Task runners / same-node experiment

Work has
an order.

Nift, Turbo and Nx run the same project workflow. What happens when tasks are clean, inputs change, branches run together—or a prerequisite fails?

20 → 100 → 1,000 tasks / workers 1, 2, 4

WORKFLOW / 012 CORE TASKSCHECKED

generate

schema→client

check

typesunitlint

produce

webhelperdocs

release

package→integration→verify→release

Branches converge on package. Client feeds docs; helper is independent. Added asset families exercise fan-out and independent lanes. Exact relationships ↓

Nift 4.12.0 finalSource-built snapshot. Programmable .f actions, tracked ordering and file invalidation.

Turbo 2.11.7Package-script tasks, explicit pipeline and native local task cache.

Nx 23.3.0Project targets, run-commands executor and native local task cache.

1,434 measured observations93 runner cells93 retained warmups1 fresh shared-CPU node

At 1,000 mixed tasks / j4, unchanged CLI medians were Nift 118.418 ms, Turbo 436.115 ms, Nx 1,736.465 ms. All-dirty medians were Nift 13.410 s, Turbo 112.412 s, Nx 17.838 s. These are descriptive observations of this workflow; the full-dirty cells have three measured samples.

01 / workflow

One workflow. Three native descriptions.

The common quantitative graph uses identical Python task bodies: content-hashed JSON outputs, text transforms, checks, one real helper compilation/execution, deterministic in-memory packaging and release validation. Additional asset tasks expand the graph. Each successful check produces a receipt in all runners.

Task names model workflow roles. Checks assert known properties of a 32-record input; web/docs transform text; the helper compiles a trivial C program; package validates an in-memory tar. No real TypeScript compiler, application test suite, framework build or deployment is measured.

Cached check receipts are not always-run, outputless checks. This is a synthetic single-project workflow—not a production JavaScript framework, package discovery benchmark, or remote-cache comparison. Outputless APIs are outside this refresh’s matched capability certification.

Core ordering; task inputs additionally declare consumed artifact bytes
Task	Prerequisites
generate-schema	none
generate-client	generate-schema
typecheck / unit-test	generate-client
lint / build-helper	none
build-web	typecheck, lint, unit-test
docs	generate-client
package	build-web, build-helper, docs
integration-test	package
verify	integration-test
release-prep	verify

02 / noop

Nothing changed. State still costs time.

Every task body must remain unexecuted. Nift checks retained tracked input/output state; Turbo and Nx report native local cache hits with output files already present. These are different skip mechanisms, measured as fresh CLI invocations over prepared state. Nx daemon and remote caching are disabled.

Unchanged mixed workflow; Nift, Turbo, Nx — Unchanged mixed workflow. Both axes logarithmic. Lines join measured points, not a fitted model. Same cohort, ordering, units and medians in chart and table. SVG / PNG

Unchanged mixed workflow; median ms; j4
Tasks	Nift ms	Turbo ms	Nx ms	n / cell
20	7.432	58.387	616.196	50
100	16.354	71.943	703.386	50
1,000	118.418	436.115	1,736.465	50

03 / leaf

One leaf changes. Selection stays explicit.

Changing one independent asset executes exactly that task while the full graph remains in view. The targeted comparisons start with dirty selected tasks, so Nift’s force-target behavior does not add unneeded work. Release selection includes its prerequisite closure; an untimed full invocation subsequently refreshes unselected consumers.

One dirty leaf; Nift, Turbo, Nx — One dirty leaf. Value axis starts at zero; task-count spacing is logarithmic. Lines join measured points, not a fitted model. Same cohort, ordering, units and medians in chart and table. SVG / PNG

One dirty leaf; median ms; j4
Tasks	Nift ms	Turbo ms	Nx ms	n / cell
20	56.099	414.070	689.275	20
100	65.364	431.787	767.165	20
1,000	172.437	747.922	1,878.946	20

20 mixed tasks; j4; median ms
Case	Nift ms	Turbo ms	Nx ms	Executed	n / cell
Targeted leaf	52.611	404.850	620.304	1	10
Targeted release closure	393.005	2,972.042	1,048.402	10	10
Two targets	54.834	435.153	641.712	2	10

Nift named targets are forced rather than a clean “ensure” operation. The graph’s worker limit is configured in the project; the frozen adapter does not use per-invocation worker overrides; this is not an exhaustive current-API capability test.

