club-3090/AGENTS.md

Agent guide

Guidance for AI coding agents (Claude Code, Cursor, Copilot, Continue, etc.) working in this repo. Focused — only conventions an agent wouldn't infer from the code itself.

One file, two names: this is AGENTS.md (canonical); CLAUDE.md is a symlink → it. Edit either — both names resolve to the same guide so any agent that looks for either finds it, and the two can't drift.

Read first

Before making non-trivial changes:

• README.md — what the repo is, two-routes framing, repo layout
• docs/README.md — the docs index (user track + contributor track); start here to find any guide
• docs/ARCHITECTURE.md — current stack state (services, ports, paths, model + KV config)
• docs/SINGLE_CARD.md / docs/DUAL_CARD.md — pick-by-workload guidance + the cliffs they reference
• docs/ADDING_MODELS.md — add a model: serve-locally vs the curated-catalog workflow (see "Adding a model" below)
• docs/UPSTREAM.md — every upstream issue / PR we depend on or have filed (see "Upstream issues" section below for why this matters)
• models/qwen3.6-27b/INTERNALS.md — DFlash forensics, AutoRound rationale, Marlin pad fork, MTP head
• CONTRIBUTING.md — what kinds of PRs land cleanly + benchmark + verify protocol

Hardware truths

The reference rig (the maintainer's, where all first-party numbers come from): 2× RTX 3090 Ampere SM 8.6, PCIe-only, no NVLink (and we won't add it). Its constraints:

• Custom all-reduce must be disabled in vLLM/SGLang configs on PCIe-only multi-GPU topologies (NVLink-assumed paths break).
• Ampere (sm_86) has no native FP8 compute — there, FP8 KV is a storage optimization only. On Ada/Hopper/Blackwell, FP8 compute paths are real (fp8-weights composes are 5090-safe via the launcher's VLLM_USE_DEEP_GEMM pass-through — an arch-conditional guard, not a blanket "everything works on 5090").
• Speculative decoding using EAGLE / DFlash inside vLLM is blocked on Qwen3-Next family on every arch (DeltaNet rollback). MTP works, and DFlash works via the beellama engine (external drafter). See docs/UPSTREAM.md — vllm#39931.

The rig you're running on may not be the reference rig. Supported hardware classes live in scripts/lib/profiles/hardware/*.yml (3060 → 5090, A5000, A100, H100, DGX Spark, …); the profile-catalog compat layer and launchers key on the detected class and inject arch-aware env. Before assuming any constraint above applies, check which class you're on (nvidia-smi + the matching hardware YAML) — and never hand-copy a reference-rig workaround (e.g. disabling all-reduce) onto an NVLink-equipped or non-Ampere rig without checking it's still warranted.

Upstream issues — single source of truth

docs/UPSTREAM.md tracks every upstream issue / PR we depend on, have filed, or use as context. Before filing a new upstream issue or referencing an existing one in code/docs, check this file. When status changes (issue closed, PR merged, pin bumped), update the row.

This rule exists because we previously had upstream links scattered across CHANGELOG, INTERNALS, FAQ, and per-compose comment headers — and they drifted. The tracker file is the canonical place; cross-link to it from anywhere else.

When filing a fresh upstream issue from this work:

1. Add the row to docs/UPSTREAM.md first
2. Link the issue back to noonghunna/club-3090 in the body (helps maintainers see affected user surface)
3. If the upstream eventually merges + propagates, update the row to ✅ Resolved and bump the relevant pin (Genesis commit, vLLM nightly, etc.) in the same commit / PR

Conventions on this repo

Bench protocol

3 warm + 5 measured runs. Canonical prompts: 800-word essay (narrative, max_tokens=1000) + quicksort code (max_tokens=800). temperature=0.6, top_p=0.95, top_k=20. Capture both wall-time TPS and engine-internal gen throughput from logs. Always capture per-card peak VRAM alongside TPS.

Genesis opt-in env vars

Status (2026-07-06): no shipped compose currently enables Genesis — the Genesis-pinned production paths were retired (their composes live under compose/_archive/). The guidance below stays because it applies verbatim if a Genesis-pinned compose is reintroduced, and the incident it encodes is the canonical example of why behavioral patches need repro-gating.

