Engineering archive

These notes describe work at the date shown. See the usage reference for current behavior.

Task-write skip and non-overlapping dashboard polls (2026-09-09)

This records Stage 5 item 3 in part: skip ordinary identical LocalTaskRecord replacements, and poll TaskDetail / useSnapshot only after the previous request settles. It is not the whole of item 3 and not the whole of Stage 5. Point lookup by task ID landed earlier (2026-09-09-performance-improvements.md). Background observation collection independent of HTTP remains open. Item 2 admission cache is unchanged: 2026-09-09-admission-performance.md.

Measured pair: e8c1bb24db30aadf3e7f54ce5933cf1fd3d1087b → 915227f354a847a1dabab453ca35d528493db567. Fully tested source is that after SHA. File identity is a count of atomic replacements (symlink_metadata dev/ino), not an fsync census and not a fleet latency claim. UI numbers are request starts and live in-flight under fake timers, not wall-clock speedup.

Full gate on 915227f

Default-parallel Rust and full UI on the isolated validation worktree io-polls-review-20260909, CARGO_TARGET_DIR=/private/tmp/mac-worker-io-polls-n5b3h3by/validation/target, pre/post HEAD 915227f. Evidence: /private/tmp/mac-worker-io-polls-n5b3h3by/validation/final-gate/. GitHub Actions and the live pool were not run.

Gate	Exit	Elapsed	Log sha256
cargo fmt --all --check	0	1.011 s	e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
cargo test --locked --offline --all-targets	0	610.342 s	73c5d42d7c4cebaf411ca7149ba8316602e2be69b3c4c63aa405b14ed896dd1c
cargo clippy --locked --offline --all-targets -- -D warnings	0	10.786 s	fb65b8654ee0c5bcaedd8e28b899c1269db7d71ca44f70994fdb5c72df7bc610
npm test	0	3.026 s	0030963d7be79dadaa50b58db0c66c72f196960573025c77bc086b9a567878c9
npm run lint	0	0.820 s	9ee40e30498acd376d133b190158cce71adba56bb3f72f89134c9072defa585a
npm run build -- --outDir …/final-gate/ui-build-out	0	2.607 s	a5dc47583c40e017e8c2814e38d2c2ee3f925da0b47b552b557acf6f16842dfe
diff -r build vs src/dashboard/static/app	0	0.004 s	e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 (empty)
git diff --check	0	0.011 s	e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 (empty)

--all-targets: 68 Running targets and 68 unfiltered top-level result rows; 1812 passed, 0 failed, 0 ignored. Three nested helper rows are excluded (two 1 passed / 345 filtered lib helpers and one 1 passed / 10 filtered project_inspection helper). Do not count 1815. UI: 12 files / 105 tests passed. Lint exit 0 with 15 existing oxlint warnings (including useTurnLog.ts and TaskDetail.tsx set-state/ref notes); no unrelated warning cleanup. Fresh build compared to tracked embed is byte-equal (empty diff -r). Tests ran 2026-09-09T20:52:43Z–21:02:53Z; UI build+diff finished 21:03:10Z. rustc 1.98.0 (88d9e12ae 2026-08-18) (Homebrew).

What landed

Ordinary update_task and update_task_if_current (pub(crate)) pass IdenticalTaskWrite::Skip. Under StateLock both still run the rollback-fault probe, canonical serialize, tasks_dir, residue recovery, secure read, canonical parse, and filename identity. Expected-record comparison applies only when update_task_if_current supplies an expected snapshot: a stale expected returns Ok(false) even if replacement == existing. After those checks, an identical replacement skips the durable exchange (replace_private_regular_exact_with_sync_hooks). The private helper returns Ok(true) on that skip; public update_task still returns unit success. Equality is the complete LocalTaskRecord after canonical parse (timestamps included). There is no timestamp or schema normalization to manufacture a match. Residue directory scans remain.

IdenticalTaskWrite::Replace still exchanges (and runs before_final_sync) for clear_submission_intent and runner-handoff locked updates even when the parsed record is already identical. Changed status / runner / status_observed_at still replace. Corrupt / unsafe / non-canonical still fail closed.

TaskDetail and useSnapshot: one in-flight request per effect generation; first call is immediate; next poll is 2 s after the previous request settles (success or non-abort error). Cleanup aborts and clears the timer so a stale generation cannot apply a result or schedule. Last-good on error of the same task, retry, task-switch blanking, and unmount abort are preserved. useTurnLog is unchanged. No request timeout was added. Collection is still on the HTTP path.

