LLM Eval Harness

Overview

Give this skill an endpoint (base_url + model + an API key in an env var) and it measures whether the endpoint actually works and whether the model is fast, stable, protocol-correct, and good enough — instead of trusting the vendor's headline numbers. Six dimensions, usually scattered across ad-hoc scripts that get rewritten (with the same bugs) every time:

Which dimensions to run

Route from what the user is actually doing:

• "接入新模型/新网关先测一下" / onboarding a provider → Availability → Fidelity, then speed/concurrency if adoption looks likely
• "这个 ID 能不能用 / 写个支持列表" → Availability (and read disciplines §9 before writing any "unavailable" verdict)
• "模型不听 system prompt / agent 行为怪但没报错" → Fidelity (system canary)
• "快不快 / tok/s 是真的吗" → Speed; "扛不扛得住" → Concurrency
• "这个兼容端点是真兼容吗" → Protocol + Fidelity's tools/auth checks
• "这个模型/渠道用着用着就 400 了 / session 断了 / continue 不下去了" (extended-thinking client, multi-turn session) → Protocol --check history-replay specifically — this symptom is the generation check passing while history-replay silently doesn't; don't stop at generation
• "换模型质量会不会掉" → Quality regression
• "要给厂商报 bug" → run the relevant probe with --output, then follow references/vendor_evidence_protocol.md
• "部署闸门 / 常驻监控 / 告警老误报" → references/production_testing_patterns.md

Don't run all six ritually — pick what answers the user's question.

Key handling (non-negotiable): every script takes the API key by env-var name (--key-env MY_KEY), never the key value on the command line — so it stays out of ps, shell history, and any saved report. Never hardcode a key into a use-case file or a wrapper. Read references/evaluation_disciplines.md for the full reasoning behind this and the other disciplines.

Your private data lives outside this bundle. Use-case libraries, model rosters, and keys belong in ~/.llm-eval/ (or wherever you keep secrets), NOT in this skill directory — the skill is generic and public; your test suite is yours. See "Use-case library" below.

Quick start

Detect what you have, then run the dimensions that apply. For an OpenAI-compatible model:

export MY_KEY=sk-...                       # the key never appears in a command below

# Speed: real-task throughput + sustained decode ceiling
uv run --with openai python scripts/speed_probe.py \
  --base-url https://api.example.com/v1 --model some-model --key-env MY_KEY --mode both

# Concurrency: ramp until it breaks
uv run --with aiohttp python scripts/concurrency_probe.py \
  --url https://api.example.com/v1/chat/completions --model some-model --key-env MY_KEY \
  --format openai --concurrency 10 20 40 60

If the endpoint is Anthropic-Messages-shaped (/v1/messages), also run the protocol probe (below). Pick dimensions by what the user actually asked — don't run all six if they only asked "is it fast?".

Dimension 0 — Availability (which model IDs actually work)

uv run --with aiohttp python scripts/availability_probe.py \
  --base-url <base> --key-env <ENV> --format both \
  --models model-a vendor/model-b @models.txt --output /tmp/avail.json

• Verdicts are 3-state per failure, not pass/fail: no-channel (no route for this ID), upstream-error (route exists, upstream failing — retry later), empty-content (usually a max_tokens artifact on reasoning models, NOT a broken model). The probe keeps --max-tokens at 8192 by default precisely so thinking models don't read as dead — do not "optimize" it downward.
• Before writing any "unavailable" list: verify each failing ID exists in official docs. Current-generation model names postdate your training data — WebSearch the vendor's / model developer's model page instead of enumerating guesses; suffixes (-preview, dated variants, tier names) decide routability. Full traps: disciplines §9.
• Never probe a client-side marker as if it were a real model ID. Bracket-suffixed strings like some-model[1m] are frequently a CLI client's own local convention (Claude Code parses and strips [1m] before the request is ever built — see claude-switch-models-setup's "Configuring Context Window Size" section for the full mechanism) — they never appear on the wire. The probe warns when it sees one, but the discipline is yours: probe the bare model ID the vendor actually documents, not a string copied out of a Claude Code ANTHROPIC_MODEL env var.
• The gateway's /v1/models listing is one input, never the verdict — real gateways route models the listing omits.

