test(agent-eval): add output-style A/B harness, cost/token analyzer, and DISALLOW/REP_START controls
Three additions to tighten the eval loop: - offload-eval-styles.sh: new 4-arm eval (raw/refs/map/src) isolating the Worker's output shape's effect on main-session tokens, latency, and accuracy. Delegation blocked by default (DISALLOW=Agent) so variance from Haiku subagent spawning doesn't contaminate the measurement. - offload-eval-cost.mjs: cost/token analyzer that reads Claude Code's own per-model accounting (modelUsage.costUSD) rather than re-deriving from raw token counts, giving a correct main(Sonnet)/sub(Haiku) split with proper per-tier pricing. - offload-eval-3arm.sh: adds DISALLOW env to block sub-agent delegation across all arms, and REP_START to append reps to an existing run without clobbering earlier jsonls (e.g. REP_START=4 REPS=3 → reps 4,5,6). Also adds CODEGRAPH_OFFLOAD_STYLE forwarding to the managed gateway so the styles eval can drive output shape end-to-end; the field is stripped before the upstream model call and never sent to BYO endpoints.
This commit is contained in:
Colby McHenry
@@ -36,9 +36,13 @@ prewarm() { # path extra-env (e.g. "FOO=bar")
|
||||
run() { # arm rep mcp-config usage-log-or-dash
|
||||
local arm="$1" rep="$2" cfg="$3" usage="$4" tag="$REPO-$1-$2"
|
||||
[ "$usage" != "-" ] && : > "$usage"
|
||||
# DISALLOW (optional): block sub-agent delegation across all arms so the A/B
|
||||
# measures the retrieval mode, not whether Sonnet decides to spawn a codegraph-blind
|
||||
# Explore subagent (which thrashes regardless and adds huge variance).
|
||||
( cd "$TARGET" && claude -p "$Q" \
|
||||
--output-format stream-json --verbose --permission-mode bypassPermissions \
|
||||
--model "${MODEL:-sonnet}" --effort "${EFFORT:-high}" --max-budget-usd 4 \
|
||||
${DISALLOW:+--disallowedTools "$DISALLOW"} \
|
||||
--strict-mcp-config --mcp-config "$cfg" \
|
||||
</dev/null > "$RUNS/$tag.jsonl" 2>"$RUNS/$tag.err" )
|
||||
node "$EXTRACT" --run "$RUNS/$tag.jsonl" --usage "$usage" --arm "$arm" --rep "$rep" \
|
||||
@@ -52,14 +56,17 @@ printf '{"mcpServers":{"codegraph":{"command":"env","args":["CODEGRAPH_WASM_RELA
|
||||
printf '{"mcpServers":{"codegraph":{"command":"env","args":["CODEGRAPH_WASM_RELAUNCHED=1","CODEGRAPH_OFFLOAD_DISABLE=1","node","%s","serve","--mcp","--path","%s"]}}}' "$BIN" "$TARGET" > "$CFG_RAW"
|
||||
printf '{"mcpServers":{}}' > "$CFG_NOCG"
|
||||
|
||||
echo "###### repo=$REPO tier=$TIER reps=$REPS model=${MODEL:-sonnet}/${EFFORT:-high}"
