Fix agents

This commit is contained in:
gsxdsm
2026-06-11 12:22:02 -07:00
parent 2add48c8b6
commit 245e1280ed
3 changed files with 54 additions and 9 deletions

View File

@@ -112,6 +112,12 @@ pnpm verify:workspace # deep opt-in verification (lint -> test:full -> build);
Never kill processes on port 4040 and never start test servers on 4040. Use `--port 0` or another free port.
### Never run an unbounded `find` against the system temp directory
Do not issue a recursive `find` (or any unbounded recursive directory walk) rooted at the OS temp directory — `$TMPDIR`, `/tmp`, or macOS `/var/folders/...` (canonical `/private/var/...`). The temp root can hold an enormous number of entries on CI and long-lived dev hosts, so a broad scan can hang for minutes and pin I/O.
When you need a Fusion temp artifact, target the known prefix directly and list a single level with a prefix filter — never walk the whole temp tree. The canonical bounded pattern is the engine's own sweep: a non-recursive `readdirSync(tmpdir())` filtered by a known prefix such as `fusion-ai-merge-` (`SelfHealingManager.cleanupStaleTempMergeWorktrees()` in `packages/engine/src/self-healing.ts`). Scoped `find` calls under a project worktree or `.fusion/` are fine; only the broad temp-root scan is forbidden.
### Engine Process Rules
#### Never use `execSync` for user-configured commands

View File

@@ -45,6 +45,7 @@ const {
mkdtempSync,
mkdirSync,
readFileSync,
readdirSync,
rmSync,
realpathSync,
existsSync,
@@ -192,20 +193,32 @@ function isProcessAlive(pid: number): boolean {
}
}
function removeTmpdirRedirectSinkForPid(ownerPid: number): void {
try {
rmSync(join(WORKER_ROOT, `redir-${ownerPid}`), { recursive: true, force: true });
} catch {
// Ignore stale-sink cleanup failures; global teardown still owns WORKER_ROOT.
}
}
function sweepDeadTmpdirRedirectSinks(): void {
if (tmpdirRedirectSweepComplete) return;
tmpdirRedirectSweepComplete = true;
let ownerPids: number[];
// Registry-backed cleanup avoids scanning the OS temp root while still
// reclaiming redirect sinks from fork-pool workers that were hard-killed.
let ownerPids: number[] = [];
try {
ownerPids = Array.from(new Set(
readFileSync(TMPDIR_REDIRECT_REGISTRY, "utf8")
.split(/\r?\n/)
.split(/
?
/)
.map((line) => Number.parseInt(line, 10))
.filter((pid) => Number.isInteger(pid) && pid > 0),
));
return;
} catch {
// The registry may not exist yet. The bounded WORKER_ROOT sweep below still
// catches legacy redirect dirs created before the registry was introduced.
}
@@ -215,15 +228,30 @@ function sweepDeadTmpdirRedirectSinks(): void {
liveOwnerPids.push(ownerPid);
continue;
}
try {
rmSync(join(WORKER_ROOT, `redir-${ownerPid}`), { recursive: true, force: true });
} catch {
// Ignore stale-sink cleanup failures; global teardown still owns WORKER_ROOT.
removeTmpdirRedirectSinkForPid(ownerPid);
}
// Preserve the local self-healing behavior for redirect dirs that predate the
// registry or whose registry append was skipped. This is a single-level scan
// of WORKER_ROOT (not the OS temp root) and only touches dead pid-owned dirs.
try {
for (const entry of readdirSync(WORKER_ROOT)) {
const match = /^redir-(\d+)$/.exec(entry);
if (!match) continue;
const ownerPid = Number.parseInt(match[1], 10);
if (ownerPid === process.pid || liveOwnerPids.includes(ownerPid) || isProcessAlive(ownerPid)) {
continue;
}
removeTmpdirRedirectSinkForPid(ownerPid);
}
} catch {
// Best-effort only; stale entries are harmless and swept by future workers.
}
writeFileSync(TMPDIR_REDIRECT_REGISTRY, liveOwnerPids.length > 0 ? `${liveOwnerPids.join("\n")}\n` : "");
try {
writeFileSync(TMPDIR_REDIRECT_REGISTRY, liveOwnerPids.length > 0 ? `${liveOwnerPids.join("
")}
` : "");
} catch {
// Best-effort only; stale entries are harmless and swept by future workers.
@@ -236,7 +264,8 @@ function ensureTmpdirRedirectSink(): string {
sweepDeadTmpdirRedirectSinks();
const sink = join(WORKER_ROOT, `redir-${process.pid}`);
mkdirSync(sink, { recursive: true });
appendFileSync(TMPDIR_REDIRECT_REGISTRY, `${process.pid}\n`);
try {
appendFileSync(TMPDIR_REDIRECT_REGISTRY, `${process.pid}
`);
} catch {
// Best-effort only; the process exit hook and global teardown still clean up.
@@ -256,6 +285,11 @@ function ensureTmpdirRedirectSink(): string {
return sink;
}
/**
* If a mkdtemp prefix points straight at the OS temp root, rewrite it into a
* swept per-process sink under WORKER_ROOT. Prefixes already nested under a
* subdirectory pass through unchanged, as do non-string prefixes (Buffer/URL).
*/
function redirectTmpdirPrefix<T>(prefix: T): T {
if (typeof prefix !== "string") return prefix;
@@ -264,6 +298,7 @@ function redirectTmpdirPrefix<T>(prefix: T): T {
if (parent !== tmpdir() && parent !== REAL_TMPDIR) return prefix;
return join(ensureTmpdirRedirectSink(), basename(prefix)) as T;
}
}
function ensureIsolatedHome(): void {

View File

@@ -11,6 +11,10 @@
"reason": "Flake: vi.mock('node:child_process') occasionally doesn't take under workspace-concurrent runs, letting real git binary leak and report staged files unrelated to test scope (trips FileScopeViolationError). Same logic covered by real-git fixture tests in reliability-interactions/workflow-and-file-scope. FN-6206.",
"quarantinedAt": "2026-06-10"
},
<<<<<<< Updated upstream
=======
>>>>>>> Stashed changes
{
"file": "packages/engine/src/__tests__/merger-ai-cleanup.test.ts",
"reason": "Flake observed during FN-6206 verification: `pruneExistingAiMergeWorktrees skips active-session paths` failed in full `pnpm --filter @fusion/engine test` runs while the file passed standalone, indicating suite-order/concurrency sensitivity. Follow-up FN-6207.",