Genesis ships ~50 env-gated patches. Some are targeted bugfixes (P64 streaming, PN8 memory savings, P3/P5/P6 KV); others are behavioral mitigations that silently rewrite the request (P68 = tool_choice → required, P69 = inject "must use tool" reminder). Behavioral mitigations need a streaming + large-prompt repro before shipping default-on. We learned this the hard way on 2026-04-29 — see docs/UPSTREAM.md → Genesis #9 row + the club-3090 #2 thread.

If you're considering enabling a new Genesis env var by default in a shipped compose:

1. Read the patch's header docstring in vllm/_genesis/wiring/. Does it modify request.tool_choice, request.messages, or rewrite output?
2. If yes (behavioral): run a streaming repro with prompt > the patch's threshold + a casual user message ("hi") + tool_choice: auto. If finish_reason=stop with empty content, don't ship default-on.
3. Pure bugfixes (no behavioral override) are fine to ship default-on once they pass verify-full.sh.

Engine image pinning

Pin engine images when either trigger holds: (1) we vendor patches into the running container, or (2) the engine's rolling tag has burned us with a regression (a crash-loop or behavior break shipped under the same tag we'd validated — llama.cpp earned this pin via #187). Otherwise track upstream's rolling tag and accept that the compose YAML may need maintenance when upstream changes flags.

Why: patches hook into specific upstream code paths — a silent upstream change drifts those hooks and breaks the patched container in production. Pinning ensures the bytes we tested against are the bytes users get. Unpatched engines have no such hook, so upstream changes are upstream's problem to fix (or the YAML's, which is cheap to maintain) — unless the tag itself has proven mutable-under-validation, which is trigger (2).

Pins are a maintenance liability — bump them on a cadence. A stale stability-pin silently costs quality, not just features: the b9246→b9967 llama.cpp A/B (2026-07-11, #680) was think-OFF-neutral but +4 think-ON with 3 scenario-level flips. scripts/engine-pin-bump.sh <engine-id> <tag> --check does the mechanical half (spec + compose defaults + a bucketed report of fixtures/claims needing hands); the judgment half (re-validation, live boot) stays with the pin-bump checklist.

Engine	Patches we vendor	Tag policy
llama.cpp	none	pinned build tag (e.g. server-cuda-b9967) — trigger (2): stability-pinned after the #187 rolling-tag crash-loop; zero patches, so bumps are cheap (engine-pin-bump.sh llama-cpp-local <tag>)
vLLM	Marlin pad, KV/loader overlays — scripts/lib/profiles/patches.yml is the source of truth	pinned release tag (e.g. v0.22.0, v0.24.0) — never nightly-*/latest: upstream purges nightly tags, orphaning the pin
beellama.cpp	none (the fork is the delta)	digest-pinned to Anbeeld's official image (engines/beellama-local.yml install.spec)
SGLang	per-compose decision (pin if we vendor a patch, rolling otherwise)	per-compose

When adding the first vendored patch to a previously-rolling engine: pin in the same commit. When dropping the last patch: unpin in the same commit — unless the engine holds a trigger-(2) stability pin, which outlives its patches. Bump pins via PR with a verify-full.sh + bench.sh re-run, never silently.

Delivery model (vLLM): patches reach the container by volume-mounting into the pinned stock vllm/vllm-openai image (python sidecars / site-package overlays / install scripts — see delivery_mechanism in scripts/lib/profiles/patches.yml), not by baking a custom image. The older baked-image path (ghcr.io/noonghunna/vllm-club3090, which shipped the release images through club-v0.8.3) is retired — no compose or engine-pin references it, and the dockerfile_bake delivery: block in patches.yml is legacy/test-only. The GHCR package is kept as historical release artifacts (users pinned to a club-v0.8.x tag can still pull); it is not deleted and not produced by anything in-repo.

File encoding in scripts — always encoding="utf-8"

Any Python read of a repo source file (compose YAML, profile YAML, baselines.yml, …) — including python heredocs inside shell scripts — MUST pass encoding="utf-8". Path.read_text() / open() default to the locale encoding, and community rigs run non-UTF-8 locales (minimal VMs / containers with LC_ALL=C → ANSI_X3.4-1968). Repo files are full of unicode (— × → ⚠ in compose headers), so a bare read that works on the dev machine crashes switch.sh/launch.sh on those rigs (#599). Corollaries:

• Writes too: a piped stdout under the C locale defaults to ASCII — printing a status note with unicode raises UnicodeEncodeError. Python emit blocks on launcher paths pin it with sys.stdout.reconfigure(encoding="utf-8").
• The launcher table path is python-STDLIB-ONLY — no PyYAML, no pip deps (community VMs ship bare python3; #584's ModuleNotFoundError: yaml). The --json contract path may require PyYAML but must fail with a Fix: hint, not a traceback. test-registry-emit-no-yaml guards both plus the locale cases.
• Repro before claiming fixed: modern Python coerces C → C.UTF-8, so plain LC_ALL=C won't reproduce — use PYTHONUTF8=0 PYTHONCOERCECLOCALE=0 LC_ALL=C.
• Don't blind-2>/dev/null launcher derive paths — swallowing the traceback hid this exact class for months; capture stderr and surface it on failure instead.

CHANGELOG

• CHANGELOG.md (cross-cutting) and models/<name>/CHANGELOG.md (per-model) are append-only history. Don't rewrite past entries even when a finding is superseded — add a new entry. The historical trail is load-bearing for "why did we do X."
• Old entries can reference files / patches that no longer exist. That's fine — leave them.

Compose variants

• Each variant ships with a header table comparing it against sibling composes in the same directory. Update both directions when adding/removing variants.
• Keep the variant set lean. Three composes that overlap badly (similar TPS + similar context + same KV) is worse than two with a clean differentiation. Removed fast-chat.yml 2026-04-29 for exactly this reason.

Compose layout — <model>/<engine>/compose/<topology>/<quant>/<serving>.yml

The directory hierarchy encodes model, engine, topology, and the weights artifact. The filename encodes only the serving stack. No level repeats information from another.

Path level	Encodes	Examples
models/<model>/	Model	qwen3.6-27b · gemma-4-31b
models/<model>/<engine>/	Inference engine	vllm · llama-cpp · sglang
compose/<topology>/	Hardware topology	single · dual · multi3 · multi4 · multi8
<quant>/	Weights artifact / weights_variant slug	autoround-int4 · ubergarm-iq4ks · awq
<serving>.yml (filename)	Serving stack	fp8-mtp.yml · turbo.yml · dflash.yml · int8.yml

Topology rule: single (TP=1) and dual (TP=2) have no count ambiguity. multi<N> requires the count because N varies (3 / 4 / 5 / 6 / 8). Aligns with docs/SINGLE_CARD.md / DUAL_CARD.md / MULTI_CARD.md doc framing.

Default rule: there is no filesystem default and no default.yml. Defaults live in scripts/lib/profiles/compose_registry.py (DEFAULTS) and are selected by registry tag (scripts/launch.sh, scripts/switch.sh, estate planner). Direct Docker use must pass -f <path>.

Default-resolver knobs (maintainer-owned, next to DEFAULTS in compose_registry.py):

• DEFAULTS[(model, engine, topology)] → slug — the <engine>/default map (club-3090's recommended config per engine; reason can evolve, edited by PR).
• ENGINE_PREFERENCE[topology] → [engine, …] — the curated <model>/default policy. The resolver walks this list and picks the first engine with a functional (status ∉ {experimental, preview, upstream-gated, deprecated}) DEFAULTS entry. Reorder a row to change a recommendation — no code change, any topology. single = [beellama, ik-llama, llamacpp, vllm]; dual/multi = [vllm, ik-llama, llamacpp, beellama]. beellama leads the single ranking and has live DEFAULTS entries — it IS the shipped single-card default for qwen3.6-27b (beellama/dflash) and gemma-4-31b (beellama/gemma-dflash).
• RECOMMENDED_DEFAULT_MODELS — a short opt-in shortlist (["qwen3.6-27b", "gemma-4-31b"]) of models eligible to be the bare-launch.sh default (first installed → its <model>/default). NOT an exhaustive ranking; absent models are runnable by name but never auto-default; new models are NOT auto-added — promote one explicitly.
• The shared resolver model_default_target(root, model, topology) (in registry-emit.sh) is the single injection point for both launchers. Precedence: --variant → user .env pin (CLUB3090_DEFAULT_<MODELID, non-alnum→_>) → community seam (community_default_target → None today) → curated walk → degradation (nearest-lower topology, else "pick explicitly"). X/default dispatch: X ∈ engine-set → engine rec; X ∈ model-set → model default; else error. Users pin/clear via switch.sh --set-default <slug> / --clear-default <model>.

