command-center/agent/memories/--home-blank-Projects-serverAppConfig--/learned.md

• frank-large (halogen) healthcheck failure signature: the unit's healthcheck is halogen-healthcheck api with a 10 s timeout. During an engine PING-silence stall (log: "halogen: the engine has not answered PING for 45s"), /health HANGS rather than 503ing, so the healthcheck exits 125 (timeout) and podman shows the container (unhealthy) with a small FailingStreak. This is the documented transient stall — it self-recovers (streak resets to 0, /health 200) within ~1-2 min; do not restart over it. A completion request landing in the stall window takes ~60 s and then returns 200. A leftover failed transient *-132a6f9850c6fbd8.service (systemd-run podman healthcheck) may linger in the user manager — reset-failed it. (context: 2026-09-17 11:28-11:30 PDT: frank-large 4x FailingStreak exit 125, self-recovered by 11:31; 65 s completion probe straddled the tail.)
• Fleet transport: blankpad has NO passwordless SSH to FrankBook-Pro — dank@192.168.0.112 rejects publickey (the 2026-09-12 key install only covered dankclaw/dankserver). Do not burn calls probing Mac SSH. Verify the *arr stack from the LAN instead: port probes at 192.168.0.112:{9696 prowlarr, 8989 sonarr, 7878 radarr, 8686 lidarr, 8787 readarr, 8080 qbittorrent} — 302 (login redirect) and 200 are both "up". Container-level inspection on the Mac needs a key installed on the Mac or a Mac-side session. (context: 2026-09-17 fleet-wide container audit; ssh dank@192.168.0.112 → Permission denied (publickey), six-port LAN probe all answered.)
• Long-context speculative-decode tuning on llama.cpp: at 250k-token contexts, INCREASING draft depth (n-max 4) and busy-wait (poll 100) made decode WORSE (230–257 t/s) than depth 3 / poll 50 (257–302 t/s) — each longer verification round re-reads the full-context attention of the full-attention layers every round, which outweighs the extra accepted token. The productive prefill lever on a 24 GiB 3090 was ubatch 512-1024 (+5% pp at 254k, better tg too); ubatch 2048 is impossible because a spec draft context allocates its OWN pp compute buffer (~3.3 GiB at ub 2048 for a 12B model) against only ~2.2 GiB headroom AFTER the main context fits — the OOM signature is common_speculative_init_result: failed to create MTP context / graph_reserve: failed to allocate compute pp buffers AFTER the main model loaded fine. Also: idle VRAM is identical across logical batch sizes (4096 vs 8192) at the same ubatch — compute buffers reserve at max graph size, so a bigger logical -b is free. Verify with salted API probes (cache_prompt:false + uuid prefix) using the timings keys of modern builds: prompt_per_second / predicted_per_second / draft_n / draft_n_accepted. (frank-small, dankclaw 3090, 2026-09-17, deployed as 8192/1024, n-max 3, poll 50) (context: Optimizing frank-small (Gemma4-12B QAT + MTP sidecar, 4x262144 slots, q8 KV, RTX 3090) for max prefill+decode at max context.)
• In the serverAppConfig checkout, concurrent peer sessions can stash YOUR uncommitted WIP and then DROP that stash (a peer did exactly this 2026-09-17 while deploying frank-small, dropping a stash that held the frank-medium env+seam changes). Dropped stash commits are dangling but reachable: git fsck --no-reflogs lists them, and the files recover cleanly with git checkout dangling-sha -- paths (restores content, no ref touched). Check git stash list AND git fsck for your files before re-doing work after a peer session has been active. Related standing rule for this repo: the deploy gate rejects a dirty tree, so a peer's uncommitted WIP blocks your deploy — stash their files (theirs, not yours), deploy, pop. (context: serverAppConfig 2026-09-17: peer's frank-small deploy stashed then dropped my frank-medium WIP; recovered via dangling stash commit cd0a554.)
• llama.cpp draft-mtp spec-decode cost law (calibrated 2026-09-17 on frank-medium, 2x R9700 layer split): the draft context shares the main model's weights (shares_model — no sidecar file, no second model copy), so the cost is one nextn layer's KV at 4096 B/position × the TARGET's ctx (the draft context is sized from the target, same law as any drafter) PLUS draft buffers that a formula-based fit model underestimates badly: at 393216 ctx the MTP total charge on the last-layer card measured 4.20 GiB vs the fit's ~1.7 GiB — 2.46 GiB hot, which moved the card's headroom from a predicted 3.54 to a real 1.08 GiB and forced the floor down to 1.0 GiB. Calibrate the fit constant to the measured rocm-smi load, not the formula. The MTP/DFlash2 trade on this host: MTP frees ~1.1 GiB on the drafter card (no 1.92 GiB sidecar + buffers) at the cost of ~11 tok/s (32.4 vs 43.5 measured) — that freed space is what made 2x196k contexts fit. (context: frank-medium 2026-09-17: switched DFlash2 - MTP drafter to fit a 393216 pool (2x196608); ROCm1 loaded 30.78 GiB vs fit 28.32; seam constant recalibrated to 2,901,231,913 B.)
• llama.cpp --fit with an assistant-style MTP sidecar (Gemma4: mtp-gemma-4-12B-it.gguf, arch gemma4-assistant) prints at boot llama_init_from_model: failed to initialize the context: Gemma4Assistant requires ctx_other to be set + failed to measure the memory of the extra model, fitting without it — this is the fit probe, self-described normal in the log, and the draft loads against the real context seconds later. Do NOT restart over it. Also: Gemma4 SWA + q8_0 KV made --fit take the full 262144 window in ~12 GiB of a 24 GiB RTX 3090 (idle 11,996 MiB with 7.4 GiB of Q4_K_M payloads), and Gemma4's chat template accepts every reasoning_effort level (all seven probed 2026-09-17: 200), unlike Qwen3.8's xhigh/medium/low gate. q8 KV on CUDA required rebuilding build-cuda with -DGGML_CUDA_FA_ALL_QUANTS=ON. Deployed as frank-small on dankclaw :7771 with --spec-type draft-mtp,ngram-mod (acceptance 0.38-0.46, mean len ~2.2). (context: frank-small Gemma4-12B QAT cutover on dankclaw 2026-09-17 (commits 4b7de59, 87caca4; runbook serverAppConfigs-docs/LLM/frank-small-gemma4-12b.md))
• User preference (Frank, 2026-09-17): never do multi-step arithmetic by hand in prose — "stop doing math without python, you're an LLM. your handling numbers is like expecting determinism out of a pachinko board." All VRAM fits, per-token costs, gap/slack computations and the like must go through the eval tool (Python) and quote the computed numbers. Hand arithmetic is acceptable only for trivial single-step checks. (context: frank-medium VRAM fit for 36:29 @ 327680 — I had computed card balances in-head before the user corrected)
• llama.cpp mainline (fb27a525, the frank-medium build) does NOT autoload mmproj*.gguf from the model directory: a server started without an explicit --mmproj flag comes up healthy and text-only, answers image requests with 500 "image input is not supported - hint: if this is unexpected, you may need to provide the mmproj", and the load log contains no mmproj line at all — so /health and the journal both look fine. Always pass --mmproj file explicitly (the frank-small pattern), and verify vision behaviorally (send an actual image) after any deploy that touches the projector, because health checks cannot see it. Cost of the trap: one full extra deploy cycle on 2026-09-17 (frank-medium 36:29/327680 cutover). (context: frank-medium 2026-09-17: NO_MMPROJ=0 + mmproj-BF16.gguf beside the model, first deploy healthy, vision 500'd; fixed with explicit --mmproj in serve-frank-medium.sh)
• Measuring batch geometry and VRAM balance on a ROCm llama.cpp host — three durable rules (Dankserver 2026-09-16, frank-medium, 2x R9700, scripts/bench-batch-ubatch.sh): 1. SETTLE GATE, or your infeasibility verdict is an artifact. Loading a config, reading rocm-smi, killing it and immediately loading the next one fails with error loading model: unable to allocate ROCm0 buffer — amdgpu does not return ~26 GiB the moment the process dies, and a 180 s settle is NOT enough (600 s was still not enough in one run; the VRAM came back after the transient unit's cgroup was torn down). That failure is indistinguishable from "this configuration does not fit". Gate each load on rocm-smi showing both cards near-empty (awk over --showmeminfo vram --csv, wait until used 1 GiB per card), and run each config in its own transient unit so systemd's cgroup teardown does the killing. 2. A "plateau" claim must be tested from BOTH sides. Sampling only 512→2048 showed a shallow basin (1135-1170 tok/s) that read as "the knob does not matter"; extending downward found the actual PEAK at ubatch 512 (1193) with real losses on both sides — 256 → 1046 and 384 → 1068 (SMs not saturated), 2048 → 1077 (10 % off). The deployed 1536 was thus both 2.4 % slower than the optimum AND 3.71 GiB more expensive. 3. THE SPLIT IS A FUNCTION OF THE BATCH, because the compute buffers are not symmetric. Shrinking ubatch 1536 → 512 freed 1.30 GiB on the plain card but 2.41 GiB on the card holding the output layers and the drafter's graphs — so the balance point moved a layer (36:29 → 35:30 on a 65-layer hybrid) and the reachable minimum gap became one DeltaNet layer (0.46 GiB), not zero. Change the batch and you must re-measure the split. Method notes that made the result trustworthy: llama-bench records n_ubatch per test in its -o json output, so "is the knob even applied" is answerable from the artifact; and the VRAM columns scale monotonically with ubatch, an independent confirmation. The falsification test for (context: Dankserver 2026-09-16: swept llama.cpp --batch-size x --ubatch-size on frank-medium (2x R9700 layer split, Qwen3.8-27B UD-Q8_K_XL + DFlash2) and applied the result. Operator had said "that sounds like something's wrong" about a flat prefill curve.)
• In serverAppConfig, other sessions commit to the same checkout while you work, so line-anchored edits rot within minutes. Twice on 2026-09-16 an edit landed in the wrong section for the same reason: (1) AGENTS.md was edited against a read taken many turns earlier, and the peer session's commits had shifted the file — the frank-large text ended up split between the Model Config section and the service section, needing a repair; (2) an append at N used an N taken from wc -l, but the file had no trailing newline so its last content line was N+1 — the new section was inserted before the paragraph's final line, orphaning it. Discipline that prevents this: re-read the target region (read tool) immediately before editing, take the tag and line numbers from that read, prefer text-anchor insertions (PUT N: where N is a line you just saw) over remembered numbers, and immediately after an edit whose result looks even slightly off, re-read the region and check the surrounding structure (headings, list continuity, closing braces) rather than trusting the edit's summary. git diff --stat plus a targeted sed -n on the changed range is the cheap verification; for multi-hunk edits in a large markdown file, verify by reading the file, not by reasoning about the diff. Also: wc -l undercounts files that lack a final newline — never use it as the append anchor. (context: Dankclaw/serverAppConfig 2026-09-16: while documenting the x4 link and Q4_K_S results, a peer session was committing frank-medium work to the same checkout; a stale-anchor AGENTS.md edit split a section, and a wc -l-based append split a paragraph in the x4 note. Both were caught by re-reading the files and repaired before commit.)
• When a serving engine keeps some weights in host RAM, the decisive question is whether it streams them across PCIe or computes them there — the link decides the first and is irrelevant to the second. Dankclaw's 3090 is on Gen4 x4 (measured H2D 6.72 GB/s, D2H 6.60 GB/s = 85 % of the 7.877 GB/s spec; idle-downshifts to 2.5 GT/s x4 and trains up under load). For an A3B MoE whose active experts are ~1.70 GiB/token (Q4_K_M, 75 GiB expert mass), streaming them costs 272 ms/token ⇒ 3.7 tok/s (4.6 with a fifth resident), while computing the same bytes in host RAM at 256 GB/s costs 7.1 ms — the link is 31× slower than the memory behind it. Measured alternatives on the same host: 20.46 tok/s for llama.cpp with experts computed on the CPU (-ot exps=CPU; the link then carries activations only), 25 for halogen, 26.5 for mainline on the iGPU. So "is the link fast enough?" is answerable in one line of arithmetic: per-token bytes the engine must transfer ÷ measured rate = seconds/token, and compare that against the compute-in-place path. Two measurement subtleties from the probe: cudaMemcpyAsync on one stream was no faster than sync, and pageable transfers ran within 1 % of pinned — i.e. the copies were link-limited rather than staging-limited, so don't assume the usual "pageable costs 2-3x" rule. Also, hardware ceilings must be read from the platform port (LnkCap of the upstream bridge) and not just the card: the 3090 is x16-capable, the Strix Halo GPP port under it is x8-capable, so the x4 lives in the adapter or cable and the realistic upgrade is x8 (2×, ≈7-9 tok/s) — still behind every alternative. And FreeToken specifically (expert LRU in host RAM = exactly the streaming shape) loads only modelopt block-FP8 HF safetensors for qwen3_5_moe, so a Q4_K_M GGUF is unloadable there before bandwidth is even the question. (context: Dankclaw 2026-09-16: user asked whether the x4 link could support running the 111 GiB Q4_K_M Flash-Next model with FreeToken on the 3090 alone; built pcie-link-test.cu, stopped both services, measured, and derived the ~3.7 tok/s streaming ceiling vs 20.46 tok/s for llama.cpp's compute-in-RAM path.)
• FreeToken v0.1.2 is an "edge-native MoE serving engine that keeps EXPERT WEIGHTS IN HOST RAM with an LRU GPU cache" — so unlike llama.cpp's -ot exps=CPU (which COMPUTES experts in host RAM and moves only activations over the bus), FreeToken STREAMS expert weights across PCIe every token. That makes it the one design where link width is the critical path, and its own Dankserver measurements prove the mechanism: Qwen3.6-35B-A3B-FP8 on a 3090 gave 56.7 tok/s at ~89% expert residency but 25.2 tok/s at 5% residency, and the runbook itself notes the bottleneck "flips to PCIe expert streaming" at low residency — 25.2 tok/s × ~2.4 GB/token of misses ≈ 60 GB/s, i.e. an x16 Gen4 link saturated. Two consequences for Dankclaw, where the 3090 sits behind Strix Halo GPP bridge 00:03.1 (LnkCap 16GT/s x8, LnkSta x4; the platform's other ports are x1/x4/x16-for-iGPU, so there is no x16 slot to move it to): (1) x4 = 7.877 GB/s and x8 = 15.75 GB/s cap that design at roughly 4.6–10 tok/s for a 111 GiB A3B model, because ~1.7 GiB of active expert bytes per token cannot cross in time; (2) it cannot even load this file — FreeToken's loader matrix is modelopt block-FP8 HF safetensors for qwen3_5_moe only (compressed-tensors FP8 fails; GGUF is not supported at all). The measured ceiling for "Q4_K_M experts-in-RAM on just the 3090" is therefore the CPU-compute path at 20.46 tok/s decode / 60–81 t/s prefill, and the host's fastest single-model arrangement remains the iGPU (25.8 tok/s Q4_K_M, 26.5 Q4_K_S at 64k depth). Also durable: FreeToken builds for CUDA only (its _pinned_tensor kernel links cudart), and torch.cuda.is_available() can report False on a box whose kernels are fine. Separately, the model-payload rules that bit me on 2026-09-16: mradermacher ships large quants as .partNofM fragments (a .gguf filter hides Q4_K_S entirely); curl -o FILE URL exits 0 on HTTP 401/403 and writes the error body (three 153-byte "shards" with identical sha256 and a cheerful DOWNLOAD_CODE=done in 20 (context: Dankclaw 2026-09-16/17: user asked whether the 3090's PCIe Gen4 x4 link is fast enough to run the 111 GiB Flash-Next Q4_K_M under FreeToken on just the 3090. Answered from the FreeToken runbook's own mechanism and measurements, the observed Dankserver residency-vs-throughput table, lspci -vv topology of the Strix Halo GPP bridge (x8 cap, x4 trained) and the earlier measured x4 bandwidth; the same )
• Dankserver's frank-medium (2x Radeon AI PRO R9700, gfx1201/Navi 48) runs ROCm/HIP, never Vulkan: scripts/build-llama-rdna4.sh builds with -DGGML_HIP=ON -DGGML_CUDA=OFF -DGGML_VULKAN=OFF -DGGML_HIP_RCCL=OFF -DCMAKE_HIP_ARCHITECTURES=gfx1201 against a private bundled ROCm 7.14 tree (lib/rocblas, lib/hipblaslt, libamdhip64.so.7). Two log-reading traps that follow: (1) a HIP build still logs ggml_cuda_init: found 3 ROCm devices because upstream implements the HIP backend INSIDE ggml-cuda (so the CUDA name persists) — the giveaways are device names coming back as ROCm0/ROCm1 and ldd showing libamdhip64/librocblas/libhsa-runtime64; (2) ldd | grep -ci vulkan = 0 is the one-line proof of no Vulkan. Vulkan exists only on Dankclaw's separate build-dual tree (CUDA+Vulkan) used for the x4 3090+iGPU experiment; it has nothing to do with Dankserver. RCCL is compiled out by design because card #2 cannot dispatch an RCCL kernel (PCIe atomics), which is why the two cards are joined by llama.cpp -sm layer, not tensor/pipeline parallelism. (context: 2026-09-16, Dankserver: user asked "are we using rocm or vulkan?" while a llama-bench batch x ubatch sweep was running on frank-medium. Verified from scripts/build-llama-rdna4.sh:85-88, llama-bench --version (0.4.1-dev, fb27a525d), and ldd on the binary.)
• In the serverAppConfig homelab repo, more than one agent session can be committing at once (I shared it with a peer working on frank-medium tonight). That breaks line-anchored edits: I edited AGENTS.md using a tag from a read several turns earlier, the peer's commits had shifted every line below my anchor, and my two hunks landed in the wrong sections — the first split the "Model Config (harness)" section from the "frank-large" heading, the second inserted a paragraph mid-sentence so the halogen description read "halogen-flash-server 0.5.8 as a rootless Quadlet again since 2026-09-16 [paragraph] (dankclaw/frank-large/quadlet/...), serving ...". The tell was in the edit response itself: neighbouring lines that had no business being adjacent. Discipline that prevents it: re-read the exact target region immediately before editing (or grep the target text for a fresh line/anchor) rather than reusing a tag from earlier in the session, and always glance at the returned context lines to confirm they are the ones you expect; repair a mis-landed edit with a range rewrite (CUT the stray lines, PUT the whole section back in order) rather than incremental patching. Second, durable purge practice learned the same night: before deleting a "model cache" on a serving host, (a) inventory what is actually there, (b) find which dirs live units mount (grep Volume=/MODEL_DIR in the units) and protect them by name, (c) verify each service payload is re-fetchable upstream before deleting — query the Hub API anonymously, because a gated repo returns 403 even with the host's token, (d) record byte-size fingerprints before deletion so the re-fetch can be compared afterwards (that is how I proved halogen's overlays had been stale in place: 2,477,677,120 → 2,572,466,560 and 2,383,306,048 → 2,478,095,488 while the checkpoint returned byte-identical), and (e) expect the bind-mount trap when clearing a container's model dir: the directory must exist but be empty, or podman fails with Error: sta (context: Dankclaw/serverAppConfig 2026-09-16 night: operator ordered the model cache cleared with one-time rm permission; I purged ~600 GB (service payloads included) and re-fetched, which revealed the halogen sidecars had been stale in place. The same session's AGENTS.md doc update mis-landed because a peer session's commits had shifted the file between my read and my write.)
