test(cli): replace fixed waits in Pi TUI coverage with observable completion by UncertaintyDeterminesYou4ndMe · Pull Request #2448 · apache/maka

UncertaintyDeterminesYou4ndMe · 2026-08-07T15:39:17Z

Part of #2389 (Pi TUI workspace; Runtime Host, desktop e2e, and alignment audit follow in separate PRs per the issue).

What changed

Three fixed-wait patterns in pi-tui-runner.test.ts (plus one helper in tui-terminal-mock.ts), replaced with the closest deterministic mechanism:

Pattern	Count	Replacement
Unconditional `await delay(20..2500)`	55	`waitFor` on the actual observable fact: first TUI paint (`waitForTuiPaint`), prompt arming ("Allow access outside the workspace?" on screen), wizard frames opening/closing, drafts appearing, driver retry counts
`Promise.race([run, delay(50/500).then(throw)])` close watchdogs	81	`CLOSE_BUDGET_MS = max(WAIT_BUDGET_MS, 500)` — a passing race still resolves the moment `run` settles; only failure reporting pays the budget (same split as `waitFor`, #2221)
Off-screen ticker test sleeping 2.5s for two real 1s ticks	1	The shell-run elapsed ticker's `now` + `schedule` are now injectable via `MakaPiTuiInput.shellRunTicker` (same test-injection pattern as `attentionLongTurnThresholdMs`); the test fires both ticks deterministically

For must-NOT-happen assertions, the suite now uses three deterministic idioms instead of sleeping and hoping:

Sentinel render: type one char and wait for it to paint — any wrongly-scheduled UI change (stale wizard failure line, dropped-recap render, rewind picker, scrollback wipe) would be in that same coalesced frame.
Close-anchored asserts: for driver-level negatives (no prompt sent, no response for a dead request, no rewind), assert after await run — every queued input has provably drained.
delay(0) macrotask barrier: for gates that run synchronously off input dispatch (double-Escape counting, Ctrl-C/Ctrl-D, submit gates) — one macrotask turn runs strictly after every queued microtask, independent of machine load.

The never-dropped fallback-enqueue test no longer sleeps 2.2s blind: the fake driver counts enqueue attempts, and the test waits for the retries to observably cross the old ~20-attempt give-up budget (the exact regression #1135-era bug shape), still at the real 100ms cadence so a reintroduced time-based budget would also trip it.

Retained real-boundary contracts (unchanged by design)

runSignalExitProbe / runFatalExitProbe: real child processes, real signals, and the real 3s PROCESS_EXIT_GRACE_MS deadline (3.5s + 2×0.5s of the remaining wall time).
The 100ms fallback retry cadence (retry-budget test above).
The 250ms hydration-retry cancellation window — a real-timer negative; now documented as derived from the retry schedule rather than an arbitrary sleep.

Production surface

One additive, optional injection point: MakaPiTuiInput.shellRunTicker?: { now?, schedule? }, threaded into createShellRunElapsedTicker (which already accepted both). Defaults unchanged.

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

	Before	After
Wall	15.5s	12.1s
CPU (user)	4.2s	4.1s

The remaining wall time is dominated by the preserved real-boundary tests listed above; every other test now completes the moment its observed condition holds, and no assertion in the suite races scheduler timing anymore. Full packages/cli suite: 492/494 pass — the 2 failures are pre-existing local noise (run-command.test.ts expects empty stderr; Node 22.17 emits a SQLite ExperimentalWarning) on paths this PR does not touch.

…pletion Part of apache#2389. The Pi TUI runner suite guessed at scheduler timing in three ways, all sensitive to CI load: - 55 unconditional sleeps (delay(20)..delay(2500)) where the awaited fact is observable: startup paints, prompt arming, wizard frames opening/closing, drafts appearing, and retry attempts are all now waited on directly via the existing waitFor helper, a new waitForTuiPaint helper, sentinel-render barriers for must-not-render negatives, and close-anchored assertions for must-never-happen driver calls. - 81 Promise.race close watchdogs with a flat 50/500ms deadline that a loaded runner can miss; they now share CLOSE_BUDGET_MS (WAIT_BUDGET_MS-scaled), so a passing race is unchanged and only failure reporting pays the budget. - The off-screen ticker test slept 2.5s for two real 1s ticks; the shell-run elapsed ticker's clock and interval scheduling are now injectable (MakaPiTuiInput.shellRunTicker, same pattern as attentionLongTurnThresholdMs) and the test drives both ticks itself. Synchronous input-dispatch negatives (double-Escape counting, Ctrl-C/Ctrl-D gating, submit gates) settle on one macrotask turn (delay(0)) instead of a 20-30ms guess, and the never-dropped fallback-enqueue test now waits for the driver to observe the retries crossing the old give-up budget instead of sleeping 2.2s blind. Preserved as real timeouts or real process boundaries: the SIGTERM/fatal exit probes (real child processes and the 3s exit grace), the 100ms fallback retry cadence, and the 250ms hydration-retry cancellation window (now documented as schedule-derived). Focused timing (node --test dist/__tests__/pi-tui-runner.test.js, 142 tests): 15.5s -> 12.1s wall locally, with the remainder dominated by the preserved real-boundary tests, and no fixed-delay races left in the suite.

