The two eval-dataset suites that download real corpora from the
HuggingFace Hub at runtime were running in the default CI lane. When the
Hub was unreachable or rate-limited they failed and reddened `main` even
though no code changed — confirmed by #506 (docs-only) failing on merge
while its own PR run passed an hour earlier. They also dominated CI
wall-time (~33 min of downloads + retry backoff on the failing run).
Add a `hub` pytest marker, apply it to both suites via module-level
`pytestmark`, and exclude it from the default CI lane
(`-m "not live and not cloud and not hub"`). The tests stay runnable on
demand with `pytest -m hub`.
The ADP provider swallows per-config download errors and returns 0
records on a network failure, so a Hub outage surfaced there as
`assert 1 <= 0` (not an exception) — it could not be made non-flaky by
exception handling alone, only by gating.
Coverage holds: removing these from CI drops total from 60.92% to
~60.73% (paranoid worst case 60.18%), still above the 60% gate.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(leaderboard): correct telemetry pipeline and outlier handling
The public leaderboard at /leaderboard showed a clear bimodal Wh/token
distribution: most users at ~3-5 J/token, ~30% inflated by 1000-4000×.
A 5-agent investigation workflow + 4-agent verification pass traced the
inflation to a cluster of related bugs across telemetry, server, and
display layers. This PR fixes the in-repo half. Backfilling existing
Supabase data is a separate follow-up.
Bug 1 — Dual telemetry recording (`server/routes.py`)
=====================================================
`_handle_direct` wrapped the engine with `instrumented_generate`
unconditionally. When `app.state.engine` was already an
`InstrumentedEngine` (the common case when telemetry is wired in),
BOTH layers published `TELEMETRY_RECORD` — once from the inner
`InstrumentedEngine.generate`, once from the outer wrapper. Every
chat-completion request was counted twice in the leaderboard pipeline.
Fix: detect the InstrumentedEngine and unwrap to `._inner` before
passing to the wrapper so only one layer fires.
Bug 2 — KV-cache fallback over-counts multi-turn (`server/savings.py`)
=====================================================================
`compute_savings` falls back from `prompt_tokens_evaluated` to
`prompt_tokens` when the KV-cache-aware count is missing. But routes.py
aggregates by summing each turn's full prompt — which counts the system
prompt N times for an N-turn conversation. The fallback inflated FLOPs
and energy by N×.
Fix: use 0 (conservative under-count) when the evaluated count is
missing rather than falling back to the inflated `prompt_tokens` sum.
Bug 3 — TelemetryRecord lacked methodology versioning
=====================================================
There was no per-record version tag, so legacy (pre-fix) and current
records were silently aggregated together in the public leaderboard
even though they used different methodologies.
Fix: add `token_counting_version: Optional[int]` field to the
`TelemetryRecord` dataclass + a nullable column to the SQLite schema
(with idempotent migration); the constant moves from `server.savings`
to `core.types` to avoid the server→telemetry layering. New records
write the current version; pre-fix rows remain NULL. The aggregator
gains a `current_methodology_only=True` flag that filters NULL rows out
of leaderboard sums — local dashboards leave it off so historical
aggregates still render.
Bug 4 — Leaderboard JS displayed missing telemetry as legit zeros
=================================================================
Rows with significant token counts but `energy_wh_saved = 0` and
`flops_saved = 0` rendered as `0.00 Wh / 0 FLOPs` — visually identical
to a user who genuinely did almost nothing. The headline totals also
included these rows.
Fix: new `isMissingTelemetry()` detector renders missing-telemetry
energy/FLOPs cells as `—` (with a tooltip explaining why); a new
`.lb-missing` CSS class differentiates the placeholder visually.
Bug 5 — Outlier filter was too generous
=======================================
`MAX_ENERGY_WH_PER_TOKEN = 10` left a 10,000× margin that admitted
every Group B row even though those values are physically impossible
(would imply a space-heater per token). Same for the FLOPs cap at
`1e17`.
Fix: tighten to `0.5` Wh/token (still 500× over a typical consumer GPU)
and `1e15` FLOPs/token (still 10,000× over typical). This removes
existing pre-fix corrupt rows from the public view without touching
Supabase.
Tests
=====
New regression tests (all pass, ruff clean):
- `tests/server/test_routes.py::test_instrumented_engine_unwrapped_to_avoid_dual_telemetry`
— pins the bug 1 fix; asserts exactly ONE TELEMETRY_RECORD event per
request when the engine is already an InstrumentedEngine.