Dimension 0.5 — Request fidelity (does your payload reach the model?)

uv run --with aiohttp python scripts/fidelity_probe.py \
  --base-url <base> --model <model> --key-env <ENV> \
  --format anthropic --check all --repeat 10 --output /tmp/fidelity.json

• Four checks: system (canary code planted in the system prompt — can the model echo it back?), tools (full round-trip including returning a tool result and verifying the final answer uses it), multiturn (plant facts across turns, recall them all), auth (x-api-key vs Bearer on both protocol endpoints — gateways commonly accept both on one path and only one on the other, and the wrong-header 401 reads exactly like a dead key).
• This is the dimension that catches the nastiest gateway failure: everything returns 200 and chats fluently, but the system prompt never reached the model — so any tool whose rules live in the system prompt silently misbehaves with no error anywhere. On one real gateway, one model alias never delivered the system prompt in 100+ samples across every legal way of sending it, while sibling aliases delivered intermittently with rates that changed by the hour.
• Delivery through routed gateways is probabilistic and time-varying — hence --repeat (default 10) and a three-state verdict (delivered / intermittent k-of-N / not-delivered). A clean single-window result is still just that window: re-sample in another time slice before publishing a number (disciplines §12).
• The probe reports token-accounting corroboration but never trusts it alone — behavioral evidence rules; see disciplines §10 (calibrate the meter) and §11 (why the canary is a neutral external fact, not an identity or obedience test).

Dimension 1 — Speed (thinking-aware)

uv run --with openai python scripts/speed_probe.py \
  --base-url <…/v1> --model <model> --key-env <ENV> --mode both --output /tmp/speed.json

• mixed runs representative tasks (what real usage feels like); decode forces one long output to find the sustained ceiling (the number to compare against a vendor's claim); both does both.
• The trap this script exists to avoid: reasoning models stream thinking in a separate reasoning_content field, but completion_tokens counts it. Collecting only content while dividing by completion_tokens produces wildly inflated numbers — a real ~750 tok/s model once measured as 4700 tok/s this way. The script captures both, takes TTFT as the first token of either kind, and reports completion_tokens / (total − TTFT).
• Read the output correctly: real-task throughput is lower than the decode ceiling because short outputs never reach steady state — that's expected, not a bug. Report both numbers, and note when the model emits thinking (its end-to-end latency includes reasoning time, not just typing).

Dimension 2 — Concurrency / stability

uv run --with aiohttp python scripts/concurrency_probe.py \
  --url <full endpoint URL> --model <model> --key-env <ENV> \
  --format openai|anthropic --concurrency 10 20 40 60 --output /tmp/conc.json

• Pass several --concurrency levels to ramp and find the ceiling — the level where success rate drops or latency explodes. A model that's fast single-threaded can still collapse at modest concurrency (real example: one provider held 50 concurrent at 0.4s while another dropped requests at just 5 concurrent).
• The script isolates from any ambient proxy (trust_env=False) and disables keep-alive pooling (force_close) — otherwise you measure the proxy's limit or one pinned upstream replica, not the model. It prints a "concurrency proof" (overlapping request pairs) so you can confirm requests really ran in parallel.
• Distinguish failure modes from the output: HTTP 429 (clean throttle, retriable) vs a TCP drop that hangs to timeout (much worse for UX) vs 5xx. They imply very different fixes.

Dimension 3 — Protocol compliance (Anthropic thinking block: generation AND history-replay)

uv run python scripts/protocol_probe.py \
  --url <…/v1/messages> --model <model> --key-env <ENV> --check all --repeat 10 --output /tmp/proto.json

• Only relevant for endpoints claiming Anthropic /v1/messages compatibility. --check all (default) runs BOTH sub-checks — they test different code paths and a vendor can pass one while hard-failing the other:
  • generation: does thinking: {type: enabled} actually produce thinking_delta / signature_delta SSE events when you request it? (the original check)
  • history-replay: does the endpoint ACCEPT a type: "thinking" block that's already sitting in a prior assistant turn, when that history is replayed back on a later turn — exactly what Claude Code and every other agentic client does on every continuation? (--check history-replay to run just this one)
• Real incident (2026-07-21) that motivated the history-replay check: a Kimi/Moonshot model via a China reseller passed generation fine (emits thinking correctly when asked) but hard-rejected history-replay — 400 "invalid part type: thinking" the instant a prior thinking block came back as input, killing the session on every subsequent turn. Passing generation told us nothing about this; they're orthogonal failure modes (response generation vs. request validation).
• Don't conclude "vendor/reseller X is broken" from one cross-axis comparison — and don't stop at the first single-axis comparison that confirms a vendor-documented parameter either. This took three rounds to get right, kept in disciplines §§17-19 as the canonical cautionary tale: round 1 compared a different model AND a different reseller at once and (wrongly) blamed the reseller. Round 2 fixed that — same reseller, only the model varied — and the result matched a documented, named vendor parameter (Moonshot's own preserve_thinking), which looked like confirmation. Round 2 was STILL wrong: the probe never left that one reseller, so it couldn't see that the vendor's own direct/native endpoint handled the "rejected" model fine — real production traffic proved it. Read §19 before treating any single-reseller-confirmed result as final; check the vendor's own endpoint or real traffic for the actual channel in question before writing an "X doesn't support thinking" conclusion into anything.
• Compliance is often probabilistic, not binary, for BOTH checks. One real vendor honored the thinking block on only ~13% of generation requests (vs 100% for two competitors). That's why --repeat defaults to 10; generation's verdict has three states (fully-implemented, intermittent (k/N), not-implemented), history-replay's has its own three-plus states (accepts-thinking-in-history, rejects-thinking-in-history, inconsistent, or inconclusive when errors look unrelated to thinking at all). Never conclude from a single sample.
• It forces Connection: close per request so a load balancer can't pin all samples to one replica and hide the real distribution (a real probe saw 0/10 with keep-alive vs 17/90 with close on the same endpoint).

