@saluzi/codegraph 0.1.0 → 0.2.0

package/scripts/agent-eval/run-all.sh

@@ -0,0 +1,75 @@
+#!/usr/bin/env bash
+# With/without A/B (and optional interactive) eval for a codegraph version on a
+# repo. Codegraph is the ONLY variable: both arms launch claude with
+# --strict-mcp-config — with = codegraph-only MCP (pointed at $CG_BIN),
+# without = empty MCP. Built-in Read/Grep/Bash stay available in both arms.
+#
+# Usage: run-all.sh <repo-path> "<question>" [headless|tmux|all]
+# Env:   CG_BIN          codegraph binary (default: command -v codegraph)
+#        AGENT_EVAL_OUT  output dir (default: /tmp/agent-eval)
+#        MODEL / EFFORT  claude model/effort (default: sonnet / high — the
+#                        standing A/B policy; see CLAUDE.md, don't raise)
+set -uo pipefail
+
+REPO="${1:?usage: run-all.sh <repo-path> \"<question>\" [headless|tmux|all]}"
+Q="${2:?question required}"
+MODE="${3:-headless}"
+CG_BIN="${CG_BIN:-$(command -v codegraph)}"
+OUT="${AGENT_EVAL_OUT:-/tmp/agent-eval}"
+HARNESS="$(cd "$(dirname "$0")" && pwd)"
+mkdir -p "$OUT"
+
+# Neutralize any ambient CodeGraph prompt-hook (~/.claude) in BOTH arms:
+# the hook injects codegraph context into every prompt, which contaminates
+# the without-arm (free structural context) and double-counts the with-arm.
+# The A/B's only variable must be the MCP server wired below.
+export CODEGRAPH_NO_PROMPT_HOOK=1
+
+[ -n "$CG_BIN" ] || { echo "no codegraph binary on PATH (set CG_BIN)"; exit 1; }
+[ -d "$REPO/.codegraph" ] || { echo "no .codegraph index at $REPO — index it first"; exit 1; }
+case "$MODE" in headless|tmux|all) ;; *) echo "mode must be headless|tmux|all (got '$MODE')"; exit 1;; esac
+
+# MCP config files (path form avoids inline-JSON quoting through tmux).
+cat > "$OUT/mcp-codegraph.json" <<JSON
+{"mcpServers":{"codegraph":{"command":"$CG_BIN","args":["serve","--mcp","--path","$REPO"]}}}
+JSON
+echo '{"mcpServers":{}}' > "$OUT/mcp-empty.json"
+
+echo "###### codegraph: $CG_BIN"
+echo "###### repo:      $REPO"
+echo "###### question:  $Q"
+echo
+
+# Headless arm: claude -p with stream-json -> exact tool sequence + tokens/cost.
+headless() {
+  local label="$1" cfg="$2"
+  echo "############################## HEADLESS [$label] ##############################"
+  ( cd "$REPO" && claude -p "$Q" \
+      --output-format stream-json --verbose \
+      --permission-mode bypassPermissions \
+      --model "${MODEL:-sonnet}" --effort "${EFFORT:-high}" \
+      --max-budget-usd 4 \
+      --strict-mcp-config --mcp-config "$cfg" \
+      > "$OUT/run-$label.jsonl" 2>"$OUT/run-$label.err" )
+  echo "exit $? -> $OUT/run-$label.jsonl ($(wc -l < "$OUT/run-$label.jsonl" | tr -d ' ') lines)"
+  tail -2 "$OUT/run-$label.err" 2>/dev/null
+  node "$HARNESS/parse-run.mjs" "$OUT/run-$label.jsonl" 2>&1 || true
+  echo
+}
+
+if [ "$MODE" = headless ] || [ "$MODE" = all ]; then
+  headless "headless-with"    "$OUT/mcp-codegraph.json"
+  headless "headless-without" "$OUT/mcp-empty.json"
+fi
+
+if [ "$MODE" = tmux ] || [ "$MODE" = all ]; then
+  echo "############################## INTERACTIVE [with] ##############################"
+  CLAUDE_EXTRA_ARGS="--model ${MODEL:-sonnet} --effort ${EFFORT:-high} --strict-mcp-config --mcp-config $OUT/mcp-codegraph.json" \
+    bash "$HARNESS/itrun.sh" "$REPO" "int-with" "$Q" 2>&1 || echo "[itrun WITH failed]"
+  echo
+  echo "############################## INTERACTIVE [without] ##############################"
+  CLAUDE_EXTRA_ARGS="--model ${MODEL:-sonnet} --effort ${EFFORT:-high} --strict-mcp-config --mcp-config $OUT/mcp-empty.json" \
+    bash "$HARNESS/itrun.sh" "$REPO" "int-without" "$Q" 2>&1 || echo "[itrun WITHOUT failed]"
+  echo
+fi
+echo "############################## RUN-ALL COMPLETE ##############################"

package/scripts/agent-eval/run-arms.sh

@@ -0,0 +1,56 @@
+#!/usr/bin/env bash
+# Tool-surface ablation — run ONE repo+question under ONE arm.
