homegraph 1.1.2 → 1.2.0

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

@@ -1,34 +1,34 @@
-#!/usr/bin/env bash
-# Headless Claude Code run against a repo with homegraph MCP, capturing the
-# full stream-json so we can see tool calls + token usage. Complements the
-# interactive itrun.sh: headless gives a clean per-tool breakdown + exact
-# tokens/cost, but defaults to the general-purpose subagent (not Explore).
-# To force the Explore path, ask for it in the prompt.
-#
-# Usage: run-agent.sh <repo-path> <label> "<prompt>"
-# Env: AGENT_EVAL_OUT (default /tmp/agent-eval), CG_BIN (homegraph dist binary)
-set -uo pipefail
-
-REPO="$1"; LABEL="$2"; PROMPT="$3"
-CG_BIN="${CG_BIN:-$(command -v homegraph || echo /usr/local/bin/homegraph)}"
-OUT_DIR="${AGENT_EVAL_OUT:-/tmp/agent-eval}"; mkdir -p "$OUT_DIR"
-OUT="$OUT_DIR/run-${LABEL}.jsonl"
-
-MCP_CONFIG=$(cat <<JSON
-{"mcpServers":{"homegraph":{"command":"${CG_BIN}","args":["serve","--mcp","--path","${REPO}"]}}}
-JSON
-)
-
-echo "→ running [$LABEL] in $REPO"
-cd "$REPO" || exit 1
-
-claude -p "$PROMPT" \
-  --output-format stream-json --verbose \
-  --permission-mode bypassPermissions \
-  --model "${MODEL:-sonnet}" --effort "${EFFORT:-high}" \
-  --max-budget-usd 2 \
-  --strict-mcp-config --mcp-config "$MCP_CONFIG" \
-  > "$OUT" 2>"$OUT_DIR/run-${LABEL}.err"
-
-echo "exit: $? | wrote $OUT ($(wc -l < "$OUT") lines)"
-node "$(cd "$(dirname "$0")" && pwd)/parse-run.mjs" "$OUT" 2>/dev/null || true
+#!/usr/bin/env bash
+# Headless Claude Code run against a repo with homegraph MCP, capturing the
+# full stream-json so we can see tool calls + token usage. Complements the
+# interactive itrun.sh: headless gives a clean per-tool breakdown + exact
+# tokens/cost, but defaults to the general-purpose subagent (not Explore).
+# To force the Explore path, ask for it in the prompt.
+#
+# Usage: run-agent.sh <repo-path> <label> "<prompt>"
+# Env: AGENT_EVAL_OUT (default /tmp/agent-eval), CG_BIN (homegraph dist binary)
+set -uo pipefail
+
+REPO="$1"; LABEL="$2"; PROMPT="$3"
+CG_BIN="${CG_BIN:-$(command -v homegraph || echo /usr/local/bin/homegraph)}"
+OUT_DIR="${AGENT_EVAL_OUT:-/tmp/agent-eval}"; mkdir -p "$OUT_DIR"
+OUT="$OUT_DIR/run-${LABEL}.jsonl"
+
+MCP_CONFIG=$(cat <<JSON
+{"mcpServers":{"homegraph":{"command":"${CG_BIN}","args":["serve","--mcp","--path","${REPO}"]}}}
+JSON
+)
+
+echo "→ running [$LABEL] in $REPO"
+cd "$REPO" || exit 1
+
+claude -p "$PROMPT" \
+  --output-format stream-json --verbose \
+  --permission-mode bypassPermissions \
+  --model "${MODEL:-sonnet}" --effort "${EFFORT:-high}" \
+  --max-budget-usd 2 \
+  --strict-mcp-config --mcp-config "$MCP_CONFIG" \
+  > "$OUT" 2>"$OUT_DIR/run-${LABEL}.err"
+
+echo "exit: $? | wrote $OUT ($(wc -l < "$OUT") lines)"
+node "$(cd "$(dirname "$0")" && pwd)/parse-run.mjs" "$OUT" 2>/dev/null || true