04 / fanout

A changed input reaches its consumers.

The 10% case changes independent asset inputs. The generated-prerequisite case changes schema input and requires the complete transitive consumer set, including schema-dependent added assets. Ordering alone is insufficient: consumed output bytes are declared as invalidation inputs.

10% task inputs dirty; Nift, Turbo, Nx — 10% task inputs dirty. Value axis starts at zero; task-count spacing is logarithmic. Lines join measured points, not a fitted model. Same cohort, ordering, units and medians in chart and table. SVG / PNG

10% task inputs dirty; median ms; j4
Tasks	Nift ms	Turbo ms	Nx ms	n / cell
20	58.449	442.417	701.240	10
100	163.845	1,345.675	928.215	5
1,000	1,449.169	11,544.650	3,372.805	3

Generated prerequisite dirty; Nift, Turbo, Nx — Generated prerequisite dirty. Value axis starts at zero; task-count spacing is logarithmic. Lines join measured points, not a fitted model. Same cohort, ordering, units and medians in chart and table. SVG / PNG

Generated prerequisite dirty; median ms; j4
Tasks	Nift ms	Turbo ms	Nx ms	n / cell
20	412.478	3,127.584	1,121.103	10
100	903.696	7,478.441	1,554.100	5
1,000	6,798.757	57,030.636	9,865.574	3

05 / full

All inputs dirty. Every task runs.

Unique bytes change in every task input. Native caching remains enabled, including cache writes, but the exact execution oracle requires all tasks to run. These light controlled bodies do not estimate savings in compiler- or application-test-dominated projects. A separate 20-task cohort explicitly disables Turbo/Nx task-cache reads and writes and forces Nift work; these modes are not pooled.

All mixed-task inputs dirty; Nift, Turbo, Nx — All mixed-task inputs dirty. Value axis starts at zero; task-count spacing is logarithmic. Lines join measured points, not a fitted model. Same cohort, ordering, units and medians in chart and table. SVG / PNG

All mixed-task inputs dirty; median s; j4
Tasks	Nift s	Turbo s	Nx s	n / cell
20	0.464	3.601	1.125	10
100	1.425	12.209	2.287	5
1,000	13.410	112.412	17.838	3

20 mixed tasks; j4; median ms
Case	Nift ms	Turbo ms	Nx ms	Executed	n / cell
Cache-disabled full	442.104	3,555.334	1,089.735	20	10

Full-dirty throughput; task count / median wall seconds; j4
Graph	Nift tasks/s	Turbo tasks/s	Nx tasks/s
mixed / 20	43.060	5.554	17.775
mixed / 100	70.172	8.191	43.723
mixed / 1000	74.568	8.896	56.059
tiny / 100	763.722	9.988	67.133
tiny / 1000	651.692	10.173	93.544

06 / parallel

Independent branches share the worker budget.

One hundred mixed tasks execute with limits of one, two and four. Task-body start/complete events independently verify ordering and overlap. These limits control task bodies; they do not imply an equal total operating-system process count because native launch layers differ.

Mixed workflow worker scaling; Nift, Turbo, Nx — Mixed workflow worker scaling. Value axis starts at zero; worker-count spacing is logarithmic. Lines join measured points, not a fitted model. Same cohort, ordering, units and medians in chart and table. SVG / PNG

Mixed workflow worker scaling; median s; 100 mixed tasks
Workers	Nift s	Turbo s	Nx s	n / cell
1	4.716	35.701	6.139	5
2	2.525	19.458	3.586	5
4	1.425	12.209	2.287	5

Maximum observed task-body overlap, 100 mixed tasks, full dirty
Worker limit	Nift	Turbo	Nx
1	1	1	1
2	2	2	2
4	3	2	3

07 / failure

A failure blocks descendants, not independent work.

Semantic certification injects a schema failure, requires a nonzero exit, blocks its descendants and checks independent branch completion. Recovery verifies all expected work and output bytes. These feature tests are separate from successful timing distributions.