+#
+# Arms vary (exposed codegraph tools, trace-first steering). Tools are trimmed
+# SERVER-SIDE via CODEGRAPH_MCP_TOOLS in the MCP config's `env` block, so an
+# ablated tool is genuinely absent from ListTools — no deferred-ToolSearch or
+# denied-call confound (which --disallowedTools would introduce). Steering is
+# injected with --append-system-prompt, so no rebuild of the shipped
+# server-instructions is needed to A/B it.
+#
+#   A control       all tools            no steering
+#   B steer         all tools            trace-first
+#   C no-explore    hide explore         trace-first
+#   D trace-centric hide explore+context trace-first
+#   E control-probe hide explore+context trace-first  (caller passes a NON-flow Q)
+#
+# Usage: run-arms.sh <repo-path> "<question>" <A|B|C|D|E> [run-id]
+set -uo pipefail
+REPO="${1:?repo path}"; Q="${2:?question}"; ARM="${3:?arm A-E}"; RID="${4:-1}"
+CG_BIN="${CG_BIN:-$(command -v codegraph)}"
+OUT="${ARMS_OUT:-/tmp/arms}/$(basename "$REPO")"
+mkdir -p "$OUT"
+[ -n "$CG_BIN" ] || { echo "no codegraph binary (set CG_BIN)"; exit 1; }
+[ -d "$REPO/.codegraph" ] || { echo "no .codegraph index at $REPO"; exit 1; }
+
+STEER='Flow questions ("how does X reach/become Y", "trace the flow", request to handler, state to render): call codegraph_trace(from,to) FIRST — one call returns the whole path. Use codegraph_context/search only to locate the two endpoint symbols if you do not know them. Do NOT reconstruct the path with repeated search/callers/explore.'
+KEEP_NO_EXPLORE="trace,search,node,context,callers,callees,impact,files,status"
+KEEP_TRACE_CENTRIC="trace,search,node,callers,callees,impact,files,status"
+
+case "$ARM" in
+  A|G|H|I) TOOLS="";            STEERING="" ;;  # no steering; H = body-trace, I = body-trace + destination callees (sufficiency)
+  B|F) TOOLS="";                STEERING="$STEER" ;;  # F = B's surface, run on the body-inlining trace build
+  C) TOOLS="$KEEP_NO_EXPLORE";  STEERING="$STEER" ;;
+  D|E) TOOLS="$KEEP_TRACE_CENTRIC"; STEERING="$STEER" ;;
+  *) echo "bad arm '$ARM' (want A|B|C|D|E)"; exit 1 ;;
+esac
+
+CFG="$OUT/mcp-$ARM.json"
+if [ -n "$TOOLS" ]; then
+  cat > "$CFG" <<JSON
+{"mcpServers":{"codegraph":{"command":"$CG_BIN","args":["serve","--mcp","--path","$REPO"],"env":{"CODEGRAPH_MCP_TOOLS":"$TOOLS"}}}}
+JSON
+else
+  cat > "$CFG" <<JSON
+{"mcpServers":{"codegraph":{"command":"$CG_BIN","args":["serve","--mcp","--path","$REPO"]}}}
+JSON
+fi
+
+LOG="$OUT/$ARM-r$RID.jsonl"; ERR="$OUT/$ARM-r$RID.err"
+ARGS=( -p "$Q" --output-format stream-json --verbose
+       --permission-mode bypassPermissions --model "${MODEL:-sonnet}" --effort "${EFFORT:-high}" --max-budget-usd 4
+       --strict-mcp-config --mcp-config "$CFG" )
+[ -n "$STEERING" ] && ARGS+=( --append-system-prompt "$STEERING" )
+
+( cd "$REPO" && claude "${ARGS[@]}" > "$LOG" 2>"$ERR" )
+echo "[$(basename "$REPO") $ARM r$RID] exit $? -> $LOG ($(wc -l < "$LOG" | tr -d ' ') lines)"

package/scripts/agent-eval/seq-matrix.mjs

@@ -0,0 +1,137 @@
+#!/usr/bin/env node
+// Mine the surviving A/B stream-json logs (/tmp/ab-matrix/<Cell>/run-headless-*.jsonl)
+// for what the aggregate matrix can't see: the call SEQUENCE and per-call output SIZE.