Feature suffix order (when stacking): interconnect → drafter → KV → vision modifier. Examples:

• dual/autoround-int4/turbo.yml — TP=2 + AutoRound INT4 weights + TurboQuant KV
• dual/autoround-int4/dflash.yml — TP=2 + AutoRound INT4 weights + DFlash drafter
• (the nvlink- interconnect prefix is reserved but currently unused — NVLink is auto-detected at boot via NVLINK_MODE, not encoded in the filename)
• dual/autoround-int4/int8.yml — TP=2 + AutoRound INT4 weights + INT8 PTH KV
• dual/awq/bf16-mtp.yml — TP=2 + AWQ weights + BF16 KV + MTP
• multi4/autoround-int4/dflash.yml — TP=4 + DFlash

Plain / default cases: a compose with the engine-default KV and no drafter is base.yml — do NOT name the default KV (no bf16.yml when bf16 is the default; that's base.yml); name only a non-default KV (int8.yml, fp8.yml, tq3.yml). Workload-tuned variants (use-case ctx/sampling, not a feature delta) keep a descriptive name (long-text.yml, tools-text.yml, bounded-thinking.yml, minimal.yml). Names predating this are grandfathered — don't rename (it re-paths the registry compose_path). Full filename + registry-slug conventions: docs/ADDING_MODELS.md → "Build the first compose."

Concrete examples (post-2026-05-09 restructure):

• models/qwen3.6-27b/vllm/compose/single/autoround-int4/tq3-mtp.yml — Qwen single-card default
• models/qwen3.6-27b/vllm/compose/single/autoround-int4/long-text.yml — Qwen single-card text-only long-ctx variant
• models/qwen3.6-27b/vllm/compose/dual/autoround-int4/fp8-mtp.yml — Qwen dual default (fp8 + MTP)
• models/qwen3.6-27b/vllm/compose/dual/autoround-int4/turbo.yml — Qwen dual + TQ3
• models/qwen3.6-27b/vllm/compose/multi4/autoround-int4/dflash.yml — Qwen 4-card + DFlash
• models/gemma-4-31b/vllm/compose/dual/autoround-int4/bf16-mtp.yml — Gemma dual default (bf16 + MTP)
• models/gemma-4-31b/vllm/compose/dual/autoround-int4/int8.yml — Gemma dual + INT8 PTH KV

Filename collisions across topology and quant dirs (e.g. dual/autoround-int4/dflash.yml vs multi4/autoround-int4/dflash.yml) are fine — the path disambiguates. Registry tags in scripts/switch.sh decouple from filesystem paths; rename only the file path in the registry / VARIANTS map and keep tags backward-compatible.

Fine-tune convention: model-specific fine-tunes that share the canonical model's compose directory get their own quant slug (for example dual/carnice-bf16mtp/bf16-mtp.yml and dual/qwopus-bf16mtp/bf16-mtp.yml under models/qwen3.6-27b/). Long-term those fine-tunes can graduate to their own model directories (models/carnice-v2-27b/, models/qwopus3.6-27b/) when their variant set warrants it.

Patches and caches stay engine-level (NOT under a topology)

<model>/<engine>/patches/ and <model>/<engine>/cache/ sit parallel to compose/, not under any topology subdirectory. Three reasons:

1. Patches are reused across topologies and quant variants. vllm-marlin-pad/ is mounted by dual and multi-card AutoRound composes. Putting it under one topology or quant dir would force the others to symlink or duplicate.
2. Patches are scoped by (model, engine), not by topology. A vLLM source override doesn't change based on TP=1 vs TP=4; it's an in-container engine-internal patch that applies the same way regardless of mount layout.
3. Caches (torch_compile/, triton/) warm-start across composes. Sharing them at the engine level means switching from single/autoround-int4/tq3-mtp.yml to single/autoround-int4/long-text.yml reuses the JIT'd kernels instead of recompiling.

Relative paths from a compose to its sibling patches/caches: ../../../patches/... and ../../../cache/... (one .. from <quant>/ to <topology>/, second to compose/, third to the engine dir, then patches/ or cache/). Repo-root mounts such as scripts/ and models-cache/ need one extra ../ for the quant layer too.

If a future patch is genuinely topology-specific (e.g., a kernel rewrite that only applies to TP=2), keep it at <engine>/patches/<patch-name>/ and document the topology constraint in the patch's README. Discoverability ("one patches/ per engine, search there") trumps the marginal benefit of a topology partition.