Observed failure and retry behavior
Runner	Failure / recovery
Nift	Names the failed custom action; independent branches complete. A normal retry reports unfinished build; build --repair is required.
Turbo	Failed npm task exits nonzero; dependencies-successful continuation permits independent work and blocks dependent tasks. Fix then rerun.
Nx	Target failure produces a failed-task summary and nonzero status. Independent targets complete; dependent targets do not run. Fix then rerun.

Actual stdout/stderr and start/complete/fail traces are retained with the semantic observations. Exit-code values differ and are not treated as a performance result. Turbo/Nx recovery is tested with full forced execution; Nift uses repair. These tests certify correct retry, not the fastest default incremental recovery path.

08 / tiny

Tiny work exposes orchestration and launch cost.

The same compiled C body hashes a small input and writes one deterministic output. Independent 100- and 1,000-task graphs exercise unchanged, one-task, 1%, 10% and full-dirty cases. Turbo retains its npm script launch layer, Nx uses run-commands, and Nift uses cmd: this is not isolated scheduler CPU.

Unchanged tiny-task graph; Nift, Turbo, Nx — Unchanged tiny-task graph. Both axes logarithmic. Lines join measured points, not a fitted model. Same cohort, ordering, units and medians in chart and table. SVG / PNG

Unchanged tiny-task graph; median ms; j4
Tasks	Nift ms	Turbo ms	Nx ms	n / cell
100	14.090	70.034	656.822	50
1,000	98.629	412.628	1,554.188	50

One dirty tiny task; Nift, Turbo, Nx — One dirty tiny task. Value axis starts at zero; task-count spacing is logarithmic. Lines join measured points, not a fitted model. Same cohort, ordering, units and medians in chart and table. SVG / PNG

One dirty tiny task; median ms; j4
Tasks	Nift ms	Turbo ms	Nx ms	n / cell
100	18.794	385.966	692.073	20
1,000	104.292	683.510	1,646.318	20

10% tiny-task inputs dirty; Nift, Turbo, Nx — 10% tiny-task inputs dirty. Value axis starts at zero; task-count spacing is logarithmic. Lines join measured points, not a fitted model. Same cohort, ordering, units and medians in chart and table. SVG / PNG

10% tiny-task inputs dirty; median ms; j4
Tasks	Nift ms	Turbo ms	Nx ms	n / cell
100	27.705	1,187.426	758.190	5
1,000	238.544	10,095.797	2,566.731	3

Fully dirty tiny-task graph; Nift, Turbo, Nx — Fully dirty tiny-task graph. Value axis starts at zero; task-count spacing is logarithmic. Lines join measured points, not a fitted model. Same cohort, ordering, units and medians in chart and table. SVG / PNG

Fully dirty tiny-task graph; median s; j4
Tasks	Nift s	Turbo s	Nx s	n / cell
100	0.131	10.012	1.490	5
1,000	1.534	98.304	10.690	3

1,000 tiny tasks; j4; median ms
Case	Nift ms	Turbo ms	Nx ms	Executed	n / cell
Unchanged	98.629	412.628	1,554.188	0	50
One leaf dirty	104.292	683.510	1,646.318	1	20
1% inputs dirty	117.400	1,432.019	1,761.650	10	3
10% inputs dirty	238.544	10,095.797	2,566.731	100	3
All inputs dirty	1,534.467	98,304.240	10,690.135	1000	3

The optional 10,000-task cohort was excluded after local feasibility because repeated launch-heavy runs were disproportionate to this bounded diagnostic. This is a runtime-budget decision, not evidence of a graph-size capability limit.

09 / inputs

Inputs, outputs and cache are distinct contracts.

The frozen adapter declares file inputs and artifact ownership separately from task completion order. Nift validates retained state; Turbo/Nx use their native local result caches. Incremental skipping is not restoring missing outputs. Remote caching and daemon-assisted execution are excluded.

Capability evidence boundaries
Question	What this refresh establishes
Artifact-producing DAG	All three execute equivalent input/output/order contracts; exact action and output oracles pass.
Outputless checks / environment inputs	Not recertified as a matched new capability comparison. Receipt-producing checks in this adapter must not be presented as proof of every current API limit.
Task-result restoration	Native peer cache hits and prepared outputs are audited; no new Nift result-restoration implementation is claimed.
Ordering vs invalidation	Completion prerequisites and declared consumed file bytes remain distinct facts.
Remote cache / project inference	Not measured; choose a tool for these requirements using its own capability contract.