+//
+// Answers three questions:
+//   1. Trace adoption — on a flow question, does the with-arm actually call codegraph_trace?
+//   2. Payload size vs repo size — is trace path-scoped (tiny, size-independent) while
+//      explore is breadth-scoped (grows with the repo / over-returns on small repos)?
+//   3. Round-trips — num_turns with vs without (the real wall-clock driver).
+//
+// Usage: node scripts/agent-eval/seq-matrix.mjs [/tmp/ab-matrix]
+import { readFileSync, readdirSync, existsSync } from 'fs';
+import { join } from 'path';
+
+const AB = process.argv[2] || '/tmp/ab-matrix';
+const MD = new URL('../../docs/benchmarks/codegraph-ab-matrix.md', import.meta.url).pathname;
+
+// repo -> {lang,size,files} from the published matrix table
+const repoMeta = {};
+if (existsSync(MD)) for (const line of readFileSync(MD, 'utf8').split('\n')) {
+  const m = line.match(/^\|\s*([^|]+?)\s*\|\s*(S|M|L)\s*\|\s*`([^`]+)`\s*\|\s*(\d+)\s*\|/);
+  if (m) repoMeta[m[3]] = { lang: m[1].trim(), size: m[2], files: +m[4] };
+}
+
+const cgShort = (n) => n.replace('mcp__codegraph__codegraph_', '').replace('mcp__codegraph__', '');
+const tag = (n) => n === 'Read' ? 'R' : n === 'Grep' ? 'G' : n === 'Glob' ? 'Gl'
+  : n === 'Bash' ? 'B' : n === 'Task' ? 'Ag' : n === 'ToolSearch' ? 'TS'
+  : n.includes('codegraph') ? cgShort(n) : n;
+
+function parse(file) {
+  if (!existsSync(file)) return null;
+  const lines = readFileSync(file, 'utf8').split('\n').filter(Boolean);
+  const calls = []; let result = null, initCg = 0;
+  for (const l of lines) {
+    let ev; try { ev = JSON.parse(l); } catch { continue; }
+    if (ev.type === 'system' && ev.subtype === 'init') initCg = (ev.tools || []).filter(t => /codegraph/.test(t)).length;
+    if (ev.type === 'assistant') for (const b of (ev.message?.content || [])) if (b.type === 'tool_use') {
+      const i = b.input || {};
+      const q = i.query ?? i.symbol ?? i.task ?? (i.from && i.to ? `${i.from}->${i.to}` : (i.file_path || i.command || ''));
+      calls.push({ id: b.id, name: b.name, q: String(q ?? '').slice(0, 38), out: 0 });
+    }
+    if (ev.type === 'user') for (const b of (ev.message?.content || [])) if (b.type === 'tool_result') {
+      const c = b.content;
+      const txt = typeof c === 'string' ? c : Array.isArray(c) ? c.map(x => x?.text || '').join('') : '';
+      const call = calls.find(k => k.id === b.tool_use_id); if (call) call.out = txt.length;
+    }
+    if (ev.type === 'result') result = ev;
+  }
+  const cg = calls.filter(c => c.name.includes('codegraph'));
+  const perTool = {};
+  for (const c of cg) { const k = cgShort(c.name); (perTool[k] ??= { n: 0, out: 0 }); perTool[k].n++; perTool[k].out += c.out; }
+  const traceIdx = cg.findIndex(c => c.name.includes('trace'));
+  const u = result?.usage || {};
+  return {
+    initCg, cg, perTool,
+    cgSeq: cg.map(c => cgShort(c.name)),
+    seq: calls.map(c => tag(c.name)),
+    reads: calls.filter(c => c.name === 'Read').length,
+    greps: calls.filter(c => c.name === 'Grep').length,
+    cgOut: cg.reduce((s, c) => s + c.out, 0),
+    traceUsed: traceIdx >= 0,
+    afterTrace: traceIdx >= 0 ? cg.slice(traceIdx + 1).map(c => cgShort(c.name)) : null,
+    turns: result?.num_turns ?? null,
+    dur: result?.duration_ms ? Math.round(result.duration_ms / 1000) : null,
+    cost: result?.total_cost_usd || 0,
+  };
+}
+
+const cells = [];
+for (const d of readdirSync(AB)) {
+  const dir = join(AB, d);
+  if (!existsSync(join(dir, 'run-headless-with.jsonl'))) continue;
+  const log = existsSync(join(AB, d + '.log')) ? readFileSync(join(AB, d + '.log'), 'utf8') : '';
+  const repo = (log.match(/repo:\s*\S*\/([^\s/]+)/) || [])[1] || d;