• You can characterize a live LLM endpoint's real performance WITHOUT stopping or benchmarking it, by mining the serving runtime's own per-request telemetry from journald: llama.cpp (mainline, verbosity 3) logs slot print_timing per task with prompt eval time, eval time, total time, graphs reused, draft acceptance = A (n accepted / n generated), mean len = L, plus newer n_gen = N, tg = X t/s, tg_3s = Y t/s; halogen-flash-server logs one line per request: serve_api: mtp N tok in Xs = Y t/s | R rounds, commit Z/round | prompt P (C cached), prefill Ss | detok Uus/tok. Root unit - sudo journalctl -u unit; rootless quadlet - journalctl --user -u unit (never sudo). This is the policy-compliant substitute on hosts whose rules forbid running a bench beside a serving GPU, and it measures the REAL workload rather than a synthetic one. To attribute PCIe cost, run nvidia-smi dmon -s t -c N -o T (columns rxpci/txpci in MB/s) in the background while issuing ONE ordinary request; measured on dankclaw frank-small mid-generation: peak ~230 MB/s tx against a 63 Gb/s (Gen4 x4) link = ~3% utilization, zero at idle - i.e. a 4-lane link costs a fully-VRAM-resident model nothing, because decode traffic is only the sampled logits stream. THE SSH TRAP: if you background the curl and your remote command exits, ssh closes the session and kills the client, which CANCELS the in-flight generation - the journal shows W srv stop: cancel task, id_task = N and NO print_timing line at all, so you get 200-looking HTTP with no telemetry. Always end the remote command with wait so the session outlives the request. Also: per-request numbers are content-dependent (measured MTP draft acceptance 0.46-0.76 on the same build, mean accepted length 2.39-3.27), so a live sample is not comparable to a deploy-time baseline without a masked llama-bench run. (context: 2026-09-16, dankclaw: user asked for "the performance of the model" after a PCIe x4 link discussion. Measured frank-small (llama.cpp + MTP on the RTX 3090) and frank-large (halogen on the iGPU) from production journal telemetry because AGENTS.md forbids benchmarking beside a live service, and proved PCIe x4 is not a factor with nvidia-smi dmon during one real 200-token generation.)
• Running llama-bench on the Dank fleet (verified 2026-09-16 on dankserver, frank-medium's UD-Q8_K_XL + 2x R9700 layer split) — four facts that decide whether a sweep measures what it claims: 1. llama-bench has NO -c/--ctx-size. Source line: cparams.n_ctx = n_prompt + n_gen + n_depth. The context is a CONSEQUENCE of the test shape, so llama-bench can never test whether a config allocates beside the deployment's real KV (frank-medium's 16 GiB at 262144 over 2 slots). That check must be done by starting the real unit. Same reason there is no --parallel and no --spec-: llama-bench measures single-sequence, drafterless numbers — its tg is the autoregressive baseline (~15.9 tok/s at 64k depth here) against the deployed 43.8 tok/s with DFlash2+ngram. -d N is a depth prefix that puts N tokens of KV in place before the measured run and is CACHED across reps via a state save/restore (llama_state_seq_), so the fill is not re-paid per rep and is not part of the measurement. 2. Separator trap: in llama-bench, / separates devices WITHIN one configuration and , separates configurations to sweep; -ts copies this. -ts 36,29 silently becomes one value (36) and puts everything on one card, benching a different machine. llama-server takes the comma form for the same pair. Also useful: list-valued args accept ranges — -ub 512-2048+256. 3. A test that cannot run calls exit(1) IMMEDIATELY (failed depth/prompt/gen run → llama_free + exit(1)), so one infeasible ubatch in a big cross-product invocation takes every later cell with it and loses the results. Drive one invocation per cell, continue on failure, and resume by skipping cells that already have a non-empty output file. 4. memlock: the deploy units set LimitMEMLOCK=infinity because the ROCm backend pins staging buffers; an SSH shell on these hosts gets soft=hard=8 MiB (ulimit -l), so a bare llama-bench run can fail the load for a reason that has nothing to do with the model. Run it inside a transient unit: sudo systemd-run --unit=n - (context: Dankserver 2026-09-16: operator asked for a batch x ubatch sweep at 64k context on frank-medium (full 4x7 cross-product, -p 65536 -n 128 -d 65536, -r 3 --no-warmup, no spec decoding). Harness written, dry-run tested with a stub llama-bench, deployed via git, run as a transient systemd unit.)
• Three traps hit in one model-payload fetch on the Dank fleet (2026-09-16), all worth knowing before the next swap. (1) mradermacher's GGUF repos ship the LARGE quants as .partNofM fragments, not .gguf: mradermacher/Qwen3.8-Flash-Next-Uncensored-i1-GGUF holds Q4_K_M as part1of3/2of3/3of3 (37.00+37.00+36.97 = 110.97 GiB — which is exactly how the deployed single-file payload was assembled) and Q4_K_S the same way (35.00+35.00+34.09 = 104.09 GiB), while the smaller rungs (Q2_K, Q3_K_, IQ3_, IQ4_XS) are single files. Listing the Hub tree and filtering for paths ending in .gguf therefore makes Q4_K_S look like it does not exist — I asserted that to the user twice before they pushed back. Filter on the quant name, not the extension. (2) A fetch script must use curl -f/--fail-with-body, compare byte counts to the Hub's declared size, and verify the LFS sha256; otherwise it reports success on an access error. curl -o FILE URL exits 0 on HTTP 401/403 and writes the error body, so a gated repo produced three 153-byte "shards" with identical sha256 values and a cheerful DOWNLOAD_DONE in 20 seconds for 104 GiB — caught only because the time and the duplicate hashes were absurd. (3) A token passed as curl -H "Authorization: Bearer …" is visible in /proc/pid/cmdline, so pgrep -a curl while a download runs prints the secret into your transcript (it happened); use curl -K mode-600 file or an HF_TOKEN environment variable, and never pgrep -a a command carrying an auth header. Related: pgrep -f in a wait loop matches the loop's own command line when the pattern appears in it, so a "wait for the download to finish" loop can spin its full iteration count without ever observing completion. Gated mirrors: orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF carries the same Q4_K_S ladder but returns 403 even with the host's HF token until that account accepts the gate — mradermacher's copy of the same quant is the ungated source. (context: Dankclaw 2026-09-16: user asked to download Q4_K_S of mradermacher/Qwen3.8-Flash-Next-Uncensored-i1 for a smaller-quant test; the fetch went through a wrong "the repo has no Q4_K_S" claim, a 403-on-gated-repo false success, and a token leak via pgrep -a, before landing a verified payload (111,769,915,680 bytes, parts hash-matched against the Hub's LFS oids).)
• Two durable traps from rebalancing frank-medium's tensor split (2026-09-16, Dankserver; revert DSpark-DFlash2): 1. A RECORDED ARTIFACT SIZE IS A CLAIM, NOT EVIDENCE — verify against the Hub's tree API. The fetch script declared the DSpark drafter as 1,359,284,737 bytes while the Hub's authoritative size was 1,455,376,576 (96 MB short, the LFS object's own size), so --verify-only failed on a payload that was in fact byte-correct (its sha256 DID match the LFS oid, because my recorded hash was taken from the host file, not the Hub — a self-consistent loop that proves nothing). The check that settles it in one call: curl -sL "https://huggingface.co/api/models/repo/tree/main?recursive=1" then read each entry's size and lfs.oid (a 128-char sha256, pointerSize:135 for the pointer file). Compare BOTH against stat -c%s and sha256sum on the host. Note the bare /api/models/repo endpoint does NOT list files; the ?blobs=true variant returned null siblings on these repos, and piping the tree into jq failed with "cannot use null as iterable" when the response was partial — parse it in python from a saved file instead. 2. WHEN A REPO CARRIES A FITTED VRAM MODEL (a test seam), FITTING AGAINST IT REQUIRES MATCHING ITS ARITHMETIC UNIT CONVENTION EXACTLY, or the constants you derive are meaningless. frank_medium_unit.bats expresses compute buffers as 1442 * 1024 * ubatch — i.e. the coefficient is ~1.4 MB per ubatch UNIT, not 1442 bytes — while my first fit (in python) used 1442 * ubatch. The result was a fitted card-0 constant of -54,541,757 B (a negative "constant", the tell) and a seam that failed the floor assertion by ~2 GiB: 144210241536 vs 1442*1536 is a 2.1 GiB difference. Read the assertion expression literally before fitting to it, and treat a negative constant as evidence of a unit error rather than something to leave in. Also worth keeping: bash -n file.bats is NOT a syntax check — bats' @test "..." { } form is not valid bash, so it fails on a valid file (line 32 in that f (context: Dankserver 2026-09-16: reverting frank-medium's spec drafter from magnitudedev DSpark to z-lab DFlash2 (DSpark's window-less draft KV spanned the full 262144 ctx = 5.0 GiB) and re-measuring the tensor split. The revert freed ~4.3 GiB on ROCm1 and moved the balance from 40:25 to 36:29. Repo: /home/blank/Projects/serverAppConfig, files scripts/frank-medium.env, tests/frank_medium_unit.bats, scripts/)
• llama.cpp multi-GPU layer-split sizing, learned on Dankserver (2x R9700, 262144 ctx targeted) 2026-09-16: 1. A SPECULATIVE DRAFTER'S KV IS SIZED FROM THE TARGET, NOT FROM ITS OWN FILE. common/speculative.cpp builds the draft context with cparams.n_ctx = llama_n_ctx(ctx_tgt) and there is NO flag to cap it independently (the --spec-draft-* surface has no ctx). So a drafter that declares no sliding window caches the WHOLE target context on whatever device --spec-draft-device names: DSpark (5 layers x 2 x 8 kv heads x 128 dim x 2 B = 20 KiB/position) costs 5.0 GiB at 262144, where DFlash2's dflash.attention.sliding_window = 2048 had held it to ~40 MiB. Read sliding_window (and block_size, n_ctx_train) out of the drafter's GGUF header BEFORE sizing. Signature of getting it wrong: the model LOADS, /health answers, and the FIRST draft decode dies with ROCm error: out of memory inside common_speculative_impl_draft_dflash::process -- an allocation failure after a successful load is a SIZING failure, not a broken drafter, and a crash-looping unit makes it look like the latter. 2. A BIGGER --ubatch-size IS ASYMMETRIC ACROSS A LAYER SPLIT. Measured 2x R9700 at 262144: ubatch 512 - 1536 cost the output/drafter card ~3.3 GiB against ~1.7 GiB on the other (that card holds the output layers AND the drafter's graphs). So a split balanced at one batch is UNBALANCED at another -- 38:27 was balanced at 2048/512 and left the draft-carrying card at 0.08 GiB free after the batch tripled. One -ts step is worth ~1.5 GiB at that size; the balance point must be found by measurement (38:27 - 3.78/0.08; 41:24 - 1.47/2.38; 40:25 - 1.88/1.88). Corollary: the balanced split is also what MAXIMISES ctx for a given per-card headroom floor, because the headroom has to come out of CTX_SIZE. Also measured: batch 8192/ubatch 2048 does NOT fit at 262144 (its 2020 MiB pp buffer fails to allocate on the busier card), while 4096/1024 fits but leaves only 1.11 GiB. 3. A GPU PAGE FAULT IS NOT AN OOM. One split (context: Dankclaw/Dankserver 2026-09-16: swapping frank-medium's drafter from z-lab DFlash2 to magnitudedev DSpark, then re-specifying batch/parallel/ctx per operator instruction. Repo: serverAppConfig, commits 7e11d24..718a283.)
• Dank-fleet VRAM accounting, learned while answering "what's the VRAM breakdown?" (2026-09-16). (1) THE TWO DEVICES ON DANKCLAW ARE NOT THE SAME KIND OF MEMORY. The RTX 3090 is discrete: nvidia-smi --query-gpu=memory.total,memory.used --format=csv,noheader (24,576 MiB total; 22,582–22,592 MiB used by the single llama-server PID serving frank-small, confirmable with nvidia-smi --query-compute-apps=pid,process_name,used_memory --format=csv,noheader). The gfx1151 iGPU is UMA: /sys/class/drm/card1/device/mem_info_gtt_total = 122,402,963,456 (114 GiB) and mem_info_gtt_used = ~36.2 GB (33.7 GiB) is the real "VRAM", while mem_info_vram_total = 2,147,483,648 (a 2 GiB carve-out window) with mem_info_vram_used ~208 MB. So iGPU VRAM IS host RAM, and a phrase like "halogen holds 95.4 GiB" is misleading: halogen's ledger shows 68 GiB of weights PINNED AS HOST PAGES (mlock/page cache, not GTT) plus ~33.7 GiB inside the device budget (11.0 GiB KV pool over 400,384 positions + working memory/scratch). Its own measurement lines are the best source: podman logs frank-large | grep -aiE "weights pinned|kv pool|left for everything|overlay". (2) THE REAL PER-BUFFER BREAKDOWN IS PRINTED BY llama.cpp AT INFO — llama-model.cpp:1827 ("%12s model buffer size = %8.2f MiB", per backend), llama-context.cpp:692 (compute buffer size), plus the KV self-size line — BUT NO ARTIFACT IN THIS FLEET CONTAINS IT: llama-bench suppresses the loader's INFO (its logs show only the device banner and the results table), and sudo journalctl -u frank-small | grep -c "model buffer size" = 0 across the unit's whole history. To actually get it, launch the model once through llama-cli or llama-server (not llama-bench) and keep the first ~200 lines; add -v/--verbose if the level filters it. Otherwise compose the breakdown by arithmetic: read the GGUF tensor table for group sizes (on Flash-Next: experts 75.0 GiB / 48 layers = 1.5625 GiB per layer, per_layer_token_embd 32.8 GiB, everything else 3.2 GiB / 48 = 0.067 GiB p (context: Dankclaw 2026-09-16: after the 3090+iGPU split experiment, reconstructing what the VRAM looked like per device (measured totals from nvidia-smi + sysfs, halogen's ledger lines, GGUF tensor arithmetic) because no log carried llama.cpp's per-buffer accounting; also pinned down llama-bench's derived context lengths (n_ctx = n_prompt + n_gen + n_depth) for the runs whose parameters were being reviewed)
• Working convention on the Dank fleet for answering "what was running when it broke": experiment artifacts must survive on disk because /tmp does NOT (/tmp is cleared by a reboot — after the 2026-09-16 Dankclaw incident the x4*.sh experiment scripts were gone while their logs in /home/dank/models/logs/ survived). So launch experiment scripts from a persistent path and log to /home/dank/models/logs/tag.log. Reconstruction then works from three independent evidence sources: (1) log file mtimes (ls -l --time-style=+%m-%d_%H:%M:%S) minus each run's own real duration from time = that run's start, which orders/nests the runs; (2) set -x lines inside those logs are the authoritative record of the exact command line and parameters — but a log can be truncated mid-write if it is cleaned while a script still holds it, so a log's mtime is not proof the run finished; (3) the user journal gives the first-failure instant for any service (journalctl --user --since … | grep -m2 'invalid internal status'), and on this fleet container healthchecks shell out to the podman CLI, which makes them a sensitive early indicator that podman itself has broken. The answer to "what was running" is whichever invocation's [start, end] window contains that first-failure instant; with several overlapping resource-heavy invocations, check whether they overlapped at all (mine did not prove it: the sampler output was in /tmp and gone). Related trap when reading such logs through an agent tool: long command lines get display-truncated, so wrap them (sed 's/(.{130})/\1\n /g') rather than concluding the parameters were short, and beware off-by-one log names (x4b.log vs x4b-experiment.log) silently skipping files in a loop. (context: Dankclaw 2026-09-16, answering "what were the test parameters before the crash?" about the rootless-podman pause-process incident: reconstructed from log mtimes + per-run real durations + the first invalid internal status at 18:58:30, which pinned the in-flight run (launched 18:54:31) — after /tmp/x4*.sh had been wiped by the repair reboot.)
• Correction to my earlier note on the rootless-podman pause-process death (Dankclaw, 2026-09-16): the cause is NOT established. I had written that memory pressure from a GPU experiment that committed 116 of the host's 121 GiB killed it, but journalctl -k --since for that window records NO OOM kill and the kernel says nothing else about the pause process — so the mechanism was a guess presented as a finding. The operational content that IS verified stands: every podman command fails with invalid internal status, try resetting the pause process with "podman system migrate": could not find any running process; /run/user/uid/libpod/tmp/alive holds a stale UUID token no process owns; container starts then fail because the rootless netns holding podman's bridges is gone (failed to bind udp listener on 10.89.x.1:53: Cannot assign requested address); podman system migrate panics in 5.7.0 (nil deref in teardownNetworkBackend); moving the stale token aside restores the CLI but not the bridges; a reboot repairs everything and all boot-persistent quadlets return on their own. The generalizable discipline: before publishing a cause, grep the evidence that would prove it (for a killed process: journalctl -k | grep -i oom), and if it is absent, publish "trigger not established" rather than a plausible story — an incident narrative that turns out to be invented is worse than an open question, because the next agent plans against it. (context: Dankclaw 2026-09-16: writing up the pause-process incident in AGENTS.md, the x4-pair-experiment note and log.md, then checking journalctl -k and finding no OOM kill for the window — which falsified the cause I had already written down in three documents.)
