bench · bez · Tangled

burrito.space / bez

This repository has no description

bez bench

6 folders · 88 files

dietrich ayala · dagu-lane: gzip the source upload and skip it when the lane has the copy · 1d ago

Merge branch 'track/repair-requests' into track/gate-landing-hooks

bench: the box test's canned replies keep is_root and pass every train case

Model variety for generation runs: qwen3.8-27b-fast-2 and deepseek-v4.1-flash

DeepSeek pair waits for its fabric address at boot; GLM disabled on all four ARM nodes

Add rig-capacity benchmark scripts and crate

Drop the GENERA_ fallbacks and refuse every old variable prefix in code

folknet-check.sh: check the rig against a private inventory, --fix repairs

absolute-mixed-batch.sh

Drop the GENERA_ fallbacks and refuse every old variable prefix in code

after-land: reinstall job-check so the submit scripts follow main

Drop the GENERA_ fallbacks and refuse every old variable prefix in code

canned-gateway.py

bench: run the display-contents job's real stages on the box with a canned model

Measure rig capacity: CPU-side benchmarks, GPU-tier formulas, capacity model

Add rig-capacity benchmark scripts and crate

Add rig-capacity benchmark scripts and crate

compile-batch.sh

Measure rig capacity: CPU-side benchmarks, GPU-tier formulas, capacity model

Run model batches only on a lane; probe candidates in one wide build

Run model batches only on a lane; probe candidates in one wide build

dagu-lane: gzip the source upload and skip it when the lane has the copy

dashboard-update-test.sh

Landing refreshes the coverage dashboard

dashboard-update.sh

Landing refreshes the coverage dashboard

e2e-gate-test.sh

e2e-gate: run the feature-job stage tests on the box side

e2e-gate: run the feature-job stage tests on the box side

e2e-gate: run verify and the scenarios at the same time; submit the scenarios four at a time

e2e-lane-clean.sh

e2e-jobs: the cleanup removes the scenarios' stage cache entries; docs say what prunes the cache

env-from-vault-test.sh

env-from-vault: log in again from an access token when the pass-cli session is unusable

env-from-vault.sh

env-from-vault: log in again from an access token when the pass-cli session is unusable

expected-features-check.py

Say after a landing when a job declared features that did not move

feature-announce-test.sh

runs branch: push to the forge and the public remote; link without the @ form

feature-announce.py

feature-announce: the job run time is the running time the wall-clock limit counts, not the calendar span

feature-job-box-test.sh

bench: the feature-job box test takes about six minutes with the workspace suite

folknet_snapshot.py

folknet_snapshot: match only real IPv6 forms in the leftover check

folknet-check.sh

folknet-check: an unreachable node is reported without saying how the login works

folknet-power-probe.py

Merge main into job-parallel-writers

folknet-probe-install.sh

Drop the GENERA_ fallbacks and refuse every old variable prefix in code

folknet-probe.sh

Drop the GENERA_ fallbacks and refuse every old variable prefix in code

folknet-status.sh

folknet-check: a model the gateway does not advertise is reported, not failed

folknet-watch-heartbeat-check.sh

folknet-watch: a heartbeat to the job event store and a lane 0 check that posts when it stops

folknet-watch-heartbeat-install.sh

folknet-watch: a heartbeat to the job event store and a lane 0 check that posts when it stops

folknet-watch-heartbeat-test.sh

folknet-watch: a heartbeat to the job event store and a lane 0 check that posts when it stops

folknet-watch.sh

folknet-watch: WEBENGINE_WATCH_MUTE=yes logs every post instead of sending it

folknet.example.json

folknet-check.sh: check the rig against a private inventory, --fix repairs

Rename neze to webengine in code and bez in prose

forge-credential.sh

Add a git credential helper for the rig forge

fragment-floor.sh

Add bench/land.sh so a tick can land approved spec tracks

gen_candidates.py

Measure rig capacity: CPU-side benchmarks, GPU-tier formulas, capacity model

generated-floors-update.sh

generated floors: reuse the gate's raise lines for the landed tree instead of scoring again

generated-raises.sh

generated floors: reuse the gate's raise lines for the landed tree instead of scoring again

generated-scores-test.sh

generated floors: reuse the gate's raise lines for the landed tree instead of scoring again