+  const question = (log.match(/question:\s*(.+)/) || [])[1] || '';
+  cells.push({ cell: d, repo, question, ...(repoMeta[repo] || {}),
+    with: parse(join(dir, 'run-headless-with.jsonl')),
+    without: parse(join(dir, 'run-headless-without.jsonl')) });
+}
+cells.sort((a, b) => (a.files || 0) - (b.files || 0));
+
+const k = (n) => (n / 1000).toFixed(1);
+const pad = (s, n) => String(s).padEnd(n);
+
+// ---- per-cell sequence table ----
+console.log('\n=== PER-CELL: with-arm codegraph sequence + payload (sorted by repo size) ===');
+console.log(pad('repo', 22), pad('files', 6), 'trace', pad('cg-call sequence', 40), pad('cgOutK', 7), 'turns(w/wo)');
+for (const c of cells) {
+  const w = c.with;
+  console.log(
+    pad(c.repo, 22), pad(c.files ?? '?', 6),
+    pad(w.traceUsed ? 'YES' : 'no', 5),
+    pad(w.cgSeq.join(',') || '(none)', 40),
+    pad(k(w.cgOut), 7),
+    `${w.turns}/${c.without?.turns}`,
+  );
+}
+
+// ---- trace adoption ----
+const flow = cells; // every matrix question is a canonical flow question by design
+const used = flow.filter(c => c.with.traceUsed);
+console.log(`\n=== TRACE ADOPTION (all ${flow.length} cells are flow questions) ===`);
+console.log(`trace called in ${used.length}/${flow.length} cells`);
+console.log('used trace:', used.map(c => c.repo).join(', ') || '(none)');
+if (used.length) console.log('after-trace follow-ups:', used.map(c => `${c.repo}[${c.with.afterTrace.join(',') || 'none'}]`).join('  '));
+
+// ---- payload size by repo-size tier ----
+const tier = (f) => f < 200 ? 'S(<200)' : f < 2000 ? 'M(<2000)' : 'L(>=2000)';
+const byTier = {};
+for (const c of cells) { (byTier[tier(c.files || 0)] ??= []).push(c.with.cgOut); }
+console.log('\n=== with-arm TOTAL codegraph payload by repo-size tier ===');
+for (const t of ['S(<200)', 'M(<2000)', 'L(>=2000)']) {
+  const a = byTier[t] || []; if (!a.length) continue;
+  const avg = a.reduce((s, x) => s + x, 0) / a.length;
+  console.log(`  ${pad(t, 10)} n=${a.length}  avg cgOut=${k(avg)}K  range ${k(Math.min(...a))}-${k(Math.max(...a))}K`);
+}
+
+// ---- per-tool usage + avg payload (breadth vs path evidence) ----
+const tot = {};
+for (const c of cells) for (const [name, v] of Object.entries(c.with.perTool)) {
+  (tot[name] ??= { n: 0, out: 0 }); tot[name].n += v.n; tot[name].out += v.out;
+}
+console.log('\n=== codegraph tool usage across all cells (n calls, avg payload/call) ===');
+for (const [name, v] of Object.entries(tot).sort((a, b) => b[1].n - a[1].n)) {
+  console.log(`  ${pad(name, 10)} calls=${pad(v.n, 4)} avg=${k(v.out / v.n)}K/call  total=${k(v.out)}K`);
+}
+
+// ---- round-trips ----
+const sum = (arr, f) => arr.reduce((s, x) => s + (f(x) || 0), 0);
+const wTurns = sum(cells, c => c.with.turns), woTurns = sum(cells, c => c.without?.turns);
+const wCalls = sum(cells, c => c.with.cg.length);
+const tsAll = cells.every(c => c.with.seq[0] === 'TS');
+console.log('\n=== ROUND-TRIPS ===');
+console.log(`turns: with=${wTurns}  without=${woTurns}  (${((1 - wTurns / woTurns) * 100).toFixed(0)}% fewer with)`);
+console.log(`avg turns/cell: with=${(wTurns / cells.length).toFixed(1)}  without=${(woTurns / cells.length).toFixed(1)}`);
+console.log(`total codegraph calls=${wCalls} (avg ${(wCalls / cells.length).toFixed(1)}/cell)`);
+console.log(`every with-arm opens with a ToolSearch round-trip (deferred tools): ${tsAll ? 'YES — 1 fixed tax/run' : 'no'}`);

package/scripts/build-bundle.sh

@@ -0,0 +1,123 @@
+#!/usr/bin/env bash
+#
+# Build a self-contained CodeGraph bundle: an official Node runtime + the
+# compiled app + its production deps, so CodeGraph runs with NO system Node and
+# NO native build — node:sqlite is built into the bundled Node. One archive per
+# platform.