|
||||
# REP_START lets a later batch ADD reps without clobbering earlier jsonls
|
||||
# (e.g. REP_START=4 REPS=3 -> reps 4,5,6; default starts at 1).
|
||||
START="${REP_START:-1}"; END=$((START + REPS - 1))
|
||||
echo "###### repo=$REPO tier=$TIER reps=$START..$END model=${MODEL:-sonnet}/${EFFORT:-high}"
|
||||
echo "###### Q=$Q"
|
||||
echo "== ARM offload =="; prewarm "$TARGET" "CODEGRAPH_OFFLOAD_USAGE_LOG=$USAGE"
|
||||
for r in $(seq 1 "$REPS"); do run offload "$r" "$CFG_OFF" "$USAGE"; done
|
||||
for r in $(seq "$START" "$END"); do run offload "$r" "$CFG_OFF" "$USAGE"; done
|
||||
pkill -9 -f "serve --mcp --path $TARGET" 2>/dev/null; rm -f "$TARGET/.codegraph/daemon.sock" 2>/dev/null; sleep 1
|
||||
echo "== ARM raw =="; prewarm "$TARGET" "CODEGRAPH_OFFLOAD_DISABLE=1"
|
||||
for r in $(seq 1 "$REPS"); do run raw "$r" "$CFG_RAW" "-"; done
|
||||
for r in $(seq "$START" "$END"); do run raw "$r" "$CFG_RAW" "-"; done
|
||||
pkill -9 -f "serve --mcp --path $TARGET" 2>/dev/null; rm -f "$TARGET/.codegraph/daemon.sock" 2>/dev/null; sleep 1
|
||||
echo "== ARM nocg =="
|
||||
for r in $(seq 1 "$REPS"); do run nocg "$r" "$CFG_NOCG" "-"; done
|
||||
for r in $(seq "$START" "$END"); do run nocg "$r" "$CFG_NOCG" "-"; done
|
||||
echo "###### DONE $REPO"
|
||||
|
||||
@@ -0,0 +1,133 @@
|
||||
#!/usr/bin/env node
|
||||
// Cost/token analysis for the 3-arm offload eval, with a MAIN-vs-SUBAGENT split.
|
||||
//
|
||||
// The explore-subagent question. With delegation ALLOWED, the nocg arm spawns a
|
||||
// Claude Code Explore subagent; the codegraph arms do all work in the main agent.
|
||||
// Two facts make naive accounting wrong:
|
||||
// 1. The Explore subagent runs on HAIKU 4.5; the main agent on SONNET 4.6.
|
||||
// So per-token cost differs ~3x between them — you cannot price both the same.
|
||||
// 2. The subagent's consumption is ~95% cache-reads. At Haiku's $0.10/MTok
|
||||
// cache-read rate, a huge TOKEN volume is a small DOLLAR cost.
|
||||
//
|
||||
// Rather than re-derive cost from raw token counts (and guess the cache TTL —
|
||||
// Claude Code uses 1-hour ephemeral cache here, 2x write, not 5-min), we read
|
||||
// Claude Code's OWN authoritative accounting from the `result` event:
|
||||
// result.modelUsage[model].costUSD — per-model cost CC itself billed
|
||||
// result.total_cost_usd — their sum (INCLUDES the Haiku subagent;
|
||||
// the handoff's "excludes subagent" was wrong)
|
||||
// The model split IS the agent split here: sonnet => main, haiku => Explore subagent
|
||||
// (only nocg spawns one, and only nocg shows haiku usage). Token volume is still
|
||||
// summed per-model from modelUsage for the separate "tokens" story.
|
||||
//
|
||||
// Usage: offload-eval-cost.mjs <runs-dir> <repo> [reps]
|
||||
// e.g. offload-eval-cost.mjs /tmp/cg-offload-eval/runs trezor 3
|
||||
import { readFileSync, existsSync } from 'fs';
|
||||
|
||||
const MAIN_TIER = /sonnet/; // main agent
|
||||
const SUB_TIER = /haiku/; // Claude Code Explore subagent
|
||||
|
||||
const [,, runsDir, repo, repsArg] = process.argv;
|
||||
if (!runsDir || !repo) { console.error('usage: offload-eval-cost.mjs <runs-dir> <repo> [reps] (env ARMS=nocg,raw,offload)'); process.exit(1); }
|
||||
const REPS = Number(repsArg || 3);
|
||||
// Arms to analyze (file stems `<repo>-<arm>-<rep>.jsonl`). Override for the style A/B:
|
||||
// ARMS=raw,refs,map,src. nocg's Haiku subagent is the only sub-tier; the rest are main-only.