Supplemental comparisons not recertified by this suite are excluded from current results.

10 / resources

Resource figures keep their measurement boundary.

Every observation retains C-supervised wall time, CPU time and Linux wait4 maximum-child RSS. RSS is a child high-water measure—not summed simultaneous memory across parallel task/Node/compiler processes. The same incremental workloads appear in timing and memory panels.

Unchanged workflow memory. Value axis starts at zero; task-count spacing is logarithmic. Lines join measured points, not a fitted model. Same cohort, ordering, units and medians in chart and table. SVG / PNG

Unchanged workflow memory; median MiB; j4
Tasks	Nift MiB	Turbo MiB	Nx MiB	n / cell
20	5.750	46.234	108.736	50
100	5.875	49.959	115.479	50
1,000	6.969	108.133	165.766	50

One dirty leaf memory. Value axis starts at zero; task-count spacing is logarithmic. Lines join measured points, not a fitted model. Same cohort, ordering, units and medians in chart and table. SVG / PNG

One dirty leaf memory; median MiB; j4
Tasks	Nift MiB	Turbo MiB	Nx MiB	n / cell
20	17.375	80.863	108.062	20
100	17.375	80.846	115.943	20
1,000	18.000	112.090	165.268	20

All dirty workflow memory. Value axis starts at zero; task-count spacing is logarithmic. Lines join measured points, not a fitted model. Same cohort, ordering, units and medians in chart and table. SVG / PNG

All dirty workflow memory; median MiB; j4
Tasks	Nift MiB	Turbo MiB	Nx MiB	n / cell
20	20.000	81.721	110.617	10
100	20.000	82.199	120.230	5
1,000	19.875	113.047	170.570	3

11 / correctness

A fast skipped task is a rejected result.

All 1,527 accepted observations pass exact task-start/completion/output, prerequisite order, concurrency, rotation, raw-metric and distribution audits. Every warmup remains. A first attempt reused preflight caches and was rejected at the first changed leaf when Turbo restored a cached result rather than executing the required action. Its 153 observations and failed command are preserved separately; none enter these distributions. The accepted run starts at the first cell in a fresh exact suite checkout with empty fixture/project caches.

The independently certified 21 semantic observations check failure, blocked descendants, independent completion and recovery. Outputless/environment behavior is outside the matched performance matrix. The unchanged refresh runner recognizes the accepted Nx cache-disabled receipt; this is a full new run, not a continuation or pooled partial series.

23 of 93 cells have a min–max span greater than 25% of their median. This is a descriptive spread flag, not a confidence interval. Sparse three-sample cells do not support tail or significance claims.

