Understand any codebase instantly. System intelligence for codebases, built for humans and AI.
TypeScript aiai-memoryclaude-codeclicode-analysiscode-mappingcode-memorycodebase-understandingcodexdeveloper-toolsllmopenclawpersistent-memoryprogram-analysissoftware-architecturesystem-intelligencesystem-mapping
fix(perf-gate): map the fixture at one fixed path so context bytes are stable
KageBinary3a6d74e8d7bb
diff --git a/ix-cli/scripts/perf-gate.mjs b/ix-cli/scripts/perf-gate.mjsindex 65cbb787..5720cb1a 100644--- a/ix-cli/scripts/perf-gate.mjs+++ b/ix-cli/scripts/perf-gate.mjs@@ -65,9 +65,23 @@ async function reset(base) {if (!res.ok) throw new Error(`POST /v1/reset: ${res.status}`);}+/**+ * Where every copy of the fixture is made: the same absolute path on every run+ * and every machine whose temp directory is /tmp, CI included.+ *+ * Node ids are hashed from the workspace's absolute path (workspaceIdForPath),+ * and context's fan-out follows the backend's order of a node's neighbours,+ * which follows those ids. A fresh random directory per run therefore drew a+ * different set of walk seeds each time, and `context` response bytes moved+ * by tens of percent between identical runs of the same commit.+ */+const CHECKOUT_ROOT = join(tmpdir(), "ix-perf-gate-checkout");+/** A fresh copy of the fixture (same directory name) and an empty IX_HOME. */-function checkout(fixture, scratch) {- const root = mkdtempSync(join(scratch, "run-"));+function checkout(fixture) {+ const root = CHECKOUT_ROOT;+ rmSync(root, { recursive: true, force: true });+ mkdirSync(root, { recursive: true });const dir = join(root, basename(fixture));cpSync(fixture, dir, { recursive: true });const ixHome = join(root, "ix-home");@@ -140,7 +154,7 @@ async function main() {let co = co0;if (sc.fresh) {await reset(base);- co = checkout(fixture, scratch);+ co = checkout(fixture);}sc.prepare?.(co, i);const r = measure(base, co, sc.args(co), log);@@ -162,13 +176,14 @@ async function main() {// Then one mapped copy shared by every other scenario, in order.await reset(base);- const shared = checkout(fixture, scratch);+ const shared = checkout(fixture);const first = measure(base, shared, ["map", shared.dir, "--silent"], log);if (first.status !== 0) throw new Error(`initial map failed (exit ${first.status}): ${first.stderr}`);for (const sc of all.filter((s) => !s.fresh)) await measureRuns(sc, shared);} finally {await reset(base).catch(() => {});rmSync(scratch, { recursive: true, force: true });+ rmSync(CHECKOUT_ROOT, { recursive: true, force: true });}if (opts.out) writeFileSync(opts.out, JSON.stringify(results, null, 2) + "\n");diff --git a/ix-cli/test/e2e/README.md b/ix-cli/test/e2e/README.mdindex f81f1f3d..9ab3d89d 100644--- a/ix-cli/test/e2e/README.md+++ b/ix-cli/test/e2e/README.md@@ -143,6 +143,13 @@ makes a command cheaper should lower its budget in the same PR (`--update`,then commit the file), so the gain can't be lost again unnoticed. In CI it is astep of job `e2e`, and its results are uploaded as `perf-results`.+Every copy of the fixture is made at one fixed path+(`$TMPDIR/ix-perf-gate-checkout/<fixture name>`). Node ids are hashed from the+workspace's absolute path, and `context` picks its walk seeds in the backend's+neighbour order, which follows those ids: from a random directory per run, its+response bytes moved by tens of percent between runs of the same commit. Run+one gate at a time per machine.+## Fixtures`fixtures/polyglot/` has 23 files: Python with cross-file calls (`app/`),diff --git a/ix-cli/test/perf/budgets.json b/ix-cli/test/perf/budgets.jsonindex e476781e..6e5dbc52 100644--- a/ix-cli/test/perf/budgets.json+++ b/ix-cli/test/perf/budgets.json@@ -22,14 +22,14 @@"wallMs": 14162},"context IxClient": {- "requests": 110,- "bytes": 5630913,+ "requests": 106,+ "bytes": 6221598,"wallMs": 2214,"tolerance": 1.25},"context --from-issue": {- "requests": 115,- "bytes": 2770336,+ "requests": 118,+ "bytes": 2959562,"wallMs": 1356,"tolerance": 1.25},