Jelajahi Sumber

docs(agent-eval): evals must run non-nested + add ab-new-vs-baseline.sh (#734)

Running scripts/agent-eval against a `claude -p` spawned from within a Claude
Code session (nested, e.g. from a Bash tool call) makes the codegraph MCP
attach unreliable: the server is healthy (full handshake ~165ms) but the
nested client marks it status:"pending"/0-tools under CPU/timing contention,
so the agent silently runs with no codegraph. NO_DAEMON + `< /dev/null` don't
fix it — it's the nested client, not the server. Documented in CLAUDE.md's
validation methodology.

Adds ab-new-vs-baseline.sh: A/Bs a retrieval/steering change as new-build vs
baseline-build (both codegraph-on, isolating the change — vs run-all.sh's
with-vs-without), on a throwaway copy of an indexed repo. Run it in a real
terminal.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Colby Mchenry 2 minggu lalu
induk
melakukan
28c5268ee3
2 mengubah file dengan 84 tambahan dan 0 penghapusan
  1. 1 0
      CLAUDE.md
  2. 83 0
      scripts/agent-eval/ab-new-vs-baseline.sh

+ 1 - 0
CLAUDE.md

@@ -137,6 +137,7 @@ For each **language × framework**, validate on **small, medium, and large** rea
 1. **Pick the canonical flow** for the framework ("how does X reach Y": state→render, request→handler→view, query→SQL, action→reducer→store…).
 2. **Deterministic probes** (`scripts/agent-eval/probe-{node,explore}.mjs` against the built `dist/`): `codegraph_explore` with the flow's symbol names connects from→to end-to-end with no break (its Flow section shows the path); **no node explosion** (`select count(*) from nodes` stable before/after re-index); synthesized-edge **precision** spot-check (`select … where provenance='heuristic'`).
 3. **Agent A/B** (`scripts/agent-eval/run-all.sh <repo> "<Q>"`): with vs without codegraph, **≥2 runs/arm** (run-to-run variance is large — never conclude from n=1). Record **duration, total tool calls, Read, Grep**. Optional forced-Read-0 sufficiency proof via the block-read hook (`scripts/agent-eval/hook-settings.json`).
+   - **Run agent-evals in a REAL terminal — NEVER nested inside a Claude Code session** (don't spawn `claude -p` from a Bash tool call). The codegraph MCP server is healthy (full `initialize`→`tools/list` handshake ~165ms, daemon and in-process modes both fine), but a nested `claude -p` marks it `status:"pending"` / 0 tools under CPU/timing contention and the agent silently runs with no codegraph — it can connect early in a session, then degrade to consistent failure as nested spawns pile up. `CODEGRAPH_NO_DAEMON=1` and `< /dev/null` do NOT fix it (it's the nested client, not the server). Confirm via `parse-run.mjs` (`codegraph tools exposed: 0` = void run). To isolate a change — **new-build vs baseline-build, both codegraph-on** (vs run-all.sh's with-vs-without) — use `scripts/agent-eval/ab-new-vs-baseline.sh <indexed-repo> "<task>" [baseline-ref]`.
 4. **Pass bar:** a normal flow question reaches **~0 Read/Grep within the repo's explore-call budget**, runs **faster** than without-codegraph, and shows **no regression on a control repo**. Record the numbers in `docs/design/dynamic-dispatch-coverage-playbook.md` (the coverage matrix).
 
 Full playbook + per-mechanism design: `docs/design/dynamic-dispatch-coverage-playbook.md` and `docs/design/callback-edge-synthesis.md`.