- `tests/server/test_savings.py` (NEW file, 3 tests) — pins the bug 2
fix (FLOPs not inflated via fallback) and the cost-side invariant
(dollar savings still use full prompt_tokens because cloud providers
bill per input token even when local KV cache hit).
- `tests/telemetry/test_aggregator.py::TestMethodologyFilter` (3 tests)
— default behaviour includes legacy rows (local dashboard parity);
`current_methodology_only=True` excludes them; the summary surface
honors the same filter.
Unrelated test note
===================
`tests/telemetry/test_energy_wiring.py::TestTelemetryStatsEnergy::test_export_includes_energy_fields`
and `::TestEndToEndPipeline::test_ask_to_export_with_energy` are flaky
on this branch but ALSO flaky on `main` (confirmed via stash + the
banner-only run). Root cause: `_version_check.py:132-135` prints the
"new version of OpenJarvis is available" banner to stdout outside the
throttle window, contaminating `json.loads(result.output)` in those two
tests. Reproducible with `OPENJARVIS_NO_UPDATE_CHECK=1` → all pass.
That's a separate bug (the banner shouldn't write to the same stream
as machine-readable output); not addressing it here to keep this PR
focused on the leaderboard pipeline.
What this PR explicitly does NOT do
====================================
- Backfill ~30% of Supabase rows that already shipped with inflated
values from pre-fix clients. That requires Supabase write access and
is the natural follow-up — see PR comments for proposed dry-run audit
query and the additive `methodology_version` column migration.
- Distinguish which past submissions came from buggy vs correct clients
retroactively (no `app_version` tag on existing submissions).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* test(evals): fix test_energy_scales_linearly under leaderboard fix
CI on #498 failed in the slow `test` job:
tests/evals/test_use_case_benchmarks.py::TestSavings::
test_energy_scales_linearly — ZeroDivisionError: float division by zero
Root cause: the test (and `test_energy_wh_matches_direct_formula`)
called `compute_savings(N, 0)` without `prompt_tokens_evaluated`. Under
the OLD buggy fallback, an unset evaluated count silently became N,
energy scaled linearly with N, and the test passed by accident. Under
this branch's conservative fix the fallback is 0, FLOPs collapse to 0,
and the `p10.energy_wh / p1.energy_wh` ratio divides 0 by 0.
The test's own docstring says "evaluated tokens (KV-cache model)" — it
was always meant to exercise the explicit-evaluated path, the API
call just didn't match. Pass `prompt_tokens_evaluated` explicitly so
the test now expresses the invariant it claims to.
Same one-line fix for `test_energy_wh_matches_direct_formula` plus an
explicit `flops > 0` sanity assertion — that test was passing trivially
with `0 == 0` after the conservative fallback fix, masking whether the
formula was actually being exercised.
No production code change in this commit. Verified locally: all 6
TestSavings tests pass; the 5 new regression tests added on this branch
still pass.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---------
Co-authored-by: krypticmouse <herumbshandilya123@gmail.com>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Update all human-readable strings (docstrings, comments, descriptions,
log messages, TOML config descriptions) to correctly say
"DeepResearchBench" instead of "LiveResearchBench" for the
deep_research_bench benchmark. Class names and file paths are
unchanged for backward compatibility.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(evals): add tool_choice=auto + fix traces thread safety
Two fixes for eval accuracy and stability:
1. TauBench agent: add tool_choice="auto" to match tau2's native
LLMAgent behavior. Without this, Qwen and GPT-5.4 score 10-14pp
below leaderboard because the models don't receive explicit
tool-calling guidance.
2. SystemBuilder: apply self._traces flag to config.traces.enabled.
Previously builder.traces(False) was a no-op — traces stayed
enabled, creating SQLite connections in the main thread that
crashed when accessed from ThreadPoolExecutor worker threads
in GAIA evals ("SQLite objects created in a thread can only be
used in that same thread").
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: enrich inference events with model response content and add Trace.messages
Add content, tool_calls, and finish_reason fields to INFERENCE_END events
published by InstrumentedEngine. For non-instrumented engines, BaseAgent._generate()
now publishes INFERENCE_START/END events with the same rich data. Add _message_to_dict
helper and messages field to Trace dataclass for full conversation capture.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: enhance TraceCollector with rich content, tool details, and messages
Capture model response content, tool_calls, and finish_reason in GENERATE
steps; store tool arguments and result text in TOOL_CALL steps; extract
conversation messages from AgentResult.metadata into Trace.messages; and
implement the last_trace property. Also adds a messages column to the
TraceStore schema so messages survive the SQLite round-trip.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: agents store conversation messages in AgentResult.metadata
Both NativeReActAgent and MonitorOperativeAgent now serialize their
internal messages list via _message_to_dict and include it in the
returned AgentResult.metadata under the "messages" key. This enables
TraceCollector to capture full conversation traces.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: wire TraceCollector into system._run_agent and JarvisAgentBackend
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: persist rich trace data in eval trace JSONL files
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add TerminalBench config for Qwen 3.5-122B
* fix: add SQLite migration for traces.messages column on existing databases
* fix: check trace_store instead of shared config for trace enablement
* feat(evals): add ToolCall-15, LiveCodeBench, LiveResearchBench + telemetry
Three new benchmark integrations and full telemetry wiring for the
NeurIPS 2026 IPW/IPJ experiments.