+#
+# Because dropping better-sqlite3 left zero native addons, the recipe is pure
+# file-packaging (download the target's Node, copy the app, archive) — so any
+# platform's bundle can be built on any OS. No cross-compile, no native runners.
+#
+# Usage:
+#   scripts/build-bundle.sh <target> [node-version]
+#     target:        darwin-arm64 | darwin-x64 | linux-x64 | linux-arm64
+#                  | win32-x64 | win32-arm64
+#     node-version:  e.g. v24.16.0 (default below; pin for reproducible builds)
+#
+# Output:
+#   unix:    release/codegraph-<target>.tar.gz   (launcher: bin/codegraph)
+#   windows: release/codegraph-<target>.zip      (launcher: bin/codegraph.cmd)
+set -euo pipefail
+
+TARGET="${1:?usage: build-bundle.sh <target> [node-version]}"
+NODE_VERSION="${2:-v24.16.0}"
+
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+OUT="$ROOT/release"
+WORK="$(mktemp -d)"
+trap 'rm -rf "$WORK"' EXIT
+
+ARCH="${TARGET##*-}"   # x64 | arm64
+OSFAM="${TARGET%-*}"   # darwin | linux | win32
+
+echo "[bundle] target=${TARGET} node=${NODE_VERSION}"
+
+# 1. Download + extract the official Node runtime for the target platform.
+if [ "$OSFAM" = "win32" ]; then
+  NODE_DIST="node-${NODE_VERSION}-win-${ARCH}"
+  NODE_URL="https://nodejs.org/dist/${NODE_VERSION}/${NODE_DIST}.zip"
+  echo "[bundle] downloading ${NODE_URL}"
+  curl -fsSL "$NODE_URL" -o "$WORK/node.zip"
+  if command -v unzip >/dev/null 2>&1; then
+    unzip -q "$WORK/node.zip" -d "$WORK"
+  else
+    tar -xf "$WORK/node.zip" -C "$WORK"   # bsdtar can read zip
+  fi
+  NODE_BIN="$WORK/${NODE_DIST}/node.exe"
+else
+  NODE_DIST="node-${NODE_VERSION}-${TARGET}"
+  NODE_URL="https://nodejs.org/dist/${NODE_VERSION}/${NODE_DIST}.tar.gz"
+  echo "[bundle] downloading ${NODE_URL}"
+  curl -fsSL "$NODE_URL" -o "$WORK/node.tar.gz"
+  tar -xzf "$WORK/node.tar.gz" -C "$WORK"
+  NODE_BIN="$WORK/${NODE_DIST}/bin/node"
+fi
+[ -f "$NODE_BIN" ] || { echo "[bundle] error: node binary not found ($NODE_BIN)" >&2; exit 1; }
+
+# 2. Build the app (compiled JS + copied wasm/schema assets).
+echo "[bundle] building app"
+( cd "$ROOT" && npm run build >/dev/null )
+
+# 3. Stage: app + production-only deps (pure JS/wasm → portable across platforms).
+STAGE="$WORK/codegraph-${TARGET}"
+mkdir -p "$STAGE/lib" "$STAGE/bin"
+cp -R "$ROOT/dist" "$STAGE/lib/dist"
+cp "$ROOT/package.json" "$ROOT/package-lock.json" "$STAGE/lib/"
+echo "[bundle] installing production dependencies"
+( cd "$STAGE/lib" && npm ci --omit=dev --ignore-scripts >/dev/null 2>&1 )
+rm -f "$STAGE/lib/package-lock.json"
+
+# 4. Vendored Node + launcher (the launcher uses the bundled Node by relative
+#    path, so no system Node is ever needed).