• On a rootless-podman fleet host, the rootless "pause" process is load-bearing for ALL container management, and its death is not recoverable in place. Symptoms seen on Dankclaw 2026-09-16: every podman command (even podman ps, podman system df) fails with invalid internal status, try resetting the pause process with "podman system migrate": could not find any running process: no such process; the alive file in /run/user/uid/libpod/tmp/ holds a stale UUID token (not a pid) that no process owns; every container start then fails with netavark: error while applying dns entries: aardvark-dns failed to start / failed to bind udp listener on 10.89.x.1:53: Cannot assign requested address because the rootless netns that held podman's bridge interfaces died with the pause process; and every Restart=always quadlet loops forever (28 of them here). Podman's own named remedy, podman system migrate, PANICS in 5.7.0 (panic: runtime error: invalid memory address or nil pointer dereference in libpod.(*Runtime).teardownNetworkBackend while tearing down a stopped container). Moving the stale alive/alive.lck aside makes podman respond again but the missing bridge interfaces remain, so containers still fail. The reliable repair is a REBOOT: /run/user clears, the pause process and rootless netns are recreated, and every boot-persistent quadlet returns by itself (verify loginctl show-user u -p Linger → Linger=yes and [Install] WantedBy=default.target present in each unit BEFORE rebooting). Plausible trigger in that incident: memory pressure from a GPU experiment that committed 116 of the host's 121 GiB — the pause process is ordinary OOM-killable host state. Also from the same recovery: a container cannot reach a Network=host service at the host's LAN IP, because netavark DNATs only PUBLISHED ports — from inside another container 192.168.0.143:published answers but 192.168.0.143:host-bind and the gateway 10.89.1.1 are refused; use host.containers.internal (169.254.1.2), which fixed a 502 o (context: Dankclaw 2026-09-16 evening: after the x4-pair GPU experiment (which committed 116 of 121 GiB) the rootless podman pause process was gone; all 28 rootless quadlets were down/looping and the Caddy front door with them. Diagnosed the pause token + missing rootless netns, found podman system migrate panicking, repaired by reboot, then fixed a hermes vhost 502 caused by the netavark published-port-onl)
• llama.cpp's llama-bench separates a device/tensor-split LIST with / and SWEEP values with , (-dev dev0/dev1/..., -ts ts0/ts1/..). So -dev CUDA0,Vulkan1 is TWO single-device configurations, never a two-device run — and -ts 0.85,0.15 is two sweep values, each becoming [0.85,0,0,...] / [0.15,0,0,...], i.e. device 0 takes every layer. Write -dev A/B -ts x/y. On 2026-09-16 this artifact produced a completely wrong fleet conclusion on Dankclaw ("CUDA + Vulkan cannot serve one model"): the "abort" was a single-device run with -ot exps=Vulkan1 pointing at a device absent from the scheduler's list, and an "OOM at 25% split" was really all 111 GiB on the 24 GiB card. Tell-tale: llama-bench's own dev column prints the devices actually used per row — read it. Corollary that generalizes: before concluding anything from an abort or a resource failure, prove the harness parsed your intent (echo the resolved config, or check the tool's own reported parameters), and for placement-dependent numbers get an independent witness (sample the other device's memory with nvidia-smi while the run is loaded) rather than trusting the label. The real blocker in that episode was different and simpler: per-tensor cross-device placement is killed by --override-tensor itself — has_tensor_overrides() is a required term of the scheduler's pipeline_parallel flag (llama-context.cpp:433, the only read site), which sets n_copies in ggml_backend_sched_new; with n_copies == 1 a pre-allocated weight on a non-primary device has no backend and ggml-backend.cpp:941 aborts with the leaf's own op name (NONE). Two-device LAYER splits (no -ot) work fine, including cross-vendor CUDA+Vulkan. (context: Dankclaw 2026-09-16: x4-pair experiment (3090 + Strix Halo iGPU under mainline llama.cpp, dual CUDA+Vulkan build 37b53fd); pure-Vulkan and cross-vendor layer splits then measured at 27.55 / 27.23 tok/s @64k depth vs 25.77 iGPU-only.)
• Authoritative per-flag documentation for a llama.cpp build = its own common/arg.cpp: every flag is defined as add_opt(common_arg({"-x", "--flag"}, "TYPE", "multi-line help", parser)) with the help text inline (modern master, e.g. fb27a525 on Dankserver, keeps ALL arg definitions in this one file; older trees split them across common/common.cpp + tools). Two traps: (1) llama-server --help on this master does NOT print help and exit — with no --model it silently starts a router-mode server on 127.0.0.1:8080 and stays there, so grep the source instead of probing the binary; (2) --help-style patterns won't match — grep for the literal quoted flag string, e.g. grep -n -A9 -e '"--split-mode"' common/arg.cpp. Recipe: ssh host 'P=$(find buildroot -maxdepth 5 -name arg.cpp -path "common" | head -1); cat "$P"' /tmp/arg.cpp (the checkout path can move under you mid-session — resolve it in the same call), then extract per flag locally. (context: frank-medium exhaustive settings review, 2026-09-16: extracted all 44 live flags' help text from common/arg.cpp in the fb27a525 tree on Dankserver after --help started a server instead of printing.)
• On the Dankclaw/Dankware fleet (and llama.cpp generally), a single model CANNOT be served by two different device backends — e.g. trunk+KV on the 3090 via CUDA and MoE experts on the Strix Halo iGPU via Vulkan — and this is a hard structural limit, not a tuning problem. Two failure modes, both measured 2026-09-16 on mainline llama.cpp 37b53fd: (1) per-tensor -ot pattern=device across device backends aborts in the scheduler with ggml-backend.cpp:941: pre-allocated tensor (blk.0.ffn_down_exps.weight) in a buffer (Vulkan1) that cannot run the operation (NONE) — a weight in one device buffer cannot be consumed by an op the scheduler placed on another device; (2) per-layer cross-backend splits (-sm layer -ts, or layer-scoped -ot) fail at model load with only the generic "failed to load model". The CPU is the ONLY side that works as half of a heterogeneous split, because the CPU backend is the universal fallback and host buffers can be copied — which is exactly why -ot exps=CPU works. So on a one-NVIDIA-plus-one-AMD host, the 3090 can be the trunk OR the expert host but never the complement of an iGPU-resident peer; every working multi-device llama.cpp deployment in this fleet is either single-vendor (frank-medium's two R9700s via -sm layer) or device+CPU. The surprising measurement that came out of the same run, and which supersedes the old "mainline llama.cpp is slow on gfx1151" assumption: a CUDA+Vulkan dual-backend build with EVERYTHING on the iGPU (-dev Vulkan1 -ngl 999 -ot per_layer_token_embd=CPU) did 32.64 tok/s at ~0 depth and 25.77 tok/s at 64k depth, 456.7 tok/s fresh prefill, with the iGPU sampled at 99-100% busy and the 3090 idle at 0%. That MATCHES halogen's deployed ~25 tok/s (the ik-era "mainline = 1.01 tok/s" figure was the CPU+3090 hybrid, never the iGPU). It is nonetheless not deployable on Dankclaw: it commits 116 of the host's 121 GiB (78 GiB of Vulkan device buffers + 32.8 GiB PLE page cache) against halogen's 95.4 GiB held with 14.7 GiB spare besi (context: Dankclaw 2026-09-16: dual-backend llama.cpp experiment testing whether the RTX 3090 (CUDA) plus the Strix Halo iGPU (Vulkan, RADV) could serve one 177B MoE (Flash-Next) together over the 3090's Gen4 x4 link. Result: not expressible; the iGPU-only configuration matches halogen but eats the host.)
• Deciding whether to split one model across two heterogeneous accelerators on a single host — the arithmetic that settles it, from the 2026-09-16 Dankclaw design pass (RTX 3090 + Strix Halo gfx1151 iGPU, RTX 3090 on a PCIe Gen4 x4 M.2 dock): 1. LINK FACTS (measured): the 3090's LnkCap is 16GT/s x16 but LnkSta is x4, and it idle-downshifts to Gen1 (2.5GT/s) x4 ≈ 1 GB/s, reporting Speed 16GT/s, Width x4 only under load (7.88 GB/s). So the link is a latency resource, not a bandwidth one: a hidden state is ~10 KiB (Qwen3.8-27B embedding_length 5120 x 2 B), i.e. 1.3 µs per hop, and 2 hops/layer × 64 layers ≈ 1 MiB/token ≈ 128 µs — ~0.3 % of a 40 ms token. Byte count is never the problem; hop COUNT is (ik_llama.cpp measured graph splits = 100/token with the GPU at 5 % SM). Design rule: partition by where the weights live, never by where activations flow; weights stay resident and only activations cross. 2. WHAT ACTUALLY PAYS: at batch-1 decode the KV cache is an O(context) read every token (Flash-Next: 28 KiB/position measured, so 5.5 GiB/token at 200k ≈ 21 ms of a 40 ms token at the LPDDR's shared ~256 GB/s vs 5.9 ms at the 3090's 936 GB/s). Expert weights are read once per token and are a rounding error beside it. So the highest-leverage move is KV + its attention compute on the fast-VRAM device, and KV must be co-located with its attention (separating them would push 5.5 GiB/token over a 7.88 GB/s link = ~700 ms/token, fatal). 3. WHAT DOESN'T: prefill is compute/amortized-weight bound, so moving weights to faster memory buys nothing and a device stage costs throughput. Calibration measured on this box, same model: moving 12.5 % of the expert mass to the GPU (-ncmoe 42) gave decode +12 % (4.03 → 4.54 tok/s) but prefill 192 → 32 tok/s. My bandwidth model predicts +10 % for that case, so it is usable for decode projections — and prefill is the casualty to budget for. 4. HARDWARE ASYMMETRY: on a UMA APU the iGPU is **not an independent memo (context: Dankclaw 2026-09-16: after re-splitting frank-large (halogen on the iGPU) and frank-small (llama.cpp on the 3090), Frank asked what a useful 3090 + Strix Halo combination would be over the 3090's x4 link, then to elaborate the one-process split. Verified with live probes: lspci LnkSta/LnkCap idle vs loaded, sysfs VRAM/GTT windows, GGUF tokenizer identity between Qwen3.8-27B and Flash-Next, frank)
• Speculative decoding is only worth deploying when you check the ARITHMETIC against the incumbent, not when it "sounds free". Durable method (verified on Dankclaw 2026-09-16): round cost = K·d_s + ~1.1·t_s (d_s = drafter per-token step, t_s = target per-token step); tokens/round A = (1−p^(K+1))/(1−p); rate = A/round; compare against the incumbent's rate (a target with its own MTP/EAGLE head is already speculating). Two structural facts that decide most cases: (1) a PLAIN autoregressive drafter spends K sequential forwards per round, so it must be 10–50× faster than the target to pay — a 27B drafting for a 177B MoE (which is sparse and therefore fast) fails: measured 12 ms/drafter-token vs 67 ms/target-step gives break-even p ≈ 0.70 at K=3; (2) trained multi-token heads (MTP/EAGLE/DFlash) draft K candidates in ~one pass, which is why they are the drafter of record. Acceptance reality: the fleet's TRAINED z-lab DFlash2 drafter accepted only ~41% at depth 7 on the matched 27B target — a bar external/cross-finetune pairs won't clear. Also durable: cross-device speculation is cheap (token ids once per round, not activations) and the drafter's VRAM on a discrete card is an ADDITIVE pool, unlike an APU iGPU whose GTT window is a mapping of host LPDDR (no extra capacity or bandwidth). Cross-model pairs require a byte-identical tokenizer — cheap to verify by hashing the GGUF tokenizer.ggml.tokens list (Dankclaw's Qwen3.8-27B and Qwen3.8-Flash-Next GGUFs matched exactly: vocab 248320, merges 247587, pre=qwen35, same sha256 over the first 2000 tokens), but tokenizer identity alone does not make a drafter cheap or accepted. (context: Dankclaw 2026-09-16: user asked whether a 27B on the 3090 could draft for Flash-Next on the iGPU, then whether a DFlash drafter could be hooked to halogen. Verified halogen's drafter surface is closed (serve_api.py DRAFTERS/DRAFTER_CAP, live 400 for dflash2), llama.cpp exposes --spec-draft-device/-devd, and computed break-even tables from measured step times (frank-small 11.96 ms/token; halogen ~6)
• llama.cpp speculative decoding, durable rules learned 2026-09-16 on Dankserver (frank-medium, Qwen3.8-27B UD-Q8_K_XL + zLayer split): (1) CHAINING IS GATED BY ctx_dft — common/speculative.cpp adds the draftless ngram family unconditionally but gates draft-mtp/draft-dflash/draft-dspark/draft-eagle3 on a draft CONTEXT existing, so a chain may hold one context-backed drafter plus any number of ngram types (draft-dflash,ngram-mod works), never two context-backed ones (draft-mtp,draft-dflash dies at load: "context type MTP requested but model doesn't contain MTP layers" — MTP builds against whatever model the draft context holds). Fixed internal order: ngram types first, then drafters, so CLI order is irrelevant. (2) MTP's head lives in the TARGET GGUF (blk.n.nextn.*) and only loads when MTP IS the draft model; with draft-dflash those tensors show as "unused tensor blk.64.nextn..." which is correct behaviour, NOT a truncated download. (3) ngram-mod prints NO startup banner in this build — a working chain looks unengaged in the journal; identify it behaviourally (draft_n0 only where text repeats) or via the per-task "draft acceptance = ... mean len = ..." line. It costs no VRAM (64 MiB host RAM). (4) Its window decides which traffic it can see: the fleet's 48/64 (n_min/n_max) wins because shorter windows propose on ordinary prose fragments, and rejected proposals still cost verification (43.9 - 36.0 tok/s); measured chain result: 2.6x the drafter alone on prompt-echo traffic, byte-identical to it on prose. (5) Measuring this is shape-driven and easy to get wrong: a "repeat verbatim" prompt with thinking ON puts the copy inside reasoning_content, so content is empty, cross-config identity checks compare '' to '', and you "verify" an echo you never looked at; disable thinking per request (chat_template_kwargs.enable_thinking=false) for echo/structured shapes. (6) Any harness that stops a production unit must restore it in a finally — mine crashed mid-run (a botched edit del (context: serverAppConfig 2026-09-16: frank-medium moved from vllm-radiance replicas to mainline llama.cpp (layer split across 2x R9700, DFlash2 drafter) and then to a chained draft-dflash,ngram-mod. Findings live in serverAppConfigs-docs/LLM/frank-medium-llamacpp-dflash2.md and AGENTS.md; harness is scripts/bench-spec-ab.py.)
• Speculative-decoding interfaces on the Dank fleet (verified 2026-09-16): HALOGEN (frank-large, iGPU) has a CLOSED drafter set — DRAFTERS = {"serial": 0, "mtp": 1, "dflash2": 2} in /halogen/tools/serve_api.py, wire values in the front-end→engine GEN protocol; DRAFTER_CAP gates mtp/dflash2; an unknown name is a deliberate 400 ("drafter 'x' unknown; use serial | mtp | dflash2"), and the entrypoint states there is no separate draft checkpoint for this model, so no external drafter can attach. Selection: per-request drafter field (omitted → engine applies the HALOGEN_DRAFTER env default), plus the image's own bench drafters max_tokens effort reps / sweep -d serial,mtp modes. Our deployment already runs the mtp drafter: engine verbose lines read mtp 63 tok in 1.31s = 47.97 t/s | 33 rounds, commit 1.88/round. LLaMA.CPP (frank-small, 3090) is the opposite: --spec-type takes a COMMA-SEPARATED LIST (none,draft-simple,draft-eagle3,draft-mtp,draft-dflash,draft-dspark,ngram-simple,ngram-map-k,...), so drafters COMPOSE rather than replace, and there is a full external-draft surface (-md/--model-draft, -ngld/--gpu-layers-draft, -otd/--override-tensor-draft, --spec-draft-n-max, -ctkd/-ctvd, draft-cpu-moe/ncmoe). The RULE that generalizes: speculative decoding only exists inside the process that owns the target's KV, because verification needs a forkable/roll-back-able cache; hence "use device A as a drafter for device B" is only expressible as one llama.cpp process with both backends (needs a CUDA+Vulkan build — ours has GGML_VULKAN=OFF, GGML_RPC=OFF), and never as an HTTP-level arrangement. Also: the fast device belongs on the TARGET (the expensive stage), not the drafter. (context: Answering "is it possible to use the 3090 as a drafter?" for Dankclaw's two endpoints (2026-09-16). Read from the halogen image's serve_api.py + entrypoint.sh and this llama.cpp build's --help.)
• Dankclaw's Strix Halo (Ryzen AI Max+ 395, Radeon 8060S/gfx1151) does NOT have a second memory pool for the iGPU — it has two windows onto the one 121.2 GiB LPDDR5X pool (8000 MT/s x 256-bit = 256 GB/s, verified via dmidecode; the 3090's 24 GB GDDR6X is 936 GB/s and genuinely additive). Probe: cat /sys/class/drm/card1/device/mem_info_vram_total = 2.0 GiB (BIOS UMA carve-out, genuinely private), .../mem_info_gtt_total = 114.0 GiB (GTT aperture mapping SYSTEM pages on demand); rocm-smi --showmeminfo all shows the same plus GTT Used, which tracks the engine's device-side footprint. Decisive proof that GPU memory here IS host RAM: halogen's own fit check is written 121.2 - 67.7 resident weights - reserve = device budget, and its ledger's "68.0 GiB locked in RAM + 11.0 KV pool + 16.5 working" shows up as GTT Used while the CPU counts the same pages. Consequences for anyone proposing a split: (1) capacity is not additive — moving layers "onto the iGPU" maps the same LPDDR pages through GTT, so a 117 GB model (84 GB experts + 33 GB PLE) still needs 117 GB; only the 2 GiB carve-out is private; (2) bandwidth is not additive — both engines share the 256 GB/s; (3) fencing more via amdgpu.gttsize (parm verified: "Size of the GTT userspace domain in megabytes") or the BIOS frame-buffer size subtracts from CPU-visible RAM, so it partitions rather than adds. This is why the split that paid on this box was device-vs-host (3090 GDDR6X for attention/dense/KV, LPDDR for experts via -ot exps=CPU), why iGPU-as-a-pipeline-stage buys nothing, and why "second memory pool" should be stated as "second compute engine over the same bus" — the iGPU's real value is as a whole-model runtime (halogen's ~4x prefill over llama.cpp on this SKU), not as a peer in a multi-device split. (context: Answering "why is the iGPU not a second memory pool?" and "why was pipeline parallelism not attempted?" for Dankclaw (2026-09-16). Verified live on the host plus dmidecode/modinfo; the substantive correction was that the iGPU has addressable windows (2 GiB carve-out + 114 GiB GTT of a 121.2 GiB pool) but no independent capacity or bandwidth.)