|
||||
const ARMS = (process.env.ARMS || 'nocg,raw,offload').split(',').map((s) => s.trim()).filter(Boolean);
|
||||
|
||||
const toks = (u) => (u.inputTokens||0)+(u.outputTokens||0)+(u.cacheReadInputTokens||0)+(u.cacheCreationInputTokens||0);
|
||||
|
||||
function analyzeRun(file) {
|
||||
let result = null, agentCalls = 0;
|
||||
const tools = {}, subPids = new Set();
|
||||
for (const line of readFileSync(file, 'utf8').split('\n')) {
|
||||
if (!line) continue;
|
||||
let e; try { e = JSON.parse(line); } catch { continue; }
|
||||
if (e.parent_tool_use_id && e.message?.usage) subPids.add(e.parent_tool_use_id);
|
||||
if (e.type === 'assistant' && Array.isArray(e.message?.content))
|
||||
for (const b of e.message.content)
|
||||
if (b.type === 'tool_use') { tools[b.name] = (tools[b.name]||0)+1; if (b.name === 'Agent') agentCalls++; }
|
||||
if (e.type === 'result') result = e;
|
||||
}
|
||||
// Authoritative cost + tokens from Claude Code's per-model accounting.
|
||||
const mu = result?.modelUsage || {};
|
||||
const main = { cost: 0, tok: 0 }, sub = { cost: 0, tok: 0 };
|
||||
for (const [model, u] of Object.entries(mu)) {
|
||||
const bucket = SUB_TIER.test(model) ? sub : main; // sonnet/anything-else => main
|
||||
bucket.cost += u.costUSD || 0;
|
||||
bucket.tok += toks(u);
|
||||
}
|
||||
return {
|
||||
main, sub, subagents: subPids.size, agentCalls,
|
||||
ccTotal: result?.total_cost_usd ?? null,
|
||||
ok: result?.subtype === 'success',
|
||||
durationSec: result?.duration_ms ? +(result.duration_ms/1000).toFixed(1) : null,
|
||||
models: Object.keys(mu), tools,
|
||||
};
|
||||
}
|
||||
|
||||
const k = (n) => (n/1000).toFixed(0).padStart(5) + 'K';
|
||||
const d = (n) => '$' + n.toFixed(3);
|
||||
const cost = (b) => b.cost;
|
||||
const tot = (b) => b.tok;
|
||||
|
||||
const byArm = {};
|
||||
for (const arm of ARMS) {
|
||||
const runs = [];
|
||||
for (let r = 1; r <= REPS; r++) {
|
||||
const f = `${runsDir}/${repo}-${arm}-${r}.jsonl`;
|
||||
if (existsSync(f)) runs.push({ rep: r, ...analyzeRun(f) });
|
||||
}
|
||||
byArm[arm] = runs;
|
||||
}
|
||||
|
||||
// Per-run detail. Cost is Claude Code's own modelUsage.costUSD (authoritative,
|
||||
// per-model pricing + correct cache TTL). MAIN=Sonnet, SUB=Haiku Explore subagent.
|
||||
// cc-check: main$+sub$ must equal result.total_cost_usd (delta should be ~0).
|
||||
console.log(`\n=== ${repo}: per-run main(Sonnet)/sub(Haiku) split — Claude Code's own cost accounting ===`);
|
||||
console.log('arm rep | subAg | MAIN(sonnet) tok / $ | SUB(haiku) tok / $ | TOTAL tok / $ | cc_total Δ | dur reads');
|
||||
for (const arm of ARMS) for (const r of byArm[arm]) {
|
||||
const mC = cost(r.main), sC = cost(r.sub), mT = tot(r.main), sT = tot(r.sub);
|
||||
const reads = r.tools['Read'] || 0, grep = (r.tools['Grep']||0)+(r.tools['Bash']||0)+(r.tools['Glob']||0);