+ 83 - 0
scripts/agent-eval/ab-new-vs-baseline.sh

@@ -0,0 +1,83 @@
+#!/usr/bin/env bash
+# A/B a codegraph retrieval/steering change: the NEW build (current HEAD) vs a
+# BASELINE build (a git ref) — BOTH with codegraph attached — on the same
+# implementation task, measuring how many Read vs codegraph calls the agent
+# makes. This ISOLATES the change (unlike run-all.sh, which is with-vs-without
+# codegraph). The agent works on a throwaway copy of the target, so its edits
+# never touch your repos.
+#
+# *** RUN THIS IN A REAL TERMINAL — NOT nested inside a Claude Code session. ***
+# A `claude -p` spawned from within another Claude session (e.g. from a Bash
+# tool call) cannot reliably attach the codegraph MCP server: the server is
+# healthy (full handshake ~165ms) but the nested client marks it
+# status:"pending" / 0 tools under CPU/timing contention, and degrades to
+# consistent failure over a long session. NO_DAEMON + `< /dev/null` do NOT fix
+# it — it's the nested client, not the server. See codegraph/CLAUDE.md
+# ("Running agent-evals — do NOT nest").
+#
+# Usage: ab-new-vs-baseline.sh <indexed-repo> "<task>" [baseline-ref]
+#   <indexed-repo>  a repo with a .codegraph index (copied per arm)
+#   "<task>"        an implementation task, e.g. "Add X to Y and wire it through"
+#   [baseline-ref]  git ref for the BEFORE build (default: HEAD~1)
+# Env: AGENT_EVAL_OUT (default: /tmp/ab-new-vs-baseline)
+set -uo pipefail
+
+TARGET="${1:?usage: ab-new-vs-baseline.sh <indexed-repo> \"<task>\" [baseline-ref]}"
+TASK="${2:?task required}"
+BASE_REF="${3:-HEAD~1}"
+ENGINE="$(cd "$(dirname "$0")/../.." && pwd)"
+BIN="$ENGINE/dist/bin/codegraph.js"
+OUT="${AGENT_EVAL_OUT:-/tmp/ab-new-vs-baseline}"
+PARSE="$ENGINE/scripts/agent-eval/parse-run.mjs"
+
+command -v claude >/dev/null || { echo "claude CLI not on PATH"; exit 1; }
+[ -d "$TARGET/.codegraph" ] || { echo "target not indexed: run 'codegraph init $TARGET' first"; exit 1; }
+if ! git -C "$ENGINE" diff --quiet || ! git -C "$ENGINE" diff --cached --quiet; then
+  echo "engine repo has uncommitted changes — commit or stash first (this script checks files out)"; exit 1
+fi
+CHANGED=$(git -C "$ENGINE" diff --name-only "$BASE_REF" HEAD -- src 2>/dev/null)
+[ -n "$CHANGED" ] || { echo "no src/ changes between $BASE_REF and HEAD — nothing to A/B"; exit 1; }
+
+# Always restore the engine to HEAD on exit, even if interrupted mid-arm.
+restore() { git -C "$ENGINE" checkout HEAD -- $CHANGED 2>/dev/null; ( cd "$ENGINE" && npm run build >/dev/null 2>&1 ); }
+trap restore EXIT
+
+mkdir -p "$OUT"
+echo "###### engine=$ENGINE  baseline=$BASE_REF"
+echo "###### changed: $(echo "$CHANGED" | tr '\n' ' ')"
+echo "###### target=$TARGET"
+echo "###### task=$TASK"
+echo
+
+# Two pristine copies so each arm starts clean (the agent edits its own copy).
+rm -rf "$OUT/t-new" "$OUT/t-base"
+rsync -a --exclude node_modules --exclude .git --exclude dist --exclude .codegraph "$TARGET/" "$OUT/t-new/"
+cp -R "$OUT/t-new" "$OUT/t-base"
+
+cfg() { printf '{"mcpServers":{"codegraph":{"command":"%s","args":["serve","--mcp","--path","%s"]}}}' "$BIN" "$1" > "$2"; }
+
+run_arm() { # label, target-copy
+  local label="$1" tgt="$2" c="$OUT/mcp-$1.json"
+  cfg "$tgt" "$c"
+  echo "############## ARM [$label] ##############"
+  ( cd "$tgt" && claude -p "$TASK" \
+      --output-format stream-json --verbose --permission-mode bypassPermissions \
+      --model opus --max-budget-usd 4 --strict-mcp-config --mcp-config "$c" \
+      < /dev/null > "$OUT/run-$label.jsonl" 2>"$OUT/run-$label.err" )
+  node "$PARSE" "$OUT/run-$label.jsonl" 2>&1 | grep -E "tools exposed|by type|Result" || echo "  (parse failed — see $OUT/run-$label.jsonl)"
+  echo
+}
+
+echo "== NEW build (HEAD) =="
+( cd "$ENGINE" && npm run build >/dev/null 2>&1 ) && echo "  built"
+node "$BIN" init "$OUT/t-new" >/dev/null 2>&1 && echo "  indexed t-new"
+run_arm new "$OUT/t-new"
+
+echo "== BASELINE build ($BASE_REF) =="
+git -C "$ENGINE" checkout "$BASE_REF" -- $CHANGED
+( cd "$ENGINE" && npm run build >/dev/null 2>&1 ) && echo "  built"
+node "$BIN" init "$OUT/t-base" >/dev/null 2>&1 && echo "  indexed t-base"
+run_arm baseline "$OUT/t-base"
+
+echo "###### DONE. Compare the [new] vs [baseline] 'by type' counts above"
+echo "###### (especially Read vs mcp__codegraph__*). Full logs in: $OUT"