Skip to content

fix(plugin-eval): load the eval handoff classifier under its own package scope #13636

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
kyle-sexton merged 2 commits into from
Nov 6, 2026
Merged

fix(plugin-eval): load the eval handoff classifier under its own package scope #13636

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
2823bd9
fix(plugin-eval): load the eval handoff classifier under its own pack…
kyle-sexton Nov 6, 2026
3000b31
Merge branch 'main' into fix/eval-handoff-package-scope
kyle-sexton Nov 6, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 10 additions & 0 deletions .github/actions/resolve-config/callers.test.mjs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9123,6 +9123,16 @@ for (const [file, gateId, [keyId, subId, fbId], marker] of [
assert.match(step.env.CLAUDE_CODE_SUBPROCESS_ENV_SCRUB, /^(?:"?1"?|\$\{\{ .+ \}\})$/, `${id} ${step.name} scrubs Bash subprocesses`);
}
assert.equal(steps.find((s) => s.id === "judge").env.LEG, leg, id);
// A leg whose credential is empty, or whose key the gate skipped, runs
// nothing past its credential check: every later step but the budget
// log has a condition, and the first ones read `run`.
const after = steps.slice(steps.findIndex((s) => s.id === "credential") + 1);
assert.deepEqual(after.filter((s) => s.if === undefined || s.if === "always()").map((s) => s.name).filter((n) => n !== "Log the API budget at the end"), [], id);
assert.match(after[0].if, /^steps\.credential\.outputs\.run == 'true'/, id);
// A classifier that cannot run stops the leg red, never as a handoff.
for (const step of steps.filter((s) => /plugin-eval\/eval-handoff\.mjs/.test(String(s.run ?? "")))) {
assert.match(step.run, /eval-handoff\.mjs\) \|\|\n\s+\{ echo "::error::the handoff classifier failed[^"]*"; exit 1; \}/, `${id} ${step.name}`);
}
const checkout = steps.find((s) => s.with?.repository === "melodic-software/ci-workflows");
assert.equal(checkout.with.ref, CLASSIFIER_PIN, `${id} reads the classifier at the run-claude-cli pin`);
assert.equal(checkout.with["persist-credentials"], false);
Expand Down
33 changes: 28 additions & 5 deletions .github/scripts/plugin-eval/eval-handoff.mjs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,7 @@
// one result message carrying the suites' summed durationSeconds and costUsd,
// nothing a run wrote).
// Prints handoff, handoff-reason, lane-class and outcome as key=value lines.
import { existsSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from "node:fs";
import { copyFileSync, existsSync, mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from "node:fs";
import { createRequire } from "node:module";
import { tmpdir } from "node:os";
import { dirname, isAbsolute, join } from "node:path";
Expand Down Expand Up @@ -103,13 +103,36 @@ function readJson(path) {
}
}

// The classifier's three files, copied under a folder of their own whose
// package.json ends Node's package-scope lookup there. Required in place, the
// lookup climbs past the sparse checkout to the workspace root, and a
// package.json there that does not parse stops the require (run 42138674762).
export const CLASSIFIER_FILES = [
".github/actions/run-claude-cli/lane.cjs",
".github/actions/report-lane-outcome/classify.cjs",
".github/scripts/resolve-npm-tool-version.cjs",
];
export function loadLauncher(root) {
const dir = mkdtempSync(join(tmpdir(), "eval-handoff-classifier-"));
try {
writeFileSync(join(dir, "package.json"), '{"type":"commonjs"}\n');
for (const path of CLASSIFIER_FILES) {
mkdirSync(dirname(join(dir, path)), { recursive: true });
copyFileSync(join(root, path), join(dir, path));
}
const require = createRequire(join(dir, CLASSIFIER_FILES[0]));
const { judgeLeg, linesToMessages } = require("./lane.cjs");
const { isCreditStop } = require("../report-lane-outcome/classify.cjs");
return { judgeLeg, linesToMessages, isCreditStop };
} finally {
rmSync(dir, { recursive: true, force: true });
}
}