Profile schema header (every compose, every time)

Every compose starts with a Profile (at-a-glance) block declaring the (Model, Topology, Drafter, KV, Vision, Max-ctx, Genesis) tuple in structured form. Free-form description follows below the schema, not in place of it.

# ===========================================================================
# Profile (at-a-glance):
#   Model:     <name + quant — e.g. "Qwen3.6-27B (Lorbus AutoRound INT4)">
#   Topology:  <e.g. "Dual 3090 PCIe (TP=2, no NVLink)">
#   Drafter:   <none | MTP n=N | DFlash n=N | ngram K=N>
#   KV:        <fp8_e5m2 | bf16 | int8_per_token_head | turboquant_3bit_nc>
#   Vision:    <yes | no>
#   Max ctx:   <e.g. 262K>
#   Genesis:   <none | v7.72.2 | N/A — Genesis is Qwen3-Next-specific>
#   Status:    <REQUIRED — exactly one of the enum values below>
#   Caveats:   <REQUIRED if Status is ⚠️ / 🐣 / 👁️ / ⏸️ / 🗑️; otherwise omit>
#   Quality:   <OPTIONAL — populated by `bash scripts/quality-test.sh --medium`>
#                e.g. "ToolCall-15 14/15 (93%) · InstructFollow-15 13/15 (87%)
#                      · StructOutput-15 15/15 (100%) · DataExtract-15 12/15 (80%)
#                      (--medium, packs v1.0.x, 2026-05-09)"
#   Best for:  <one short phrase — what workload this serves; ⭐ for canonical>
# ---------------------------------------------------------------------------
# (existing free-form description continues below)

Status enum — pick exactly one:

Value	Meaning	Validation gate
✅ Production	Recommended for users.	verify-full 8/8 + verify-stress 7/7 + bench (BENCHMARKS row) + soak-continuous PASS + quality-test.sh --quick PASS (no ≥10pp regression on ToolCall / InstructFollow vs the pre-change baseline). Quality numbers on the compose's Quality: schema field when --medium has been run.
⚠️ Production w/ caveats	Works under documented constraints; not the same as broken.	Same gates as Production, but a known-and-disclosed limitation exists (e.g., Cliff 2b at >50K, or a >10pp drop on a specific quality pack). Caveats line MUST list the constraint.
🧪 Experimental	Under active validation; may not boot or pass all tests.	Typically untracked in git. No production guarantee.
🐣 Incubating	Pre-experimental: works, but not ready for the actionable list — a niche specialist or one that fails the standard gate by design (e.g. an always-reasoning model with no tool-calling).	HIDDEN from switch.sh --list by default (revealed by --list --all), and launch requires --force (non-functional). Caveats line MUST state why it's not gate-passing. Promote to 🧪/✅ when it earns the actionable list.
👁️ Preview	Known quality issues; tracked but not for production.	E.g., quality regressions in soak / NIAH. Caveats line MUST list specific issues.
⏸️ Upstream-gated	Exists but blocked by external action (PR merge, driver fix, hardware ceiling).	Boots only with vendored override OR doesn't boot until external dep lands. Caveats line MUST point at the external dep.
🗑️ Deprecated	Kept for historical reference; will be removed.	N/A — flagged for cleanup.

Why this enum exists: the previous "Status optional, only when not production" convention left readers guessing whether absence-of-status meant "validated production" or "author forgot to fill it in." Making Status required + enumerated removes that ambiguity. Users picking a config can scan to one field and know the lifecycle stage instantly; new contributors must consciously declare it when authoring.

The Caveats: line is REQUIRED whenever Status is ⚠️ / 🐣 / 👁️ / ⏸️ / 🗑️, OMITTED for ✅ / 🧪. Format: a single-line summary or a short bullet list, with links to issues / discussions / upstream PRs where relevant.

This rule applies to shipped composes AND local-only test composes — apply the convention even before deciding whether to ship; it avoids a rename later if the experiment graduates.

When testing a new model, create the directory hierarchy from the start: models/<new-model>/<engine>/compose/<topology>/<quant-slug>/<serving>.yml. The quant slug must match the weights_variant key in scripts/lib/profiles/models/<model>.yml and scripts/lib/profiles/weights.py. When the model isn't Qwen3-Next, write Genesis: N/A — Genesis is Qwen3-Next-specific in the profile schema so readers don't expect Genesis-style perf folds where they don't apply.

