agentera 3.0.0-dev.27 → 3.0.0-dev.28

package/bundle/references/analysis/worker-policy-2026-07-21.yaml

@@ -1,207 +0,0 @@
-schemaVersion: agentera.workerPolicyBenchmark.v1
-status: selected
-captured: 2026-07-21
-
-runner:
-  baseline_reference: references/analysis/verification-baseline-2026-07-20.yaml
-  name: local-16-logical-cpu-node22
-  plan_label: named 16-core Node 22 runner
-  plan_label_mapping: >-
-    "16-core" means the 16 logical scheduling CPUs reported by nproc and
-    os.cpus(), not physical cores. The named runner is an AMD Ryzen 7 9800X3D
-    with 8 physical cores and 16 threads, running Linux 7.1.4-1-cachyos,
-    Node v22.23.1, and pnpm 10.30.3. This preserves the historical baseline's
-    terminology without claiming 16 physical cores.
-
-method:
-  command: pnpm run test:source
-  working_directory: packages/cli
-  selection: >-
-    Each candidate was supplied only through VITEST_MAX_WORKERS, after which
-    the source owner selected its complete inventory through verify-lane.mjs.
-    No file filters, retries, reporter overrides, or concurrent invocations
-    were used.
-  protocol: >-
-    Run one unretained warmup followed immediately by three retained runs for
-    each safe candidate, serially in ascending worker order. Record Bash
-    process wall time and Vitest's reported duration, file count, and test
-    count. The shell emitted comma decimal separators; table values normalize
-    them to decimal points.
-  reproduction: >-
-    From packages/cli, for each workers value and phase in warmup sample-1
-    sample-2 sample-3, run: TIMEFORMAT="workers=$workers phase=$phase
-    wall_seconds=%3R"; time env VITEST_MAX_WORKERS="$workers" pnpm run
-    test:source. Run candidates serially; retain only the three samples after
-    each warmup.
-  pass_conditions:
-    - source command exits zero
-    - 193 test files pass and 2365 tests pass with one declared Darwin skip
-    - no test exceeds the configured 30000 ms timeout
-    - no retry, shared-output failure, or generated-output race is reported
-
-candidates:
-  - workers: 6
-    warmup:
-      wall_seconds: 66.085
-      vitest_duration_seconds: 64.47
-    retained:
-      - wall_seconds: 55.297
-        vitest_duration_seconds: 54.30
-      - wall_seconds: 50.028
-        vitest_duration_seconds: 49.09
-      - wall_seconds: 49.322
-        vitest_duration_seconds: 48.38
-    median:
-      wall_seconds: 50.028
-      vitest_duration_seconds: 49.09
-    result: pass
-  - workers: 8
-    warmup:
-      wall_seconds: 43.899
-      vitest_duration_seconds: 42.94
-    retained:
-      - wall_seconds: 43.968
-        vitest_duration_seconds: 43.03
-      - wall_seconds: 43.598
-        vitest_duration_seconds: 42.66
-      - wall_seconds: 43.926
-        vitest_duration_seconds: 42.99
-    median:
-      wall_seconds: 43.926
-      vitest_duration_seconds: 42.99
-    result: pass
-
-selection:
-  precedence:
-    - VITEST_MAX_WORKERS explicit numeric override
-    - AGENTERA_VITEST_RUNNER_POLICY=local-16-logical-cpu-node22
-    - unmeasured fallback
-  measured_local:
-    selector: AGENTERA_VITEST_RUNNER_POLICY=local-16-logical-cpu-node22
-    workers: 8
-    evidence_status: measured
-  rationale: >-
-    Eight is the faster of the two safe, repeatably passing candidates: its
-    43.926-second retained median is 6.102 seconds (12.197 percent) below the
-    six-worker median. It is also the plan's stated maximum safe candidate.
-    It is selected only through the named runner-policy selector, never from
-    CPU count or host fingerprinting.
-  unmeasured_fallback:
-    selector: AGENTERA_VITEST_RUNNER_POLICY=unmeasured or no recognized selector
-    workers: 4
-    evidence_status: unmeasured_pending_benchmark
-    evidence: >-
-      references/analysis/verification-baseline-2026-07-20.yaml retains only
-      the named local runner's measurements and requires each CI runner to be
-      benchmarked. No retained GitHub-hosted runner measurement is available.
-    rationale: >-
-      Four workers is a conservative bounded fallback, not a GitHub-hosted
-      runner measurement. Every CI runner needs its own retained benchmark
-      before a tuned policy may replace it.