function main(env) {
const leg = LEGS[env.LEG];
if (!leg) throw new Error(`LEG must be key, subscription or fallback, not ${JSON.stringify(env.LEG)}`);
const root = env.CI_WORKFLOWS ?? "";
const require = createRequire(join(root, ".github/actions/run-claude-cli/lane.cjs"));
const { judgeLeg, linesToMessages } = require("./lane.cjs");
const { isCreditStop } = require("../report-lane-outcome/classify.cjs");
const { judgeLeg, linesToMessages, isCreditStop } = loadLauncher(env.CI_WORKFLOWS ?? "");
const streams = [];
if (env.PROBE && existsSync(env.PROBE)) {
const exitCode = Number.parseInt(env.PROBE_EXIT ?? "", 10);
Expand Down
54 changes: 45 additions & 9 deletions .github/scripts/plugin-eval/eval-handoff.test.mjs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,11 +9,10 @@
// records (CLI 2.1.302).
import assert from "node:assert/strict";
import { mkdirSync, readFileSync, writeFileSync } from "node:fs";
import { createRequire } from "node:module";
import { join } from "node:path";
import { test } from "node:test";
import { tempDir } from "../../../lib/test-tmp.mjs";
import { erroredTraces, handoffReason, judgeLegStreams } from "./eval-handoff.mjs";
import { CLASSIFIER_FILES, erroredTraces, handoffReason, judgeLegStreams, loadLauncher } from "./eval-handoff.mjs";

test("a key leg hands off on a credit stop or a 401 or 402, never a 403 or a limit stop", () => {
assert.equal(handoffReason("api-key", "credit-stop", 400), "credit-stop");
Expand Down Expand Up @@ -66,7 +65,6 @@ test("the first stream that hands off decides, else the first stop class is kept
// a pin bump that reshapes judgeLeg's return (failureClass, resets, the last
// message's api_error_status) or isCreditStop fails here. The three paths are
// the legs' sparse checkout.
const CLASSIFIER_FILES = [".github/actions/run-claude-cli/lane.cjs", ".github/actions/report-lane-outcome/classify.cjs", ".github/scripts/resolve-npm-tool-version.cjs"];
const workflowPins = ["eval-hill-climb-pass.yml", "eval-browser-dispatch.yml"].flatMap((name) => [
...readFileSync(new URL(`../../workflows/${name}`, import.meta.url), "utf8").matchAll(/repository: melodic-software\/ci-workflows\n\s+ref: ([0-9a-f]{40})/g),
].map((m) => m[1]));
Expand All @@ -79,12 +77,50 @@ for (const path of CLASSIFIER_FILES) {
mkdirSync(join(ROOT, path, ".."), { recursive: true });
writeFileSync(join(ROOT, path), await response.text());
}
const real = (() => {
const require = createRequire(join(ROOT, ".github/actions/run-claude-cli/lane.cjs"));
const { judgeLeg, linesToMessages } = require("./lane.cjs");
const { isCreditStop } = require("../report-lane-outcome/classify.cjs");
return { judgeLeg, linesToMessages, isCreditStop };
})();
const real = loadLauncher(ROOT);

test("the files eval-handoff loads are the eval legs' sparse checkout", () => {
for (const name of ["eval-hill-climb-pass.yml", "eval-browser-dispatch.yml"]) {
const sparse = [...readFileSync(new URL(`../../workflows/${name}`, import.meta.url), "utf8").matchAll(/sparse-checkout: \|\n((?: {12}\S+\n)+)/g)]
.filter(([block]) => block.includes("run-claude-cli/lane.cjs"))
.map(([, block]) => block.trim().split(/\s+/));
assert.equal(sparse.length, 3, name);
for (const paths of sparse) assert.deepEqual(paths, CLASSIFIER_FILES, name);
}
});