UncertaintyDeterminesYou4ndMe · 2026-08-07T15:39:17Z

Part of #2389 (Pi TUI workspace; Runtime Host, desktop e2e, and alignment audit follow in separate PRs per the issue).

What changed

Three fixed-wait patterns in pi-tui-runner.test.ts (plus one helper in tui-terminal-mock.ts), replaced with the closest deterministic mechanism:

Pattern	Count	Replacement
Unconditional `await delay(20..2500)`	55	`waitFor` on the actual observable fact: first TUI paint (`waitForTuiPaint`), prompt arming ("Allow access outside the workspace?" on screen), wizard frames opening/closing, drafts appearing, driver retry counts
`Promise.race([run, delay(50/500).then(throw)])` close watchdogs	81	`CLOSE_BUDGET_MS = max(WAIT_BUDGET_MS, 500)` — a passing race still resolves the moment `run` settles; only failure reporting pays the budget (same split as `waitFor`, #2221)
Off-screen ticker test sleeping 2.5s for two real 1s ticks	1	The shell-run elapsed ticker's `now` + `schedule` are now injectable via `MakaPiTuiInput.shellRunTicker` (same test-injection pattern as `attentionLongTurnThresholdMs`); the test fires both ticks deterministically

For must-NOT-happen assertions, the suite now uses three deterministic idioms instead of sleeping and hoping:

Sentinel render: type one char and wait for it to paint — any wrongly-scheduled UI change (stale wizard failure line, dropped-recap render, rewind picker, scrollback wipe) would be in that same coalesced frame.
Close-anchored asserts: for driver-level negatives (no prompt sent, no response for a dead request, no rewind), assert after await run — every queued input has provably drained.
delay(0) macrotask barrier: for gates that run synchronously off input dispatch (double-Escape counting, Ctrl-C/Ctrl-D, submit gates) — one macrotask turn runs strictly after every queued microtask, independent of machine load.

The never-dropped fallback-enqueue test no longer sleeps 2.2s blind: the fake driver counts enqueue attempts, and the test waits for the retries to observably cross the old ~20-attempt give-up budget (the exact regression #1135-era bug shape), still at the real 100ms cadence so a reintroduced time-based budget would also trip it.

Retained real-boundary contracts (unchanged by design)

runSignalExitProbe / runFatalExitProbe: real child processes, real signals, and the real 3s PROCESS_EXIT_GRACE_MS deadline (3.5s + 2×0.5s of the remaining wall time).
The 100ms fallback retry cadence (retry-budget test above).
The 250ms hydration-retry cancellation window — a real-timer negative; now documented as derived from the retry schedule rather than an arbitrary sleep.

Production surface

One additive, optional injection point: MakaPiTuiInput.shellRunTicker?: { now?, schedule? }, threaded into createShellRunElapsedTicker (which already accepted both). Defaults unchanged.

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

	Before	After
Wall	15.5s	12.1s
CPU (user)	4.2s	4.1s

The remaining wall time is dominated by the preserved real-boundary tests listed above; every other test now completes the moment its observed condition holds, and no assertion in the suite races scheduler timing anymore. Full packages/cli suite: 492/494 pass — the 2 failures are pre-existing local noise (run-command.test.ts expects empty stderr; Node 22.17 emits a SQLite ExperimentalWarning) on paths this PR does not touch.

…pletion Part of apache#2389. The Pi TUI runner suite guessed at scheduler timing in three ways, all sensitive to CI load: - 55 unconditional sleeps (delay(20)..delay(2500)) where the awaited fact is observable: startup paints, prompt arming, wizard frames opening/closing, drafts appearing, and retry attempts are all now waited on directly via the existing waitFor helper, a new waitForTuiPaint helper, sentinel-render barriers for must-not-render negatives, and close-anchored assertions for must-never-happen driver calls. - 81 Promise.race close watchdogs with a flat 50/500ms deadline that a loaded runner can miss; they now share CLOSE_BUDGET_MS (WAIT_BUDGET_MS-scaled), so a passing race is unchanged and only failure reporting pays the budget. - The off-screen ticker test slept 2.5s for two real 1s ticks; the shell-run elapsed ticker's clock and interval scheduling are now injectable (MakaPiTuiInput.shellRunTicker, same pattern as attentionLongTurnThresholdMs) and the test drives both ticks itself. Synchronous input-dispatch negatives (double-Escape counting, Ctrl-C/Ctrl-D gating, submit gates) settle on one macrotask turn (delay(0)) instead of a 20-30ms guess, and the never-dropped fallback-enqueue test now waits for the driver to observe the retries crossing the old give-up budget instead of sleeping 2.2s blind. Preserved as real timeouts or real process boundaries: the SIGTERM/fatal exit probes (real child processes and the 3s exit grace), the 100ms fallback retry cadence, and the 250ms hydration-retry cancellation window (now documented as schedule-derived). Focused timing (node --test dist/__tests__/pi-tui-runner.test.js, 142 tests): 15.5s -> 12.1s wall locally, with the remainder dominated by the preserved real-boundary tests, and no fixed-delay races left in the suite.

UncertaintyDeterminesYou4ndMe · 2026-08-07T15:39:17Z

Part of #2389 (Pi TUI workspace; Runtime Host, desktop e2e, and alignment audit follow in separate PRs per the issue).

