You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
daemon inputs: reads through a file descriptor (open, read, FileHandle) are no input yet #141
Family: reads through a file descriptor.fs.openSync/fs.open/fs.promises.open, followed by readSync/read/readv/fstat, FileHandle.read/readFile/stat, or readFileSync(fd), reads a file without naming its path in the read. None of these reads is an input today. The README lists them.
Gate kind fd: file b opens tests/gate.json with openSync, reads it with readSync into a buffer, closes it, and parses the text. A second history, fdPromise, can use fs.promises.open and handle.readFile().
Fix: wrap openSync, open and fs.promises.open. An open with a reading flag (r, r+, a+, w+, or the default) records fs:<path>, hashed by the whole file, so every later fd read of it is covered (sound over-approximation). An open with only a writing flag (w, a, wx) records a write, which joins #133's own-writes rule (ratified M 2026-11-17). The fd argument forms of readFileSync and readFile need no key of their own: the open already named the path. Inherited fds (stdin, a pipe) have no path and are a stated limit.
Authority: maintainer decision M 2026-11-16 "Every read is a daemon input" (the user ruled it with no exceptions), v21.2 R46, G42, I8. The watch family (watch, watchFile) stays no input (C-19).
Context:#133 records the six #55 fs calls on any path, node_modules included, plus environment variables, plugin inputs and directory listings (readdir/opendir, dir:/dirs: keys). The recorder is packages/oracles/oracle-gaps/inputs.mjs (watchReads), shared by the worker (capture-setup.mjs) and the CLI process (check.mjs). Keys are hashed by inputHash and looked up through servedHash. This family is still unbuilt, so I8 does not hold for it.
Add a value to inputRead (drawn, appended; update the canary's inputRead line) and one constructed history appended at the end, with steps [inputGate, recheck]. It enters edit inputGate changed the cold result, an edit to the <kind> gate input changed the cold result and an unchanged check through the daemon computed nothing. Add it to scripts/pr-shards.mjs shard 4 (and a Windows shard) and to scripts/e2e-shards.mjs.
RED: run pnpm exec vitest --run --config oracle-gaps/vitest.config.mjs gap-report-oracle -t "fixed seed 20261105: <name>" on the oracle-only commit. Expect incremental equals cold: "after inputGate, the daemon's report differs from the cold run's", with the daemon reusing killed where the cold run says survived.
Fix in inputs.mjs (see below), then show GREEN on shard 4.
Mutant: remove the new wrapper or key from watchReads. Only this history should fail.
Activity
fix(oracle-gaps): a file a test reads through a stream is a daemon input
docs(oracle-gaps): every plugin's reads, a full recompute on any env …