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

@@ -1,69 +1,69 @@
-#!/usr/bin/env bash
-# With/without A/B (and optional interactive) eval for a homegraph version on a
-# repo. Codegraph is the ONLY variable: both arms launch claude with
-# --strict-mcp-config — with = homegraph-only MCP (pointed at $CG_BIN),
-# without = empty MCP. Built-in Read/Grep/Bash stay available in both arms.
-#
-# Usage: run-all.sh <repo-path> "<question>" [headless|tmux|all]
-# Env:   CG_BIN          homegraph binary (default: command -v homegraph)
-#        AGENT_EVAL_OUT  output dir (default: /tmp/agent-eval)
-#        MODEL / EFFORT  claude model/effort (default: sonnet / high — the
-#                        standing A/B policy; see CLAUDE.md, don't raise)
-set -uo pipefail
-
-REPO="${1:?usage: run-all.sh <repo-path> \"<question>\" [headless|tmux|all]}"
-Q="${2:?question required}"
-MODE="${3:-headless}"
-CG_BIN="${CG_BIN:-$(command -v homegraph)}"
-OUT="${AGENT_EVAL_OUT:-/tmp/agent-eval}"
-HARNESS="$(cd "$(dirname "$0")" && pwd)"
-mkdir -p "$OUT"
-
-[ -n "$CG_BIN" ] || { echo "no homegraph binary on PATH (set CG_BIN)"; exit 1; }
-[ -d "$REPO/.homegraph" ] || { echo "no .homegraph 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-homegraph.json" <<JSON
-{"mcpServers":{"homegraph":{"command":"$CG_BIN","args":["serve","--mcp","--path","$REPO"]}}}
-JSON
-echo '{"mcpServers":{}}' > "$OUT/mcp-empty.json"
-
-echo "###### homegraph: $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"    "$OUT/mcp-homegraph.json"
-  headless "headless-without" "$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 $OUT/mcp-homegraph.json" \
-    bash "$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 ##############################"
+#!/usr/bin/env bash
+# With/without A/B (and optional interactive) eval for a homegraph version on a
+# repo. Codegraph is the ONLY variable: both arms launch claude with
+# --strict-mcp-config — with = homegraph-only MCP (pointed at $CG_BIN),
+# without = empty MCP. Built-in Read/Grep/Bash stay available in both arms.
+#
+# Usage: run-all.sh <repo-path> "<question>" [headless|tmux|all]
+# Env:   CG_BIN          homegraph binary (default: command -v homegraph)
+#        AGENT_EVAL_OUT  output dir (default: /tmp/agent-eval)
+#        MODEL / EFFORT  claude model/effort (default: sonnet / high — the
+#                        standing A/B policy; see CLAUDE.md, don't raise)
+set -uo pipefail
+
+REPO="${1:?usage: run-all.sh <repo-path> \"<question>\" [headless|tmux|all]}"
+Q="${2:?question required}"
+MODE="${3:-headless}"
+CG_BIN="${CG_BIN:-$(command -v homegraph)}"
+OUT="${AGENT_EVAL_OUT:-/tmp/agent-eval}"
+HARNESS="$(cd "$(dirname "$0")" && pwd)"
+mkdir -p "$OUT"
+
+[ -n "$CG_BIN" ] || { echo "no homegraph binary on PATH (set CG_BIN)"; exit 1; }
+[ -d "$REPO/.homegraph" ] || { echo "no .homegraph 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-homegraph.json" <<JSON
+{"mcpServers":{"homegraph":{"command":"$CG_BIN","args":["serve","--mcp","--path","$REPO"]}}}
+JSON
+echo '{"mcpServers":{}}' > "$OUT/mcp-empty.json"
+
+echo "###### homegraph: $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"    "$OUT/mcp-homegraph.json"
+  headless "headless-without" "$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 $OUT/mcp-homegraph.json" \
+    bash "$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