What changed

Three fixed-wait patterns in pi-tui-runner.test.ts (plus one helper in tui-terminal-mock.ts), replaced with the closest deterministic mechanism:

Pattern	Count	Replacement
Unconditional `await delay(20..2500)`	55	`waitFor` on the actual observable fact: first TUI paint (`waitForTuiPaint`), prompt arming ("Allow access outside the workspace?" on screen), wizard frames opening/closing, drafts appearing, driver retry counts
`Promise.race([run, delay(50/500).then(throw)])` close watchdogs	81	`CLOSE_BUDGET_MS = max(WAIT_BUDGET_MS, 500)` — a passing race still resolves the moment `run` settles; only failure reporting pays the budget (same split as `waitFor`, #2221)
Off-screen ticker test sleeping 2.5s for two real 1s ticks	1	The shell-run elapsed ticker's `now` + `schedule` are now injectable via `MakaPiTuiInput.shellRunTicker` (same test-injection pattern as `attentionLongTurnThresholdMs`); the test fires both ticks deterministically

For must-NOT-happen assertions, the suite now uses three deterministic idioms instead of sleeping and hoping:

Sentinel render: type one char and wait for it to paint — any wrongly-scheduled UI change (stale wizard failure line, dropped-recap render, rewind picker, scrollback wipe) would be in that same coalesced frame.
Close-anchored asserts: for driver-level negatives (no prompt sent, no response for a dead request, no rewind), assert after await run — every queued input has provably drained.
delay(0) macrotask barrier: for gates that run synchronously off input dispatch (double-Escape counting, Ctrl-C/Ctrl-D, submit gates) — one macrotask turn runs strictly after every queued microtask, independent of machine load.

The never-dropped fallback-enqueue test no longer sleeps 2.2s blind: the fake driver counts enqueue attempts, and the test waits for the retries to observably cross the old ~20-attempt give-up budget (the exact regression #1135-era bug shape), still at the real 100ms cadence so a reintroduced time-based budget would also trip it.

Retained real-boundary contracts (unchanged by design)

runSignalExitProbe / runFatalExitProbe: real child processes, real signals, and the real 3s PROCESS_EXIT_GRACE_MS deadline (3.5s + 2×0.5s of the remaining wall time).
The 100ms fallback retry cadence (retry-budget test above).
The 250ms hydration-retry cancellation window — a real-timer negative; now documented as derived from the retry schedule rather than an arbitrary sleep.

Production surface

One additive, optional injection point: MakaPiTuiInput.shellRunTicker?: { now?, schedule? }, threaded into createShellRunElapsedTicker (which already accepted both). Defaults unchanged.

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

	Before	After
Wall	15.5s	12.1s
CPU (user)	4.2s	4.1s

The remaining wall time is dominated by the preserved real-boundary tests listed above; every other test now completes the moment its observed condition holds, and no assertion in the suite races scheduler timing anymore. Full packages/cli suite: 492/494 pass — the 2 failures are pre-existing local noise (run-command.test.ts expects empty stderr; Node 22.17 emits a SQLite ExperimentalWarning) on paths this PR does not touch.

…pletion Part of apache#2389. The Pi TUI runner suite guessed at scheduler timing in three ways, all sensitive to CI load: - 55 unconditional sleeps (delay(20)..delay(2500)) where the awaited fact is observable: startup paints, prompt arming, wizard frames opening/closing, drafts appearing, and retry attempts are all now waited on directly via the existing waitFor helper, a new waitForTuiPaint helper, sentinel-render barriers for must-not-render negatives, and close-anchored assertions for must-never-happen driver calls. - 81 Promise.race close watchdogs with a flat 50/500ms deadline that a loaded runner can miss; they now share CLOSE_BUDGET_MS (WAIT_BUDGET_MS-scaled), so a passing race is unchanged and only failure reporting pays the budget. - The off-screen ticker test slept 2.5s for two real 1s ticks; the shell-run elapsed ticker's clock and interval scheduling are now injectable (MakaPiTuiInput.shellRunTicker, same pattern as attentionLongTurnThresholdMs) and the test drives both ticks itself. Synchronous input-dispatch negatives (double-Escape counting, Ctrl-C/Ctrl-D gating, submit gates) settle on one macrotask turn (delay(0)) instead of a 20-30ms guess, and the never-dropped fallback-enqueue test now waits for the driver to observe the retries crossing the old give-up budget instead of sleeping 2.2s blind. Preserved as real timeouts or real process boundaries: the SIGTERM/fatal exit probes (real child processes and the 3s exit grace), the 100ms fallback retry cadence, and the 250ms hydration-retry cancellation window (now documented as schedule-derived). Focused timing (node --test dist/__tests__/pi-tui-runner.test.js, 142 tests): 15.5s -> 12.1s wall locally, with the remainder dominated by the preserved real-boundary tests, and no fixed-delay races left in the suite.

UncertaintyDeterminesYou4ndMe · 2026-08-07T15:39:17Z

Part of #2389 (Pi TUI workspace; Runtime Host, desktop e2e, and alignment audit follow in separate PRs per the issue).

