`agentic-agents-core-models` carried the 36s scaled estimate while measuring
56.3s (n=6, p90 58.6s) across 260 compact jobs on 2026-08-16. Packed beside
`agentic-agents-core-runtime-hosted-1` (60.4s) it built the only bin running
>=1.25x its prediction: compact-large-19 ran 122s of work priced at 88s, which
made it the most frequent slowest job in the workflow.
Add the measured value to the hybrid hint map, which already exists for exactly
these Blacksmith-specific observations. The packer then separates the two, and
the tallest actual compact bin drops 122s -> 109s with no bin above 1.25x and
no change to row counts (48 push / 56 pull-request), so this costs no extra
runner registrations.
Two things deliberately left alone. The `agentic-gateway-core-3` 140s pin looks
like a 2x over-prediction against its 68.5s median, but run 31938297538 really
did take 138.0s (109.03s + 28.98s across its two configs) on a run whose fleet
slowdown factor was 0.98 -- a genuine tail on a healthy run, so unpinning it
would rebuild a >200s bin whenever the tail lands. Every other bin already sits
within 1.24x of its prediction.
The replaced guard pinned one bin arrangement (`runtime-hosted-1` not sharing a
job with `agents-core-tools`) with no stated failure mode, so any honest refit
broke it. It now asserts the property that comment was reaching for -- both
weight sources survive rebalancing under the body ceiling -- plus the specific
regression this fixes. Mutation-checked: dropping the hint fails the new guard.
Measurement note for the next refit: sum a shard's per-config Duration lines
before taking a median. Pooling them reads as a large over-prediction that is
not there.
* perf(ci): cut hosted CI critical path toward five-minute walls
Stripe the serial core test-type graphs across two hosted jobs and drop the
duplicated tsgo:test:root pass; gate the six-part QA Smoke matrix off pull
requests unless a QA-owned surface changed; split the fat multi-config Node
shards (cli/cli-process, unit-fast isolated/fake-timers, infra
logging-process/runtime-config) and lower the hosted split ceiling to 150
predicted seconds so no compact lane owns a ~280s wall; expand tooling to
seven stripes.
* perf(ci): widen hosted test-type striping to three jobs
Run 31825922122 measured ~40s per core test-type graph on loaded hosted
runners (282s worst stripe body of the two-way split); three stripes keep
each lane near 150s body under load.
* ci: run push compacts without tooling lanes and stripe giants three-way
* fix(ci): keep CompactNodeTestPlanMode internal
knip --production flags the exported type as unused; no external module
imports it.
* ci: shard hosted lint and node tests
Split core oxlint across hosted runners and use backend-specific compact Node admission caps while preserving Blacksmith behavior.
* test(ci): guard plugin prerelease planner isolation
* refactor(agents): move announce family into subagents/announce/
* refactor(agents): move completion and main-session recovery families
* refactor(agents): move registry family into subagents/registry
* refactor(agents): update registry state type imports
* style(agents): format moved subagent imports
* test(agents): preserve isolated registry test routing
* fix(scripts): recognize relocated subagent announce seams
* fix(tui): cancel buffered submissions on shutdown
* refactor(tui): preserve callable submit coalescer
* test(tui): run local PTY smoke against built CLI
The core-runtime-media-ui shard runs its ui config non-isolated for speed, but the
full chat-pane lifecycle tests instantiate the pane component, which relies on
chat-thread/chat-message module-level singletons (thread-state map, confirmation
dismisser WeakMap, module-scoped document context-menu listeners) and spies on those
modules. Under the shared non-isolated graph a stateful predecessor file can leave
those modules duplicated, so the pane binds to a different instance than the test's
spy/registry -- producing order-dependent flakes: `removeEventListener`-not-called
teardown assertions or 120s session-lifecycle hangs. Reproduced deterministically on
a Linux Node 24 Testbox (MAX_WORKERS=1); isolating a single file only shifted the
failure to a sibling pane test, so the whole full-pane family shares the fragility.