+#
+# `--liftoff-only`: keep tree-sitter's large WASM grammars on V8's Liftoff
+# baseline compiler so they never reach the turboshaft optimizing tier, whose
+# per-compilation Zone arena OOMs the whole process (`Fatal process out of
+# memory: Zone`) on Node >= 22 — even with tens of GB free. The flag is read at
+# V8 engine init so it must be on node's command line; the parse worker inherits
+# it. See issues #293/#298 and src/extraction/wasm-runtime-flags.ts. (The CLI
+# also self-relaunches with this flag when launched without it, so non-bundled
+# runs are covered too; passing it here avoids that extra spawn.)
+if [ "$OSFAM" = "win32" ]; then
+  cp "$NODE_BIN" "$STAGE/node.exe"
+  printf '@"%%~dp0..\\node.exe" --liftoff-only "%%~dp0..\\lib\\dist\\bin\\codegraph.js" %%*\r\n' \
+    > "$STAGE/bin/codegraph.cmd"
+else
+  cp "$NODE_BIN" "$STAGE/node"
+  cat > "$STAGE/bin/codegraph" <<'LAUNCH'
+#!/bin/sh
+# Resolve symlinks (e.g. the ~/.local/bin/codegraph link install.sh creates) so
+# we find the real bundle dir, not the symlink's location.
+SELF="$0"
+while [ -L "$SELF" ]; do
+  target="$(readlink "$SELF")"
+  case "$target" in
+    /*) SELF="$target" ;;
+    *) SELF="$(dirname "$SELF")/$target" ;;
+  esac
+done
+DIR="$(cd "$(dirname "$SELF")/.." && pwd)"
+# Thread the MCP host's pid to the server's orphan watchdog (issue #1185).
+# $PPID is our parent — the host itself when it launched this script directly;
+# an already-threaded value (the npm shim sets the true host pid) wins.
+CODEGRAPH_HOST_PPID="${CODEGRAPH_HOST_PPID:-$PPID}"
+export CODEGRAPH_HOST_PPID
+# --liftoff-only: avoid the V8 turboshaft WASM Zone OOM (issues #293/#298).
+exec "$DIR/node" --liftoff-only "$DIR/lib/dist/bin/codegraph.js" "$@"
+LAUNCH
+  chmod +x "$STAGE/bin/codegraph"
+fi
+
+# 5. Archive (.zip for Windows, .tar.gz otherwise).
+mkdir -p "$OUT"
+if [ "$OSFAM" = "win32" ]; then
+  ARCHIVE="$OUT/codegraph-${TARGET}.zip"
+  rm -f "$ARCHIVE"
+  ( cd "$WORK" && zip -rqX "$ARCHIVE" "codegraph-${TARGET}" )
+else
+  ARCHIVE="$OUT/codegraph-${TARGET}.tar.gz"
+  # --no-xattrs: don't embed macOS xattrs that make GNU tar warn on Linux.
+  tar --no-xattrs -czf "$ARCHIVE" -C "$WORK" "codegraph-${TARGET}"
+fi
+echo "[bundle] wrote ${ARCHIVE} ($(du -h "$ARCHIVE" | cut -f1))"

package/scripts/extract-release-notes.mjs

@@ -0,0 +1,130 @@
+#!/usr/bin/env node
+/**
+ * Extract a release-notes block from CHANGELOG.md for a given version
+ * (or unwrap text supplied on stdin), then join hard-wrapped paragraphs.
+ *
+ * Why: GitHub renders release-note Markdown with GFM hard breaks, so
+ * every `\n` becomes `<br>`. The CHANGELOG is hard-wrapped at ~75
+ * chars for readable diffs, which then renders as awkward visible
+ * line breaks on the release page. This script joins indented
+ * continuation lines into a single line per bullet so the GFM
+ * renderer produces clean paragraphs.
+ *
+ * Repo-level CHANGELOG.md viewing is unaffected (CommonMark treats
+ * newlines as spaces there).