What changed

Three fixed-wait patterns in pi-tui-runner.test.ts (plus one helper in tui-terminal-mock.ts), replaced with the closest deterministic mechanism:

Pattern	Count	Replacement
Unconditional `await delay(20..2500)`	55	`waitFor` on the actual observable fact: first TUI paint (`waitForTuiPaint`), prompt arming ("Allow access outside the workspace?" on screen), wizard frames opening/closing, drafts appearing, driver retry counts
`Promise.race([run, delay(50/500).then(throw)])` close watchdogs	81	`CLOSE_BUDGET_MS = max(WAIT_BUDGET_MS, 500)` — a passing race still resolves the moment `run` settles; only failure reporting pays the budget (same split as `waitFor`, #2221)
Off-screen ticker test sleeping 2.5s for two real 1s ticks	1	The shell-run elapsed ticker's `now` + `schedule` are now injectable via `MakaPiTuiInput.shellRunTicker` (same test-injection pattern as `attentionLongTurnThresholdMs`); the test fires both ticks deterministically

For must-NOT-happen assertions, the suite now uses three deterministic idioms instead of sleeping and hoping:

Sentinel render: type one char and wait for it to paint — any wrongly-scheduled UI change (stale wizard failure line, dropped-recap render, rewind picker, scrollback wipe) would be in that same coalesced frame.
Close-anchored asserts: for driver-level negatives (no prompt sent, no response for a dead request, no rewind), assert after await run — every queued input has provably drained.
delay(0) macrotask barrier: for gates that run synchronously off input dispatch (double-Escape counting, Ctrl-C/Ctrl-D, submit gates) — one macrotask turn runs strictly after every queued microtask, independent of machine load.

The never-dropped fallback-enqueue test no longer sleeps 2.2s blind: the fake driver counts enqueue attempts, and the test waits for the retries to observably cross the old ~20-attempt give-up budget (the exact regression #1135-era bug shape), still at the real 100ms cadence so a reintroduced time-based budget would also trip it.

Retained real-boundary contracts (unchanged by design)

runSignalExitProbe / runFatalExitProbe: real child processes, real signals, and the real 3s PROCESS_EXIT_GRACE_MS deadline (3.5s + 2×0.5s of the remaining wall time).
The 100ms fallback retry cadence (retry-budget test above).
The 250ms hydration-retry cancellation window — a real-timer negative; now documented as derived from the retry schedule rather than an arbitrary sleep.

Production surface

One additive, optional injection point: MakaPiTuiInput.shellRunTicker?: { now?, schedule? }, threaded into createShellRunElapsedTicker (which already accepted both). Defaults unchanged.

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

	Before	After
Wall	15.5s	12.1s
CPU (user)	4.2s	4.1s

The remaining wall time is dominated by the preserved real-boundary tests listed above; every other test now completes the moment its observed condition holds, and no assertion in the suite races scheduler timing anymore. Full packages/cli suite: 492/494 pass — the 2 failures are pre-existing local noise (run-command.test.ts expects empty stderr; Node 22.17 emits a SQLite ExperimentalWarning) on paths this PR does not touch.

…pletion Part of apache#2389. The Pi TUI runner suite guessed at scheduler timing in three ways, all sensitive to CI load: - 55 unconditional sleeps (delay(20)..delay(2500)) where the awaited fact is observable: startup paints, prompt arming, wizard frames opening/closing, drafts appearing, and retry attempts are all now waited on directly via the existing waitFor helper, a new waitForTuiPaint helper, sentinel-render barriers for must-not-render negatives, and close-anchored assertions for must-never-happen driver calls. - 81 Promise.race close watchdogs with a flat 50/500ms deadline that a loaded runner can miss; they now share CLOSE_BUDGET_MS (WAIT_BUDGET_MS-scaled), so a passing race is unchanged and only failure reporting pays the budget. - The off-screen ticker test slept 2.5s for two real 1s ticks; the shell-run elapsed ticker's clock and interval scheduling are now injectable (MakaPiTuiInput.shellRunTicker, same pattern as attentionLongTurnThresholdMs) and the test drives both ticks itself. Synchronous input-dispatch negatives (double-Escape counting, Ctrl-C/Ctrl-D gating, submit gates) settle on one macrotask turn (delay(0)) instead of a 20-30ms guess, and the never-dropped fallback-enqueue test now waits for the driver to observe the retries crossing the old give-up budget instead of sleeping 2.2s blind. Preserved as real timeouts or real process boundaries: the SIGTERM/fatal exit probes (real child processes and the 3s exit grace), the 100ms fallback retry cadence, and the 250ms hydration-retry cancellation window (now documented as schedule-derived). Focused timing (node --test dist/__tests__/pi-tui-runner.test.js, 142 tests): 15.5s -> 12.1s wall locally, with the remainder dominated by the preserved real-boundary tests, and no fixed-delay races left in the suite.

UncertaintyDeterminesYou4ndMe · 2026-08-07T15:39:17Z

Part of #2389 (Pi TUI workspace; Runtime Host, desktop e2e, and alignment audit follow in separate PRs per the issue).

