Sfoglia il codice sorgente

test(agent-eval): block the codegraph CLI outright — hiding it from PATH was not enough (CG-7)

An agent denied `codegraph` on PATH ran `find / -maxdepth 4 -iname "*codegraph*"`,
found the binary, and invoked it by ABSOLUTE PATH — 12 times in one without-arm
run. So block the invocation itself with a PreToolUse hook on Bash, written into
the run's output dir as an artifact alongside the MCP configs rather than as a
repo file.

The pattern matches command positions only, so looking is still allowed and only
using is denied: `grep codegraph src/`, `ls .codegraph` and `which codegraph`
pass through, while `codegraph explore`, `/abs/path/codegraph …`, `cd x &&
codegraph …` and `VAR=1 codegraph …` are refused. run-all.sh proves both
directions at startup and refuses to run if either fails. parse-run.mjs's
detector uses the same rule, so prevention and detection cannot drift — and it no
longer false-positives on the corpus path, which contains the word codegraph.

Verified end-to-end: the without-arm now probes with `ls .codegraph; which
codegraph`, finds nothing usable, and falls back to Read/Bash.
Colby McHenry 1 mese fa
parent
commit
e35d4861e0

+ 35 - 0
docs/benchmarks/residual-context-occupancy.md

@@ -126,6 +126,41 @@ question.
 
 ---
 
+## The without-arm was never actually without codegraph
+
+Establishing this baseline turned up a contamination channel that had been open
+the whole time, and it invalidates any number this harness produced for an arm
+that had Bash.
+
+The without-arm gets an empty MCP config, so it has no codegraph tool. It still
+has **Bash** — and the target repo still carries the `.codegraph/` index the
+with-arm needs, with the `codegraph` binary on PATH. Agents find that. In the
+first clean-looking 7-repo pass, **14 of 15 without-arm runs ran `codegraph
+explore` through Bash**, one of them by way of `ls .codegraph && codegraph
+explore …`. That arm was measuring codegraph-over-CLI against codegraph-over-MCP,
+not codegraph against its absence.
+
+It cuts the other way too. When the *with*-arm shells out, the output arrives as
+a Bash result and is attributed to Bash — understating what codegraph itself
+occupies. One of 15 with-arm runs did this.
+
+The fix is in `run-all.sh`: both arms now run on a PATH where the CLI is hidden,
+so the MCP server is the only way to reach codegraph and stays the A/B's single
+variable. The binary usually shares a directory with tools the run needs — here
+`claude` sits right next to it — so the directory is substituted in place by one
+of symlinks to every entry except `codegraph`, which keeps PATH order and
+precedence intact. The run aborts if `claude` or `node` did not survive.
+
+Prevention alone would fail silently the next time the binary lands somewhere
+new, so there is detection as well: `parse-run.mjs` flags any Bash command naming
+codegraph, and `parse-bench-readme.mjs` drops contaminated without-arm runs from
+the aggregate (`CG_INCLUDE_CONTAMINATED=1` keeps them).
+
+**Anyone re-reading older A/B results from this harness should assume the
+without-arm may have been using codegraph.**
+
+---
+
 ## Baseline: the 7 README repos
 
 <!-- RESULTS -->

+ 6 - 1
scripts/agent-eval/parse-run.mjs

@@ -59,6 +59,11 @@ const FAMILIES = ['codegraph', 'read', 'search', 'bash', 'other'];
 // The without-arm's way of getting the same bytes: reading and searching files.
 const FILE_ACCESS = ['read', 'search', 'bash'];
 