Adding a model — full workflow → docs/ADDING_MODELS.md

Read that doc before catalog work; the at-a-glance for agents:

• Just serving, not cataloging? Safetensors → scripts/pull.sh <org/Model> --profile-like vllm/minimal; a self-grabbed GGUF → copy an ik/llama compose and point --model at it (no registry/profile needed — see ADDING_MODELS "Run a local GGUF without the catalog"). The steps below are only for promoting a model into the curated catalog.
• Adding a new QUANT/SLUG of an existing model? That's the most common catalog change and it has its own checklist: docs/ADDING_MODELS.md → "The lightweight path". Follow it verbatim — the three items agents skip (and shouldn't) are the FULL test suite (a new slug IS a catalog-shape change), diagnose-profile.sh <slug>, and booting the ACTUAL compose via switch.sh + verify-full (an equivalent hand docker run does not count).
• Catalog steps the compose alone doesn't cover: (1) scripts/lib/profiles/models/<id>.yml — weights: is a map keyed by quant-slug, not a list; (2) a compose_registry.py entry (weights_variant=slug · kvcalc_key — vLLM "<model>:<profile>", ik/llama "SKIP" · default_port == the compose's ${PORT:-NNNN}); (3) launchers auto-derive from the registry — never edit launch.sh/switch.sh; promote a default via the DEFAULTS map.
• Profile-catalog compatibility (easy to miss — hotfix #236): the new (model, engine, KV-format) combo must validate or test-profiles-compat / diagnose-profile go red. Add the model's family to the engine's supported_model_families (scripts/lib/profiles/engines/*.yml), the KV format to the hardware profiles' supported_kv_formats (scripts/lib/profiles/hardware/*.yml); register any vendored chat-template in scripts/lib/profiles/patches.yml (with the symmetric-protocol drift_guard); bump the test-compose-registry-disk size-count.
• Run the FULL catalog test suite for catalog-shape changes (new model / engine / slug / profile-schema change), not just the serving tests in Tests: for t in scripts/tests/*.sh; do bash "$t"; done. Key gates: test-compose-registry-disk, test-compose-mounts-resolve (the ../ depth), test-model-weights-registry, test-switch-registry-parity + test-launch-registry-parity, test-profiles-compat, test-patch-attribution, plus tools/kv-calc.py --calibration. A narrow subset shipped a model with two real catalog gaps (#236) — and some failures are pre-existing/env, so baseline against the last release tag before treating one as a blocker. Scoping rule: the full sweep is for catalog-shape changes only — for a scoped change (one compose's flags, a doc, a single script) run just the guards that touch what you changed; most gates are irrelevant to it and the sweep wastes the signal.

Where do experimental / unvalidated composes live?

Same directory as shipped composes, but kept untracked until validation passes. Don't create a separate experimental/ subdirectory — the relative paths to ../patches/... and ../cache/... are calibrated to the compose dir, and promoting an experiment from a sub-folder would require re-pathing every mount.

Workflow:

1. Author the compose in models/<model>/<engine>/compose/<topology>/<quant-slug>/<serving>.yml with the standard profile schema header. Mark Status: ⚠️ EXPERIMENTAL (or ⚠️ PREVIEW if quality issues are known) so readers know it's not validated.
2. Don't git add until validation passes. The file shows up in git status as ?? — that's the signal. git ls-tree -r HEAD lists only shipped composes; the gap between that and ls compose/*.yml tells you what's pending validation.
3. Validation gates before promoting: verify-full.sh 8/8, verify-stress.sh 7/7 (or documented failures with rationale), bench.sh (numbers added to BENCHMARKS.md), soak-test.sh SOAK_MODE=continuous (catches Cliff 2b), quality-test.sh --quick (no major regression on ToolCall / InstructFollow). For pin bumps and new quants, run quality-test.sh --medium and add the result line to the compose's Quality: schema field.
4. Promote: drop the Status: ⚠️ EXPERIMENTAL line from the profile schema, git add, commit. Cross-rig validation can come later via the numbers-from-your-rig issue template.

For entirely new models under validation (e.g. "let's try MiniMax-M2.7"): keep the whole models/<new-model>/ directory untracked until at least one compose validates. Avoid pushing models/<new-model>/README.md etc. before there's a working compose to back it up — empty model directories on master signal capability we don't actually have.