What changed

Three fixed-wait patterns in pi-tui-runner.test.ts (plus one helper in tui-terminal-mock.ts), replaced with the closest deterministic mechanism:

Pattern	Count	Replacement
Unconditional `await delay(20..2500)`	55	`waitFor` on the actual observable fact: first TUI paint (`waitForTuiPaint`), prompt arming ("Allow access outside the workspace?" on screen), wizard frames opening/closing, drafts appearing, driver retry counts
`Promise.race([run, delay(50/500).then(throw)])` close watchdogs	81	`CLOSE_BUDGET_MS = max(WAIT_BUDGET_MS, 500)` — a passing race still resolves the moment `run` settles; only failure reporting pays the budget (same split as `waitFor`, #2221)
Off-screen ticker test sleeping 2.5s for two real 1s ticks	1	The shell-run elapsed ticker's `now` + `schedule` are now injectable via `MakaPiTuiInput.shellRunTicker` (same test-injection pattern as `attentionLongTurnThresholdMs`); the test fires both ticks deterministically

For must-NOT-happen assertions, the suite now uses three deterministic idioms instead of sleeping and hoping:

Sentinel render: type one char and wait for it to paint — any wrongly-scheduled UI change (stale wizard failure line, dropped-recap render, rewind picker, scrollback wipe) would be in that same coalesced frame.
Close-anchored asserts: for driver-level negatives (no prompt sent, no response for a dead request, no rewind), assert after await run — every queued input has provably drained.
delay(0) macrotask barrier: for gates that run synchronously off input dispatch (double-Escape counting, Ctrl-C/Ctrl-D, submit gates) — one macrotask turn runs strictly after every queued microtask, independent of machine load.

The never-dropped fallback-enqueue test no longer sleeps 2.2s blind: the fake driver counts enqueue attempts, and the test waits for the retries to observably cross the old ~20-attempt give-up budget (the exact regression #1135-era bug shape), still at the real 100ms cadence so a reintroduced time-based budget would also trip it.

Retained real-boundary contracts (unchanged by design)

runSignalExitProbe / runFatalExitProbe: real child processes, real signals, and the real 3s PROCESS_EXIT_GRACE_MS deadline (3.5s + 2×0.5s of the remaining wall time).
The 100ms fallback retry cadence (retry-budget test above).
The 250ms hydration-retry cancellation window — a real-timer negative; now documented as derived from the retry schedule rather than an arbitrary sleep.

Production surface

One additive, optional injection point: MakaPiTuiInput.shellRunTicker?: { now?, schedule? }, threaded into createShellRunElapsedTicker (which already accepted both). Defaults unchanged.

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

	Before	After
Wall	15.5s	12.1s
CPU (user)	4.2s	4.1s

The remaining wall time is dominated by the preserved real-boundary tests listed above; every other test now completes the moment its observed condition holds, and no assertion in the suite races scheduler timing anymore. Full packages/cli suite: 492/494 pass — the 2 failures are pre-existing local noise (run-command.test.ts expects empty stderr; Node 22.17 emits a SQLite ExperimentalWarning) on paths this PR does not touch.

…pletion Part of apache#2389. The Pi TUI runner suite guessed at scheduler timing in three ways, all sensitive to CI load: - 55 unconditional sleeps (delay(20)..delay(2500)) where the awaited fact is observable: startup paints, prompt arming, wizard frames opening/closing, drafts appearing, and retry attempts are all now waited on directly via the existing waitFor helper, a new waitForTuiPaint helper, sentinel-render barriers for must-not-render negatives, and close-anchored assertions for must-never-happen driver calls. - 81 Promise.race close watchdogs with a flat 50/500ms deadline that a loaded runner can miss; they now share CLOSE_BUDGET_MS (WAIT_BUDGET_MS-scaled), so a passing race is unchanged and only failure reporting pays the budget. - The off-screen ticker test slept 2.5s for two real 1s ticks; the shell-run elapsed ticker's clock and interval scheduling are now injectable (MakaPiTuiInput.shellRunTicker, same pattern as attentionLongTurnThresholdMs) and the test drives both ticks itself. Synchronous input-dispatch negatives (double-Escape counting, Ctrl-C/Ctrl-D gating, submit gates) settle on one macrotask turn (delay(0)) instead of a 20-30ms guess, and the never-dropped fallback-enqueue test now waits for the driver to observe the retries crossing the old give-up budget instead of sleeping 2.2s blind. Preserved as real timeouts or real process boundaries: the SIGTERM/fatal exit probes (real child processes and the 3s exit grace), the 100ms fallback retry cadence, and the 250ms hydration-retry cancellation window (now documented as schedule-derived). Focused timing (node --test dist/__tests__/pi-tui-runner.test.js, 142 tests): 15.5s -> 12.1s wall locally, with the remainder dominated by the preserved real-boundary tests, and no fixed-delay races left in the suite.

UncertaintyDeterminesYou4ndMe · 2026-08-07T15:39:17Z

Part of #2389 (Pi TUI workspace; Runtime Host, desktop e2e, and alignment audit follow in separate PRs per the issue).