|
||||
const explore = r.tools['mcp__codegraph__codegraph_explore'] || 0;
|
||||
const delta = (mC + sC) - (r.ccTotal || 0); // should be ~0
|
||||
console.log(
|
||||
`${arm.padEnd(8)} #${r.rep} | ${String(r.subagents).padStart(2)} | ${k(mT)} ${d(mC).padStart(7)} | ${k(sT)} ${d(sC).padStart(7)} | ${k(mT+sT)} ${d(mC+sC).padStart(7)} | ${d(r.ccTotal||0).padStart(7)} ${(delta>=0?'+':'')+delta.toFixed(4)} | ${String(r.durationSec).padStart(5)} r=${reads} g=${grep} x=${explore}`
|
||||
);
|
||||
}
|
||||
|
||||
// Per-arm means
|
||||
const mean = (arr, f) => arr.length ? arr.reduce((s,x)=>s+f(x),0)/arr.length : 0;
|
||||
console.log(`\n=== ${repo}: per-arm MEANS (n per arm) ===`);
|
||||
console.log('arm n | main $ sub $ TOTAL $ | main tok sub tok TOTAL tok | %$ in sub | %tok in sub');
|
||||
for (const arm of ARMS) {
|
||||
const runs = byArm[arm]; if (!runs.length) continue;
|
||||
const mC = mean(runs, r=>cost(r.main)), sC = mean(runs, r=>cost(r.sub));
|
||||
const mT = mean(runs, r=>tot(r.main)), sT = mean(runs, r=>tot(r.sub));
|
||||
const pctSubC = (mC+sC) ? (100*sC/(mC+sC)) : 0;
|
||||
const pctSubT = (mT+sT) ? (100*sT/(mT+sT)) : 0;
|
||||
console.log(
|
||||
`${arm.padEnd(8)} ${runs.length} | ${d(mC).padStart(7)} ${d(sC).padStart(7)} ${d(mC+sC).padStart(7)} | ${k(mT)} ${k(sT)} ${k(mT+sT)} | ${pctSubC.toFixed(0).padStart(3)}% | ${pctSubT.toFixed(0).padStart(3)}%`
|
||||
);
|
||||
}
|
||||
|
||||
// Headline ladders — cost, tokens, duration, all vs a baseline (nocg if present, else first arm).
|
||||
console.log(`\n=== Ladders (mean, incl. subagent) ===`);
|
||||
const totals = ARMS.map(a => ({ a, c: mean(byArm[a], r=>cost(r.main)+cost(r.sub)), t: mean(byArm[a], r=>tot(r.main)+tot(r.sub)) })).filter(x=>byArm[x.a].length);
|
||||
const base = totals.find(x=>x.a==='nocg') ?? totals[0];
|
||||
const bn = base?.a ?? '?';
|
||||
console.log(` COST (vs ${bn}):`);
|
||||
for (const x of totals) {
|
||||
const vs = base && base.c ? ` (${((x.c/base.c-1)*100>=0?'+':'')}${((x.c/base.c-1)*100).toFixed(0)}%)` : '';
|
||||
console.log(` ${x.a.padEnd(8)} ${d(x.c)}${vs}`);
|
||||
}
|
||||
console.log(` TOKENS (vs ${bn}):`);
|
||||
for (const x of totals) {
|
||||
const vs = base && base.t ? ` (${((x.t/base.t-1)*100>=0?'+':'')}${((x.t/base.t-1)*100).toFixed(0)}%)` : '';
|
||||
console.log(` ${x.a.padEnd(8)} ${k(x.t)}${vs}`);
|
||||
}
|
||||
console.log(` DURATION (wall-clock, vs ${bn}):`);
|
||||
const durs = ARMS.map(a => ({ a, s: mean(byArm[a].filter(r=>r.durationSec!=null), r=>r.durationSec) })).filter(x=>byArm[x.a].length);
|
||||
const dbase = durs.find(x=>x.a==='nocg') ?? durs[0];
|
||||
for (const x of durs) {
|
||||
const vs = dbase && dbase.s ? ` (${((x.s/dbase.s-1)*100>=0?'+':'')}${((x.s/dbase.s-1)*100).toFixed(0)}%)` : '';
|
||||
console.log(` ${x.a.padEnd(8)} ${x.s.toFixed(0)}s${vs}`);