All 93 cells: timing distributions, CPU, memory and executed counts

All observations retained; p95 only n ≥ 20; interpolated percentiles
Case	Runner	n	Median ms	Min–max ms	p95 ms	CPU ms	RSS MiB	Executed
mixed / 20 / j4 / 10% inputs dirty	Nift	10	58.449	55.854–67.660	—	113.265	17.441	2
mixed / 20 / j4 / 10% inputs dirty	Nx	10	701.240	661.003–720.747	—	1,388.121	107.732	2
mixed / 20 / j4 / 10% inputs dirty	Turbo	10	442.417	425.656–456.883	—	975.722	80.605	2
mixed / 20 / j4 / All inputs dirty	Nift	10	464.470	439.748–612.463	—	1,071.330	20.000	20
mixed / 20 / j4 / All inputs dirty	Nx	10	1,125.195	1,050.379–1,216.618	—	2,456.672	110.617	20
mixed / 20 / j4 / All inputs dirty	Turbo	10	3,601.172	3,539.753–3,866.850	—	8,738.272	81.721	20
mixed / 20 / j4 / One leaf dirty	Nift	20	56.099	53.278–88.290	62.342	61.782	17.375	1
mixed / 20 / j4 / One leaf dirty	Nx	20	689.275	643.684–737.027	716.816	1,333.051	108.062	1
mixed / 20 / j4 / One leaf dirty	Turbo	20	414.070	398.948–547.051	499.485	509.730	80.863	1
mixed / 20 / j4 / Two targets	Nift	10	54.834	52.653–59.430	—	103.413	17.375	2
mixed / 20 / j4 / Two targets	Nx	10	641.712	624.986–660.636	—	1,323.990	108.129	2
mixed / 20 / j4 / Two targets	Turbo	10	435.153	421.275–476.955	—	953.070	80.471	2
mixed / 20 / j4 / Unchanged	Nift	50	7.432	6.480–10.988	9.794	12.273	5.750	0
mixed / 20 / j4 / Unchanged	Nx	50	616.196	574.520–777.118	650.593	819.171	108.736	0
mixed / 20 / j4 / Unchanged	Turbo	50	58.387	54.027–77.297	69.050	74.827	46.234	0
mixed / 20 / j4 / Generated prerequisite dirty	Nift	10	412.478	386.153–445.961	—	732.500	17.562	14
mixed / 20 / j4 / Generated prerequisite dirty	Nx	10	1,121.103	1,083.608–1,166.944	—	2,118.177	109.047	14
mixed / 20 / j4 / Generated prerequisite dirty	Turbo	10	3,127.584	3,034.554–3,424.482	—	6,176.991	81.705	14
mixed / 20 / j4 / Targeted leaf	Nift	10	52.611	51.333–65.399	—	51.831	17.375	1
mixed / 20 / j4 / Targeted leaf	Nx	10	620.304	610.238–667.823	—	1,234.264	106.910	1
mixed / 20 / j4 / Targeted leaf	Turbo	10	404.850	391.509–578.525	—	493.143	81.090	1
mixed / 20 / j4 / Targeted release closure	Nift	10	393.005	382.501–436.912	—	493.073	17.500	10
mixed / 20 / j4 / Targeted release closure	Nx	10	1,048.402	1,032.050–1,149.503	—	1,769.548	106.852	10
mixed / 20 / j4 / Targeted release closure	Turbo	10	2,972.042	2,874.247–3,027.004	—	4,455.878	81.518	10
mixed / 20 / j4 / Cache-disabled full	Nift	10	442.104	421.346–460.946	—	1,016.784	20.000	20
mixed / 20 / j4 / Cache-disabled full	Nx	10	1,089.735	1,055.418–1,139.032	—	2,366.720	108.924	20
mixed / 20 / j4 / Cache-disabled full	Turbo	10	3,555.334	3,491.139–4,059.638	—	8,648.149	81.598	20
mixed / 100 / j1 / All inputs dirty	Nift	5	4,716.445	4,688.667–4,831.884	—	4,718.760	20.000	100
mixed / 100 / j1 / All inputs dirty	Nx	5	6,139.473	6,090.862–6,537.112	—	6,794.891	120.109	100
mixed / 100 / j1 / All inputs dirty	Turbo	5	35,700.830	35,268.800–36,438.313	—	43,379.208	82.461	100