What changed

Three fixed-wait patterns in pi-tui-runner.test.ts (plus one helper in tui-terminal-mock.ts), replaced with the closest deterministic mechanism:

Pattern	Count	Replacement
Unconditional `await delay(20..2500)`	55	`waitFor` on the actual observable fact: first TUI paint (`waitForTuiPaint`), prompt arming ("Allow access outside the workspace?" on screen), wizard frames opening/closing, drafts appearing, driver retry counts
`Promise.race([run, delay(50/500).then(throw)])` close watchdogs	81	`CLOSE_BUDGET_MS = max(WAIT_BUDGET_MS, 500)` — a passing race still resolves the moment `run` settles; only failure reporting pays the budget (same split as `waitFor`, #2221)
Off-screen ticker test sleeping 2.5s for two real 1s ticks	1	The shell-run elapsed ticker's `now` + `schedule` are now injectable via `MakaPiTuiInput.shellRunTicker` (same test-injection pattern as `attentionLongTurnThresholdMs`); the test fires both ticks deterministically

For must-NOT-happen assertions, the suite now uses three deterministic idioms instead of sleeping and hoping:

Sentinel render: type one char and wait for it to paint — any wrongly-scheduled UI change (stale wizard failure line, dropped-recap render, rewind picker, scrollback wipe) would be in that same coalesced frame.
Close-anchored asserts: for driver-level negatives (no prompt sent, no response for a dead request, no rewind), assert after await run — every queued input has provably drained.
delay(0) macrotask barrier: for gates that run synchronously off input dispatch (double-Escape counting, Ctrl-C/Ctrl-D, submit gates) — one macrotask turn runs strictly after every queued microtask, independent of machine load.

The never-dropped fallback-enqueue test no longer sleeps 2.2s blind: the fake driver counts enqueue attempts, and the test waits for the retries to observably cross the old ~20-attempt give-up budget (the exact regression #1135-era bug shape), still at the real 100ms cadence so a reintroduced time-based budget would also trip it.

Retained real-boundary contracts (unchanged by design)

runSignalExitProbe / runFatalExitProbe: real child processes, real signals, and the real 3s PROCESS_EXIT_GRACE_MS deadline (3.5s + 2×0.5s of the remaining wall time).
The 100ms fallback retry cadence (retry-budget test above).
The 250ms hydration-retry cancellation window — a real-timer negative; now documented as derived from the retry schedule rather than an arbitrary sleep.

Production surface

One additive, optional injection point: MakaPiTuiInput.shellRunTicker?: { now?, schedule? }, threaded into createShellRunElapsedTicker (which already accepted both). Defaults unchanged.

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

	Before	After
Wall	15.5s	12.1s
CPU (user)	4.2s	4.1s

The remaining wall time is dominated by the preserved real-boundary tests listed above; every other test now completes the moment its observed condition holds, and no assertion in the suite races scheduler timing anymore. Full packages/cli suite: 492/494 pass — the 2 failures are pre-existing local noise (run-command.test.ts expects empty stderr; Node 22.17 emits a SQLite ExperimentalWarning) on paths this PR does not touch.

…pletion Part of apache#2389. The Pi TUI runner suite guessed at scheduler timing in three ways, all sensitive to CI load: - 55 unconditional sleeps (delay(20)..delay(2500)) where the awaited fact is observable: startup paints, prompt arming, wizard frames opening/closing, drafts appearing, and retry attempts are all now waited on directly via the existing waitFor helper, a new waitForTuiPaint helper, sentinel-render barriers for must-not-render negatives, and close-anchored assertions for must-never-happen driver calls. - 81 Promise.race close watchdogs with a flat 50/500ms deadline that a loaded runner can miss; they now share CLOSE_BUDGET_MS (WAIT_BUDGET_MS-scaled), so a passing race is unchanged and only failure reporting pays the budget. - The off-screen ticker test slept 2.5s for two real 1s ticks; the shell-run elapsed ticker's clock and interval scheduling are now injectable (MakaPiTuiInput.shellRunTicker, same pattern as attentionLongTurnThresholdMs) and the test drives both ticks itself. Synchronous input-dispatch negatives (double-Escape counting, Ctrl-C/Ctrl-D gating, submit gates) settle on one macrotask turn (delay(0)) instead of a 20-30ms guess, and the never-dropped fallback-enqueue test now waits for the driver to observe the retries crossing the old give-up budget instead of sleeping 2.2s blind. Preserved as real timeouts or real process boundaries: the SIGTERM/fatal exit probes (real child processes and the 3s exit grace), the 100ms fallback retry cadence, and the 250ms hydration-retry cancellation window (now documented as schedule-derived). Focused timing (node --test dist/__tests__/pi-tui-runner.test.js, 142 tests): 15.5s -> 12.1s wall locally, with the remainder dominated by the preserved real-boundary tests, and no fixed-delay races left in the suite.

UncertaintyDeterminesYou4ndMe · 2026-08-07T15:39:17Z

Part of #2389 (Pi TUI workspace; Runtime Host, desktop e2e, and alignment audit follow in separate PRs per the issue).

What changed

Three fixed-wait patterns in pi-tui-runner.test.ts (plus one helper in tui-terminal-mock.ts), replaced with the closest deterministic mechanism:

Pattern	Count	Replacement
Unconditional `await delay(20..2500)`	55	`waitFor` on the actual observable fact: first TUI paint (`waitForTuiPaint`), prompt arming ("Allow access outside the workspace?" on screen), wizard frames opening/closing, drafts appearing, driver retry counts
`Promise.race([run, delay(50/500).then(throw)])` close watchdogs	81	`CLOSE_BUDGET_MS = max(WAIT_BUDGET_MS, 500)` — a passing race still resolves the moment `run` settles; only failure reporting pays the budget (same split as `waitFor`, #2221)
Off-screen ticker test sleeping 2.5s for two real 1s ticks	1	The shell-run elapsed ticker's `now` + `schedule` are now injectable via `MakaPiTuiInput.shellRunTicker` (same test-injection pattern as `attentionLongTurnThresholdMs`); the test fires both ticks deterministically

