EXPECTED_TOOLS} tools)"

DATA_DIR="$WORK/data-${ATTEMPT}"

start_proxy "$DATA_DIR"

if COUNT="$(wait_for_tools)"; then

log "All ${COUNT} tools indexed"

break

if [ "$ATTEMPT" -ge 2 ]; then

if [ "$COUNT" -eq 0 ]; then

echo "error: no tools appeared after ${ATTEMPT} attempts — proxy log tail:" >&2

tail -30 "$WORK/proxy.log" >&2

exit 1

echo "::warning:: only ${COUNT}/${EXPECTED_TOOLS} tools after ${ATTEMPT} attempts — proceeding with the REDUCED corpus. Record the reduced count + the missing server in datasets/README.md." >&2

break

echo "::warning:: ${COUNT}/${EXPECTED_TOOLS} tools after ${WAIT_SECS}s — restarting proxy once (transient npx/uvx download failures)" >&2

stop_proxy

ATTEMPT=$((ATTEMPT + 1))

done

# Stop the proxy BEFORE the export: the dump helper opens the Bleve index

# directly (single-writer lock) and reads the authoritative ParamsJSON.

stop_proxy

log "Exporting + canonicalizing corpus_v2 (${COUNT} tools) from the index at ${DATA_DIR}"

CAPTURE_DATE="$(date -u +%Y-%m-%d)"

mkdir -p "$(dirname "$OUT")"

go run ./scripts/gen-corpus-v2-dump -data-dir "$DATA_DIR" \

| jq -S --arg version "corpus_v2@${CAPTURE_DATE}" --arg date "$CAPTURE_DATE" '{

version: $version,

generated_from: {

note: "Spec 083 D4 schema-bearing frozen corpus: the 7 no-auth Spec-065 reference servers (filesystem, git, memory, sqlite, fetch, time, sequential-thinking) via snapshot-servers.config.json, exported WITH full JSON input schemas.",

source: "Bleve index ParamsJSON via scripts/gen-corpus-v2-dump (GET /api/v1/tools serves stub schemas — see header comment)",

config: "specs/065-evaluation-foundation/datasets/snapshot-servers.config.json",

generator: "scripts/gen-corpus-v2.sh",

date: $date

tools: .

}' > "$OUT"

TOOLS_IN_FILE="$(jq '.tools | length' "$OUT")"

SCHEMALESS="$(jq '[.tools[] | select((.schema // empty | type) != "object" or (.schema | length) == 0)] | map(.tool_id)' "$OUT")"

log "Verification"

echo " tools exported: ${TOOLS_IN_FILE}"

echo " schema-less tools: ${SCHEMALESS}"

if [ "$SCHEMALESS" != "[]" ]; then

echo "error: corpus_v2 must be schema-bearing — the tools above have empty/missing schemas" >&2

exit 1

if [ "$TOOLS_IN_FILE" -lt "$EXPECTED_TOOLS" ]; then

echo "::warning:: reduced corpus (${TOOLS_IN_FILE}/${EXPECTED_TOOLS} tools) — note this explicitly in specs/083-discovery-profiler/datasets/README.md" >&2

log "Wrote ${OUT} (${TOOLS_IN_FILE} tools, capture date ${CAPTURE_DATE})"

echo "Next: record tool count + provenance in specs/083-discovery-profiler/datasets/README.md and commit both."

Sunbelt Computer Software

PL/B Language Development and Support

Navigation Menu

Repository navigation

gen-corpus-v2.sh

gen-corpus-v2.sh

Sunbelt Computer Software

PL/B Language Development and Support

FilesExpand file tree

gen-corpus-v2.sh

Latest commit

History

gen-corpus-v2.sh

File metadata and controls