// Run 42138674762: the classifier required in place climbed to the workspace
// root's package.json, which did not parse, and the leg failed before judging.
test("the classifier loads under a workspace whose root package.json does not parse", async () => {
const { spawnSync } = await import("node:child_process");
const ws = tempDir("eval-handoff-ws-");
writeFileSync(join(ws, "package.json"), "not json {");
for (const path of CLASSIFIER_FILES) {
mkdirSync(join(ws, ".ci-workflows", path, ".."), { recursive: true });
writeFileSync(join(ws, ".ci-workflows", path), readFileSync(join(ROOT, path)));
}
writeFileSync(join(ws, "probe.jsonl"), jsonl(SPEND_LIMIT));
const r = spawnSync(process.execPath, [new URL("./eval-handoff.mjs", import.meta.url).pathname], {
cwd: ws,
encoding: "utf8",
env: { PATH: process.env.PATH, LEG: "key", CI_WORKFLOWS: join(ws, ".ci-workflows"), PROBE: join(ws, "probe.jsonl"), PROBE_EXIT: "1", RESULTS: "" },
});
assert.equal(r.status, 0, r.stderr);
assert.equal(r.stdout, "handoff=true\nhandoff-reason=credit-stop\nlane-class=credit-stop\noutcome=failure\n");
});

// A classifier that cannot load fails the run with exit 2 and no verdict, so
// the workflow step stops the leg red rather than guessing a handoff.
test("a classifier that cannot load prints no verdict and exits 2", async () => {
const { spawnSync } = await import("node:child_process");
const r = spawnSync(process.execPath, [new URL("./eval-handoff.mjs", import.meta.url).pathname], {
encoding: "utf8",
env: { PATH: process.env.PATH, LEG: "key", CI_WORKFLOWS: tempDir("eval-handoff-empty-"), RESULTS: "" },
});
assert.equal(r.status, 2);
assert.equal(r.stdout, "");
assert.match(r.stderr, /^eval-handoff: /);
});