@@ -1,56 +1,56 @@
-#!/usr/bin/env bash
-# Tool-surface ablation — run ONE repo+question under ONE arm.
-#
-# Arms vary (exposed homegraph tools, trace-first steering). Tools are trimmed
-# SERVER-SIDE via HOMEGRAPH_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 homegraph)}"
-OUT="${ARMS_OUT:-/tmp/arms}/$(basename "$REPO")"
-mkdir -p "$OUT"
-[ -n "$CG_BIN" ] || { echo "no homegraph binary (set CG_BIN)"; exit 1; }
-[ -d "$REPO/.homegraph" ] || { echo "no .homegraph index at $REPO"; exit 1; }
-
-STEER='Flow questions ("how does X reach/become Y", "trace the flow", request to handler, state to render): call homegraph_trace(from,to) FIRST — one call returns the whole path. Use homegraph_context/search only to locate the two endpoint symbols if you do not know them. Do NOT reconstruct the path with repeated search/callers/explore.'
-KEEP_NO_EXPLORE="trace,search,node,context,callers,callees,impact,files,status"
-KEEP_TRACE_CENTRIC="trace,search,node,callers,callees,impact,files,status"
-
-case "$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":{"homegraph":{"command":"$CG_BIN","args":["serve","--mcp","--path","$REPO"],"env":{"HOMEGRAPH_MCP_TOOLS":"$TOOLS"}}}}
-JSON
-else
-  cat > "$CFG" <<JSON
-{"mcpServers":{"homegraph":{"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 < "$LOG" | tr -d ' ') lines)"
+#!/usr/bin/env bash
+# Tool-surface ablation — run ONE repo+question under ONE arm.
+#
+# Arms vary (exposed homegraph tools, trace-first steering). Tools are trimmed
+# SERVER-SIDE via HOMEGRAPH_MCP_TOOLS in the MCP config's `env` block, so an
+# ablated tool is genuinely absent from ListTools — no deferred-ToolSearch or
+# denied-call confound (which --disallowedTools would introduce). Steering is
+# injected with --append-system-prompt, so no rebuild of the shipped
+# server-instructions is needed to A/B it.
+#
+#   A control       all tools            no steering
+#   B steer         all tools            trace-first
+#   C no-explore    hide explore         trace-first
+#   D trace-centric hide explore+context trace-first
+#   E control-probe hide explore+context trace-first  (caller passes a NON-flow Q)
+#
+# Usage: run-arms.sh <repo-path> "<question>" <A|B|C|D|E> [run-id]
+set -uo pipefail
+REPO="${1:?repo path}"; Q="${2:?question}"; ARM="${3:?arm A-E}"; RID="${4:-1}"
+CG_BIN="${CG_BIN:-$(command -v homegraph)}"
+OUT="${ARMS_OUT:-/tmp/arms}/$(basename "$REPO")"
+mkdir -p "$OUT"
+[ -n "$CG_BIN" ] || { echo "no homegraph binary (set CG_BIN)"; exit 1; }
+[ -d "$REPO/.homegraph" ] || { echo "no .homegraph index at $REPO"; exit 1; }
+
+STEER='Flow questions ("how does X reach/become Y", "trace the flow", request to handler, state to render): call homegraph_trace(from,to) FIRST — one call returns the whole path. Use homegraph_context/search only to locate the two endpoint symbols if you do not know them. Do NOT reconstruct the path with repeated search/callers/explore.'
+KEEP_NO_EXPLORE="trace,search,node,context,callers,callees,impact,files,status"
+KEEP_TRACE_CENTRIC="trace,search,node,callers,callees,impact,files,status"
+
+case "$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":{"homegraph":{"command":"$CG_BIN","args":["serve","--mcp","--path","$REPO"],"env":{"HOMEGRAPH_MCP_TOOLS":"$TOOLS"}}}}
+JSON
+else
+  cat > "$CFG" <<JSON
+{"mcpServers":{"homegraph":{"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

@@ -1,137 +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 homegraph_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/homegraph-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__homegraph__homegraph_', '').replace('mcp__homegraph__', '');
-const tag = (n) => n === 'Read' ? 'R' : n === 'Grep' ? 'G' : n === 'Glob' ? 'Gl'
-  : n === 'Bash' ? 'B' : n === 'Task' ? 'Ag' : n === 'ToolSearch' ? 'TS'
-  : n.includes('homegraph') ? 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 => /homegraph/.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('homegraph'));
-  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 homegraph 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 homegraph 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=== homegraph 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=${(woTurns / cells.length).toFixed(1)}`);
-console.log(`total homegraph calls=${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'}`);
+#!/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 homegraph_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/homegraph-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__homegraph__homegraph_', '').replace('mcp__homegraph__', '');
+const tag = (n) => n === 'Read' ? 'R' : n === 'Grep' ? 'G' : n === 'Glob' ? 'Gl'
+  : n === 'Bash' ? 'B' : n === 'Task' ? 'Ag' : n === 'ToolSearch' ? 'TS'
+  : n.includes('homegraph') ? 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 => /homegraph/.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('homegraph'));
+  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 homegraph 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 homegraph 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=== homegraph 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=${(woTurns / cells.length).toFixed(1)}`);
+console.log(`total homegraph calls=${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'}`);