+// A Bash command that INVOKES the codegraph CLI, in any command position and by
+// any path. Mentions are not invocations: `grep codegraph src/`, `ls .codegraph`
+// and `which codegraph` all pass. Kept in step with run-all.sh's blocking hook.
+const CG_CLI_RE = /(^|[;&|(]|&&|\|\||\$\(|`)\s*(?:[A-Za-z_]\w*=\S*\s+)*[\w./~-]*codegraph(\s|$)/;
+
 const textOf = (content) =>
   Array.isArray(content) ? content.map((c) => c.text ?? (typeof c === 'string' ? c : JSON.stringify(c))).join('')
     : typeof content === 'string' ? content
@@ -131,7 +136,7 @@ export function parseSession(files) {
             // An arm with no codegraph MCP can still shell out to the CLI — the
             // target repo carries the .codegraph/ index and the binary is on
             // PATH. That silently turns a "without" arm into codegraph-over-CLI.
-            if (/\bcodegraph\b/.test(b.input?.command ?? '')) cliCalls++;
+            if (CG_CLI_RE.test(b.input?.command ?? '')) cliCalls++;
           }
           else if (b.name === 'Read') detail = ` ${(b.input?.file_path ?? '').split('/').slice(-1)[0]}`;
           toolCalls.push(`${b.name}${detail}`);

+ 30 - 0
scripts/agent-eval/run-all.sh

@@ -83,6 +83,35 @@ for t in claude node; do
   PATH="$ARM_PATH" command -v "$t" >/dev/null || { echo "sanitized PATH lost '$t' — refusing to run"; exit 1; }
 done
 
+# Hiding it from PATH is not enough. An agent denied `codegraph` ran
+# `find / -maxdepth 4 -iname "*codegraph*"`, found the binary, and invoked it by
+# ABSOLUTE PATH — so block the invocation itself with a PreToolUse hook. Written
+# into $OUT as a run artifact rather than a repo file, same as the MCP configs.
+# The pattern deliberately matches only COMMAND positions: `grep codegraph x`,
+# `ls .codegraph` and `which codegraph` are looking, not using, and pass through.
+CG_CMD_RE='(^|[;&|(]|&&|\|\||\$\(|`)[[:space:]]*([A-Za-z_][A-Za-z0-9_]*=[^[:space:]]*[[:space:]]+)*[A-Za-z0-9_./~-]*codegraph([[:space:]]|$)'
+cat > "$OUT/no-cli-hook.sh" <<HOOK
+#!/usr/bin/env bash
+# Deny Bash invocations of the codegraph CLI so the MCP server stays the A/B's
+# single variable. Looking for it is fine; running it is not.
+set -uo pipefail
+cmd="\$(cat | jq -r '.tool_input.command // empty' 2>/dev/null)"
+if printf '%s' "\$cmd" | grep -Eq '$CG_CMD_RE'; then
+  msg="The codegraph CLI is not available in this session. Answer using the tools you have."
+  jq -n --arg m "\$msg" '{reason:\$m, hookSpecificOutput:{hookEventName:"PreToolUse",permissionDecision:"deny",permissionDecisionReason:\$m}}'
+fi
+exit 0
+HOOK
+chmod +x "$OUT/no-cli-hook.sh"
+cat > "$OUT/hook-settings.json" <<JSON
+{"hooks":{"PreToolUse":[{"matcher":"Bash","hooks":[{"type":"command","command":"bash $OUT/no-cli-hook.sh"}]}]}}
+JSON
+command -v jq >/dev/null || { echo "jq is required for the CLI-block hook — install it or the arms will be contaminated"; exit 1; }
+# Prove the hook denies a real invocation and lets a mere mention through.
+_probe() { printf '{"tool_input":{"command":%s}}' "$1" | bash "$OUT/no-cli-hook.sh" | grep -c deny; }
+[ "$(_probe '"/Users/x/.local/bin/codegraph explore \"q\""')" = 1 ] || { echo "hook fails to block an absolute-path invocation"; exit 1; }
+[ "$(_probe '"grep -rn codegraph src/"')" = 0 ] || { echo "hook over-blocks a plain mention"; exit 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 '$MODE')"; exit 1;; esac
@@ -129,6 +158,7 @@ headless() {
         --model "${MODEL:-sonnet}" --effort "${EFFORT:-high}" \
         --max-budget-usd 4 \
         --strict-mcp-config --mcp-config "$cfg" \
+        --settings "$OUT/hook-settings.json" \
         ${resume[@]+"${resume[@]}"} \
         </dev/null > "$out" 2>>"$OUT/run-$label.err" )
     echo "exit $? -> $out ($(wc -l < "$out" | tr -d ' ') lines) [turn $seg/${#TURNS[@]}]"