Dimension 4 — Quality / use-case regression (blind judge)

This is two halves on purpose: collect, then judge independently.

Step 1 — collect the model's answers to your use-case library:

uv run --with openai python scripts/usecase_runner.py \
  --base-url <…/v1> --model <model> --key-env <ENV> \
  --usecases ~/.llm-eval/usecases.json --output-dir ~/.llm-eval/runs/<model>

Step 2 — judge with independent blind judges (orchestrate inline — do NOT let the model grade itself). For each answer in the run directory, spawn 3 independent Task agents (or fewer for a quick pass). Each judge gets ONLY: the prompt, the answer, and the case's rubric — and is explicitly told it is judging in isolation, with no knowledge of other judges' scores or any prior evaluation (this prevents anchoring). Then aggregate:

• A case passes only on majority agreement among judges.
• Compute precision per category (using each case's tags): a category where judges systematically disagree with the rubric is a real weakness — on one real eval, a whole category scored 12.5% precision and exposed a systematic misclassification that a single grader would have missed.
• Count only explicit judgments. A judge that didn't return a verdict is not a pass — silence ≠ consent. This guards against automation bias.

For the rubric-scoring mechanics (LLM-as-judge thresholds, llm-rubric), you can also compose with the promptfoo-evaluation skill — point its providers at the same endpoint. This harness's blind-judge method and promptfoo's rubric assertions are complementary: use promptfoo for fast per-case pass/fail gating, blind judges for precision on a category you suspect is weak. Full method: references/quality_blind_judge.md.

Use-case library

Keep it OUTSIDE this bundle (e.g. ~/.llm-eval/usecases.json) so it survives skill updates and never lands in a public repo. It's a plain JSON list — version it in a private repo to accumulate a regression suite over time:

[
  {"id": "refund-window", "prompt": "A customer asks for a refund 20 days after purchase. Reply as support.",
   "rubric": "1.0 if it correctly cites the 30-day refund window; 0.0 if it refuses or invents a different window.",
   "tags": ["support", "policy"]},
  {"id": "lru-cache", "prompt": "Implement an LRU cache in Python with O(1) get/put.",
   "rubric": "1.0 if get and put are both O(1) via dict + doubly linked list and the self-test passes.",
   "tags": ["code"]}
]

assets/example_usecases.json is a starter you can copy. Only id and prompt are required; rubric, expected, and tags make judging sharper.

Running a full evaluation

When the user says "evaluate / benchmark this model", the typical flow is:

1. Identify the shape — OpenAI-compatible (/v1/chat/completions) or Anthropic-Messages (/v1/messages)? Hit GET /v1/models or read the vendor docs; don't assume. This decides which probes apply (protocol probe is Anthropic-only). Remember the listing is incomplete evidence either way (disciplines §9).
2. For a new endpoint, availability and fidelity come first — speed numbers for a model whose system prompt never arrives are answering the wrong question.
3. Run the dimensions the user cares about — speed and concurrency for "is it fast/stable", add protocol for an Anthropic vendor, add quality when they have a use-case suite. Write each probe's --output JSON to a run directory.
4. Report honestly, separating measured from inferred. Lead with the headline the user asked about (e.g. "sustained decode ceiling exceeds the vendor's claimed tok/s, while real-task throughput runs lower"). If a number looks impossible (e.g. throughput far above the vendor claim, or a single-sample protocol verdict), treat it as a measurement artifact to investigate, not a result — that skepticism is the whole point of this harness. Rates on routed gateways are additionally time-varying: re-sample another window before freezing any number into a document (disciplines §12).
5. Comparing two models? Run the identical probes against each with the same flags, and put the two JSON outputs side by side. Keep the test conditions identical (same concurrency levels, same use cases) or the comparison is meaningless.
6. Found a vendor bug worth reporting? Don't paste raw probe output at their support channel — build the evidence package per references/vendor_evidence_protocol.md (self-audit first, observation wording, request-id table, pre-registered thresholds, adversarial counter-review).

For tests that will run repeatedly against a live system — deployment gates, resident canaries, fault-injection mocks, and the monitoring statistics that lie — see references/production_testing_patterns.md.

Next step

After a run, offer the natural follow-ups:

Evaluation complete for <model>.

Options:
A) Render an HTML dashboard of the results — compose with a visualization skill (Recommended if sharing)
B) Compare against another model — same probes, side-by-side
C) Add the failing cases to ~/.llm-eval/usecases.json as a permanent regression guard
D) Done — the numbers answer the question