+ *
+ * Usage:
+ *   extract-release-notes.mjs <version>     # read CHANGELOG.md
+ *   extract-release-notes.mjs --stdin       # read from stdin (any text)
+ */
+
+import { readFileSync } from 'fs';
+
+const arg = process.argv[2];
+if (!arg) {
+  console.error('usage: extract-release-notes.mjs <version> | --stdin');
+  process.exit(1);
+}
+
+let block;
+if (arg === '--stdin') {
+  block = readFileSync(0, 'utf8').replace(/\r\n?/g, '\n').split('\n');
+} else {
+  const version = arg;
+  const escaped = version.replace(/\./g, '\\.');
+  const headerRe = new RegExp(`^## \\[${escaped}\\]`);
+  const anyHeaderRe = /^## \[/;
+  const lines = readFileSync('CHANGELOG.md', 'utf8').split('\n');
+  const start = lines.findIndex((l) => headerRe.test(l));
+  if (start === -1) {
+    console.error(`no '## [${version}]' entry found in CHANGELOG.md`);
+    process.exit(1);
+  }
+  const after = lines.findIndex((l, i) => i > start && anyHeaderRe.test(l));
+  block = lines.slice(start, after === -1 ? lines.length : after);
+}
+
+// Track a stack of `{ indent: number }` frames so a continuation line
+// can attach to the right ancestor. Handles the post-nested-list
+// continuation pattern:
+//
+//     - top-level
+//         - nested
+//       back to top-level  <- 2-space indent, joins the top-level bullet
+const out = [];
+let buf = '';
+let stack = [];
+
+function flushBuf() {
+  if (buf !== '') {
+    out.push(buf);
+    buf = '';
+  }
+}
+
+function leadingSpaces(s) {
+  const m = s.match(/^(\s*)/);
+  return m ? m[1].length : 0;
+}
+
+// Bullets: `-`, `*`, `digit.` only. `+` is intentionally excluded — the
+// CHANGELOG uses literal `+` inline (`config + instructions`) and we
+// don't want to misread those as nested bullets.
+const listItemRe = /^(\s*)([-*]|\d+\.)\s+/;
+const fenceRe = /^\s*```/;
+
+let inFence = false;
+
+for (const line of block) {
+  // Fenced code blocks: pass through verbatim, no joining.
+  if (fenceRe.test(line)) {
+    flushBuf();
+    stack = [];
+    out.push(line);
+    inFence = !inFence;
+    continue;
+  }
+  if (inFence) {
+    out.push(line);
+    continue;
+  }
+  if (/^\s*$/.test(line)) {
+    flushBuf();
+    out.push('');
+    continue;
+  }
+  if (/^#/.test(line)) {
+    flushBuf();
+    stack = [];
+    out.push(line);
+    continue;
+  }
+  const itemMatch = line.match(listItemRe);
+  if (itemMatch) {
+    flushBuf();
+    const indent = itemMatch[1].length;
+    while (stack.length > 0 && stack[stack.length - 1].indent >= indent) {
+      stack.pop();
+    }
+    stack.push({ indent });
+    buf = line;
+    continue;
+  }
+  if (/^\s/.test(line)) {
+    const indent = leadingSpaces(line);
+    while (stack.length > 1 && stack[stack.length - 1].indent >= indent) {
+      flushBuf();
+      stack.pop();
+    }
+    const trimmed = line.replace(/^\s+/, '');
+    buf = buf === '' ? trimmed : `${buf} ${trimmed}`;
+    continue;
+  }
+  flushBuf();
+  stack = [];
+  out.push(line);
+}
+flushBuf();
+
+process.stdout.write(out.join('\n'));
+if (!out[out.length - 1]?.endsWith('\n')) process.stdout.write('\n');

package/scripts/local-install.sh

@@ -0,0 +1,41 @@
+#!/usr/bin/env bash
+# Build the current branch and link it as the global `codegraph` for
+# hands-on testing. Replaces any existing global install for as long
+# as the symlink is in place.
+#
+# Usage:
+#   ./scripts/local-install.sh           # build + link
+#   ./scripts/local-install.sh --undo    # unlink + restore the published version
+
+set -euo pipefail
+
+cd "$(dirname "$0")/.."