|
||||
}
|
||||
Executable
+72
@@ -0,0 +1,72 @@
|
||||
#!/usr/bin/env bash
|
||||
# Offload reasoning-OUTPUT-STYLE A/B — all codegraph-on, isolating the Worker's
|
||||
# output shape's effect on main-session tokens / latency / accuracy:
|
||||
# raw : CODEGRAPH_OFFLOAD_DISABLE=1 (verbatim explore source, the floor)
|
||||
# refs : managed offload, default (Cerebras map re-expanded to verbatim, ~24K)
|
||||
# map : managed offload, STYLE=map (compact reasoned map + file:line anchors, ~1-3K)
|
||||
# src : managed offload, STYLE=src (map + cited line ranges only, ~1-5K)
|
||||
# Delegation BLOCKED by default (DISALLOW=Agent) so we measure the offload payload's
|
||||
# effect on the main Sonnet agent, not whether it spawns a Haiku Explore subagent.
|
||||
#
|
||||
# Usage: offload-eval-styles.sh <indexed-repo> <reps> "<question>"
|
||||
# Env: RESULTS=<file> AGENT_EVAL_OUT=<dir> REP_START=1 DISALLOW=Agent MODEL/EFFORT
|
||||
set -uo pipefail
|
||||
HERE="$(cd "$(dirname "$0")" && pwd)"
|
||||
ENGINE="$(cd "$HERE/../.." && pwd)"
|
||||
BIN="$ENGINE/dist/bin/codegraph.js"
|
||||
OUT="${AGENT_EVAL_OUT:-/tmp/cg-offload-eval}"
|
||||
TARGET="${1:?usage: offload-eval-styles.sh <indexed-repo> <reps> \"<question>\"}"
|
||||
REPS="${2:?reps}"; Q="${3:?question}"
|
||||
RUNS="$OUT/runs"; EXTRACT="$HERE/offload-eval-metrics.mjs"
|
||||
RESULTS="${RESULTS:-$OUT/results-styles.jsonl}"
|
||||
REPO=$(basename "$TARGET")
|
||||
DISALLOW="${DISALLOW-Agent}" # default: block delegation. `DISALLOW= ` to allow.
|
||||
START="${REP_START:-1}"; END=$((START + REPS - 1))
|
||||
mkdir -p "$RUNS"
|
||||
command -v claude >/dev/null || { echo "no claude on PATH"; exit 1; }
|
||||
[ -d "$TARGET/.codegraph" ] || { echo "not indexed: $TARGET"; exit 1; }
|
||||
TARGET=$(cd "$TARGET" && pwd -P)
|
||||
|
||||
prewarm() { # path extra-env
|
||||
pkill -9 -f "serve --mcp --path $1" 2>/dev/null; rm -f "$1/.codegraph/daemon.sock" 2>/dev/null; sleep 0.6
|
||||
env ${2:-} CODEGRAPH_DAEMON_IDLE_TIMEOUT_MS=1800000 node "$BIN" serve --mcp --path "$1" </dev/null >/dev/null 2>&1 &
|
||||
node -e 'const fs=require("fs");let n=0;const t=setInterval(()=>{if(fs.existsSync(process.argv[1]+"/.codegraph/daemon.sock")){clearInterval(t);process.exit(0)}if(n++>150){clearInterval(t);process.exit(1)}},100)' "$1" \
|
||||
&& echo " daemon warm" || echo " WARN daemon never bound"
|
||||
}
|
||||
kill_daemon() { pkill -9 -f "serve --mcp --path $TARGET" 2>/dev/null; rm -f "$TARGET/.codegraph/daemon.sock" 2>/dev/null; sleep 1; }
|
||||
|
||||
run() { # arm rep mcp-config usage-log-or-dash
|
||||
local arm="$1" rep="$2" cfg="$3" usage="$4" tag="$REPO-$1-$2"
|
||||
[ "$usage" != "-" ] && : > "$usage"
|
||||
( cd "$TARGET" && claude -p "$Q" \
|
||||
--output-format stream-json --verbose --permission-mode bypassPermissions \
|
||||
--model "${MODEL:-sonnet}" --effort "${EFFORT:-high}" --max-budget-usd 4 \
|
||||
${DISALLOW:+--disallowedTools "$DISALLOW"} \
|
||||
--strict-mcp-config --mcp-config "$cfg" \
|
||||
</dev/null > "$RUNS/$tag.jsonl" 2>"$RUNS/$tag.err" )
|
||||
node "$EXTRACT" --run "$RUNS/$tag.jsonl" --usage "$usage" --arm "$arm" --rep "$rep" \
|
||||
--repo "$REPO" --tier styles --q "$Q" >> "$RESULTS"
|
||||
node -e 'const o=JSON.parse(require("fs").readFileSync(process.argv[1],"utf8").trim().split("\n").pop());console.log(` [${o.arm} #${o.rep}] ${o.durationSec}s | ${o.tokBillable} billable tok | read=${o.read} grep=${o.grep} explore=${o.explore} offload=${o.offloadFired} | AI ${o.ai.calls}c/${o.ai.totalTokens}t | ok=${o.ok}`)' "$RESULTS"
