ci(tests): restore upstream slice-matrix — Python wall fix (4,847 files vs 30-min budget)

[POWERFULMOVES]

Summary

Restores upstream's pre-August slice-matrix design for Python tests / Run tests — one file (.github/workflows/tests.yml, ~140-line diff), restored verbatim from upstream's own history (fce30d818, parent of the 2026-08-22 removal commit 10f99bc15 "ci: run the work lanes on larger runners and merge the split jobs"). Zero editorial changes: this is what upstream itself ran in July, when this exact suite partitioned 8 ways ran green on identical ubuntu-latest runners.

Why

Upstream removed slicing when it moved the suite to a 96-core runner class. This fork cannot service that runner class (receipted in the label audits of PRs #9–#15): our ubuntu-latest is a 4-core host, and the unsliced suite (4,847 test files) cannot finish inside the 30-minute job budget — every CI head since the v2026.9.21 absorb died completed/cancelled at ~30m40s. Full write-up filed upstream: NousResearch#120148.

The change

• Restores the generate job (LPT slicing via scripts/run_tests_parallel.py --generate-slices, seeded from the duration cache), the 8-way slice matrix (run_tests.sh --files '<matrix.slice.files>'), and the save-durations merge job. All machinery is still alive in the repo's runner scripts (stdlib-only) — receipted at fork main 467624c.
• The base commit carries upstream's pre-removal hardenings: retry-action wrappers on env setup, and the corrected per-directory duration merge (no merge-multiple same-path race).
• ci.yaml needs no edit: it calls this workflow without with:, and slice_count defaults to 8.
• classify_changes.py runs the full lane set on any .github/ change — this PR validates everything automatically.

Evidence before push

• Generator dry-run at fork main 467624c: rc=0 → 8 slices, balanced 606×7 + 605 files.
• July upstream check-runs at 68391930d: Run tests slice 1/8 … 8/8 all success on this design.

Validation plan

This PR's own run is the empirical verdict. Key metric: per-slice wall time on 4-core. If slices approach the 30-minute per-job budget, slice_count is a one-line tunable (workflow input default; ci.yaml gains a with:) — iterate by push.

[github-actions]

૮ >ﻌ< ა ci review

_{ran on f7de6f2 — fix(tests): overlay triage — dedupe fireworks shadow, drop-i}

❌ Job failures

Python tests / Run tests slice 1/8 · View job

Job Python tests / Run tests slice 1/8 failed.

---

Python tests / Run tests slice 2/8 · View job

Job Python tests / Run tests slice 2/8 failed.

---

Python tests / Run tests slice 5/8 · View job

Job Python tests / Run tests slice 5/8 failed.

---

Python tests / Run tests slice 6/8 · View job

Job Python tests / Run tests slice 6/8 failed.

---

Python tests / Run tests slice 7/8 · View job

Job Python tests / Run tests slice 7/8 failed.

---

Python tests / Run tests slice 8/8 · View job

Job Python tests / Run tests slice 8/8 failed.

---

ℹ️ Info

CI-sensitive file review · View job

PR touches sensitive files, but the ci-reviewed label has been added, approving them.

Sensitive files changed:

• .github/workflows/tests.yml

---

debug info

CI timings

CI timings · View report · View job

Wall time 17m18s vs 31m35s (-45.2%). 6 job(s) slower, 11 faster, 1 unchanged.

• Rust tests / cargo test (bootstrap installer): +88.0s
• Python tests / e2e: -71.0s
• OS-specific tests / Windows-only tests: -59.0s
• JS & TS checks / JS & TS checks: -56.0s
• Check no committed infographics / check-no-committed-infographics: -49.0s

[POWERFULMOVES]

Label audit: ci-reviewed — slice-restore PR verification table

Record, per the #9–#15 pattern: the Review label gate ran at push time (pre-label). This comment is the verification record.

Item	Status
Branch provenance	ci/test-slices cut at fork main 467624c (post-#15 merge); single commit a44ebf58 — tests.yml restored verbatim from upstream parent-of-removal fce30d818
Content integrity	contents-API read-back byte-identical to the staged source; single-file commit asserted (commit touched only tests.yml); YAML parse OK (PyYAML) before push
Design provenance	upstream's own pre-August slice matrix (removal: 10f99bc15, 2026-08-22, "run the work lanes on larger runners"); July upstream runs green on this design at ubuntu-latest (slices 1/8–8/8 at 68391930d)
Machinery alive	run_tests_parallel.py --generate-slices rc=0 at fork main; 8 slices balanced 606×7+605 = 4,847 files; stdlib-only imports
Root cause	unsliced 4,847-file suite vs 30-min job budget on 4-core — receipted every-head cancelled at ~30m40s; filed upstream as NousResearch#120148
What validates here	classify_changes.py runs all lanes on .github/ changes — this PR's run exercises the full matrix including the restored workflow itself
Expected noise	JS may red intermittently (receipted flaky virtualHistoryOffsetCache, upstream NousResearch#120138 — non-blocking by classification); nix expected green (hermes_platform fix is in the base); per-slice wall time is the metric of interest — cold duration cache means blind LPT on this first run

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository: POWERFULMOVES/PMOVES-hermes-agent/.coderabbit.yaml

Review profile: CHILL

Plan: Advanced

Run ID: 128170be-dc36-4d23-84d1-a3eebd2321c5

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[POWERFULMOVES]

CI verdict — slice-matrix restore: the wall is dead; first real Python verdicts; 7 run-1 failures triaged (3 fork-owned, 4 upstream-class across 3 files)

The fix worked. 8 slices, Generate slices success, per-slice job walls 6m55s–10m53s against the old unconditional 30m40s kill — the suite now finishes with ~3× headroom on ubuntu-latest. Totals across all 8 slices: 55,589 tests passed, 7 failed (6 files) — green slices (1/3/4/7): 29,706 passed, 0 failed; red slices (2/5/6/8): 25,883 passed, 7 failed.