• frank-medium on dankserver (2x Radeon AI PRO R9700, gfx1201) is now one mainline-llama.cpp process, layer-split across both cards, and the reason is worth remembering: vLLM cannot use both cards for one model here in ANY of its parallelisms — TP needs an allreduce and PP's stage handoff is GroupCoordinator.send_tensor_dict, which sends metadata over the gloo CPU group but the hidden-state TENSORS over the RCCL device group (vllm/distributed/parallel_state.py, the comment "run on the device backend where is_completed() is reliable"). That is the same p2p primitive this host cannot dispatch, because ROCr refuses a kernel that declares a hidden hostcall buffer on a device without PCIe atomics (gate at rocclr/device/rocm/rocvirtual.cpp, case KernelParameterDescriptor::HiddenHostcallBuffer), and card #2 sits behind the B650 chipset (AtomicOpsCtl: ReqEn-, bridges Routing-). So "use PP to dodge the atomics problem" is exactly backwards. llama.cpp's --split-mode layer is the path that works: layers are distributed and the boundary activation is a per-backend copy, once per forward pass — no collective is ever launched. Measured here: 1991 t/s prefill / 18.8 tok/s decode layer-split vs 272 / 26.3 tensor (butterfly fallback). The ~$30 M.2→x4 riser into the free CPU-attached M2_2 is still the only way to get TENSOR parallelism (≈+40% decode), not the way to "use both cards" — that battle is already won in software. Three operational traps from the same cutover, all measured: 1. rocBLAS/hipBLASLt bundle depth. Bundling a ROCm runtime beside a llama.cpp binary with cp -r SRC DST/ where DST does not exist creates DST as a copy of SRC and flattens one level; the engine then loads 29 GiB of weights, answers /health, and dies on the FIRST request with rocBLAS error: Cannot read .../lib/rocblas/library/TensileLibrary.dat: No such file or directory for GPU arch : gfx1201. Device code must live at $ORIGIN/../lib/rocblas/library and .../hipblaslt/library/arch. 2. **An env fi (context: 2026-09-16 frank-medium cutover on dankserver: vLLM (radiance MXFP4 replicas, then an FP8 + PP attempt) → mainline llama.cpp fb27a525 layer split + DFlash2, 262144 ctx/1 slot. Produced scripts/build-llama-rdna4.sh, scripts/download-frank-medium-gguf.sh, scripts/frank-medium.{env,service}, scripts/serve-frank-medium.sh, scripts/bench-spec-ab.py, tests/frank_medium_unit.bats, serverAppConfigs-docs/L)
• Dankclaw (GMKtec EVO-X2, Strix Halo) has a hard 4-lane budget for BOTH M.2 2280 PCIe 4.0 x4 slots, and the RTX 3090 lives in one of them on an M.2-PCIe riser: kernel boot logs "pci 0000:c4:00.0: 63.012 Gb/s available PCIe bandwidth, limited by 16.0 GT/s PCIe x4 link at 0000:00:03.1 (capable of 252.048 Gb/s with 16.0 GT/s PCIe x16 link)". The GPU's root port 00:03.1 advertises LnkCap Width x8 (APU GPP bridge IP max) but only 4 lanes are wired at the socket, so the link trains x4; port 00:02.5 (NVMe) is x4 too. Therefore swapping the SSD and the GPU between the two M.2 slots cannot widen the GPU link - both are x4, and the GPU is already behind the wider-capable bridge. THE TRAP: lspci shows the idle NVIDIA GPU as "LnkSta: Speed 2.5GT/s (downgraded), Width x4 (downgraded)" because the NVIDIA driver downshifts the link to Gen1 for power when the GPU is idle - this is NOT a fault. Prove the real link by sampling NVML under load: nvidia-smi --query-gpu=pcie.link.gen.current,pcie.link.width.current --format=csv reads 4,4 while a generation request is in flight (and 1,4 at idle). Corroborating evidence: LnkCtl2 Target Link Speed: 16GT/s on both ends, and nvidia-smi hostmax/gpumax = 4. Also: this box's SMBIOS type-9 slot table is junk (it labels the RTL8125 Ethernet port and the SD reader as "M.2 Socket 3" slots and lists a phantom free slot at bus address 0200:00:03.2, a PCI domain that does not exist) - trust live lspci, not dmidecode -t slot. There is no OCuLink and no add-in slot on the EVO-X2; two M.2 2280 slots is the whole expansion story. (context: 2026-09-16, user asked whether swapping the dankclaw SSD and the RTX 3090 between M.2 slots would let the GPU run at full speed. Verified live over ssh with lspci -vv, journalctl -k boot log, dmidecode -t slot, and NVML sampling during a real frank-small generation.)
• Dankclaw (GMKtec EVO-X2, NucBox_EVO-X2, Strix Halo APU) has exactly TWO M.2 2280 slots, both PCIe 4.0 x4 on the SoC root complex (no chipset), SATA M.2 unsupported, 8TB max per slot. Both are OCCUPIED: slot 1 = 2TB AirDisk NVMe (MAXIO MAP1608 DRAM-less) at full Gen4 x4 (16GT/s), holding the root FS and all model weights; slot 2 = RTX 3090 (frank-small) in a dock/riser, and that dock negotiates only Gen1 x4 (2.5GT/s, downgraded from the card's x16 Gen4 LnkCap) — the dock/riser is the likely cap, ~$30 Gen4 riser class of fix. Consequence: unlike dankserver (free CPU-attached M.2_2 = R9700 riser target), dankclaw has NO free M.2 slot, so there is no multi-GPU-via-M.2 path on this box — the only other GPU is the integrated gfx1151. PCI addresses: NVMe c3:00.0, 3090 c4:00.0. (context: 2026-09-16, live ssh dankclaw: lsblk/lspci/dmidecode; user asked "tell me about the m.2 slots on dankclaw".)
• On a host where the second GPU cannot dispatch RCCL (no PCIe atomic routing on its bridge chain), vLLM cannot use both cards for one model in ANY of its parallelisms: tensor parallelism needs a per-layer allreduce, and pipeline parallelism needs the same class of primitive — vLLM's PP stage handoff sends metadata over the gloo CPU group but the hidden-state tensors over the device backend (vllm/distributed/parallel_state.py, isend_tensor_dict/recv_tensor_dict; the comment at the tensor-send path says these "run on the device backend where is_completed() is reliable"). So "use PP to dodge the atomics problem" is exactly backwards. The working software path for two cards on one model is llama.cpp --split-mode layer, which launches no collective at all: layers are distributed in order and only the boundary activation crosses PCIe, once per forward pass. Also note the kyuz0/amd-r9700-vllm-toolboxes images are stale for this purpose (Docker Hub latest = 2026-06-13, dev = 2026-08-12 vLLM 0.27.1 with dflash/ but no dflash2/), and that switching to such an image changes nothing about the atomics gate because the gate lives in ROCclr (rocclr/device/rocm/rocvirtual.cpp, case KernelParameterDescriptor::HiddenHostcallBuffer → !pcie_atomics_ → hipErrorIllegalState) and the shipped RCCL build is the same one that already failed. Measured on dankserver 2026-09-16: layer split 1991 t/s prefill / 18.8 tok/s decode vs tensor 272 / 26.3 (butterfly fallback). (context: serverAppConfig 2026-09-16: frank-medium cutover from a vllm-radiance MXFP4 replica pair to mainline llama.cpp layer-split across two R9700s (the user asked first for vLLM + PP, then approved the llama.cpp pivot).)
• frank-medium (dankserver, 2x R9700 gfx1201) is now ONE mainline-llama.cpp process layer-split across both cards — and the decisive insight is that vLLM's pipeline parallelism is not an escape from the PCIe-atomics wall: PP sends its hidden-state handoff on the device backend (RCCL send/recv, vllm/distributed/parallel_state.py ~1117-1119), the same primitive card #2 cannot dispatch (AtomicOpsCtl: ReqEn- behind Routing- chipset bridges). llama.cpp's --split-mode layer needs no collective at all (layers distributed, boundary activation is a per-backend copy), so it is the one software path that puts both cards on one model: 1991 t/s prefill layer-split vs 272 tensor (butterfly fallback), and it delivered 1.0-1.4k tok/s prefill at 4k-140k tokens with 43.8 tok/s decode. Switching to the kyuz0 R9700 toolbox image changes nothing about this — its librccl is byte-identical in version to the radiance image's (RCCL 2.30.4 compiled with ROCm 7.14.0.0-9999-2b22ab01), and the toolbox's published tags are stale for DFlash2 (latest=2026-06-13, dev=2026-08-12 = vLLM 0.27.1, has spec_decode/dflash/ but no dflash2/); DFlash2 landed in llama.cpp via PR #27342 and is in main at fb27a525. Three durable traps from this deployment: (1) Bundling a ROCm runtime next to a llama.cpp binary: cp -r SRC DST/ with a missing DST creates DST as a copy of SRC and drops one directory level — rocBLAS/hipBLASLt look for device code at a fixed depth (../lib/rocblas/library, ../lib/hipblaslt/library/arch), so the flattened copy loads the model, answers /health, and dies on the FIRST request with rocBLAS error: Cannot read .../TensileLibrary.dat: No such file or directory for GPU arch : gfx1201. Declare the destination then copy contents. (2) DFlash2 in llama.cpp: --spec-type draft-dflash + --spec-draft-model DFlash2 gguf; the init line block_size=8, n_extract=5, sample_from_anchor=true is the proof it took the DFlash2 path, n_max must be = block_size-1 = 7, and with draft-dflash the target (context: 2026-09-16, dankserver frank-medium cutover: user asked for kyuz0 R9700 toolbox vLLM + d0xin FP8 + z-lab DFlash2 + pipeline parallelism; investigation showed PP needs the same RCCL p2p that the atomics wall blocks and that a 29.8 GB FP8 checkpoint cannot fit one 31.86 GiB card, so the user pivoted to llama.cpp + ajgazin UD-Q8_K_XL + the DFlash2 GGUF drafter + layer split as the sole LLM runtime. R)
• Dankclaw/Dank fleet serving conventions established 2026-09-16 (serverAppConfig), all verified live: 1. PORT RELEASE RACE (deployer convention). Starting an LLM replacement while the predecessor still holds the published port fails: pasta (rootless podman's userspace net stack) binds the published port at container CREATE, and the ik/LLM server logged "Deactivated successfully" ~1s before pasta still reported Address already in use (unit exit 126); only Restart=on-failure recovered it. Repo fix: wait_port_free port in scripts/deploy-remote.sh polls ss -Htan "sport = :$port" (listener + accepted + TIME_WAIT) and fails loudly with the socket table. Ordering matters: stop this service's OWN quadlet BEFORE waiting, or a redeploy deadlocks against itself. 2. THINKING-LEVEL TRAP. A model's own chat template can reject the harness's reasoning level: Qwen3.8's template accepts only xhigh|medium|low for reasoning_effort and raises — HTTP 500 naming Jinja, not the model. Subagents run at high, so every subagent on that endpoint died until compat.reasoningEffortMap folded minimal/low→low, medium→medium, high/max→xhigh in models.yml. Halogen accepts the wider set (verified minimal/low/medium/high/xhigh all 200), so per-provider entries legitimately differ. Corollary: probe each reasoning level against a new endpoint before routing roles at it. 3. TOOLS DEPEND ON THE RUNTIME, NOT THE MODEL. The ik_llama.cpp runtime rejected every request carrying tools (500 tools param requires --jinja flag), silently breaking every tool-using subagent on that endpoint; mainline llama.cpp defaults --jinja on, and halogen accepts tools and returns tool_calls. Always test a tool-call round trip before pointing task/smol roles at an endpoint. 4. TWO ENDPOINTS, ONE HOST = two providers in .omp/agent/models.yml (one provider per HOST:PORT; the file cannot express same-host-different-port), plus a separate modelRoles entry per role. The harness config lives with the harness on the dev box (/home/blank (context: Dankclaw re-split 2026-09-16: halogen restored as frank-large (pool 400384 = 2x200192, 2 slots, MAX_TOK 24576 for the co-tenant) and frank-small created (mainline llama.cpp on the RTX 3090, Qwen3.8-27B GSQ-RCO IQ3_S -mtp + BF16 mmproj, port 7771). Failures: pasta bind race on cutover, subagent tools 500 from ik, reasoning_effort high rejected by the Qwen3.8 template.)
• Dankclaw now serves TWO LLM endpoints and the split produced three durable traps. (1) SHAPE: frank-large = halogen-flash-server 0.5.8 rootless Quadlet on :7770 (gfx1151 iGPU), retuned 2026-09-16 to HALOGEN_CTX=200192 + KV_POOL_POSITIONS=400384 (pool == 2 × ctx, so two 200k conversations are exactly resident) + KV_SLOTS=2 + MAX_TOK=24576 + HOST_RESERVE_GIB=20, VISION_TOWER unset; frank-small = mainline llama.cpp 37b53fd root systemd unit on :7771 (RTX 3090), 131072 ctx / 1 slot / f16 KV, alias frank-small. Measured: halogen budget 121−67.7−20=33.5 vs need 27.2 (no clamp), 95.4 GiB held, 14.7 GiB left, co-tenant stall NOTE prints again with 4,350→2,199 compaction stalls; frank-small 22,362 MiB of 24,576. (2) PORT RACE: rootless podman publishes through pasta, which binds the published port at container CREATE — a just-stopped predecessor still holds it (exit 126, "Address already in use"), so the deployer must stop the service's own unit, then poll ss -Htan 'sport = :PORT' empty (listener + accepted + TIME_WAIT) before installing/starting, and must stop its own quadlet first or a redeploy deadlocks against the gate. (3) TEMPLATE/TOOL TRAPS that break subagents invisibly: ik_llama.cpp serving frank-large returned 500 tools param requires --jinja flag for every request carrying tools (every tool-using agent on that endpoint was broken and it read like a model fault; mainline llama.cpp 0.4.1 defaults --jinja on, halogen accepts tools); and Qwen3.8's own GGUF template accepts ONLY xhigh|medium|low for reasoning_effort and raises a 500 otherwise — harness subagents run at high, so models.yml needs compat.reasoningEffortMap {minimal:low, low:low, medium:medium, high:xhigh, max:xhigh} for any endpoint whose template is that strict (halogen accepts all five, so the two provider entries legitimately differ). One provider per HOST:PORT in models.yml, and one role per endpoint: smol/tiny/task/vision → frank-small, slow/plan/designer/advisor → frank-large. (context: 2026-09-16: re-split Dankclaw's inference (halogen restored as frank-large; new llama.cpp frank-small on the 3090), live cutover + docs; three defects found by driving the fleet rather than reading config.)
• dankserver's two R9700s cannot cooperate through RCCL, and the escape is a ~$30 piece of hardware. Causes, all verified: (1) ROCclr gates the hostcall buffer RCCL collective kernels need on a queried hardware property — Device::checkAtomicSupport() in rocdevice.cpp:181 reads LinkAttribute::kLinkAtomicSupport from the CPU fine-grain pool and only then sets info_.pcie_atomics_ = true; there is no else branch and no env/driver/sysfs override anywhere near the gate (grepped ROCclr for getenv/HSA_/ROCR_), and the value is consumed at rocdevice create time so setpci writes are useless. (2) The board's second x16 slot is chipset-attached (AtomicOpsCap: Routing- on both bridges, endpoint ReqEn-). ASUS TUF GAMING B650-E WIFI lane wiring from ASUS's own techspec: PCIEX16_1 = CPU x16; PCIEX16_2 = B650 chipset, x16 slot in x1 mode; M.2_1 = CPU PCIe 5.0 x4 (boot NVMe, occupied); M.2_2 = CPU PCIe 4.0 x4 and free; M.2_3 = chipset x4. No x8/x8 bifurcation on this board. So an M.2→PCIe x4 riser into M.2_2 is the cheapest route to two CPU-attached cards and therefore to real TP=2; a CPU x8/x8 board (X670E/X870E class) is the better-bandwidth alternative. Bandwidth is not the blocker: a 27B TP=2 allreduce at batch 4 moves ~10 MB per token-step, ~1.4 ms at Gen4 x4. (3) Empirically, RCCL launches NOTHING involving the chipset card: rank0 (CPU card) completes broadcast/all_reduce/barrier while rank1 fails, and dist.send/recv — the primitive pipeline parallelism would use — fails too, so PP=2 is dead as well, not just TP=2. (4) vLLM's one collective-free path (IPC custom allreduce, custom_ar=True compiled in the radiance image) is gated off for ROCm RDNA4 purely by a hardcoded CUDA-sm-keyed table CUSTOM_ALL_REDUCE_MAX_SIZES = {9.0, 10.0, 10.3, 10.7} while RocmPlatform.get_device_capability() returns DeviceCapability(major=12, minor=0) → '12.0' not in table → False; patching that table is NOT sufficient, because broadcast/barrier/all_gather on the TP init path have no non-RCCL (context: 2026-09-16 dual-R9700 TP=2 research on dankserver (ASUS TUF GAMING B650-E WIFI, Ryzen AM5, 2x Radeon AI PRO R9700). Produced the ranked options and the source-level proofs.)
• NEVER run RCCL/NCCL collective experiments on a live GPU host that is serving production. On 2026-09-16 I ran a 2-rank RCCL probe (torch.distributed, backend=nccl, both R9700s) on dankserver — the production LLM host — while frank-medium was serving, reasoning that the probe needed only ~200 MB of free VRAM and that a failure would just error out. It hit the known PCIe-atomics wall ('rank1 broadcast/all_reduce/barrier/send-recv FAIL', 'rank0 OK'), the process hung until my 200-300 s timeout killed it, and within ~15-20 minutes the whole host was hard down: no ARP reply (ip neigh FAILED from a peer), 100% ICMP loss from two subnets, every port closed, and it did not come back over 25+ minutes — i.e. powered off or kernel-wedged, needing a physical power cycle (consumer board, no BMC). A hung AQL dispatch / failed hostcall on the atomics-less card is a plausible cause of a GPU-driver wedge that takes the box down, and this host has a documented precedent (the 2026-08 pinned-host-buffer incident which 'starved co-tenants into unreachability until a power cycle'). The correct procedure for any multi-GPU collective experiment is the benchmarking policy: sudo systemctl stop frank-medium && sudo systemctl mask frank-medium first (mask, never disable), then experiment, then unmask and verify /health — and ideally experiment on a host that is not serving users at all. Two secondary lessons from the same hour: (1) when a host goes unreachable, ip neigh show ip from a peer distinguishes 'powered off / NIC dead' (FAILED) from 'kernel alive but wedged' — ARP usually still answers on a merely wedged box; (2) validate the naming of your own probes before reporting an outage: I briefly believed Gitea was down on dankclaw because systemctl --user is-active gitea returned 'inactive', but the unit is gitea-server and the container was 'Up 9 hours (healthy)' — a check against a non-existent unit name produced a false outage report. (context: dankserver hard-down for 25+ min on 2026-09-16 during the dual-R9700 TP=2 research pass, ~20 min after running a bare 2-rank RCCL collective/P2P probe on its GPUs while frank-medium was live.)