|
||||
}
|
||||
|
||||
# MCP configs: env baked into the daemon-spawn command claude uses.
|
||||
USAGE="$RUNS/$REPO-usage.jsonl"
|
||||
mkcfg() { # file extra-env-pairs(JSON array entries, comma-led or empty)
|
||||
printf '{"mcpServers":{"codegraph":{"command":"env","args":["CODEGRAPH_WASM_RELAUNCHED=1"%s,"node","%s","serve","--mcp","--path","%s"]}}}' "$1" "$BIN" "$TARGET"
|
||||
}
|
||||
CFG_RAW="$RUNS/mcp-sty-raw-$REPO.json"; mkcfg ',"CODEGRAPH_OFFLOAD_DISABLE=1"' > "$CFG_RAW"
|
||||
CFG_REFS="$RUNS/mcp-sty-refs-$REPO.json"; mkcfg ",\"CODEGRAPH_OFFLOAD_USAGE_LOG=$USAGE\"" > "$CFG_REFS"
|
||||
CFG_MAP="$RUNS/mcp-sty-map-$REPO.json"; mkcfg ",\"CODEGRAPH_OFFLOAD_USAGE_LOG=$USAGE\",\"CODEGRAPH_OFFLOAD_STYLE=map\"" > "$CFG_MAP"
|
||||
CFG_SRC="$RUNS/mcp-sty-src-$REPO.json"; mkcfg ",\"CODEGRAPH_OFFLOAD_USAGE_LOG=$USAGE\",\"CODEGRAPH_OFFLOAD_STYLE=src\"" > "$CFG_SRC"
|
||||
|
||||
echo "###### repo=$REPO reps=$START..$END model=${MODEL:-sonnet}/${EFFORT:-high} disallow=${DISALLOW:-<none>}"
|
||||
echo "###### Q=$Q"
|
||||
echo "== ARM raw =="; prewarm "$TARGET" "CODEGRAPH_OFFLOAD_DISABLE=1"
|
||||
for r in $(seq "$START" "$END"); do run raw "$r" "$CFG_RAW" "-"; done; kill_daemon
|
||||
echo "== ARM refs =="; prewarm "$TARGET" "CODEGRAPH_OFFLOAD_USAGE_LOG=$USAGE"
|
||||
for r in $(seq "$START" "$END"); do run refs "$r" "$CFG_REFS" "$USAGE"; done; kill_daemon
|
||||
echo "== ARM map =="; prewarm "$TARGET" "CODEGRAPH_OFFLOAD_USAGE_LOG=$USAGE CODEGRAPH_OFFLOAD_STYLE=map"
|
||||
for r in $(seq "$START" "$END"); do run map "$r" "$CFG_MAP" "$USAGE"; done; kill_daemon
|
||||
echo "== ARM src =="; prewarm "$TARGET" "CODEGRAPH_OFFLOAD_USAGE_LOG=$USAGE CODEGRAPH_OFFLOAD_STYLE=src"
|
||||
for r in $(seq "$START" "$END"); do run src "$r" "$CFG_SRC" "$USAGE"; done; kill_daemon
|
||||
echo "###### DONE $REPO — judge: node $HERE/offload-eval-judge.mjs --results $RESULTS --truth $HERE/offload-eval-ground-truth.json --out $OUT/judged-styles.jsonl"
|
||||
Reference in New Issue
Block a user