(The 2026-05-09 orphan set has since shipped or been removed — don't expect specific untracked files; the git status ?? gap is the live signal.)

Retired composes → compose/_archive/. A compose that's fully superseded but still referenced by history (CHANGELOG entries, learnings, old discussions) moves to models/<model>/<engine>/compose/_archive/<topology>/... instead of being deleted: it keeps old links resolving while staying out of the registry (no slug, not launchable, invisible to switch.sh --list). This is one step beyond 🗑️ Deprecated (which keeps the registry entry — see the Status enum): deprecate when users may still reference the slug; archive when nothing but history points at it.

Documentation

• Don't create new docs proactively. Most non-obvious things belong in INTERNALS.md, FAQ.md, SINGLE_CARD.md, or DUAL_CARD.md. New top-level files only when there's a recurring search miss.
• Charts: source .svg + exported .png at retina resolution (≥1500px wide). Markdown embeds use .png (clicking opens a viewable image; SVG opens raw XML). Re-generate with python3 tools/charts/gen-perf.py and gen-vram.py after editing data.
• For any change that adds a footnote to "this depends on upstream X" — the answer is to link the row in docs/UPSTREAM.md, not to inline-cite the upstream URL.

Tests

• verify.sh — fast smoke (~15s). Confirms the stack is responding; runs after setup.sh.
• verify-full.sh — functional (~1-2 min). Runs on every compose change.
• verify-stress.sh — boundary cases (longctx ladder + tool-prefill OOM ~5-10 min). Runs on cliff-related changes.
• bench.sh — canonical TPS bench (~3-5 min). Run when you change anything that could move TPS (compose flags, Genesis env vars, vLLM pin).
• quality-test.sh — behavioral quality (~10-30 min depending on --quick / --medium / --full). Wraps benchlocal-cli — runs verifier-backed bench packs (ToolCall-15, InstructFollow-15, StructOutput-15, etc) against the running endpoint. Catches what operational tests miss: a compose can pass verify + stress + bench + soak and still ship with degraded tool-call accuracy or instruction-follow drift from quantization or Genesis env-flips. Run before promoting Status: ✅ Production and before any pin bump that could shift behavior. Run via this wrapper, not raw benchlocal-cli — it auto-detects endpoint/model and, for localhost URLs, sets BENCHLOCAL_HERMES_RESOLVE_LOCALHOST=1 so the sandboxed HermesAgent can reach the host model (direct benchlocal-cli skips that → hermes silently scores ~0/20). Live per-scenario [N/M] progress is on by default (the wrapper forwards --progress to benchlocal-cli) so long runs don't go dark; pass --no-progress only for CI / log-volume contexts. Timeout sizing: the wrapper auto-scales per-scenario timeouts (startup decode-TPS probe × thinking-token multiplier), deliberately over-budgeting — the fix for the thinking-on spurious-timeout class (benchlocal-cli #54/#59). Don't hand-set --timeout-per-case to "fix" a slow run unless you've confirmed the probe measured wrong. A planned opt-in tier will size from a soak-derived per-depth TPS curve (#114). Full precedence + flags → docs/QUALITY_TEST.md "Per-scenario timeouts".
• soak-test.sh — stability (30-60 min). Run before shipping config / Genesis / memory-policy changes — catches Cliff 2b.
• rebench-full.sh — the canonical full-eval orchestrator: use this instead of hand-sequencing the scripts above. verify-full preflight (fail-fast) + 5 measured steps in the docs/QUALITY_TEST.md pipeline order, with the recurring manual-run mistakes guarded (wrong cwd, missing --save-json, forgotten MODEL= override, missing hermes localhost env, wrong port) and --resume idempotency so an interrupt doesn't redo the matrix.

Long-running tests: redirect full output to a log file — NEVER pipe through tail/head/grep. A pipe buffers until the process exits, so a --full quality run piped to tail -12 is a ~2 h black box: no live [N/M] progress, no partial scores, and an interrupt leaves nothing readable (benchlocal also writes its results JSON only at completion — noonghunna/benchlocal-cli#82 tracks scenario-level resume). Do bash scripts/quality-test.sh --full ... > /path/run.log 2>&1 (or | tee) and summarize from the file; tail -f the file for live progress. Learned 2026-07-11 on a template A/B.

Gotcha (all serving tests): verify-*/bench.sh default MODEL=qwen3.6-27b-autoround — against any other served model that's a silent HTTP 404. Pass MODEL=<served-name> explicitly (or use rebench-full.sh, which autodetects).