test("every eval leg checks out the classifier at one pin", () => {
assert.equal(workflowPins.length, 6, "three legs in each eval workflow");
Expand Down
18 changes: 12 additions & 6 deletions .github/workflows/eval-browser-dispatch.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -594,7 +594,8 @@ jobs:
echo "::error::the probe's result holds the credential, so no pull request case runs"
exit 1
fi
verdict=$(env -u ANTHROPIC_API_KEY -u CLAUDE_CODE_OAUTH_TOKEN PROBE="$out" PROBE_EXIT="$status" node .base/.github/scripts/plugin-eval/eval-handoff.mjs)
verdict=$(env -u ANTHROPIC_API_KEY -u CLAUDE_CODE_OAUTH_TOKEN PROBE="$out" PROBE_EXIT="$status" node .base/.github/scripts/plugin-eval/eval-handoff.mjs) ||
{ echo "::error::the handoff classifier failed, so this leg cannot tell its credential's stop from a case failure and stops"; exit 1; }
printf '%s\n' "$verdict" >>"$GITHUB_OUTPUT"
if grep -qx 'handoff=true' <<<"$verdict"; then
echo "::notice::the probe ended on the credential's own stop; the next leg runs the eval"
Expand Down Expand Up @@ -768,7 +769,8 @@ jobs:
[ -n "$dir" ] || continue
results+="$OUT/$(basename "$dir")/results.json"$'\n'
done <<<"$SUITES"
verdict=$(RESULTS="$results" node .base/.github/scripts/plugin-eval/eval-handoff.mjs)
verdict=$(RESULTS="$results" node .base/.github/scripts/plugin-eval/eval-handoff.mjs) ||
{ echo "::error::the handoff classifier failed, so this leg cannot tell its credential's stop from a case failure and stops"; exit 1; }
if [ "$PROBE_HANDOFF" = true ]; then
verdict=$(printf 'handoff=true\nhandoff-reason=%s\nlane-class=%s\noutcome=failure' "$PROBE_REASON" "$PROBE_CLASS")
fi
Expand Down Expand Up @@ -1397,7 +1399,8 @@ jobs:
echo "::error::the probe's result holds the credential, so no pull request case runs"
exit 1
fi
verdict=$(env -u ANTHROPIC_API_KEY -u CLAUDE_CODE_OAUTH_TOKEN PROBE="$out" PROBE_EXIT="$status" node .base/.github/scripts/plugin-eval/eval-handoff.mjs)
verdict=$(env -u ANTHROPIC_API_KEY -u CLAUDE_CODE_OAUTH_TOKEN PROBE="$out" PROBE_EXIT="$status" node .base/.github/scripts/plugin-eval/eval-handoff.mjs) ||
{ echo "::error::the handoff classifier failed, so this leg cannot tell its credential's stop from a case failure and stops"; exit 1; }
printf '%s\n' "$verdict" >>"$GITHUB_OUTPUT"
if grep -qx 'handoff=true' <<<"$verdict"; then
echo "::notice::the probe ended on the credential's own stop; the next leg runs the eval"
Expand Down Expand Up @@ -1571,7 +1574,8 @@ jobs:
[ -n "$dir" ] || continue
results+="$OUT/$(basename "$dir")/results.json"$'\n'
done <<<"$SUITES"
verdict=$(RESULTS="$results" node .base/.github/scripts/plugin-eval/eval-handoff.mjs)
verdict=$(RESULTS="$results" node .base/.github/scripts/plugin-eval/eval-handoff.mjs) ||
{ echo "::error::the handoff classifier failed, so this leg cannot tell its credential's stop from a case failure and stops"; exit 1; }
if [ "$PROBE_HANDOFF" = true ]; then
verdict=$(printf 'handoff=true\nhandoff-reason=%s\nlane-class=%s\noutcome=failure' "$PROBE_REASON" "$PROBE_CLASS")
fi
Expand Down Expand Up @@ -2154,7 +2158,8 @@ jobs:
echo "::error::the probe's result holds the credential, so no pull request case runs"
exit 1
fi
verdict=$(env -u ANTHROPIC_API_KEY -u CLAUDE_CODE_OAUTH_TOKEN PROBE="$out" PROBE_EXIT="$status" node .base/.github/scripts/plugin-eval/eval-handoff.mjs)
verdict=$(env -u ANTHROPIC_API_KEY -u CLAUDE_CODE_OAUTH_TOKEN PROBE="$out" PROBE_EXIT="$status" node .base/.github/scripts/plugin-eval/eval-handoff.mjs) ||
{ echo "::error::the handoff classifier failed, so this leg cannot tell its credential's stop from a case failure and stops"; exit 1; }
printf '%s\n' "$verdict" >>"$GITHUB_OUTPUT"
if grep -qx 'handoff=true' <<<"$verdict"; then
echo "::notice::the probe ended on the credential's own stop; the next leg runs the eval"
Expand Down Expand Up @@ -2328,7 +2333,8 @@ jobs:
[ -n "$dir" ] || continue
results+="$OUT/$(basename "$dir")/results.json"$'\n'
done <<<"$SUITES"
verdict=$(RESULTS="$results" node .base/.github/scripts/plugin-eval/eval-handoff.mjs)
verdict=$(RESULTS="$results" node .base/.github/scripts/plugin-eval/eval-handoff.mjs) ||
{ echo "::error::the handoff classifier failed, so this leg cannot tell its credential's stop from a case failure and stops"; exit 1; }
if [ "$PROBE_HANDOFF" = true ]; then