• In serverAppConfig, the highest-yield overnight work (4 of 4 passes found something real) came from one audit family: surfaces whose signals were already wrong before I arrived. Concrete recurring defects and their fixes: (1) Homepage status cards can be red by construction — the checker runs on Dankclaw, so it reaches only that host's loopback, its container names, and published LAN addresses; mariadb/valkey/Postgres probed container-internal DB ports and could never be up. Rule now in Infrastructure/Homepage.md: a card must be probeable from the monitor's host or be wip (Gluetun was the pre-existing example). (2) Prometheus jobs comfyui, services, gitea scrape HTML or a 404 /metrics — a scrape cannot express "is this HTTP service answering?"; remedy is the blackbox exporter (/probe?target=&module=), which also gives guarded vhosts an uptime signal since Prometheus cannot solve an Anubis challenge. (3) Retiring a file makes every assertion that greps it pass vacuously — with dankclaw/compose.yaml gone, four bats tests (observability_quadlet, comfyui_rocm, caddyfile×2, authentik dead var) passed by absence (a missing path makes grep exit 2 and ! succeed). Fix pattern: assert the invariant directly ("no dankclaw compose file competes with the quadlet units") and keep a conditional live check for a re-added entry — then prove teeth by recreating the file and watching the tests go red. (4) A linter must be validated against the resolver it models before acting on its output: my vault-link checker produced 45 false positives (Obsidian wikilinks are case-insensitive, folder/Note resolves by path suffix not stem, %20 in markdown links is URL-encoding) before reaching the 2 real breaks. It now lives at scripts/check-vault-links.py with a docs_completeness seam. (5) Your own change invalidates prose and unit comments: check the repo's rule files first — AGENTS.md claimed "ComfyUI runs CPU-only" (false since the GPU unit was restored) and "observability RUN (context: serverAppConfig overnight session 2026-09-16: found a permanently-red homepage board, three unprobeable Prometheus jobs, four hollow tests after retiring dankclaw/compose.yaml, three dangling vault links, and a flatly wrong AGENTS.md lesson about ComfyUI. Suite 315 green, e2e 32.)
• In serverAppConfig, grep-style bats seams and doc claims silently rot whenever a file they reference is retired — the audit that keeps finding real defects is "which signal here cannot be correct?" Three concrete traps, all hit on 2026-09-16: (1) VACUOUS ASSERTIONS — after dankclaw/compose.yaml was retired, four tests that ended in ! grep -qE '^\s+svc:' "$REPO_ROOT/dankclaw/compose.yaml" passed because grep exits non-zero on a MISSING FILE, so the negation succeeded. Fix pattern: assert the real invariant ([ ! -f "$cf" ]) in ONE place plus keep a live conditional check ([ ! -f "$cf" ] || ! grep ...) for a re-added entry, and TEETH-TEST by recreating the file and watching the tests go red. (2) PERMANENTLY-RED PROBES: homepage cards and Prometheus jobs probing container-internal or HTML endpoints can never report up (mariadb 192.168.0.235:3306, valkey/postgres 127.0.0.1:6379/:5432 unpublished; comfyui/services scrape / = HTML; gitea /metrics 404). Rule: a monitor signal must be probeable from the monitor's host or be marked wip/none — an always-red signal trains the operator to ignore the surface. (3) STALE CLAIMS AFTER YOUR OWN CHANGE: retiring a file or adding a feature invalidates prose — AGENTS.md was telling agents "ComfyUI runs CPU-only, no ROCm wheels for gfx1151" after ComfyUI had been restored with GPU access (an agent trusting it would re-break it), and it still said observability was "RUNNING BUT UNCONFIGURED" plus named halogen as the active frank-large runbook. Grep the repo for every reference to anything you retire/change, not just code. When auditing a docs vault, validate the LINTER against the resolver it models before acting: a naive Obsidian link checker reported 45 breaks of which only 3 were real (case-insensitivity, path-suffix resolution for folder/Note, and %20 URL-decoding in markdown links were the three false-positive classes). scripts/check-vault-links.py (docstring records all three) plus the no unresolved links in the docs vault seam (context: Overnight 2026-09-16 passes on serverAppConfig: homepage board audit (3 impossible cards), vault link integrity + checker, four hollow bats assertions re-armed, AGENTS.md/doc stale-claim sweep after retiring dankclaw/compose.yaml and adding alert rules.)
• Fleet convention: a monitoring/status signal must be probeable from the vantage point of whatever produces it, or it must be marked intentional (wip) — an always-red signal teaches the operator to ignore the surface it lives on. Instances found on 2026-09-16: (1) the homepage status checker runs on Dankclaw, so it can reach that host's loopback, its container names, and published LAN addresses — but NOT container-internal ports of another host; three cards (mariadb probing 192.168.0.235:3306, valkey 127.0.0.1:6379, Postgres 127.0.0.1:5432) were red by construction while the databases were healthy, and publishing a DB port on the LAN to make a badge green is not a trade this fleet makes, so they became wip with the reason in their description (Gluetun/WireGuard-UDP was the pre-existing example of the same call); (2) four Prometheus jobs read down and three always will — comfyui and the two services probes scrape HTML pages (Prometheus refuses to parse HTML, error 'unsupported Content-Type text/html'), and gitea's /metrics is a 404 — none is an outage, and the tempting fix (deleting the job) is refused because an absent scrape and a healthy one look identical; the documented remedy is a blackbox exporter in probe mode, and Gitea's metrics enabled or the job repointed; (3) a bats test failed on a clean tree for many sessions ('comfyui Containerfile ...') because its selector did find | head -1 and read the wrong file — a permanently red test trains everyone to ignore red, so it was fixed at the root rather than excused. Diagnostic pattern that works: for every signal that is not green, ask what produces it and from where; if the producer cannot reach the target, it is a config defect, not an outage — classify before touching. When fixing, prefer honest state (wip / documented) over widening exposure or deleting the signal. (context: Dankclaw/Dankserver 2026-09-16 night: audited the homepage board (31 cards, 3 impossible-red DB cards), the Prometheus target list (7/11 up, 3 impossible), and the bats suite (one chronic red test); fixed each and documented the rule in the Homepage and Observability pages.)
• Migrating a compose-only container to a boot-persistent Quadlet on this fleet has three mechanical traps that bit on 2026-09-16, all invisible until something later breaks — a skill for the general procedure already exists (compose-to-quadlet-cutover), so these are the additions to remember. (1) THE CONTAINER, NOT THE SERVICE, IS WHAT FAILS: restart: unless-stopped (or Restart=always) restarts the existing container with its old config; when the compose service later gains GPU/devices/volumes, the running container keeps the stale config forever. ComfyUI died 9,146 times this way — created before devices:/group_add: were added, so it started with no /dev/kfd, no /dev/dri, no keep-groups and failed on "No HIP GPUs are available" while the YAML on disk looked perfect. Diagnosis: podman inspect c --format '{{.Config.CreateCommand}}' (the argv actually used) and {{.Created}} — NOT .HostConfig.Devices, which reported [] for a container that demonstrably had the devices (verified: torch saw the GPU, --device /dev/kfd was in the create command). Fix: remove the stale container so the unit's own can take the name. (2) A HAND-WRITTEN UNIT COLLIDES BY NAME with the one the generator produces (~/.config/systemd/user/comfyui.service shelling out to podman compose up -d vs the generated comfyui.service); retire it — stop, disable, move to ~/.AGENTTRASH — and never let both exist. (3) ONE OWNER PER CONTAINER NAME: delete the compose service, and rewire the deploy path in the same commit — planner allowlist (DANKCLAW_SERVICES/DANKSERVER_SERVICES) + a path rule for target/svc/**, the dispatcher's service_is_allowed entry, a handler that installs/restarts/health-waits, and install_quadlet pointed at "$TARGET/observability/quadlet"-style per-target trees rather than a hardcoded host. When a target ends up with zero compose services, retire the file rather than leaving a services: key with nothing under it (the compose-validity seam asserts ≥1 service). (4) SEAMS THAT READ COMPOSE (context: Dankclaw/Dankserver 2026-09-16: migrated Dankserver node-exporter and Dankclaw ComfyUI from compose to Quadlet (ComfyUI had 9,146 restarts from a stale device-less container), retired Dankclaw's now-empty compose.yaml, and found two services whose test seams read the stale compose while the unit diverged.)
• Before tightening a container's bind address, enumerate EVERY consumer of it — a bind is a contract with something, not just an exposure choice. On 2026-09-16 I set HERMES_WEBUI_HOST=127.0.0.1 on dankclaw's hermes-webui (reasoning: the compose service it replaced published loopback, and the repo policy favours internal-only), and broke hermes.franksdank.cloud with a 502: that unit is HOST-NETWORKED (it shares the loopback with the chrome-devtools CDP daemon), so Caddy cannot reach it by container name and fronts the guarded vhost with reverse_proxy 192.168.0.143:8787 — the host's LAN address, recorded in dankclaw/Caddyfile with the same explanation. With network_mode host there is no Podman port forwarder, so the app's bind IS the only path in: the wildcard serves both consumers (Caddy on the LAN address, the SSH tunnel on the host's localhost), and loopback serves only the latter. The doc's own gotcha already said this was "an exposure change to decide deliberately, not a silent fix" — read the service's doc before changing its bind. Two process lessons from the same near-miss: (1) the e2e suite stayed 100% green through it because playwright.config.ts matches specs BY NUMBER PREFIX, not glob, so a new spec file that is not registered in a project silently never runs — register the prefix and let tests/e2e_scaffold.bats fail if any spec matches no project; (2) for every guarded vhost, at least one gate should assert that the vhost reaches its app through the guard chain (status 500 is the load-bearing check: a dead upstream is a 502 from Caddy), especially for services whose upstream is a host address rather than a container name. (context: Dankclaw 2026-09-16: tightened hermes-webui's bind to loopback, broke the hermes.franksdank.cloud vhost (502) with the whole e2e suite green; reverted, added specs/17-hermes-vhost.spec.ts, registered it, and added a guard that every spec matches a playwright project.)
• On this fleet, "which container will not come back after a reboot?" has a one-line answer: every running container should carry a PODMAN_SYSTEMD_UNIT label; anything without one is compose-only or bare and will die silently at the next boot/shutdown. Audit with for c in $(podman ps --format '{{.Names}}'); do printf '%s %s\n' "$c" "$(podman inspect $c --format '{{index .Config.Labels "PODMAN_SYSTEMD_UNIT"}}')"; done and compare against the host's ~/.config/containers/systemd/.container and the repo's host/svc/quadlet/. THREE separate instances of this failure mode were found on 2026-09-16 alone: (1) Dankserver's node-exporter — compose-only, Exited (143) for 37 h, no unit, Prometheus's node-dankserver job scraping nothing, and the compose form bound the wildcard address; fixed by adding dankserver/observability/quadlet/node-exporter.container (loopback + LAN 9100, the deliberate LAN listener posture). (2) ComfyUI on Dankclaw — its container was created BEFORE the GPU declarations were added to the compose service and restart: unless-stopped then kept restarting that stale, device-less container for 9,146 restarts (RuntimeError: No HIP GPUs are available); the service was fine, the container was old. Fixed by a Quadlet unit + deleting the compose service. (3) hermes/hermes-dashboard/hermes-webui were ALREADY quadlet-owned (labels proved it) while dankclaw/compose.yaml still defined them — a second owner for the same container names, which is how a stray podman compose up creates a duplicate that fights the unit. Generalisations worth keeping: editing a compose service does not replace a running container (restart policy will keep the old config alive forever — recreate explicitly); podman inspect --format '{{.HostConfig.Devices}}' is not a reliable device-passthrough check (it read devices=0 on a container that demonstrably had --device /dev/kfd — use {{.Config.CreateCommand}} for the real argv, and confirm from inside with e.g. podman exec c python3 -c "impo (context: 2026-09-16 night session on the serverAppConfig fleet (Dankclaw + Dankserver): restored Dankserver's node-exporter (37 h dark) and ComfyUI (9,146 restarts, 7 h down) as boot-persistent Quadlets, audited every running container's systemd owner, and found that all three hermes services were already units while compose.yaml still duplicated them. Verified fixes live: node targets up, ComfyUI health)_
• The recurring outage class on this fleet is a container whose RUNNING config has drifted from its DEFINITION, kept alive by a restart policy: compose restart: unless-stopped (or a hand-written systemd unit calling podman compose up -d) restarts the stale container forever instead of recreating it from the edited definition, so a service can sit broken for days looking merely "unhealthy". Found twice on 2026-09-16: Dankserver's node-exporter had been Exited (143) for 37h with no unit at all (compose-only, dead since a reboot), and ComfyUI had restarted 9,146 times because its container was created 2026-09-15 14:12 BEFORE the /dev/kfd //dev/dri/keep-groups declarations were added to the compose service — so torch reported "No HIP GPUs are available" on every start while the definition was correct. Diagnostic trap that nearly produced a wrong root cause: podman inspect --format '{{.HostConfig.Devices}}' reported EMPTY for a container that demonstrably had --device flags; the trustworthy view is --format '{{.Config.CreateCommand}}' (the actual argv). Also: find who started a container via the PODMAN_SYSTEMD_UNIT label — a podman-compose@… label means compose owns it, regardless of what unit files exist. The durable fix in this repo is the convention already written down: one owner per container name (delete the compose service), a repo-tracked Quadlet under host/svc/quadlet/ installed by the deployer, and retirement of any legacy compose-era unit file into ~/.AGENTTRASH (a leftover ~/.config/systemd/user/svc.service COLLIDES by name with the unit the Quadlet generator produces). Remediation ordering that worked: build image - install unit - stop/disable+move the legacy unit - podman rm -f the stale container - daemon-reload - restart unit - health-wait - then verify the real capability INSIDE the container (e.g. torch.cuda.is_available()), not just that it is Up. (context: serverAppConfigs, 2026-09-16 night: restored Dankserver node-exporter (37h dark, compose-only) and ComfyUI (9,146 restarts, stale device-less container) by migrating both to repo-tracked Quadlet units; also documented that three Prometheus jobs (comfyui + two services probes scraping HTML, gitea /metrics 404) can never be green.)
• Two durable facts from the 2026-09-16 night (Dankclaw/Dankserver fleet). (1) TRAP: podman inspect c --format "{{len .HostConfig.Devices}}" reports 0 and .HostConfig.GroupAdd reports [] even for a container that WAS started with --device /dev/kfd --device /dev/dri --group-add keep-groups. I nearly misdiagnosed ComfyUI as "no GPU devices configured" from that field. The ground truth is the actual argv: podman inspect c --format "{{.Config.CreateCommand}}" (or podman exec c python3 -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))" for the behavioural truth — which printed True | Radeon 8060S Graphics once fixed). (2) ROOT-CAUSE PATTERN for "a service crash-loops after a reboot while its config looks right": a container created by an EARLIER revision of the compose service (before devices:/group_add: were added) is kept alive by restart: unless-stopped; podman compose up -d does not recreate it, so it restarts forever with the OLD config (ComfyUI: RuntimeError: No HIP GPUs are available, 9,146 restarts, ~7h down). Decisive check: compare the container's creation timestamp against when the config gained the missing lines, and read .Config.CreateCommand for the flags it actually has. Fix on this fleet is not "recreate the compose container" but migrating the service to a repo-tracked Quadlet (the repo's own rule: every container is a boot-persistent, self-healing unit with repo provenance). Result: ComfyUI now runs from dankclaw/comfyui/quadlet/comfyui.container — image comfyui-rocm:gfx1151, /dev/kfd + /dev/dri, GroupAdd=keep-groups, four comfyui-data volumes, loopback-only publish 127.0.0.1:8188 (unguarded service = internal-only; homepage card probes loopback), Restart=always, WantedBy=default.target, healthcheck via a mounted file (the podman 5.x v1 generator splits HealthCmd on spaces, so inline python3 -c "..." is a quoting trap). Unit is generated, healthy, restarts=0, and curl 127.0.0.1:8188 → 200. Also fixed the same night: Dankser (context: Dankclaw ComfyUI outage found while auditing Prometheus targets during the 2026-09-16 frank-medium night shift: 9,146 restarts, stale device-less compose container, migrated to a Quadlet; verified hip available True. Dankserver node-exporter restored the same way.)
• Fleet state after the 2026-09-16 night session on serverAppConfigs (frank-medium + observability work): (1) DANKSERVER NODE-EXPORTER RESTORED — it had run only from dankserver/compose.yaml, died Exited (143) at a reboot and left Prometheus' node-dankserver job scraping nothing for ~37 h; it is now dankserver/observability/quadlet/node-exporter.container (generated, Restart=always, binds exactly 127.0.0.1:9100 + 192.168.0.235:9100, no wildcard — verified with ss -ltn), the compose service is deleted (one owner per container name), and deploy_observability now installs from "$TARGET/observability/quadlet" instead of a hardcoded Dankclaw path. Dankclaw keeps five units there, Dankserver one. (2) PENDING FOR FRANK — ComfyUI on Dankclaw is crash-looping: 9,146 restarts, exit 0, log says RuntimeError: No HIP GPUs are available, unit shows active (exited) while the container respawns; deliberately not touched because the fix depends on which posture Frank wants (its documented posture is CPU-only --cpu, which contradicts that error). (3) THREE PROMETHEUS TARGETS ARE CONFIGURATION DEFECTS, NOT OUTAGES: gitea returns 404 on /metrics (service itself fine), and the two services targets (SearXNG :8081, OpenWebUI :3005) answer 200 with Content-Type text/html, which Prometheus 3 refuses to parse — "blackbox-style probes" without a blackbox exporter are inert; the observability doc's own blackbox-exporter note is the fix. So "7 of 11 targets up" is the expected reading, and an alert on up==0 across all jobs would be permanent noise (the LLM rules are scoped to frank-* and are unaffected). (4) Also this session: tests are GREEN (313, ALL TESTS PASSED) — the long-standing "comfyui test 73 fails on a clean tree, ignore it" belief is now false and must not be repeated; the homepage card for Dankserver :7770 was mislabelled frank-small (retired) and is now frank-medium; Grafana has a provisioned Prometheus datasource + an LLM fleet dashboard; Prometheus evaluates 5 LLM alert rules but (context: serverAppConfigs, 2026-09-16 pre-dawn session: frank-medium (2x R9700, vllm-radiance, two TP=1 replicas behind an adaptive-routing proxy) plus an observability audit that found the node-exporter gap, the mislabelled homepage card, a red test that had been excused for sessions, and a crash-looping ComfyUI.)