## New Benchmarks
### ToolCall-15
15-scenario tool calling accuracy benchmark across 5 categories.
All scenarios defined inline with deterministic scoring (0/1/2 per
scenario). Fast to run (~5min/model) — ideal for optimization loops.
### LiveCodeBench
Competitive programming from LeetCode/AtCoder/CodeForces via
HuggingFace dataset. Sandboxed code execution with per-test
timeouts. Single-turn generation via jarvis-direct backend.
### LiveResearchBench
100 expert-curated deep research tasks. LLM-as-judge scoring
across 4 dimensions (comprehensiveness, insight, instruction
following, readability). Uses web_search tool for live research.
## Telemetry Wiring
- FLOPs estimation: 2 * active_params * total_tokens (MoE-aware)
- Energy/power capture flows from InstrumentedEngine through
backends to EvalResult and RunSummary
- New telemetry_summary section in output JSON with IPW/IPJ
- JarvisDirectBackend now propagates gpu_metrics flag
- TauBench forwards telemetry flags to SystemBuilder
## Experiment Plan
Added docs/experiments/neurips-2026-plan.md tracking the full
experiment matrix: 9 models x 7 benchmarks across NVIDIA, AMD,
and Apple hardware stacks.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Jon Saad-Falcon <jonsaadfalcon@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(evals): PinchBench harness fixes — scores from 26% to 84%
Multiple infrastructure bugs prevented PinchBench from producing
accurate scores. This commit fixes the eval harness so model scores
match the official leaderboard (Qwen3.5-397B: 26% → 84%, GPT-5.4:
5% → 58%).
Key fixes:
- EvalRunner: wrap generation in PinchBenchTaskEnv context so workspace
files persist through grading (was deleting before scorer ran)
- EvalRunner: detect task_env datasets and force sequential processing
(CWD changes aren't thread-safe)
- EvalRunner: fix episode_mode auto-detection to check for real
iter_episodes() override instead of hasattr() (always True)
- Scorer: use "params" field in transcripts to match PinchBench grade()
functions (was "arguments")
- Scorer: add None guard in _trace_to_transcript for tool_calls
- Scorer: capture final assistant text response in transcript so
text-only tasks (like sanity check) can be graded
- Scorer: add _tool_results_to_transcript() helper for EvalRunner path
- native_openhands: add native function-calling support (tools=
parameter) with text-based fallback, matching monitor_operative
- Tools: add Python fallbacks for file_read, file_write, shell_exec,
calculator, think, http_request when openjarvis_rust unavailable
- Security: add Python fallback for is_sensitive_file()
- Config: add "pinchbench" to KNOWN_BENCHMARKS
- Add PinchBench eval configs for Qwen3.5-397B and GPT-5.4
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(evals): fix hybrid grading crash when grading_weights is None
The 4 tasks with grading_type=hybrid (task_10, task_13, task_16_market,
task_22) crashed because grading_weights was explicitly None in task
metadata. dict.get("key", default) returns None (not the default) when
the key exists with value None. Use `or` to coalesce None to defaults.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(evals): handle MagicMock datasets in runner type checks
The iter_episodes and create_task_env type identity checks fail with
AttributeError when the dataset is a MagicMock (used in tracker tests).
Wrap in try/except to default to False for non-DatasetProvider objects.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds a `tool_calls` list field to `TurnTrace` capturing name, arguments,
and result for each tool invocation, enabling PinchBench transcript integration.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The FLOPs formula changed from quadratic (P*N*(N+1)) to linear
(2*P*T_evaluated) with KV-cache awareness. Update the test
expectation from ~100x to ~10x for 10x token increase.
The energy_wh_saved and flops_saved values were orders of magnitude too
high because a scaling factor that grows linearly with N was applied to
the energy calculation, making it scale as O(N³) instead of O(N²).
Replace the buggy scale-factor approach with a direct FLOP-to-energy
conversion using each provider's per-token constants. Add regression
tests to prevent recurrence.
Closes#95
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