Route the 7 full-pane lifecycle test files through a new isolated jsdom lane
(vitest.ui-isolated.config.ts, isolate: true) for a fresh module graph; the other
~370 ui tests stay fast and non-isolated. Registered in both shard registries
(ci-node-test-plan.mjs, vitest.test-shards.mjs) and excluded from vitest.ui.config.ts.
Verified on Testbox: the deterministic single-worker media-ui shard goes from failing
to 5182 passed / 0 failed. Test-infrastructure only; no product code changes.
Group walls measured from compact run 29564411446 (post-#109769):
core-unit-fast-isolated runs 78.6s on 4 vCPU because fork-per-file
isolation parallelizes poorly there; pin it to the 8 vCPU class where
the same segment runs ~50s. Unit-fast stripes measure ~25-37s (hinted
100/60) and tooling stripes 71-87s; refresh both so the packer stops
overfilling small bins (the 299s compact tail this run produced).
Cheap stripes may now co-locate in one bin, so the compact test only
requires their presence.
* perf(sdk): load Claude CLI identity from a narrow plugin artifact
src/plugin-sdk/anthropic-cli.ts snapshots CLAUDE_CLI_BACKEND_ID at module
scope through the sync facade loader, which jiti-evaluates the full
anthropic api.js barrel on source checkouts: 130.86s of self time per
cold worker on CI (Testbox import profile), silently stalling every job
whose graph reaches cli-runner/prepare.ts. A narrow cli-api.js artifact
carries the two static facts; Testbox proof: the reliability+helpers
agents-core pair drops from 157.6s to 14.3s.
* perf(ci): stripe unit-fast and tooling node tests on the full plan
core-fast ran the import-bound unit-fast graph as one job (247s vitest,
181s module evaluation) and core-tooling as one serial job (241s); both
now stripe on the full plan like the compact plan, so the compact-only
expansion is deleted and the docker helper config rides with the
isolated shard on both plans. Group hints refreshed from main run
29551077288; stale per-file commands hints dropped (the consolidation
landed); cache-writer selection and the warm workflow track the stripe
names.
* fix(state): tolerate vanished sqlite sidecars in agent-db permission sweep
existsSync+chmodSync raced SQLite's own WAL/SHM cleanup: a checkpoint or
close between the two calls throws ENOENT (observed from the transcript
reconcile worker in server-startup-web-fetch-bind on CI). chmod directly
and swallow only ENOENT, which removes the TOCTOU window.
The agents-core graph can stall emitting nothing after warming Vitest's
module cache; a cli-runner stripe sat silent for the full 300s watchdog
window before its retry passed in seconds (run 29499091157, +5min on a
green PR). Give the family the same 60s no-output timeout the gateway
startup-health group already uses for this stall class.
Fleet-contended runners inflate measured group runtimes ~20-25% over
the serial-run hints (bins hit 300-358s on a loaded 12:1xZ run against
a 260s packed estimate). A 220s cap keeps the slowest bin near the
5-minute PR wall-clock budget under load at the cost of two extra
jobs (24 -> 26).
* perf(ci): scale Vitest workers with runner class and share serial module cache
Compact node-test jobs pinned OPENCLAW_VITEST_MAX_WORKERS=2 regardless of
runner size, so import-bound suites (30-45s of module-graph import per
file) crawled: runner-cli spent 284s on 5s of tests, commands-1 296s on
10s. With serial plans the budget now scales with the class (6 on 8vcpu,
3 on 4vcpu); timing-sensitive groups (tooling, tui-pty, infra-process)
stay pinned to 2 via plan-level env. Serial bins share one Vitest
fs-module-cache path so later plans reuse the first plan's transforms.
The import-bound commands stripes and security suite move to the 8vcpu
class, and packing hints refresh from serial-run measurements (run
29481835688).
* perf(ci): stripe cli-runner suite and pin media-ui worker budget
The agents-core config runs files serially (fileParallelism false guards
shared module state), so raising the worker budget cannot help its
import-heavy cli-runner suite (~35s module import per file, 213s serial).
Stripe it three ways so bins parallelize the imports instead. media-ui
hosts browser locator tests that timed out at 6 workers on the first
validation run; pin it to the proven 2-worker budget. commands-1's hint
drops to its measured 6-worker runtime.
Peter Steinberger