mixed / 100 / j2 / All inputs dirty	Nift	5	2,525.478	2,482.060–2,554.872	—	4,815.844	20.000	100
mixed / 100 / j2 / All inputs dirty	Nx	5	3,586.496	3,460.881–3,703.912	—	7,051.815	118.227	100
mixed / 100 / j2 / All inputs dirty	Turbo	5	19,458.066	19,399.598–20,268.609	—	43,702.884	82.145	100
mixed / 100 / j4 / 10% inputs dirty	Nift	5	163.845	161.577–173.394	—	521.474	17.484	10
mixed / 100 / j4 / 10% inputs dirty	Nx	5	928.215	859.667–929.475	—	1,984.184	115.484	10
mixed / 100 / j4 / 10% inputs dirty	Turbo	5	1,345.675	1,324.245–1,440.487	—	4,510.022	81.320	10
mixed / 100 / j4 / All inputs dirty	Nift	5	1,425.074	1,407.273–1,527.202	—	4,853.412	20.000	100
mixed / 100 / j4 / All inputs dirty	Nx	5	2,287.110	2,190.814–2,390.148	—	6,975.336	120.230	100
mixed / 100 / j4 / All inputs dirty	Turbo	5	12,208.815	12,139.558–12,467.575	—	42,457.598	82.199	100
mixed / 100 / j4 / One leaf dirty	Nift	20	65.364	61.579–76.691	75.230	87.292	17.375	1
mixed / 100 / j4 / One leaf dirty	Nx	20	767.165	725.741–1,001.403	837.243	1,433.155	115.943	1
mixed / 100 / j4 / One leaf dirty	Turbo	20	431.787	401.255–545.077	461.535	555.150	80.846	1
mixed / 100 / j4 / Unchanged	Nift	50	16.354	14.115–19.268	18.611	38.486	5.875	0
mixed / 100 / j4 / Unchanged	Nx	50	703.386	642.050–877.500	746.153	938.307	115.479	0
mixed / 100 / j4 / Unchanged	Turbo	50	71.943	67.381–88.782	80.284	109.675	49.959	0
mixed / 100 / j4 / Generated prerequisite dirty	Nift	5	903.696	875.765–1,159.209	—	2,648.059	17.500	54
mixed / 100 / j4 / Generated prerequisite dirty	Nx	5	1,554.100	1,541.022–1,578.031	—	4,299.430	116.789	54
mixed / 100 / j4 / Generated prerequisite dirty	Turbo	5	7,478.441	7,316.147–7,528.788	—	23,104.817	82.449	54
mixed / 1,000 / j4 / 10% inputs dirty	Nift	3	1,449.169	1,437.979–1,479.934	—	5,471.860	18.000	100
mixed / 1,000 / j4 / 10% inputs dirty	Nx	3	3,372.805	3,359.855–3,460.728	—	8,743.753	166.848	100
mixed / 1,000 / j4 / 10% inputs dirty	Turbo	3	11,544.650	11,287.011–11,549.681	—	44,118.034	112.102	100
mixed / 1,000 / j4 / All inputs dirty	Nift	3	13,410.497	13,227.737–13,545.120	—	51,619.212	19.875	1000
mixed / 1,000 / j4 / All inputs dirty	Nx	3	17,838.287	17,368.879–18,100.818	—	61,772.930	170.570	1000
mixed / 1,000 / j4 / All inputs dirty	Turbo	3	112,412.417	111,816.614–112,537.266	—	432,836.734	113.047	1000
mixed / 1,000 / j4 / One leaf dirty	Nift	20	172.437	161.334–188.555	184.440	385.038	18.000	1
mixed / 1,000 / j4 / One leaf dirty	Nx	20	1,878.946	1,834.386–2,124.401	2,038.956	2,990.431	165.268	1
mixed / 1,000 / j4 / One leaf dirty	Turbo	20	747.922	703.977–814.099	792.800	1,576.822	112.090	1
mixed / 1,000 / j4 / Unchanged	Nift	50	118.418	107.218–173.430	131.746	332.211	6.969	0
mixed / 1,000 / j4 / Unchanged	Nx	50	1,736.465	1,634.803–2,030.220	1,859.865	2,342.836	165.766	0
mixed / 1,000 / j4 / Unchanged	Turbo	50	436.115	399.335–773.195	504.104	1,146.745	108.133	0
mixed / 1,000 / j4 / Generated prerequisite dirty	Nift	3	6,798.757	6,754.315–6,869.772	—	25,901.754	18.125	504