• Durable fact for this repo (serverAppConfigs): the measurement harnesses for the CURRENT LLM serving stack now live in git — do not rewrite them in /tmp. Four files, added 2026-09-16 with their traps documented in each header: scripts/bench-vllm-corpus.py (builds a code-shaped corpus from the repo's own sources; synthetic --dataset-name random prompts misreport speculative draft acceptance badly — measured 2.9-3.6 tokens/update on random against ~5.8 on code, so tuning SPEC depth on random data picks the wrong value), scripts/bench-vllm-serve.sh (concurrency sweep via vllm bench serve from inside the serving image; it exists mainly to encode two traps — the client needs --tokenizer /models/... with /models mounted, because with only --model frank-medium it tries to resolve that alias on HuggingFace and dies before sending a request, producing a run that looks like a benchmark and reports nothing; and --ignore-eos output is not comparable with runs that lack it), scripts/bench-turn-latency.py (per-turn TTFT for an agent-loop shape — the instrument that decided the deployed prefix-affinity routing policy, and the fastest way to demonstrate cold-vs-warm prefill: measured 12.5 s cold vs 1.4 s warm on the same ~50k-token prompt), and scripts/bench-llm-soak.py (sustained multi-family load; this is what found the two-card capacity wall). Prompts in every harness carry a salt as their leading bytes, because that changes both the router's identity window and vLLM's prefix-cache key — without it a run silently inherits a warm cache from the previous run and reports fantasy numbers. Also durable: scripts/.py are now compiled by the scripts_syntax seam (previously only scripts/.sh were checked), and the whole suite is green (311 tests) with no known-failure exceptions to explain away. Older llama.cpp-era harnesses (bench-decode*, bench-spec-decode.py, bench-threads-batch*, generate-prompts.py) are still in scripts/ and target a different runtime's endpoint — check which stack (context: Night session 2026-09-16 on Dankserver frank-medium (vllm-radiance MXFP4, two R9700 replicas behind an adaptive-routing proxy): after building and using throwaway measurement scripts for hours, the useful four were committed to scripts/ and smoke-tested against the live fleet, and the script-syntax seam gained its missing python half.)
• CORRECTION to the long-standing belief recorded in this project's memory that "./tests/run.sh exits 1 on a clean tree because of a pre-existing comfyui test-73 failure, and that is expected, so don't treat a red suite as blocking." That is now FALSE as of 2026-09-16: the test was fixed properly and the suite is GREEN (310 tests, ALL TESTS PASSED). Root cause of the old failure was a selector bug in tests/comfyui_rocm.bats — find dankclaw -name 'Containerfile*' | head -1 returned whichever Containerfile the directory listing happened to put first (dankclaw/alloy/Containerfile), so the assertion read a file that never claimed to be ComfyUI's and failed; the same selector would have passed on a wrong file. The fix resolves the Containerfile the way the build does — build.context joined with build.dockerfile, read from dankclaw/compose.yaml — so the test checks the file the deploy actually builds. The generalisable lesson: in this repo a test that guesses a path with find ... | head -1 is a bug waiting to be believed as "a known failure"; read the contract (compose, unit file, config) instead of the filesystem. Also: a permanently red test trains every operator and every future agent to ignore red — treat "known failure" as a defect with an owner, not as a baseline. A dead duplicate was retired alongside it (dankclaw/comfyui/Containerfile was referenced by nothing — compose builds Containerfile.comfyui), staged in .AGENTTRASH/ awaiting Frank's purge decision. (context: serverAppConfig repo, 2026-09-16 night session: fixed tests/comfyui_rocm.bats test "comfyui Containerfile is based on the gfx1151 torch image" after it had failed across many sessions; suite went from 285 ok + 1 not ok to 310 ok, ALL TESTS PASSED.)
• Dankserver frank-medium as of 2026-09-16 is NOT a single tensor-parallel process: it is two single-GPU TP=1 replicas of the vllm-radiance MXFP4 stack (frank-medium-a on R9700 #1 :7771, frank-medium-b on #2 :7772) behind a prefix-affinity proxy (frank-medium-proxy :7770, routing.py policy + affinity-proxy.py server), each at 114688 context / 4 sequences / CHUNK=8192 / KV_MEM=0, with the DFlash2 drafter at SPEC=7. TP=2 is impossible on this host for a hardware reason: RCCL's collective kernels carry a hidden hostcall-buffer parameter and ROCr refuses to dispatch them on a device without PCIe atomics, and the second R9700 sits behind chipset bridges 05:00.0/06:09.0 advertising AtomicOpsCap: Routing- (read-only) — so no software enables atomics there while the first card, on CPU root port 00:01.1, has them. The practical capacity ceiling per card is 122,636 KV tokens (the engine's own ledger), i.e. about two 30k-token conversations; three concurrent 50-80k contexts thrash the prefix cache (turns alternate ~1.4 s warm / 22-45 s cold) and neither load-balanced placement nor per-card admission control fixes it, because a running sequence and its cached prefix draw on the same pool. CHUNK is not a lever: available KV is 5.44 GiB at 8192 versus 4.22 at 16384 and 4.16 at 4096 (which refuses to boot), so 8192 is a sweet spot. Long-prefix caching itself is healthy — the same ~50k-token prompt sent twice to an idle card is 22.3 s then 1.4 s — and the engine's lifetime "Prefix cache hit rate" line is an average over all traffic ever scraped, not a verdict on any request. The fleet's Prometheus scrapes both LLM endpoints (frank-medium and frank-large) and five rules watch them; nothing is notified because there is no Alertmanager, so alerts only appear on Prometheus' /alerts and the dashboard's ALERTS panel. (context: Night-long session on Dankserver frank-medium: TP=2 ruled out by PCIe atomic routing, deployment reshaped as two TP=1 replicas + adaptive-routing proxy, capacity measured, four optimization hypotheses tested (routing and failover shipped; placement, admission, CHUNK variants rejected with numbers), then Prometheus scraping + Grafana provisioning + five alert rules added.)
• On the Dankclaw fleet, Grafana is configured as CODE, not by clicking: shared/grafana/provisioning/{datasources,dashboards} is installed to /opt/grafana/provisioning by copy_grafana_provisioning() in scripts/deploy-remote.sh (the grafana quadlet already mounts that path read-only), and scripts/deploy-plan.py maps any shared/grafana/** path to the grafana service, so ./scripts/deploy.sh dankclaw --service grafana is the whole apply step. VERIFICATION TRAP (cost me a false negative): a provisioned dashboard does NOT appear in the legacy dashboard table of grafana.db on the current Grafana build — dashboards live in unified storage behind the dashboard.grafana.app API group — so querying that table returns empty and looks like the dashboard failed to load. The datasource DOES appear in the data_source table, which makes the asymmetry easy to misread. Correct verification, in order: podman logs --since 10m grafana | grep -i "inserting datasource|provision dashboards" (the provisioning paths log), then the API — curl -u admin:admin http://127.0.0.1:3002/api/search?type=dash-db for dashboards and /api/datasources/proxy/uid/uid/api/v1/query?query=... to prove the datasource actually reaches Prometheus. Grafana on this host publishes only on loopback :3002 (the off-host path is the guarded vhost), and its admin login is still the default admin:admin — worth changing, and worth remembering because every metric/dashboard assertion needs it. One more habit that paid off twice in this repo: before writing any dashboard panel or config comment, grep the metric names out of the live /metrics endpoint (curl -s host:port/metrics | awk '{print $1}' | sort -u | grep word); a comment naming vllm:prefix_cache_hit_rate shipped once and that series does not exist (the real pair is prefix_cache_hits_total / prefix_cache_queries_total). (context: Dankclaw 2026-09-16: provisioned the Prometheus datasource plus an "LLM fleet — capacity and reuse" dashboard from git (shared/grafana/provisioning/*), after discovering Grafana had zero datasources so the newly scraped vLLM series were only reachable through the Prometheus API. Verified via the Grafana API after the legacy dashboard table gave a false empty.)
• frank-medium on Dankserver (192.168.0.235:7770) is no longer one tensor-parallel process: since 2026-09-16 it is TWO single-GPU vLLM replicas (frank-medium-a on :7771, frank-medium-b on :7772) behind a prefix-affinity proxy on :7770, because TP=2 cannot launch a collective on this host (RCCL kernels need PCIe atomics; the second R9700 sits behind chipset bridges advertising AtomicOpsCap: Routing-). Each replica is TP=1 at 114688 context / 4 sequences / CHUNK=8192 / SPEC=7 with vLLM-profiled KV. Concrete consequences a future session must not re-derive: the harness entry in /home/blank/.omp/agent/models.yml is contextWindow 114688 / maxTokens 32768 and omp --model frank-medium -p "..." works end to end; per-card KV is 122,636 tokens, so two concurrent large contexts are comfortable and three thrash; the deploy handler installs a/b/proxy and moves any stale TP=2 unit into ~/.AGENTTRASH; Prometheus now scrapes both LLM endpoints (frank-medium vLLM, frank-large llama.cpp) on their LAN addresses; and the trio is proven boot-persistent by an actual unattended reboot (active 59 s after boot, healthy ~3.5 min, no manual step). Full evidence lives in serverAppConfigs-docs/LLM/frank-medium-radiance-mxfp4.md — read it before assuming TP=2 or the retired ROCmFPX deployment. (context: frank-medium deployment and fleet wiring, 2026-09-15/16: TP=2 → two TP=1 replicas + adaptive proxy, harness/models.yml update, Prometheus scrape jobs, reboot verification.)
• Benchmarking an LLM serving deployment honestly (used repeatedly on frank-medium 2026-09-16, applies to any future swap incl. frank-large): (1) vllm bench serve needs an explicit --tokenizer local path with /models mounted — passing only --model alias makes the client resolve the alias on HuggingFace and die before sending a request, a green-looking run that produces no numbers; (2) synthetic --dataset-name random under-reports speculative decoding badly (2.9-3.6 tok/update measured vs 5.8 on code) so tuning SPEC depth on it picks the wrong default for a code workload — build the corpus from the repo's own sources instead (JSONL of {"prompt": instruction + file body}); (3) vLLM's Prefix cache hit rate: log line is a LIFETIME average over every prompt that replica ever saw, not a per-request verdict — to learn whether long-prefix reuse works, send the same prompt twice to an idle card and compare TTFTs (measured: 22.3 s cold → 1.4 s warm on a 50k-token prefix); (4) A/B arms must use different leading bytes (a per-arm salt at the front of every prompt), otherwise the second arm silently reuses the first arm's warm cache; (5) repeat each arm at least twice — I saw 5-10% run-to-run variance, and a single measurement was enough to make me nearly ship the wrong SPEC depth; (6) when a proxy sits in front of several replicas, measure the workload shape the fleet actually runs (sequential conversation turns) separately from batch benchmarks: a cold batch of distinct prompts favors spreading across cards, sequential turns favor pinning to one, and the two conclusions pull in opposite directions. (context: frank-medium (2x R9700, two TP=1 vLLM replicas behind a proxy) optimization pass, 2026-09-16: corpus design from repo sources, salted per-arm corpora, repeated arms, sequential-turn client vs batch benchmarks, engine counter semantics.)
• frank-medium on Dankserver (2x R9700 gfx1201, since 2026-09-16) serves as TWO single-GPU TP=1 replicas behind a proxy on :7770 (adaptive routing + failover), because TP=2 is impossible on this AM5 box: RCCL collective kernels need PCIe atomics and card #2's chipset path advertises AtomicOpsCap: Routing- (read-only). Per-replica shape: MAXLEN=114688, MAXSEQS=4, CHUNK=8192, KV_MEM=0 (passed through literally — blanking it makes vLLM size KV from GPU_UTIL alone and OOM), SPEC=7. Engine ledger facts: GPU KV cache size: 122,636 tokens per card (5.44 GiB), 1.07x concurrency at full context. Measured and REJECTED tonight, do not retry without new evidence: (1) load-balanced family placement — no improvement (heavy 25.4→46.7 s median); (2) large-request admission control, one 16k-token request per card — no clear win (heavy 25.4→48.2 s, mediums improved on small samples); (3) CHUNK=4096 or 16384 to free KV — available KV is NON-monotonic in chunk size (4.16 GiB at 4096 where the engine refuses to boot, 5.44 at 8192, 4.22 at 16384), so 8192 is a sweet spot; (4) SPEC=5 for batch throughput — 7 stays (46.2 vs 41.5 tok/s single-stream). The wall is capacity arithmetic: live sequences and cached prefixes share one pool, so a card holding two 50k-token families must carry 2x50k cached + 50k live = 150k against 122,636 — co-tenants evict each other regardless of placement or serialization. Caching itself is healthy (a 50k-token prefix: 22.3 s cold then 1.4 s warm, with a stable or varying tail). Trap: the engine's "Prefix cache hit rate: N%" log line is a LIFETIME average over every prompt that replica has seen, not a per-request verdict. Verified elsewhere tonight: the trio survives an unattended reboot (units generated, Linger=yes, active 59 s after boot, replicas healthy ~3.5 min in), and the harness path works (omp --model frank-medium -p --no-tools → correct reply in ~8-10 s) with models.yml contextWindow 114688 / maxTokens 32768. (context: Dankserver frank-medium, 2026-09-16 overnight session: two-replica MXFP4 deployment, routing/failover shipped, four optimization hypotheses measured (two shipped, two rejected), capacity ceiling quantified, reboot-tested.)
• On a single-GPU vLLM replica, a request's LIVE sequence and its cached prefix draw on the same KV pool — so placement/routing policy cannot rescue an overloaded working set. Measured on 2x R9700 radiance MXFP4 (122,636 KV tokens per card, 1.07x concurrency at 114k context): one 80k-token family plus two 50k-token families, each turn re-sending its document, thrash the cache no matter how families are placed — counter placement: heavy median 25.4 s (23% cached), mediums 34.7/45.2 s (14/15% cached); load-balanced placement (new families to the card carrying less context, continuations still pinned): heavy 46.7 s (20% cached), mediums 23.4/34.3 s (0% cached). No improvement, change reverted. Why: a card running two 50k-token generations has already spent its 122,636 tokens on live sequences before any cache entry is considered, so placement only decides WHICH card is over capacity. Lever for more concurrent large contexts = fewer in flight (context compaction) or another card, never a routing tweak. Two measurement disciplines that made this decision possible, both worth repeating: (1) design the A/B so the two policies actually differ — with tiny "light" families nothing is under pressure and both arms look identical, so size the families to straddle the card's KV ceiling (e.g. 80k + 50k + 50k against 122k); (2) before trusting an arm, unit-check that the mechanism separates as intended (here: heavy-card0, both mediums-card1, confirmed by calling the routing function directly) — otherwise "no improvement" may just mean the feature never fired. Same discipline caught two other hypotheses in the same session: CHUNK=4096 also reduced available KV (4.16 GiB) rather than freeing it, so 8192 is a sweet spot, not a compromise. (context: 2026-09-16, Dankserver frank-medium: A/B of load-balanced vs counter placement for two radiance MXFP4 replicas behind an adaptive-routing proxy; hypothesis rejected on measurement and reverted, documented in serverAppConfigs-docs/LLM/frank-medium-radiance-mxfp4.md under "Placement does not rescue an overloaded working set".)
• frank-medium (Dankserver, 2x R9700) runs as TWO single-GPU vLLM-radiance replicas behind an adaptive proxy — TP=2 is impossible on this AM5 box (RCCL needs PCIe atomics; card 2's chipset path has AtomicOpsCap Routing-, read-only). Durable numbers measured 2026-09-16: each replica holds 5.44 GiB KV = 122,636 tokens at KV_MEM=0/MAXLEN=114688/CHUNK=8192/MAXSEQS=4/GPU_UTIL=0.98 (1.07x concurrency at full context). The arena/KV budget is NOT monotonic in CHUNK: available KV is 5.44 GiB at 8192 but 4.22 GiB at 16384 and 4.16 GiB at 4096 (where the engine refuses to boot: "5.0 GiB KV cache is needed"), so 8192 is a sweet spot and CHUNK is NOT a lever for freeing KV — more concurrent large contexts need another card. A 25-min soak of three concurrent large-context agent loops (each turn re-sending a ~30k-token document) gave 155/155 requests answered, zero errors/failovers, but only 25% cache-hit turns (median 1190 ms) against 75% re-prefills (median 13.9 s): the ceiling shows up as latency, not failure. Boot persistence is PROVEN by rebooting dankserver unattended: units active 59 s after boot, both replicas healthy ~3.5 min after boot, harness (omp --model frank-medium -p) working, no manual step — though SSH took ~9 min to return while systemd reports a 2m14s boot. The stack's /v1/completions returns raw continuations (no chat template), so api: openai-completions entries work only because the harness builds its own prompt; verify harness paths with omp --model alias -p --no-tools rather than assuming. OPERATIONAL TRAP hit twice: shipping files straight into the host checkout (/home/dank/serverAppConfigs) makes the next ./scripts/deploy.sh fail with "Your local changes ... would be overwritten by merge" — restore that path (git checkout -- path, or move the hand-copied file aside) before deploying, or the deploy silently stops at the pull. (context: Dankserver frank-medium 2026-09-16 (overnight session): two TP=1 radiance MXFP4 replicas behind an adaptive-routing proxy on :7770; bench/soak/reboot verification; repo commits up to 99ed1ac.)
• Two counterintuitive, measured facts about vLLM KV capacity on the frank-medium radiance stack (Dankserver, 2x R9700, TP=1 replicas, 27B MXFP4, 114688 ctx) — both cost hours if re-derived from reasoning: 1. Available KV is NOT monotonic in the prefill chunk size. Measured per card: CHUNK=4096 - 4.16 GiB available KV and the engine REFUSES TO BOOT ("To serve at least one request with the model's max seq len (114688), 5.0 GiB KV cache is needed, which is larger than the available KV cache memory (4.16 GiB)"); CHUNK=8192 - 5.44 GiB (the sweet spot, and the deployed value); CHUNK=16384 - 4.22 GiB. So shrinking the chunk to "free memory for KV" is wrong in both directions. Get the real capacity from the engine's own boot ledger, never from arithmetic: podman logs container | grep -E "GPU KV cache size|Maximum concurrency|Available KV cache memory" - here "GPU KV cache size: 122,636 tokens" and "Maximum concurrency for 114,688 tokens per request: 1.07x". 2. 122,636 KV tokens per card is about TWO 30k-token conversations with their cached prefixes. A 25-minute soak with three concurrent agent loops (each turn re-sending a ~30k-token document, growing history) came back perfectly stable — 155 requests, 155 answers, 0 failures, 0 proxy failovers — but only 25% of turns hit the prefix cache; the other 75% paid a 13.9 s median re-prefill because a third family on the same card evicts. Capacity exhaustion on this fleet shows up as latency, not errors. Related: the boot warning "Add 3 padding layers, may waste at most 60.00% KV cache memory" is real — the mamba/attention hybrid wastes KV, so the usable budget is well under the nominal pool. Also: the harness path is verifiable directly with omp --model frank-medium -p --no-tools --no-session "prompt" (9.8 s round trip here) — that exercises models.yml provider/baseUrl/alias end to end, which reading YAML cannot. Note the provider uses api: openai-completions, i.e. /v1/completions, where this stack returns a RAW continuation (no (context: Dankserver frank-medium 2026-09-16, after the two-replica + adaptive-routing deployment: soak of three concurrent agent loops, CHUNK sweep (4096/8192/16384), engine boot-ledger capacity read, and a real omp harness call.)