For must-NOT-happen assertions, the suite now uses three deterministic idioms instead of sleeping and hoping:

Sentinel render: type one char and wait for it to paint — any wrongly-scheduled UI change (stale wizard failure line, dropped-recap render, rewind picker, scrollback wipe) would be in that same coalesced frame.
Close-anchored asserts: for driver-level negatives (no prompt sent, no response for a dead request, no rewind), assert after await run — every queued input has provably drained.
delay(0) macrotask barrier: for gates that run synchronously off input dispatch (double-Escape counting, Ctrl-C/Ctrl-D, submit gates) — one macrotask turn runs strictly after every queued microtask, independent of machine load.

The never-dropped fallback-enqueue test no longer sleeps 2.2s blind: the fake driver counts enqueue attempts, and the test waits for the retries to observably cross the old ~20-attempt give-up budget (the exact regression #1135-era bug shape), still at the real 100ms cadence so a reintroduced time-based budget would also trip it.

Retained real-boundary contracts (unchanged by design)

runSignalExitProbe / runFatalExitProbe: real child processes, real signals, and the real 3s PROCESS_EXIT_GRACE_MS deadline (3.5s + 2×0.5s of the remaining wall time).
The 100ms fallback retry cadence (retry-budget test above).
The 250ms hydration-retry cancellation window — a real-timer negative; now documented as derived from the retry schedule rather than an arbitrary sleep.

Production surface

One additive, optional injection point: MakaPiTuiInput.shellRunTicker?: { now?, schedule? }, threaded into createShellRunElapsedTicker (which already accepted both). Defaults unchanged.

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

	Before	After
Wall	15.5s	12.1s
CPU (user)	4.2s	4.1s

The remaining wall time is dominated by the preserved real-boundary tests listed above; every other test now completes the moment its observed condition holds, and no assertion in the suite races scheduler timing anymore. Full packages/cli suite: 492/494 pass — the 2 failures are pre-existing local noise (run-command.test.ts expects empty stderr; Node 22.17 emits a SQLite ExperimentalWarning) on paths this PR does not touch.

…pletion Part of apache#2389. The Pi TUI runner suite guessed at scheduler timing in three ways, all sensitive to CI load: - 55 unconditional sleeps (delay(20)..delay(2500)) where the awaited fact is observable: startup paints, prompt arming, wizard frames opening/closing, drafts appearing, and retry attempts are all now waited on directly via the existing waitFor helper, a new waitForTuiPaint helper, sentinel-render barriers for must-not-render negatives, and close-anchored assertions for must-never-happen driver calls. - 81 Promise.race close watchdogs with a flat 50/500ms deadline that a loaded runner can miss; they now share CLOSE_BUDGET_MS (WAIT_BUDGET_MS-scaled), so a passing race is unchanged and only failure reporting pays the budget. - The off-screen ticker test slept 2.5s for two real 1s ticks; the shell-run elapsed ticker's clock and interval scheduling are now injectable (MakaPiTuiInput.shellRunTicker, same pattern as attentionLongTurnThresholdMs) and the test drives both ticks itself. Synchronous input-dispatch negatives (double-Escape counting, Ctrl-C/Ctrl-D gating, submit gates) settle on one macrotask turn (delay(0)) instead of a 20-30ms guess, and the never-dropped fallback-enqueue test now waits for the driver to observe the retries crossing the old give-up budget instead of sleeping 2.2s blind. Preserved as real timeouts or real process boundaries: the SIGTERM/fatal exit probes (real child processes and the 3s exit grace), the 100ms fallback retry cadence, and the 250ms hydration-retry cancellation window (now documented as schedule-derived). Focused timing (node --test dist/__tests__/pi-tui-runner.test.js, 142 tests): 15.5s -> 12.1s wall locally, with the remainder dominated by the preserved real-boundary tests, and no fixed-delay races left in the suite.

UncertaintyDeterminesYou4ndMe · 2026-08-07T15:39:17Z

Part of #2389 (Pi TUI workspace; Runtime Host, desktop e2e, and alignment audit follow in separate PRs per the issue).

What changed

Three fixed-wait patterns in pi-tui-runner.test.ts (plus one helper in tui-terminal-mock.ts), replaced with the closest deterministic mechanism:

Pattern	Count	Replacement
Unconditional `await delay(20..2500)`	55	`waitFor` on the actual observable fact: first TUI paint (`waitForTuiPaint`), prompt arming ("Allow access outside the workspace?" on screen), wizard frames opening/closing, drafts appearing, driver retry counts
`Promise.race([run, delay(50/500).then(throw)])` close watchdogs	81	`CLOSE_BUDGET_MS = max(WAIT_BUDGET_MS, 500)` — a passing race still resolves the moment `run` settles; only failure reporting pays the budget (same split as `waitFor`, #2221)
Off-screen ticker test sleeping 2.5s for two real 1s ticks	1	The shell-run elapsed ticker's `now` + `schedule` are now injectable via `MakaPiTuiInput.shellRunTicker` (same test-injection pattern as `attentionLongTurnThresholdMs`); the test fires both ticks deterministically

For must-NOT-happen assertions, the suite now uses three deterministic idioms instead of sleeping and hoping:

Sentinel render: type one char and wait for it to paint — any wrongly-scheduled UI change (stale wizard failure line, dropped-recap render, rewind picker, scrollback wipe) would be in that same coalesced frame.
Close-anchored asserts: for driver-level negatives (no prompt sent, no response for a dead request, no rewind), assert after await run — every queued input has provably drained.
delay(0) macrotask barrier: for gates that run synchronously off input dispatch (double-Escape counting, Ctrl-C/Ctrl-D, submit gates) — one macrotask turn runs strictly after every queued microtask, independent of machine load.