mixed / 1,000 / j4 / Generated prerequisite dirty	Nx	3	9,865.574	9,591.053–10,119.385	—	32,350.729	168.477	504
mixed / 1,000 / j4 / Generated prerequisite dirty	Turbo	3	57,030.636	56,937.077–58,053.630	—	216,071.479	118.062	504
tiny / 100 / j4 / 10% inputs dirty	Nift	5	27.705	25.560–28.753	—	64.979	7.625	10
tiny / 100 / j4 / 10% inputs dirty	Nx	5	758.190	710.772–774.041	—	1,417.798	115.301	10
tiny / 100 / j4 / 10% inputs dirty	Turbo	5	1,187.426	1,144.975–1,193.217	—	3,949.316	81.082	10
tiny / 100 / j4 / All inputs dirty	Nift	5	130.938	116.224–140.168	—	395.488	7.625	100
tiny / 100 / j4 / All inputs dirty	Nx	5	1,489.571	1,393.462–1,550.249	—	2,218.322	117.496	100
tiny / 100 / j4 / All inputs dirty	Turbo	5	10,011.869	9,783.629–10,165.612	—	38,337.891	81.883	100
tiny / 100 / j4 / One leaf dirty	Nift	20	18.794	17.499–22.907	21.856	32.292	7.500	1
tiny / 100 / j4 / One leaf dirty	Nx	20	692.073	669.927–738.196	732.285	1,141.396	112.203	1
tiny / 100 / j4 / One leaf dirty	Turbo	20	385.966	362.545–416.346	411.659	506.441	80.934	1
tiny / 100 / j4 / Unchanged	Nift	50	14.090	12.752–20.642	16.361	27.761	5.875	0
tiny / 100 / j4 / Unchanged	Nx	50	656.822	626.428–827.572	742.879	894.064	111.928	0
tiny / 100 / j4 / Unchanged	Turbo	50	70.034	65.318–84.789	79.224	108.013	49.906	0
tiny / 100 / j4 / 1% inputs dirty	Nift	5	18.318	17.086–18.894	—	31.294	7.500	1
tiny / 100 / j4 / 1% inputs dirty	Nx	5	691.039	670.109–716.734	—	1,152.569	112.059	1
tiny / 100 / j4 / 1% inputs dirty	Turbo	5	382.282	360.968–394.894	—	501.962	81.070	1
tiny / 1,000 / j4 / 10% inputs dirty	Nift	3	238.544	237.419–251.027	—	692.323	8.863	100
tiny / 1,000 / j4 / 10% inputs dirty	Nx	3	2,566.731	2,476.270–2,566.864	—	3,636.693	158.535	100
tiny / 1,000 / j4 / 10% inputs dirty	Turbo	3	10,095.797	10,090.992–10,294.454	—	38,709.316	108.469	100
tiny / 1,000 / j4 / All inputs dirty	Nift	3	1,534.467	1,494.854–1,631.212	—	4,807.637	9.113	1000
tiny / 1,000 / j4 / All inputs dirty	Nx	3	10,690.135	10,148.098–10,740.128	—	12,647.889	163.527	1000
tiny / 1,000 / j4 / All inputs dirty	Turbo	3	98,304.240	97,869.821–98,576.084	—	383,624.039	115.102	1000
tiny / 1,000 / j4 / One leaf dirty	Nift	20	104.292	96.783–130.515	121.849	241.938	8.361	1
tiny / 1,000 / j4 / One leaf dirty	Nx	20	1,646.318	1,565.497–1,912.412	1,778.445	2,471.122	156.811	1
tiny / 1,000 / j4 / One leaf dirty	Turbo	20	683.510	645.701–805.022	733.598	1,464.522	110.053	1
tiny / 1,000 / j4 / Unchanged	Nift	50	98.629	89.516–126.818	109.456	241.268	6.484	0
tiny / 1,000 / j4 / Unchanged	Nx	50	1,554.188	1,486.365–1,650.379	1,617.413	2,151.386	157.449	0
tiny / 1,000 / j4 / Unchanged	Turbo	50	412.628	360.601–1,243.480	574.317	1,093.295	107.902	0
tiny / 1,000 / j4 / 1% inputs dirty	Nift	3	117.400	114.357–124.691	—	293.366	8.863	10
tiny / 1,000 / j4 / 1% inputs dirty	Nx	3	1,761.650	1,744.474–1,762.258	—	2,771.659	156.129	10
tiny / 1,000 / j4 / 1% inputs dirty	Turbo	3	1,432.019	1,428.106–1,463.275	—	4,789.735	114.340	10