-  explicit_override:
-    selector: VITEST_MAX_WORKERS=<Vitest-supported worker count>
-    evidence_status: operator_selected
-    rationale: >-
-      This takes precedence over runner policy and is not presented as a
-      benchmarked setting unless separately measured and recorded.
-  historical_baseline: >-
-    The preserved four-worker baseline is 81.466 seconds median at source
-    revision af9cf773 (references/analysis/verification-baseline-2026-07-20.yaml).
-    Its 191 files and 2230 tests differ from this checkout's 193 files and
-    2365 tests, so it provides historical context rather than a like-for-like
-    speedup calculation.
-  unsafe_upper_bound: >-
-    Twelve workers previously failed under contention. It was not rerun and
-    is not a supported candidate, a selection target, or evidence superseded
-    by this measurement.
-
-superseded_confirmation:
-  command: pnpm run test:source
-  environment: VITEST_MAX_WORKERS unset; previous global default selected eight workers
-  consecutive_runs:
-    - wall_seconds: 43.869
-      vitest_duration_seconds: 42.90
-      files_passed: 194
-      tests_passed: 2366
-      tests_skipped: 1
-    - wall_seconds: 43.972
-      vitest_duration_seconds: 43.02
-      files_passed: 194
-      tests_passed: 2366
-      tests_skipped: 1
-    - wall_seconds: 44.210
-      vitest_duration_seconds: 43.27
-      files_passed: 194
-      tests_passed: 2366
-      tests_skipped: 1
-  result: >-
-    These three serial source runs passed without timeout, retry, shared-output
-    race, or generated-output failure, but they predate the runner-specific
-    correction and must not be treated as evidence for the selector below.
-
-corrected_final_confirmation:
-  command: pnpm run test:source
-  environment: AGENTERA_VITEST_RUNNER_POLICY=local-16-logical-cpu-node22; VITEST_MAX_WORKERS unset
-  consecutive_runs:
-    - wall_seconds: 43.508
-      vitest_duration_seconds: 42.56
-      files_passed: 194
-      tests_passed: 2369
-      tests_skipped: 1
-    - wall_seconds: 43.806
-      vitest_duration_seconds: 42.85
-      files_passed: 194
-      tests_passed: 2369
-      tests_skipped: 1
-    - wall_seconds: 43.330
-      vitest_duration_seconds: 42.39
-      files_passed: 194
-      tests_passed: 2369
-      tests_skipped: 1
-  result: >-
-    All three serial runs selected the named measured-local policy and passed
-    without timeout, retry, shared-output race, or generated-output failure.
-
-unmeasured_path_verification:
-  environment: AGENTERA_VITEST_RUNNER_POLICY=unmeasured
-  runner: local-16-logical-cpu-node22
-  workers: 4
-  wall_seconds: 71.296
-  vitest_duration_seconds: 70.34
-  files_passed: 194
-  tests_passed: 2369
-  tests_skipped: 1
-  interpretation: >-
-    This is one local integration check of the unmeasured selector, not a
-    GitHub-hosted benchmark or evidence that CI can be tuned to four workers.
-
-policy_composition:
-  authority: references/analysis/verification-policy.yaml
-  verification: >-
-    laneOwnership.test.ts requires the exact Task 2 policy compositions:
-    targeted/precommit/fast/local=[source], merge=[source, package],
-    scheduled=[source, stress, performance], and
-    release=[source, stress, performance, package]. verify-lane validates
-    exactly one primary owner per inventory file before composing a policy.
-  precommit: >-
-    scripts/precommit-vitest.sh routes ordinary staged source work to source
-    and conservative authority surfaces to release; it delegates composition
-    to verify-lane rather than declaring a second owner policy.
-  ci: >-
-    .github/workflows/ci.yml explicitly supplies
-    AGENTERA_VITEST_RUNNER_POLICY=unmeasured to source verification. This
-    selects the four-worker pending-benchmark fallback; it does not claim a
-    GitHub-hosted measurement. CI separately invokes source and package
-    owners; stress and performance remain scheduled/release policy owners.
-    No new composition or owner was added for worker selection.
-
-caveats:
-  - >-
-    references/analysis/verification-final-proof-2026-07-21.yaml reran the
-    classified final tree on this selector after measured fixture deduplication.
-    All owner and policy evidence passed; the 42.908-second source median is
-    47.330 percent below the historical four-worker baseline and satisfies the
-    45-second task limit.
-  - Decisions retrieval was degraded to 5 of 50 records; missing detail was not reconstructed.
-  - Satisfaction records remain unavailable for llavvlpcmb and djmjbmruyk.
-  - Vision and profile state were absent from the original build context; no approval was inferred.