Observed counts (not fleet RTT)

Artifact root: /private/tmp/mac-worker-io-polls-n5b3h3by. Isolated harness crate, --locked --offline. ClientStateSyncCounts is not a task-write census. update_task_if_current is not driven from the public inode harness (crate tests cover matching skip and stale CAS).

Frozen harness (identical before/after):

Item	sha256
validation/frozen/task_update_baseline.rs	afe0ae9220a29c395c82a7db1726336bcaf2798b703baa45f6e74848c4fb2071
validation/frozen/Cargo.toml	805cfd169036549f13370931d754a6038e76e565cf8a93991bb9057e57f6786e
validation/frozen/Cargo.lock	bd37361f282d714ac9c355aa4dce5e2832a134dd48144d941cacdc586ae941ef
validation/frozen/poll-baseline.test.tsx	5de0da4140301ebb22729748691b686a64ecf9da869ac727f071392f457cedd2
ui/vite.config.ts	0095f5fc291aa282d3ba06c26e03ce890b353406fb7f142da96266eb13097f31

Rust (public update_task / clear_submission_intent; 5 identical or 5 changed reps where noted):

export CARGO_TARGET_DIR=/private/tmp/mac-worker-io-polls-n5b3h3by/validation/target
export BASELINE_KIND=before   # or after
export BASELINE_OUT=/private/tmp/mac-worker-io-polls-n5b3h3by/validation/raw/${BASELINE_KIND}-task-update.json
cargo run --locked --offline --release --manifest-path /private/tmp/mac-worker-io-polls-n5b3h3by/validation/harness/Cargo.toml

UI (Vitest 5.0.0 from worktree ui/node_modules/.bin/vitest; existing ui/vite.config.ts; temporary untracked ui/src/poll-baseline.test.tsx copy, removed before the full gate):

export BASELINE_KIND=before   # or after
export BASELINE_OUT=/private/tmp/mac-worker-io-polls-n5b3h3by/validation/raw/${BASELINE_KIND}-ui-poll.json
cd ui && ./node_modules/.bin/vitest run --config ./vite.config.ts src/poll-baseline.test.tsx

All four measurement commands ec=0. JSON:

Run	HEAD	json sha256
Rust before	e8c1bb2	d73fb6940589003293465ee6e1c3038e9d5461d9edc105f8b78662bddc9fdc7e
UI before	e8c1bb2	8a63c21c2941e12513c83c9e168341df4eb469d2229611a7cc87f81f7302bf1f
Rust after	915227f	73c0a437253f4f8ab6750c7503468decd6bdfc042fbcb7b3c782f278e074cede
UI after	915227f	010582d0e1e45f7be1be9982943d6d4eb9c832b698312283768f2f0dae946e4a

Scenario	Before	After
Ordinary identical update_task ×5	5 calls, 5 inode changes, residue 0	5 calls, 0 inode changes, residue 0
Changed status_observed_at ×5	5 / 5 / 0	5 / 5 / 0
Changed inject then identical retry of now-current	2 inode changes; plant residue 1 then 0	1 (plant only; retry inode stable, residue 0)
Corrupt bytes then update_task	inode unchanged; error task record is corrupt	same reject
clear_submission_intent then repeat already-cleared	2 inode changes	2 (Replace kept)
TaskDetail pending 6000 ms	requests 4, live peak 4	1 / 1
useSnapshot pending 6000 ms	4 / 4	1 / 1
TaskDetail error recovery	3 requests, live peak 1	3 / 1 (last-good recovered)
TaskDetail task switch	2 starts, 1 aborted, live peak 1	unchanged
StrictMode at mount	2 starts, 1 aborted, live peak 1	unchanged
StrictMode after 6000 ms	5 starts, live peak 4, 1 aborted	2 / 1 / 1 aborted

abortedBySignal is the StrictMode / task-switch predecessor. It is not live network concurrency. After: a pending first request does not start more polls; that aborted predecessor does not schedule another 2 s timer.

Changed-fault error string both sides: injected local state failure after task replacement exchange before first directory sync.

Verification history (not the successful pair)

A production skip prototype was discarded before assertion RED on compilable exactbase. Archived harness setup failures are preserved and excluded from the pair: validation/raw/superseded-pre-correction-20260909T203245/ and validation/raw/superseded-ui-timeout-20260909T204231/.

Still open

Background observations independent of HTTP, status/log transport, SSH multiplexing, origin delivery, and the rest of Stage 5. QUEUE_TIME_REGRESSION on ungated concurrent enqueue remains a separate Stage 5.2 leftover. This increment does not close item 3 or Stage 5 as a whole.