12 / practitioner

AI-DX changes the choice—but not the evidence.

Qualitative practitioner judgement / observed contracts kept separate

For human-plus-agent and agent-primary bespoke authored workflows, I would choose Nift. The same applies to mixed website/assets/helpers/tasks projects owned by Nift. All 31 measured configurations have the lowest Nift wall and maximum-child RSS medians here; programmable .f policy and explicit ownership, ordering, invalidation and lifecycle also matter.

I would choose Turbo/Nx for specific task-result restoration, remote-cache/platform or framework/workspace integration requirements. Those are concrete needs, not an architectural reason to default away from Nift for new bespoke work. The timed adapter does not establish the absence of every alternative current API.

AI-DX evidence classes
Evidence	Conclusion
Demonstrated	Exact actions, output bytes, completion order, bounded overlap, targeting and failure/recovery checks pass.
Source / representation	Separate artifact/input/order declarations make semantic roles inspectable.
Qualitative inference	That explicit structure can help agents make bounded, reviewable edits; helpers should remove duplicated boilerplate without collapsing semantics.
Not measured	Autonomous authoring success, repair rate, token/time cost, arbitrary-edit safety or a complete product-native structured error contract.

No numeric productivity or safety score is assigned. Harness JSON is not product-native structured feedback. Missing comparative probes do not establish competitor incapability.

13 / method

Frozen sources. Retained samples. One node.

One fresh 4-logical-CPU / 8-GB Ubuntu 24.04 Linode, us-east. CPU affinity includes all four CPUs. Runner order rotates within rounds; scenario order follows a fixed frozen plan. Fresh CLI invocation, prepared local state, no remote cache, Nx daemon disabled. Setup/input mutations are outside the timer; native dispatch, task work and bookkeeping are inside. Stdout/stderr are captured: native status formatting and cache-log costs are included; terminal painting is excluded. Turbo uses errors-only task logs and Nx uses static output. These CLI modes are retained with each command, not presented as equal pure scheduler CPU. No outlier removal.

The mixed body reads the canonical graph for each invocation; that identical cost is included. The suite covers 20/100/1,000 mixed tasks, 100/1,000 tiny tasks, j4 scaling, 100-task j1/j2 full-dirty work and 20-task cache-disabled/targeting cohorts. Failure and unsupported semantics are separate capability tests. Shared-CPU variance, sparse expensive cells and this synthetic single-project model limit generalization.

Nift source: c14d06835bca515a5281e41da0e40219c2d60ef0
Measured suite: f37b63bd13cf71256607ba4652bcc911e270f94e
Toolchain: Node 22.22.1 / npm 9.2.0 / Turbo 2.11.7 / Nx 23.3.0 / g++ (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
Machine: AMD EPYC 7642 48-Core Processor / 6.8.0-134-generic / ext4; prepared local fixtures
Canonical evidence: Raw samples, traces, provenance, semantic probes, audit and lifecycle
Core follow-up: Reproducible semantic/DX gaps; Nift core was not changed.
Reproduce: Fairness contract · Sample plan · Independent auditor

Actual invocations and reusable Nift action

Representative 20-task unchanged command; worker count four. Nx daemon/cloud disabled; remote cache disabled. Each observation retains its complete invocation.

Nift: /opt/nift/nift build
Turbo: /opt/suites/task-runners-official/work/tools/node_modules/.bin/turbo run generate-schema generate-client lint typecheck unit-test build-web build-helper docs package integration-test verify release-prep asset-00000 asset-00001 asset-00002 asset-00003 asset-00004 asset-00005 asset-00006 asset-00007 --concurrency=4 --cache=local:rw --continue=dependencies-successful --output-logs=errors-only
Nx: /opt/suites/task-runners-official/work/tools/node_modules/.bin/nx run-many --projects=workflow --targets=generate-schema,generate-client,lint,typecheck,unit-test,build-web,build-helper,docs,package,integration-test,verify,release-prep,asset-00000,asset-00001,asset-00002,asset-00003,asset-00004,asset-00005,asset-00006,asset-00007 --parallel=4 --outputStyle=static

The Nift tracked graph uses one reusable supported .f hook:

r := cmd("python3", "task_body.py", getenv("NIFT_HOOK_TARGET")).cwd(getenv("NIFT_HOOK_ROOT")).run()
if(!r.launched || r.exit_code != 0) { throw error(r.stderr, "user.task.failed") }

Direct lint/test/package/release .f examples are separately certified over prepared prerequisites; they do not implement an adapter scheduler.

Work hasan order.