generated-scores.sh

generated-scores.sh: shrink each function's cases after its cache is built

job-archive-test.sh

job.yaml: keep job output and archive each run's work tree and candidates to the share

job-events-lane-test.sh

job-events-lane: the boot wrapper reads ~/.bashrc before putting node on the PATH, and the test reads the store PATH from /proc

job-events-lane.sh

job-events-lane: the boot wrapper reads ~/.bashrc before putting node on the PATH, and the test reads the store PATH from /proc

job-start-announce.py

run records and start posts name the models given at submit time

land: an ignored, untracked roster is edited in place, not committed; land-test tracks its roster again

land: an ignored, untracked roster is edited in place, not committed; land-test tracks its roster again

lane-apply-test.sh

tests: suites set the share and refresh switches themselves, never inherit them

bench/nas-refresh.sh: fill the share's npm-cache/ from each lockfile, and install offline on a lane

lane-create-lxc.sh

Build lane: distributed compilation and a shared cache across the EPYC nodes

Build lane: distributed compilation and a shared cache across the EPYC nodes

lane-mirror-cache-test.sh

lane: mirror a rebuilt ground-truth cache onto the shared copy

lane-mirror-cache.sh

lane: mirror a rebuilt ground-truth cache onto the shared copy

docs: the one retention rule for job-run folders replaces the nothing-prunes text

lane-probe-test.sh

on-lane: probe the lane up to 3 times before calling it unreachable

on-lane: probe the lane up to 3 times before calling it unreachable

lane-provision.sh

Rename neze to webengine in code and bez in prose

lane-sweep-test.sh

retention: one rule, newest two per group, for every folder a job run writes

retention: one rule, newest two per group, for every folder a job run writes

lane-writer-fetch.sh

Roster: lane-writer launch and fetch scripts, config gains a lanes[] array

lane-writer-launch-test.sh

Drop the GENERA_ fallbacks and refuse every old variable prefix in code

lane-writer-launch.sh

Drop the GENERA_ fallbacks and refuse every old variable prefix in code

lane.example.json

Roster: lane-writer launch and fetch scripts, config gains a lanes[] array

Announce and ledger: skip a landing or batch already announced or recorded

nas-refresh-cron.sh

bench/nas-refresh-install.sh: run the NAS refresh from a lane's crontab

nas-refresh-install-test.sh

bench/nas-refresh-install.sh: run the NAS refresh from a lane's crontab

nas-refresh-install.sh

nas-refresh: no child process inherits the refresh lock

nas-refresh-test.sh

tests: suites set the share and refresh switches themselves, never inherit them

nas-refresh: no child process inherits the refresh lock

on-lane-test.sh

on-lane: probe the lane up to 3 times before calling it unreachable

on-lane: probe the lane up to 3 times before calling it unreachable

oracle-install.sh

bench/nas-refresh.sh: fill the share's npm-cache/ from each lockfile, and install offline on a lane

oracle-scaling.sh

Measure rig capacity: CPU-side benchmarks, GPU-tier formulas, capacity model

overflow-mixed-run.sh

Drop the GENERA_ fallbacks and refuse every old variable prefix in code

pages-ground-truth-test.sh

generation: pages_ground_truth_cache measures a feature's selected WPT pages into ground-truth cases

probe-replay-measure.sh

Rename neze to webengine in code and bez in prose

lane-sweep: keep only the most recently used copy of a gone track per lane

repair-loop-test.sh

jobs: a feature job's repair below the bar starts another round from its survivors

jobs: a feature job's repair below the bar starts another round from its survivors

rig_throughput.py

Drop the GENERA_ fallbacks and refuse every old variable prefix in code

rig-inventory.example.json

Announce landings that move a web feature forward, with a per-track ledger

rig-snapshot.sh

Model variety for generation runs: qwen3.8-27b-fast-2 and deepseek-v4.1-flash

run-inner-loop.sh

Add rig-capacity benchmark scripts and crate

score-generated-oracle-test.sh

bench: the box test's canned replies keep is_root and pass every train case

generated floors: reuse the gate's raise lines for the landed tree instead of scoring again

web-features-update.sh

Add the web-features map generator, diff tool and landing step