• A reverse proxy in front of N LLM replicas that trusts a periodically-refreshed health table will hand clients 5xx errors during every restart window: with a 10 s health TTL, requests routed to a card that just went down fail before the table notices — and a deploy restarts all replicas. Fix: on connection failure (nothing written to the client yet), demote that replica immediately and retry the request on another live replica; log the failover, because a silent one hides a downed card. Safe for stateless OpenAI-compatible endpoints (chat/completions carries no session state) but NEVER retry after streaming has begun. Verify behaviourally: stop one replica and fire a burst of requests in the same second — all must return 200 with exactly one failover line in the proxy log. Measured on frank-medium (2x R9700, adaptive routing): 6/6 requests 200, one failover log line, before the change they returned 502. Also observed: a stopped radiance replica takes ~150 s to come back with a warm compile cache, so the deploy health wait must budget for it. (context: Dankserver frank-medium 2026-09-16: closed the 502-during-restart window in the two-replica proxy (routing.py + affinity-proxy.py), verified by stopping card 1 and firing a burst; deployed green with the e2e gate 30/30.)
• Benchmarking a speculative-decoding LLM endpoint on this fleet taught a discipline that changed three conclusions in one night: (1) Each configuration under test needs its OWN salted corpus — prefix caching persists inside a running engine, so reusing a document across configs silently measures a warm cache; put a unique token sequence at the very start of every prompt so both the router hash and vLLM's cache key change. (2) Synthetic random prompts under-report speculative decoding badly: the same model/server measured 2.9-3.6 tok/update on --dataset-name random and 3.86 tok/update on a corpus built from the repo's own source files, so a code/RAG/agent workload must be measured with a code-shaped corpus (vllm bench --dataset-name custom with JSONL {"prompt": ...}). (3) Run each setting at least twice: the same config measured 118.7 and 127.0 tok/s (conc 4) and 136.1 / 129.4 (the other config) — ~5-7% run-to-run variance, so a single-run difference under ~10% is not a decision, and the two settings' ranges overlapped while the ordering held. (4) Conversation workloads need a sequential-turn client measuring per-turn TTFT, not a batch benchmark: a batch of distinct prompts parallelises prefill across replicas and even makes round-robin look better, while the agent-loop shape (same prefix, growing history) is what exposes cache behaviour. Separately, a sizing fact for the radiance/vLLM stack on a 32 GiB card: the prefill arena scales with --max-num-batched-tokens and shares one device budget with the KV cache, so CHUNK=16384 leaves 4.22 GiB of KV where 114688 context needs 5.21 GiB and the engine refuses to start — CHUNK and MAXLEN are a matched pair, not independent knobs. (context: Dankserver frank-medium 2026-09-16/17: tuning two single-GPU radiance MXFP4 replicas (SPEC depth 7 vs 5, CHUNK 8192 vs 16384) and their adaptive-routing proxy; every number in the runbook came from these corpora and repeats.)
• On the vllm-radiance MXFP4 stack (frank-medium on Dankserver's R9700s), two KV-sizing traps cost deploy cycles and are counterintuitive: (1) KV_MEM=0 must be passed THROUGH to the engine, not blanked. The entrypoint expands ${KV_MEM:+--kv-cache-memory "$KV_MEM"}; the image's patched KV sizing reads an explicit 0 as "profile this shape". If the variable is blanked instead (so the flag is omitted), vLLM sizes the KV cache from --gpu-memory-utilization alone, overshoots, and the allocation dies with torch.OutOfMemoryError right after the weights load — measured: "Tried to allocate 2.16 GiB. GPU has 116 MiB free, 29.34 GiB allocated by PyTorch". Symptom shape: the quadlet crash-loops every ~2 min ("Engine core initialization failed", status=1) while the same config runs fine when started by hand with the older entrypoint — compare entrypoint versions when a hand-run works and the unit does not. Do not "fix" the entrypoint by making the expansion empty. (2) CHUNK (--max-num-batched-tokens) is coupled to the context: the prefill ARENA grows with the chunk size and shares one device budget with the KV cache. On a single 32 GiB card at MAXLEN=114688, CHUNK=16384 leaves 4.22 GiB of KV against the 5.21 GiB needed and the engine refuses to start ("To serve at least one request with the model's max seq len (114688), 5.21 GiB KV cache is needed, which is larger than the available KV cache memory (4.22 GiB)"). Raising CHUNK therefore means cutting MAXLEN. Per-card single-card prefill at CHUNK=8192 measured ~2.5k tok/s (~15k-token prompt, 5992 ms TTFT), so the trade was not worth taking. Also: 262144 context needs 9.61 GiB of KV and 131072 needs 5.58 GiB, which is why one card tops out at 114688 in this shape. (context: Dankserver frank-medium, 2026-09-16: two TP=1 radiance MXFP4 replicas, 114688 context each. These two traps are documented in serverAppConfigs-docs/LLM/frank-medium-radiance-mxfp4.md; capturing them in memory because both look like engine bugs rather than sizing rules.)
• When fronting multiple single-GPU vLLM replicas with a proxy, plain round-robin destroys prefix-cache hits: an agent loop re-sends the same system prompt + file context every turn, so alternating replicas makes every other turn pay a cold prefill. Measured on 2x R9700 (90k-token contexts): round-robin warm-turn TTFT 2676 ms vs pinned 1096 ms. But pinning everything is also wrong — a cold batch then loses a card's prefill parallelism (16 requests: 18.6 s spread vs 30.1 s pinned). The right rule is workload-shaped: pin continuations (requests carrying 1 message = an agent loop resending context) to the replica that served the prefix, and spread everything else (one-shot questions, batch members, parallel subagents) by counter. Measured result of that hybrid on the same host: 1095 ms warm turns AND 19.97 s burst. Two traps found by measuring, not reasoning: (1) hashing a fresh prefix to choose a card sends every member of a same-prefix burst to the same card (29.8 s) — spread must come from a counter, only continuations may be pinned; (2) the conversation identity is the first N KiB of the prompt (4096 by default) — if the prefix is shorter than that window, turns hash differently and never pin. Also keep the routing policy in its own stdlib-only module: a proxy that imports aiohttp at module scope cannot be unit-tested in an environment without it, and the first version of the routing code shipped a bug (stale index - HTTP 500 on every continuation) that only a policy-level test catches cheaply. (context: Dankserver frank-medium 2026-09-16: two TP=1 radiance MXFP4 replicas behind a proxy on :7770 (TP=2 impossible — PCIe atomics). Implemented routing.py + affinity-proxy.py, tests/frank_medium_proxy.bats, deployed via deploy.sh with the e2e gate.)
• Multi-GPU RCCL collectives on AMD require PCIe atomics on EVERY participating GPU, and on AMD desktop platforms the second x16 card usually sits behind chipset bridges that advertise AtomicOpsCap: Routing- (read-only). RCCL's collective kernels carry a hidden hostcall-buffer parameter; ROCr refuses to give them one when the device lacks atomics (rocclr/device/rocm/rocvirtual.cpp:3963, gated on pcie_atomics_ set once at driver probe from the CPU-agent link info), and the launch dies as hipErrorIllegalState ("the operation cannot be performed in the present state") with NO mention of atomics. Diagnostics: AMD_LOG_LEVEL=3 prints the truth (Pcie atomics not enabled, hostcall not supported → AQL dispatch failed! → hipLaunchKernel: Returned hipErrorIllegalState); lspci -vv -s gpu on the endpoint AND every upstream bridge must show AtomicOpsCap: Routing+; the endpoint's Device Control 2 (setpci -s gpu CAP_EXP+28.w) must have bit 6 set. Setting that bit by hand does NOT help — the driver consumes it at probe time. Traps: a single-rank (world_size=1) collective succeeds on either card because NCCL short-circuits, so one-card tests prove nothing; the failing kernel version was ruled out across 7.0.0-28/-30/-31; a from-source RCCL rebuild reproduces the identical failure, so the library is never the bug. Workaround when only one card is atomic-capable: two single-GPU TP=1 replicas behind a round-robin proxy (vLLM's OpenAI API is stateless per request). Related KV trap found the same night: passing --kv-cache-memory 0 explicitly means "profile this shape" (radiance-patched vLLM); BLANKING the variable so the flag is omitted makes vLLM size KV from --gpu-memory-utilization alone and the allocation OOMs right after weights load. (context: Dankserver frank-medium (2x R9700, AM5) 2026-09-16: TP=2 radiance MXFP4 could not launch any collective; root cause traced to PCIe atomic routing on the second card's chipset path; deployment changed to two TP=1 replicas + round-robin proxy, deployed green with benches (conc1 42.8, conc4 147.6, conc8 230.8 tok/s; TPOT 23.6-28.8 ms; 32k prefill ~2.8k tok/s per card).)
• serverAppConfig deploy.sh pre-flight gotcha (2026-09-16): it requires a WHOLE-TREE clean git status --porcelain (untracked included) AND the main checkout on branch main — a linked git worktree fails the "must run from the main branch" check, so worktrees cannot bypass a dirty tree. When another agent's uncommitted files block a needed deploy, do NOT commit their WIP; use a targeted stash cycle: git stash push -m "deploy-window: park peer WIP" -- peer files && ./scripts/deploy.sh target ... && git stash pop. The stash survives a conflicting pop (kept on failure), the peer's edits are restored byte-identical, and the window is seconds. Check git status --short afterward to confirm the peer's files are back. (context: 2026-09-15 frank-large deploy: deploy.sh failed twice (dirty tree from peer files; worktree rejected by main-branch check), then succeeded via targeted stash cycle; peer WIP verified restored (git status identical, stash dropped cleanly).)
• CORRECTION (2026-09-16): the "e2e 11-admin.spec fails on 192.168.0.112:8989 from blankpad — environmental, expect 29/30" lesson is STALE. On 2026-09-16 tests/e2e/run.sh passed 30/30 from blankpad, including all three 11-admin specs, with no known intervention. Do not pre-declare the admin-SSO specs environmentally doomed or treat 29/30 as the expected baseline — run the suite and read the actual result. The old diagnosis (blankpad TCP-path to the arr ports blocked while peers succeeded) no longer reproduces; if an admin spec fails again, re-diagnose fresh rather than citing the old blocker. (context: 2026-09-16 frank-large ik cutover wrap-up: tests/e2e/run.sh from blankpad returned 30/30 passed including specs/11-admin.spec.ts against 192.168.0.112 raw ports; repo commits 795b0b3/ad7a9c4.)
• frank-large ik_llama.cpp performance campaign (2026-09-16): decode ~3.8-4.3 t/s is gated by a SERIAL per-token critical path — ~100 CPU↔GPU backend splits/token with the 3090 at 5% SM utilization (launch/sync-bound across the PCIe x4 dock); thread count is irrelevant (T1≈T16); prefill ~192 t/s sits at the CPU AVX512-VNNI compute ceiling for CPU-resident experts. The 30 t/s halogen bar is unreachable on this model+hardware: ik's CUDA-graph capture aborts on qwen4exp's per-layer ggml_repeat_4d hc-broadcast ops (capture-illegal op), mainline llama.cpp 37b53fd runs 1.0 t/s (CPU MoE kernels 4x slower than ik's iqk), the GGUF has no MTP head (mradermacher stripped it), and 84 GiB of experts cannot fit 24 GiB VRAM. Re-evaluate when mainline's hc-ops + graph handling mature (ggml-org PR #28901). DANGERS learned: (1) systemd Restart=always respawns a pkill'd llama-server ~4s later — manual experiments MUST sudo systemctl stop the unit first or measurements get phantom instances; (2) ik --defer-experts does NOT keep experts off the GPU under -ngl 999 (only PLE defers) → cudaMalloc OOM crash at 79.7 GiB; (3) ik -ot/-ncmoe overrides disable mmap for the whole load (create_tensors returns false); (4) -ncmoe N without GGML_CUDA_NO_PINNED_WEIGHTS=1 page-locks ~108 GiB (the co-tenant killer); (5) never run -j16 builds concurrently with decode measurements. (context: 2026-09-15/16 overnight frank-large decode campaign on dankclaw: goal was halogen's ~30 t/s decode; measured evidence chain (perf pause-spin analysis, T1=T16 sweep, nvidia-smi 5% SM during decode, graph splits=100, mainline comparison 1.0 t/s, GGUF MTP-head check) recorded in serverAppConfigs-docs/LLM/frank-large-ik-uncensored.md, commit 795b0b3.)
• serverAppConfig repo convention (established 2026-09-16, commit f227786): AGENTS.md is evergreen rules + key lessons only. Fleet-swap history, retired-runtime detail, and per-flag sizing rationale do NOT go back into AGENTS.md — they live in serverAppConfigs-docs/LLM/ runbooks, and AGENTS.md points at them (Runbook Index table). The "Key Lessons Learned" section holds only durable, cross-runtime operational facts (ROCm layer-split, HIP env collisions, mmap/PLE, swap ordering, quadlet/podman gotchas). When a runtime swap or retirement happens: update the ACTIVE topology table and the Runbook Index row, add any new durable lesson, and write the detail into the runbook — do not grow AGENTS.md history sections. Known pre-existing test failure: tests/run.sh not ok 73 (comfyui Containerfile selector bug, test-side) is unrelated to doc changes and does not block commits. (context: AGENTS.md rewrite session 2026-09-16; Frank approved "lessons + pointer table" for retired runtimes and "condense to operational-critical" for halogen detail via grilling. Commit f227786 (734→412 lines).)
• blankpad skill directories (2026-09-16): the pi-era user skills at ~/.pi/agent/skills and ~/.agents/skills were wiped; omp discovers user skills at /.omp/agent/skills (global) and .omp/skills (project), both of which now exist again. Restored 9 skills from source repos: 8 from github.com/mattpocock/skills (grilling, tdd, implement, prototype, teach, diagnosing-bugs, code-review, handoff — code-review and handoff were mislabeled "Local" in the old omp-docs/pi/skills.md; they are from mattpocock) + find-skills from github.com/vercel-labs/skills. Copied as plain dirs (skill/name/SKILL.md, spec-compliant frontmatter) — the npx skills CLI was bypassed because it auto-detects the wrong agent (claude) and doesn't know omp. UNRECOVERABLE until dankware (192.168.0.226, offline, no SSH alias) is back online: obsidian-vault, writing-great-skills, writing-guidelines, zim-knowledge-base. New skills are discovered on the NEXT session, not the current one. _(context: blankpad skill restoration session 2026-09-16: user noticed /skill:grilling unresolved. Traced pi-era skills (/.pi/agent/skills + ~/.agents/skills, installed via vercel-labs skills CLI) as wiped; recovered 9 of 13 from source repos on GitHub into ~/.omp/agent/skills.)_
• The AGENTS.md fleet-topology snapshot embedded in session context can lag the repo: on 2026-09-15 the serverAppConfig tree at HEAD (commits d53962f/4183e31) described a NEWER layout than the context file — frank-medium (Dankserver 2x Radeon AI PRO R9700, gfx1201) replacing frank-small, and frank-large moved to ik_llama.cpp serving orcarouter/Qwen3.8-Flash-Next-Uncensored i1-Q4_K_M on Dankclaw's RTX 3090. Before reasoning about the LLM fleet from the AGENTS.md context (service names, models, runtimes, ports), verify against git log --oneline + the tracked serve/env files at HEAD; the context file is a snapshot of a past date. Related trap from the same session: the working tree carried UNCOMMITTED radiance-era edits (scripts/deploy-plan.py, scripts/deploy-remote.sh, tests/deployment_scoping.bats) plus a DELETED test (tests/frank_large_quadlet.bats — deletion must go through .AGENTTRASH per policy, never rm); reconcile git status before any deploy step. (context: 2026-09-15 frank-medium (dankserver 2x R9700, vllm-radiance MXFP4) cutover session: interrupted while reconciling working-tree state before host-side trash staging. The system-prompt AGENTS.md describes the 2026-09-11/13 fleet; commits d53962f/6085a99 ("serve the new R9700s, retire frank-small") plus scripts/frank-large.env (2026-09-15 header, ik_llama.cpp, dankclaw RTX 3090) show the fleet was re)
• FLEET STATE UPDATE (2026-09-15): frank-small (FreeToken/CUDA, RTX 3090) on dankserver is RETIRED. The 3090 was removed and replaced with two AMD Radeon AI PRO R9700 (Navi 48, gfx1201, 32,624 MiB each; no NVIDIA device remains; host iGPU gfx1036 hidden from ROCm via HIP_VISIBLE_DEVICES=0,1). The :7770 dankserver endpoint is now frank-medium: charlie12345/ROCmFPX llama.cpp (c49ebdb, built gfx1200;gfx1201 with the fork's self-contained ROCm 7.14 TheRock toolchain, RPATH-bundled — host /opt/rocm-6.3.3 was purged and never participates) serving kingjones777/Qwen3.8-27B-ROCmFPX-Q8_0-AGENT-GGUF (26.28 GiB Q8_0 + MTP Q4_0 draft pinned ROCm1) with layer split across both cards. Deploy path: ./scripts/deploy.sh dankserver --service frank-medium (unit is a root systemd unit symlinked to the repo file; first deploy retires the old frank-small unit). Clients already cut over: blankpad ~/.omp/agent/{models.yml,config.yml} and Hermes live config on dankclaw (provider key stays 'dankserver-small' as label-only). Any memory/doc still describing dankserver as FreeToken/CUDA frank-small is stale. Trash staged awaiting purge approval: repo scripts/.AGENTTRASH/{frank-small.env,frank-small.service,serve-frank-small.sh}, dankserver ~/.AGENTTRASH/{20260915-frank-small-retire,20260915-rocm633-cleanup}, dankclaw hermes models_cache.json-20260915. (context: 2026-09-15 GPU swap on dankserver (RTX 3090 out, 2x Radeon AI PRO R9700 gfx1201 in); frank-medium deployed via repo git flow, commits 6085a99/d53962f/d4024a9/db29c4d; runbook serverAppConfigs-docs/LLM/frank-medium-rocmfpx.md; supersedes the fleet-topology lines in older memory (frank-small FreeToken/CUDA on dankserver).)
