Repository navigation
fix(server): one in-flight temporary-folder probe per path - #100
Conversation
…ort result cache A probe on a dead network mount never settles and pins a libuv thread. A retry now joins the unanswered probe instead of starting another, and a settled answer is reused for 30 s. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 388d3f91a3
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| // Map order is insertion order, so the first key is the oldest. | ||
| if (entries.size > MAX_CACHED_PATHS) entries.delete(entries.keys().next().value!); |
There was a problem hiding this comment.
When more than 256 distinct paths are probed while an older probe is still unresolved, this eviction removes that in-flight entry. A later request for the evicted path then starts another filesystem call instead of joining the existing one, so repeated traffic across a sufficiently large project set can again queue unbounded calls against a dead mount—the resource exhaustion this change is meant to prevent. Evict only settled cached results, or track unresolved probes separately until they settle.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in the next commit. Unanswered probes now live in their own set, apart from the bounded cache of settled answers, so eviction cannot forget them. A retry while a probe is unanswered gets no answer at once and goes by the name, so nothing adds a reaction to the pending promise.
| known !== undefined && | ||
| (known.settledAt === null || now() - known.settledAt < resultTtlMs) | ||
| ) { | ||
| return known.answer; |
There was a problem hiding this comment.
When the probe for one dead path never settles and that project is checked repeatedly, returning the same pending native promise makes every isTemporaryFolder call attach another .catch/await reaction to it. The Effect timeout abandons the caller but cannot detach those native promise reactions, and this cache deliberately retains the pending promise forever, so a single frequently retried path can grow memory without bound even though it now consumes only one libuv thread. Later calls should fall back without subscribing again, or share a single timeout wrapper that is allowed to settle.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fixed in the next commit. Unanswered probes now live in their own set, apart from the bounded cache of settled answers, so eviction cannot forget them. A retry while a probe is unanswered gets no answer at once and goes by the name, so nothing adds a reaction to the pending promise.
…result cache Eviction could forget an unanswered probe, and every retry added a reaction to its pending promise. A retry during an unanswered probe now gets no answer at once, which sends it to the name check. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
a59003e
into
integration/mcp-events-20261109
…ol-reliability t3_thread_configure reads its configuration only when none of modelSelection, runtimeMode, applyAt and acknowledgeOnly is given. Budget re-measured on the tip. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Follow-up to #86 (delta review finding): the temporary-folder probe times out after 500 ms, but the underlying
access/realpath/statfscalls keep running on a libuv thread. On a dead network mount (NFS, SSHFS) they never settle, so every project create or launch into that path would pin another thread.dedupeProbewraps the filesystem probe: one probe per path at a time (a retry joins the unanswered probe), and a settled answer is reused for 30 s. A rejected probe is not cached. The cache holds at most 256 paths. Message-getter callers still use the name-only check and never probe.temporaryFolder.test.ts: a never-settling probe is joined by a second call for the same path (one probe call); another path probes on its own; a settled answer is reused until the TTL passes.temporaryFolder.test.ts(15 tests). No migration.🤖 Generated with Claude Code