+
+PKG=$(node -p "require('./package.json').name")
+VERSION=$(node -p "require('./package.json').version")
+BRANCH=$(git rev-parse --abbrev-ref HEAD)
+
+if [ "${1:-}" = "--undo" ]; then
+  echo "→ unlinking ${PKG}"
+  npm unlink -g "${PKG}" >/dev/null 2>&1 || true
+  echo "→ reinstalling published ${PKG}"
+  npm install -g "${PKG}"
+  echo "done: global codegraph -> $(command -v codegraph)"
+  exit 0
+fi
+
+echo "→ building ${PKG} ${VERSION} (${BRANCH})"
+npm run build
+
+echo "→ linking globally"
+npm link
+
+LINKED=$(command -v codegraph || echo "(not on PATH)")
+echo
+echo "✓ global codegraph now points to this branch"
+echo "  binary:  ${LINKED}"
+echo "  branch:  ${BRANCH}"
+echo "  version: ${VERSION}"
+echo
+echo "To restore the published version:"
+echo "  ./scripts/local-install.sh --undo"

package/scripts/npm-sdk.js

@@ -0,0 +1,75 @@
+'use strict';
+//
+// Programmatic / embedded SDK entry for @colbymchenry/codegraph (issue #354).
+//
+// The CLI/MCP `bin` (npm-shim.js) execs the per-platform bundle's OWN Node 24 so
+// the tool never depends on the user's runtime. Embedded library consumers are
+// the opposite case: they already run their own Node and just want the compiled
+// API — `require("@colbymchenry/codegraph")` returning the CodeGraph class et al.
+//
+// The compiled library + its production dependencies (web-tree-sitter,
+// tree-sitter-wasms, …) ship INSIDE the per-platform bundle, at
+//   @colbymchenry/codegraph-<platform>-<arch>/lib/dist/index.js
+// (with the deps in the sibling lib/node_modules). Re-exporting that bundle keeps
+// the main package thin — no second 50 MB copy of the grammars — while making the
+// SDK work in the consumer's process. Types are a separate concern: the main
+// package ships its own dist/**/*.d.ts tree (pointed at by `types`), built from
+// the same release so it can never skew from the runtime it re-exports.
+//
+// node:sqlite (Node >= 22.5) is required to OPEN a graph, but only lazily inside
+// the SQLite adapter — so loading this module is safe on older Node, and the
+// node:sqlite requirement surfaces with an actionable error only when a DB is
+// actually opened. Heavy extraction additionally wants the bundled launcher's
+// --liftoff-only flag (the WASM Zone-OOM guard, issues #293/#298); an embedded
+// host that drives large indexing should pass that flag to its own Node.
+
+var path = require('path');
+var os = require('os');
+var fs = require('fs');
+
+var target = process.platform + '-' + process.arch; // e.g. darwin-arm64, linux-x64
+var pkg = '@colbymchenry/codegraph-' + target;
+
+module.exports = require(resolveLibrary());
+
+// Locate the compiled library entry inside the installed per-platform bundle.
+// Throws an actionable error (rather than a bare MODULE_NOT_FOUND) when no bundle
+// is present, so an embedded consumer knows exactly what to install.
+function resolveLibrary() {
+  // 1) The npm-installed optional dependency — the normal case.
+  try {
+    return require.resolve(pkg + '/lib/dist/index.js');
+  } catch (e) {
+    /* fall through to the self-healed cache */
+  }
+
+  // 2) A bundle the CLI shim self-healed from GitHub Releases into the cache
+  //    (issue #303). Same node/lib/bin layout as the npm package. We only REUSE a
+  //    cached bundle here — unlike the CLI shim we never trigger a network
+  //    download from inside require(), which must stay synchronous and cheap.
+  var cached = cachedLibrary();
+  if (cached) return cached;
+
+  throw new Error(
+    'codegraph: the programmatic API is unavailable because the platform bundle\n' +
+    '(' + pkg + ') is not installed.\n' +
+    'The compiled library ships inside that per-platform optional dependency.\n' +
+    'Fixes:\n' +
+    '  - install from the official npm registry so the matching bundle is fetched:\n' +
+    '      npm i @colbymchenry/codegraph --registry=https://registry.npmjs.org\n' +
+    '  - or run the CLI once (e.g. `npx @colbymchenry/codegraph status`) to\n' +
+    '    self-heal the bundle into ~/.codegraph, then require() will find it.'
+  );
+}
+
+function cachedLibrary() {
+  try {
+    var version = require(path.join(__dirname, 'package.json')).version;
+    var base = process.env.CODEGRAPH_INSTALL_DIR || path.join(os.homedir(), '.codegraph');
+    var lib = path.join(base, 'bundles', target + '-' + version, 'lib', 'dist', 'index.js');
+    if (fs.existsSync(lib)) return lib;
+  } catch (e) {
+    /* no readable cache → caller reports the install guidance */
+  }
+  return null;
+}