verdict=$(printf 'handoff=true\nhandoff-reason=%s\nlane-class=%s\noutcome=failure' "$PROBE_REASON" "$PROBE_CLASS")
fi
Expand Down
18 changes: 12 additions & 6 deletions .github/workflows/eval-hill-climb-pass.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -347,7 +347,8 @@ jobs:
echo "::error::the probe's result holds the credential, so no case runs"
exit 1
fi
verdict=$(env -u ANTHROPIC_API_KEY -u CLAUDE_CODE_OAUTH_TOKEN PROBE="$out" PROBE_EXIT="$status" node .base/.github/scripts/plugin-eval/eval-handoff.mjs)
verdict=$(env -u ANTHROPIC_API_KEY -u CLAUDE_CODE_OAUTH_TOKEN PROBE="$out" PROBE_EXIT="$status" node .base/.github/scripts/plugin-eval/eval-handoff.mjs) ||
{ echo "::error::the handoff classifier failed, so this leg cannot tell its credential's stop from a case failure and stops"; exit 1; }
printf '%s\n' "$verdict" >>"$GITHUB_OUTPUT"
if grep -qx 'handoff=true' <<<"$verdict"; then
echo "::notice::the probe ended on the credential's own stop; the next leg runs the pass"
Expand Down Expand Up @@ -411,7 +412,8 @@ jobs:
PROBE_REASON: ${{ steps.probe.outputs.handoff-reason }}
PROBE_CLASS: ${{ steps.probe.outputs.lane-class }}
run: |
verdict=$(node .base/.github/scripts/plugin-eval/eval-handoff.mjs)
verdict=$(node .base/.github/scripts/plugin-eval/eval-handoff.mjs) ||
{ echo "::error::the handoff classifier failed, so this leg cannot tell its credential's stop from a case failure and stops"; exit 1; }
if [ "$PROBE_HANDOFF" = true ]; then
verdict=$(printf 'handoff=true\nhandoff-reason=%s\nlane-class=%s\noutcome=failure' "$PROBE_REASON" "$PROBE_CLASS")
fi
Expand Down Expand Up @@ -738,7 +740,8 @@ jobs:
echo "::error::the probe's result holds the credential, so no case runs"
exit 1
fi
verdict=$(env -u ANTHROPIC_API_KEY -u CLAUDE_CODE_OAUTH_TOKEN PROBE="$out" PROBE_EXIT="$status" node .base/.github/scripts/plugin-eval/eval-handoff.mjs)
verdict=$(env -u ANTHROPIC_API_KEY -u CLAUDE_CODE_OAUTH_TOKEN PROBE="$out" PROBE_EXIT="$status" node .base/.github/scripts/plugin-eval/eval-handoff.mjs) ||
{ echo "::error::the handoff classifier failed, so this leg cannot tell its credential's stop from a case failure and stops"; exit 1; }
printf '%s\n' "$verdict" >>"$GITHUB_OUTPUT"
if grep -qx 'handoff=true' <<<"$verdict"; then
echo "::notice::the probe ended on the credential's own stop; the next leg runs the pass"
Expand Down Expand Up @@ -802,7 +805,8 @@ jobs:
PROBE_REASON: ${{ steps.probe.outputs.handoff-reason }}
PROBE_CLASS: ${{ steps.probe.outputs.lane-class }}
run: |
verdict=$(node .base/.github/scripts/plugin-eval/eval-handoff.mjs)
verdict=$(node .base/.github/scripts/plugin-eval/eval-handoff.mjs) ||
{ echo "::error::the handoff classifier failed, so this leg cannot tell its credential's stop from a case failure and stops"; exit 1; }
if [ "$PROBE_HANDOFF" = true ]; then
verdict=$(printf 'handoff=true\nhandoff-reason=%s\nlane-class=%s\noutcome=failure' "$PROBE_REASON" "$PROBE_CLASS")
fi
Expand Down Expand Up @@ -1084,7 +1088,8 @@ jobs:
echo "::error::the probe's result holds the credential, so no case runs"
exit 1
fi
verdict=$(env -u ANTHROPIC_API_KEY -u CLAUDE_CODE_OAUTH_TOKEN PROBE="$out" PROBE_EXIT="$status" node .base/.github/scripts/plugin-eval/eval-handoff.mjs)
verdict=$(env -u ANTHROPIC_API_KEY -u CLAUDE_CODE_OAUTH_TOKEN PROBE="$out" PROBE_EXIT="$status" node .base/.github/scripts/plugin-eval/eval-handoff.mjs) ||
{ echo "::error::the handoff classifier failed, so this leg cannot tell its credential's stop from a case failure and stops"; exit 1; }
printf '%s\n' "$verdict" >>"$GITHUB_OUTPUT"
if grep -qx 'handoff=true' <<<"$verdict"; then
echo "::notice::the probe ended on the credential's own stop; the next leg runs the pass"
Expand Down Expand Up @@ -1148,7 +1153,8 @@ jobs:
PROBE_REASON: ${{ steps.probe.outputs.handoff-reason }}
PROBE_CLASS: ${{ steps.probe.outputs.lane-class }}
run: |
verdict=$(node .base/.github/scripts/plugin-eval/eval-handoff.mjs)
verdict=$(node .base/.github/scripts/plugin-eval/eval-handoff.mjs) ||
{ echo "::error::the handoff classifier failed, so this leg cannot tell its credential's stop from a case failure and stops"; exit 1; }
if [ "$PROBE_HANDOFF" = true ]; then
verdict=$(printf 'handoff=true\nhandoff-reason=%s\nlane-class=%s\noutcome=failure' "$PROBE_REASON" "$PROBE_CLASS")
fi
Expand Down
Loading