The never-dropped fallback-enqueue test no longer sleeps 2.2s blind: the fake driver counts enqueue attempts, and the test waits for the retries to observably cross the old ~20-attempt give-up budget (the exact regression #1135-era bug shape), still at the real 100ms cadence so a reintroduced time-based budget would also trip it.

Retained real-boundary contracts (unchanged by design)

runSignalExitProbe / runFatalExitProbe: real child processes, real signals, and the real 3s PROCESS_EXIT_GRACE_MS deadline (3.5s + 2×0.5s of the remaining wall time).
The 100ms fallback retry cadence (retry-budget test above).
The 250ms hydration-retry cancellation window — a real-timer negative; now documented as derived from the retry schedule rather than an arbitrary sleep.

Production surface

One additive, optional injection point: MakaPiTuiInput.shellRunTicker?: { now?, schedule? }, threaded into createShellRunElapsedTicker (which already accepted both). Defaults unchanged.

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

	Before	After
Wall	15.5s	12.1s
CPU (user)	4.2s	4.1s

The remaining wall time is dominated by the preserved real-boundary tests listed above; every other test now completes the moment its observed condition holds, and no assertion in the suite races scheduler timing anymore. Full packages/cli suite: 492/494 pass — the 2 failures are pre-existing local noise (run-command.test.ts expects empty stderr; Node 22.17 emits a SQLite ExperimentalWarning) on paths this PR does not touch.

…pletion Part of apache#2389. The Pi TUI runner suite guessed at scheduler timing in three ways, all sensitive to CI load: - 55 unconditional sleeps (delay(20)..delay(2500)) where the awaited fact is observable: startup paints, prompt arming, wizard frames opening/closing, drafts appearing, and retry attempts are all now waited on directly via the existing waitFor helper, a new waitForTuiPaint helper, sentinel-render barriers for must-not-render negatives, and close-anchored assertions for must-never-happen driver calls. - 81 Promise.race close watchdogs with a flat 50/500ms deadline that a loaded runner can miss; they now share CLOSE_BUDGET_MS (WAIT_BUDGET_MS-scaled), so a passing race is unchanged and only failure reporting pays the budget. - The off-screen ticker test slept 2.5s for two real 1s ticks; the shell-run elapsed ticker's clock and interval scheduling are now injectable (MakaPiTuiInput.shellRunTicker, same pattern as attentionLongTurnThresholdMs) and the test drives both ticks itself. Synchronous input-dispatch negatives (double-Escape counting, Ctrl-C/Ctrl-D gating, submit gates) settle on one macrotask turn (delay(0)) instead of a 20-30ms guess, and the never-dropped fallback-enqueue test now waits for the driver to observe the retries crossing the old give-up budget instead of sleeping 2.2s blind. Preserved as real timeouts or real process boundaries: the SIGTERM/fatal exit probes (real child processes and the 3s exit grace), the 100ms fallback retry cadence, and the 250ms hydration-retry cancellation window (now documented as schedule-derived). Focused timing (node --test dist/__tests__/pi-tui-runner.test.js, 142 tests): 15.5s -> 12.1s wall locally, with the remainder dominated by the preserved real-boundary tests, and no fixed-delay races left in the suite.

jackwener merged commit 14df804 into apache:mainAug 7, 2026
12 checks passed

Uh oh!

Conversation

UncertaintyDeterminesYou4ndMe commented Aug 7, 2026

What changed

Retained real-boundary contracts (unchanged by design)

Production surface

Timing (local, node --test dist/__tests__/pi-tui-runner.test.js, 142 tests)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

UncertaintyDeterminesYou4ndMe commented Aug 7, 2026

What changed

Retained real-boundary contracts (unchanged by design)

Production surface

Timing (local, node --test dist/__tests__/pi-tui-runner.test.js, 142 tests)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

UncertaintyDeterminesYou4ndMe commented Aug 7, 2026

What changed

Retained real-boundary contracts (unchanged by design)

Production surface

Timing (local, node --test dist/__tests__/pi-tui-runner.test.js, 142 tests)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

UncertaintyDeterminesYou4ndMe commented Aug 7, 2026

What changed

Retained real-boundary contracts (unchanged by design)

Production surface

Timing (local, node --test dist/__tests__/pi-tui-runner.test.js, 142 tests)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

UncertaintyDeterminesYou4ndMe commented Aug 7, 2026

What changed

Retained real-boundary contracts (unchanged by design)

Production surface

Timing (local, node --test dist/__tests__/pi-tui-runner.test.js, 142 tests)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)

Timing (local, `node --test dist/tests/pi-tui-runner.test.js`, 142 tests)