Fork-owned defects — fixing on this branch (each push re-runs all slices)

1. tests/plugins/model_providers/test_fireworks_profile.py (slice 5): the sync(fork): absorb NousResearch/hermes-agent@main (release v2026.9.21) #15 overlay's +4 on this file created a duplicate definition — TestFireworksHeaders.test_user_agent_identifies_hermes() at :60 shadows :56. Dedupe.
2. tests/test_schema_read_probe.py (slice 6): overlay-added probe's fixture drops last_activity_at while index idx_sessions_effective_activity still references it — sqlite3.OperationalError at :65. Test-fixture fix only (drop the index first / build the column-less table); no production schema code touched.
3. tests/test_managed_runtime_resolution.py (slice 8): upstream's allowlist meta-test names our overlay site — pmoves_bootstrap/tools_bridge.py::_resolve_cli_tool.<locals>._call (bare_which). Adding the justified row per the test's own allowlist mechanism.

Under probe (upstream content; upstream tip runs the suite green)

4. test_bedrock_adapter.py (slice 2): the one real-botocore import test died No module named 'botocore' while 83 sibling tests passed. Probe complete: file byte-identical to upstream tip (f6af83df both sides); botocore appears in the lock only as a transitive dep of boto3/s3transfer (not proof it's in the installed extras set); the file's own module guard skips cleanly when the import fails. Identical content is upstream-green → hermetic-env / parallelism-interaction class, not fork content; run 2 re-verdicts.
5. test_session_hygiene.py (slice 5): two async fence/hygiene tests failed — test_hygiene_does_not_wait_ceiling_after_fence_cancel (assert 2.275127418000011 < 2.0 at :1723 — the post-fence wait overran its 2s ceiling under load) and test_hygiene_unwind_records_cooldown (await asyncio.to_thread(worker_started.wait, 2) → assert False at :1820 — worker-start wait timed out under 8-worker parallelism). Both timing/parallelism-class; byte-identical content is upstream-green; run 2 re-verdicts.
6. test_gateway_shutdown.py (slice 8): drain-grace TimeoutError computed 0.0s — timing-sensitive shutdown window. The job-level rerun probe was refused by the API (Only jobs from the current attempt can be re-run); run 2 is the re-verdict.

Disclosed trade-off, on the record: this restores upstream's own July design, but the parallelism profile differs from today's upstream (8 workers × 4-core vs 1 × 96-core). Timing-sensitive tests get a fresh verdict on every slice run; the two timing candidates (4, 6) will re-verdict on the fix pushes before this merges. No merge until the three fork-owned fixes land green.

Run 2 (f7de6f22f, triage commit) — verdict

All three fork-owned fixes held: fireworks dedupe, schema-probe drop-index guard, and the bare_which allowlist row appear in zero run-2 FAILED lists (run 1's failures of exactly those vanished). The wall stays dead: all 8 slices finished in ~7–11 min against the old unconditional 30m40s kill.

Residual ledger (6 final failures / 6 files — none fork content):

Repeaters (2/2):

• test_bedrock_adapter.py::TestReasoningReplaySchema — real-botocore import dies ModuleNotFoundError. Probe closed: bedrock = ["boto3==1.42.89"] exists in pyproject but is in neither fork's nor upstream's CI sync line (fork :320/:508 ≡ upstream :321/:509); hermetic-env class. Upstream-green claim scoped: at 731e99d15.
• test_gateway_shutdown.py::test_unexpected_signal_starts_teardown_after_bounded_interrupt_grace — drain-grace computed 0.0s, TimeoutError (2/2).
• test_session_hygiene.py both fence tests — ceiling overruns under 8-worker load (assert 2.27… < 2.0 then assert 2.93… < 2.0 at :1723; worker_started.wait, 2 → False at :1820).

One-offs (run 2 only): test_status.py::TestReadProcessCmdlinePsFallback, test_profiles_sidebar_cache.py, test_compute_host_borrowed_lease.py::test_isolated_turn_runs_against_parent_leased_session. test_hosted_rooms.py shows a FAILED line but recovered on retry — final summary 0 failed.

Partition integrity (spot-verified, not global): s1 files-list hash f23f7b759db7 identical across both runs; test_inventory_pricing exists only in s7's list — an earlier cross-log sighting was a grep artifact of retry lines, corrected here.

Drift watch-item for the next absorb: upstream main dropped the hindsight extra post-release (73c598e31) while the restored fork tests.yml still carries it — expected keep-ours conflict at next sync, the known manifest-drift class, to be resolved at merge time.

Bottom line: zero fork-content failures remain. The residual set is documented environment/parallelism behavior of upstream's own tests, receipted above.