The pipeline is layered: each script has a different question it answers ("does it serve / work / survive / fast / behave correctly / stay healthy"). Skipping any layer can mask regressions.

Running a full eval — two non-overlapping passes

Behavioral quality (the 8-pack) and operational health (verify / stress / soak / bench / agentic) split cleanly. Run one of each — together they cover everything with nothing run twice:

One-time setup (quality): the suite wraps benchlocal-cli; three of the eight packs (bugfind-15, cli-40, hermesagent-20) run inside Docker sandboxes that build once:

pip install git+https://github.com/noonghunna/benchlocal-cli.git
git clone https://github.com/noonghunna/benchlocal-cli
bash benchlocal-cli/tools/build-sandboxes.sh   # ~30 GB free; `docker system prune` if tight

(Without the images the 5 deterministic packs still run; the 3 sandboxed ones skip with a warning.)

1. Behavioral quality — the 8-pack, both reasoning modes:

   bash scripts/quality-test.sh --full --no-thinking       # reasoning OFF
   bash scripts/quality-test.sh --full --enable-thinking   # reasoning ON
   

   ⚠️ For the reasoning-ON leg on a thinking model, boot the compose with reasoning parsing on (REASONING=on for llama.cpp composes, --reasoning-parser for vLLM) so <think> lands in reasoning_content, not the graded answer.
2. Operational health: bash scripts/report.sh --full (~43 min; redacted, paste-ready bundle — verify + stress + soak + bench + agentic).

Don't pair rebench-full.sh with report.sh --full — rebench re-runs the same operational gates (verify/bench/stress/soak), so it replaces report.sh --full rather than complementing it. Pick by goal: rebench-full --with-8pack-thinking=both when you want one synthesized REPORT.md (quant A/B, BENCHMARKS row); the two-pass split above when you want the paste-ready cross-rig bundle. If you ran rebench-full and also want the agentic curve, add only report.sh --agentic. The same guidance ships user-facing in docs/ANNOUNCEMENT_TEMPLATE.md §7 "Run the evals".

serve-cockpit (c3)

tools/serve-cockpit/ is the Textual TUI cockpit — a separate Python app with its own venv and pytest suite, NOT covered by scripts/tests/*.sh. See its README.md. For agents:

• Run tests with tools/serve-cockpit/.venv/bin/python -m pytest tools/serve-cockpit/tests/ -q. test_services.py + test_registry_parser.py are fast — run them on every c3 change; test_app_headless.py boots the full app and is slow — prefer targeted -k selection while iterating.
• c3 consumes the registry-emit.sh --json contract. Adding a field to the emit means threading it through services.py (_variant_row_from_dict) and, if displayed, app.py — and emit changes also need the test-switch-registry-parity / test-launch-registry-parity guards green.
• DataTable cells render in terminals: avoid U+FE0F variation-selector emoji (⚠️ 👁️ ⏸️ 🗑️) in fixed-width columns — Rich reserves 2 cells but many terminals draw 1, misaligning every column after it. Use Emoji_Presentation=Yes glyphs (see _STATUS_GLYPH in app.py).

Commits

• New commit per logical change. Don't amend published commits.
• Commit messages: subject ≤72 chars, imperative ("Disable P68/P69..." not "Disabled..."), optional body for "why."
• Don't push without local verify-full + verify-stress passing for the affected compose.

Hooks / verification

• Pre-flight checks live in scripts/preflight.sh and run from setup.sh + launch.sh. They check docker / GPU / disk before any heavy work. If a check fails, print an actionable Fix: hint, never a cryptic mid-run crash.

Things to NOT do

• Don't add NVLink suggestions. The user has explicitly declined.
• Don't recommend EAGLE / DFlash spec-decode on Qwen3-Next single-card. It's blocked by DeltaNet rollback (see docs/UPSTREAM.md → vllm#39931). MTP works.
• Don't enable Genesis behavioral patches (P68/P69 class) by default. They override user intent. If a user wants them, they can flip the env var.
• Don't claim a TPS number you didn't measure. "Should be ~80" labeled as estimate is fine; "is 80" needs a bench.
• Don't compress historical CHANGELOG entries. Append-only.
• Don't scatter upstream issue links across multiple docs. Link to the row in docs/UPSTREAM.md instead.