• Dankserver frank-medium (2x Radeon AI PRO R9700, gfx1201) ops facts, measured 2026-09-15: (1) llama-bench multi-GPU gotcha — '-dev ROCm0,ROCm1' runs two SEPARATE single-card benches; a real 2-GPU bench needs '-dev ROCm0/ROCm1'. (2) Parallelism on this RCCL-less ROCmFPX build: layer split (1991 t/s prefill, ~17.6-18.8 t/s tg) beats tensor split (~272 t/s prefill, 7x worse) because tensor mode falls back to the meta-backend butterfly allreduce; with MTP serving, layer also wins decode (35.7-46.2 vs 28.7 tok/s). Draft head must be pinned to ROCm1 — unpinned draft under tensor mode aborts KV build (GGML_ASSERT(buffer)). (3) First-request-after-restart on the engine can go silent ~45-60s (health engine_unresponsive, 503s) then self-recovers — do not restart over it. (4) Request-level speculative:{n_max:0} does NOT disable drafting. (5) Hermes provider KEY 'dankserver-small' is label-only and deliberately NOT renamed (DB references) even though its model id is now frank-medium; repo example dankclaw/hermes/config.yaml.example synced in db29c4d. (6) Trash still staged awaiting purge approval: repo scripts/.AGENTTRASH/{frank-small.env,frank-small.service,serve-frank-small.sh}, dankserver ~/.AGENTTRASH/{20260915-frank-small-retire,20260915-rocm633-cleanup}, dankclaw hermes models_cache.json-20260915. (context: 2026-09-15 frank-medium bring-up on Dankserver (2x Radeon AI PRO R9700, charlie12345/ROCmFPX llama.cpp, kingjones Q8_0_ROCMFPX-AGENT); runbook at serverAppConfigs-docs/LLM/frank-medium-rocmfpx.md, commits 6085a99/d53962f/d4024a9/db29c4d.)
• frank-large (halogen-flash-server, Dankclaw) pool history and post-restart behavior, measured 2026-09-13: pool trimmed 524288 → 401408 (deployed, repo commits 91406af + 27e6335) to restore co-tenant headroom. Measured boot ledger at 401408/MAX_TOK=32768/reserve 15: 68.0 weights + 11.0 KV + 21.0 working = 100.0 GiB held, host left 9.1 GiB / 1396 contiguous 2 MiB blocks (vs 7.1 / 310 at 524288), engine printed need ~31.5 GiB vs modelled ~33.4, and the co-tenant stall NOTE no longer prints (threshold is somewhere below 9.1 GiB / 1396 blocks). TWO new facts: (1) On the FIRST request after any (re)start the engine can go silent ~45-60 s — /health flips to engine_unresponsive "no PONG within 30s" and the front-end 503s, then self-recovers. Do not restart the unit over it; the quadlet HealthStartPeriod=300s covers it. First completion attempt timed out during this window; a retry 20 s later returned instantly. (2) Host-left gain (+2.0 GiB) came in below the modelled +3.4 because the engine holds freed KV space as working slack (working 21.0 at 401k vs 20.6-ish at 524k with arena). The omp models.yml on blankpad (/home/blank/.omp/agent/models.yml) had frank-large maxTokens: 262144 — above halogen's 65536 cap and a 503-queue hazard after the pool cut; fixed to 65536, and the stale flm (dankclaw-npu) provider + its modelRoles (vision/tiny/task) were removed per the AGENTS.md route map (vision→frank-large, tiny→frank-small, task→frank-large:medium). (context: frank-large KV pool trim to 401408 on Dankclaw, 2026-09-13, followed by OMP models.yml/config.yml cleanup on blankpad.)
• Dankclaw host swap (as of 2026-09-12): 64 GiB file at /swap.img, ext4 on /dev/nvme0n1p2, fallocated hole-free, driven by /etc/fstab line 12 (/swap.img none swap sw 0 0) which generates the swap.img.swap systemd unit — fstab is the only control point, there is no hand-written unit and no repo file, so deploy.sh cannot see swap state. vm.swappiness=60 (distro default, UNTESTED against this workload), zswap disabled. Raised from 8 GiB; 748 GiB still free on /. The old 8.1 GiB file is staged at dankclaw:~/.AGENTTRASH/20260912-swapfile-8g/swap.img awaiting Frank's approval to purge. Important framing: swap on this host protects the ~28 co-tenant containers, NOT frank-large — its ~68 GiB of pinned weights are not relieved by swap, and paging the KV pool (which IS anonymous) would be worse than the OOM kill it prevents, so if the box thrashes the knob is lower swappiness or a cgroup cap, not more swap. (context: Raised swap on Dankclaw 2026-09-12 after the frank-large reserve-15 config left only ~7.1 GiB for co-tenants. Documented in serverAppConfigs-docs/Dankclaw/Swap Configuration.md and AGENTS.md memory-model section.)
• In the serverAppConfig repo, ./tests/run.sh exits 1 on a clean tree because of one pre-existing failure: tests/comfyui_rocm.bats test 73 "comfyui Containerfile is based on the gfx1151 torch image". It is a test-side selector bug, not a product regression — its find ... | head -1 picks dankclaw/alloy/Containerfile instead of the ComfyUI one, so the assertion reads the wrong file. Verify before blaming your own change: a good run reports 285 ok plus this single not ok 73. Confirmed present at HEAD across several sessions (2026-09-12 frank-large work), independent of frank-large/Caddy/quadlet edits. Do NOT treat a green deploy as blocked by it, and do NOT "fix" it by weakening the assertion — the correct fix is to target the actual ComfyUI Containerfile path. Note deploy.sh gates on its own per-service health plus the Playwright e2e suite, not on tests/run.sh, so this failure does not block deploys. (context: Every ./tests/run.sh run during the 2026-09-12 frank-large KV-pool/reserve work returned exit=1 with exactly not ok 73 comfyui Containerfile is based on the gfx1151 torch image while 285 tests passed, including on commits where nothing ComfyUI-related changed.)
• Frank resolves contested tuning calls by running the experiment, not by reading the argument. On 2026-09-12 he overruled documentation-based pushback twice in a row with "lets just try it". Working with this, rather than against it: when he pushes back on a knob, state the prediction numerically up front, then run it and report measured vs predicted — he engages with that and it twice produced a better model than the docs (the arena linearity, and the vision-tower term). Corollaries: (1) get the rollback staged and named BEFORE the risky deploy, because the experiment may take production down — the 65536 trial OOMed frank-large for ~2 minutes; (2) use deploy.sh --skip-e2e for risky single-service experiments so the 4-minute e2e gate is not sitting on top of an outage; (3) after a failure, git revert sha + redeploy restores in ~10s, because the weights are already page-cached. Also learned the hard way: when a config experiment ends in a state that quietly reverses the user's original request (32768 + KV_POOL_FIT clamp meant the pool stayed 262144, i.e. the doubling never happened), say so plainly — do not let a green health check imply the goal was met. Separately, on log forensics: journalctl --user -u frank-large.service did NOT reach back to the 2026-09-11 boot even when asked with --since 2026-09-11 (rootless user journal retention is short). To compare against a container's ORIGINAL startup ledger, capture podman logs BEFORE the first deploy of the session — that is the only reliable copy of pre-change boot state. (context: frank-large KV-pool doubling work on 2026-09-12: Frank overruled doc-based pushback twice ("let's just try it") — once raising HALOGEN_MAX_TOK to 65536 (OOMed the production service for ~2 min) and once back to 32768, which quietly un-did the doubling he had originally asked for. Also: journalctl --user retention for frank-large.service did not cover the 2026-09-11 boot even when queried with --si)
• CORRECTION to the earlier "e2e 11-admin.spec.ts failing on 192.168.0.112:8989 means the Mac is asleep — wake it and re-run" lesson. That diagnosis is incomplete and can send you in circles. On 2026-09-12 the admin SSO spec failed on page.goto http://192.168.0.112:8989/ with ERR_CONNECTION_TIMED_OUT, and waking the Mac did NOT fix it: the Mac was verifiably awake (ping 33-54ms), all six socat forwarders were bound to 192.168.0.112, macOS Application Firewall reported State = 0 (disabled), netstat showed 8989 listening, and from dankclaw the ports answered 8989→200, 7878→302, 9696→302. Yet from blankpad (the dev box where Playwright actually runs) the same ports returned curl rc=28 / 000, while ICMP ping to .112 succeeded. So the blocker is a blankpad↔192.168.0.112 path problem — ICMP allowed, TCP to the arr ports not — not sleep and not the Mac's own firewall. Distinguish the two causes by probing from the box the browser runs on, not from a fleet peer: curl -s -o /dev/null -w '%{http_code}' --max-time 7 http://192.168.0.112:8989/ on blankpad vs ssh dankclaw 'curl ... 192.168.0.112:8989'. Peer 200 + local timeout = path/firewall issue, and no amount of waking or re-running will make that spec pass from this box. Consequence: the mandated e2e gate currently cannot go fully green from blankpad; 29/30 pass, 11-admin fails on Sonarr's raw IP. It is environmental and unrelated to what is being deployed — verify the deployed service directly (ledger, /health, real completion) rather than treating the red gate as a regression, and say which gate was skipped and why. (context: 2026-09-12 frank-large KV-pool work: e2e admin SSO spec failed on Sonarr raw IP while the Mac was provably awake and dankclaw reached it fine; blankpad could not open the TCP ports at all.)
• halogen-flash-server (frank-large, Dankclaw) device-RAM sizing is ONE shared budget, and it is now measured rather than estimated: device budget = host RAM - resident weights - HALOGEN_HOST_RESERVE_GIB (121 - 67.7 - 24.0 = 29.5 GiB here, plus a 1.5 GiB vision margin). The KV pool costs ~28 KiB/position (524,288 = 14.4 GiB) and the PREFILL ARENA IS LINEAR IN HALOGEN_MAX_TOK at 3.80 GiB base + 8.40 GiB per 16,384 tokens — verified against three restarts to 0.1 GiB. So the ceiling is max_tok = (budget - margin - pool - 3.8)/8.4 * 16384; with the doubled pool that is max_tok = 19,115, making the deployed 16384 the CEILING, not a conservative setting. Corner check confirms: max_tok 32768 leaves 7.4 GiB = ~269k positions, exactly the old 262144 config. TWO DIFFERENT FAILURE MODES when you violate this: (a) under the ceiling, HALOGEN_KV_POOL_FIT silently LOWERS the pool and deploy.sh still reports healthy in 12s — you think you doubled it, you did not; (b) OVER the ceiling the arena alone outgrows the budget and the server WILL NOT BOOT at all — measured at max_tok=65536: "WARNING even 262144 positions ... needs ~44.5 GiB", then "dmalloc: FAILED requesting 0.250 GiB after 39.970 GiB (out of memory)". Mode (b) is the dangerous one because systemctl stays active, the container reads Up (starting), and deploy.sh waits out the full LLM_WAIT instead of failing fast — an outage that looks like a slow deploy. Always grep podman logs frank-large for kv pool: and confirm the reserved position count equals the requested one; /health 200 proves nothing about pool size.
• The e2e suite's 11-admin.spec.ts navigates to the *arr services at their raw LAN IPs (192.168.0.112:8989 Sonarr, 7878 Radarr, 9696 Prowlarr), and Playwright runs from the DEV BOX (blankpad), not from Dankclaw. There are therefore TWO independent ways that spec fails, and they need different fixes: (1) The known one: FrankBook-Pro asleep - arr ports unreachable from everywhere. Wake it, re-run. (2) NEW, found 2026-09-12: blankpad cannot reach those forwarded ports AT ALL, even with the Mac fully awake. Measured: from blankpad, ping 192.168.0.112 succeeds (33-54 ms, 0% loss) but TCP to 8989/7878/9696 all return curl rc=28 (connection timeout). From dankclaw the same ports return 200/302/302. On the Mac itself: socat forwarders ARE running and bound to bind=192.168.0.112, curl 127.0.0.1:8989 - 200, and socketfilterfw --getglobalstate reports "Firewall is disabled (State = 0)". So the Mac is not the blocker — something on the blankpad-FrankBook-Pro path drops TCP to those ports while passing ICMP. Diagnostic order that separates them, before assuming "the Mac is asleep": a. From the dev box: curl -s -o /dev/null -w '%{http_code}' --max-time 7 http://192.168.0.112:8989/ AND ping -c2 192.168.0.112. b. From dankclaw: same curl. If dankclaw gets 200 but the dev box times out while ping works, it is failure (2), not sleep — re-running the spec will never go green from the dev box. c. On the Mac (via ssh dankclaw "ssh dank@192.168.0.112 ..."): pgrep -fl socat, curl 127.0.0.1:8989, firewall state. Implication: a red 11-admin.spec.ts is NOT automatically evidence of a fleet outage, and it is NOT caused by whatever was just deployed. Confirm which failure mode by comparing the dev-box probe against the dankclaw probe before attributing the failure to your change. (context: 2026-09-12, after doubling frank-large's KV pool on Dankclaw. The e2e gate's 11-admin.spec.ts failed on http://192.168.0.112:8989/ (ERR_CONNECTION_TIMED_OUT) on both the initial run and retry. Waking the Mac did NOT fix it, which is why this qualifies the older sleep lesson.)
• halogen-flash-server (frank-large on Dankclaw): raising HALOGEN_KV_POOL_POSITIONS can SILENTLY NOT TAKE EFFECT, while the deploy still reports healthy. HALOGEN_KV_POOL_FIT=1 (default) measures the device budget at startup and lowers the pool itself if the configured value will not fit — and it never raises. The budget is ONE pool covering the KV AND the prefill arena together: host RAM − resident weights − HALOGEN_HOST_RESERVE_GIB. Measured on Dankclaw 2026-09-12: 121 GiB − 67.7 weights − 24.0 reserve = 29.5 GiB device budget. Pool 524288 with HALOGEN_MAX_TOK=32768 needs ~35.0 GiB (+1.5 margin) → refused, silently lowered to 262144, deploy reported "healthy after 12s". Fix: HALOGEN_MAX_TOK=16384 halves the arena (~20.6 → ~12.3 GiB working mem), bringing the 524288 need to ~26.6 GiB, which fits with ~2.9 GiB margin at ~8-9% prefill cost. Net result was CHEAPER than the old config: 94.6 GiB held / 13.8 GiB host-left vs 96.3 / 12.1. RULE: after ANY halogen restart, read podman logs frank-large | grep 'kv pool:' and confirm the pool it reserved matches what the unit asked for — a green /health and a green deploy.sh say nothing about pool size. The engine prints "LOWERING THE POOL TO N" in the refusal; that line is the tell.
• frankbook-pro (FrankBook-Pro.local, macOS 192.168.0.112) is reachable from this dev box ONLY via the dankclaw jump host: ssh dankclaw "ssh dank@192.168.0.112 'bash -l -s'" script.sh (fleet inter-host key works for dank@112; frank@/admin@ are denied; NO passwordless sudo on the Mac — sudo -n fails). The *arr stack (sonarr 8989, radarr 7878, readarr 8686, lidarr 8787, prowlarr 9696, qbittorrent, gluetun) runs inside podman-machine-default (applehv); on the macOS host, socat forwarders bind 192.168.0.112:port → 127.0.0.1:port (gvproxy → VM). macOS Application Firewall is DISABLED and /etc/pf.conf is stock. THE FLAKE: when the Mac sleeps, the arr ports go unreachable from the LAN (curl connect 0.000s refused / ERR_CONNECTION_TIMED_OUT in Playwright), which fails the e2e 11-admin.spec.ts SSO test. Waking the Mac (SSH/ping) restores all ports to 200/302 within seconds. So an admin-SSO e2e failure on a *arr link is usually "the Mac is asleep" — wake it and re-run before treating it as a real outage. Diagnostic order: (1) ping 112 (high latency ~150ms = asleep/waking), (2) from the Mac host curl 127.0.0.1:8989 and curl 192.168.0.112:8989 (both 200 = chain healthy, block is external/sleep), (3) podman machine list + podman ps -a (Up = services fine). (context: Diagnosing the e2e admin SSO test failing on Sonarr raw-IP (112:8989) during the 2026-09-12 marreta work — turned out to be frankbook-pro sleeping, not a real outage.)
• authentik embedded-outpost proxy providers: a ProxyProvider is INVISIBLE to the Rust embedded outpost unless it is linked to an authentik Application. The outpost's config API (ProxyOutpostConfigViewSet in authentik/providers/proxy/api.py) filters ProxyProvider.objects.filter(application__isnull=False). So a provider created via ProxyProvider.objects.create(...) with no Application never reaches the outpost, and requests to its external_host fall through to raw flow serving on the app's own host — where the IdP session cookie (scoped to auth.domain) is invisible, so the user is stuck on a login wall (e.g. /if/flow/default-authentication-flow/ or /flows/-/default/authentication/). FIX: create Application(name=..., slug=..., provider=proxy_provider, launch_url=...) and restart authentik-server + the outpost. Also required for the same provider to work: invalidation_flow must be a real UUID (null breaks the Rust outpost's serde with "invalid type: null, expected a formatted UUID string" and takes down ALL providers on that outpost), and external_host must exactly match the Host header (trailing-slash mismatch = no match). Diagnose by: (1) ak shell → list providers with application/invalidation_flow; (2) journalctl --user -u authentik-outpost.service | grep -iE "error|invalid" for the serde error; (3) compare a working provider's full field set against the broken one. (context: Fixed marreta.franksdank.cloud login-wall on Dankclaw 2026-09-12: the marreta-proxy ProxyProvider was created without a linked Application, so the embedded outpost never loaded it. Root cause chain: null invalidation_flow (serde crash) - missing flows - missing Application link.)
• NEVER send SIGHUP to PID 1 inside a podman container to reload a service. In podman containers PID 1 is the container supervisor/init, not the app — SIGHUP to it TERMINATES the whole container. On 2026-09-12 I tried podman exec pihole kill -HUP 1 to reload dnsmasq addn-hosts; it killed the pihole container and took LAN DNS down for ~15s until the quadlet unit's Restart=always (plus an explicit systemctl --user restart pihole.service) brought it back. The app process (pihole-FTL) was PID 54, not 1. Correct ways to reload a service in a podman container: (a) systemctl --user restart svc.service (the quadlet unit — clean, intended), or (b) signal the ACTUAL service PID found via podman exec c ps -e -o pid,comm, never PID 1. For Pi-hole specifically, the repo flow is deploy.sh dankclaw --service pihole which does the addn-hosts cp + unit restart. (context: 2026-09-12, closing the marreta addn-hosts gap on the live Pi-hole. pihole reload is not a valid subcommand (printed usage, exit 0), so the || kill -HUP 1 fallback fired and killed the container.)
• On the Dankware/Dankclaw homelab, a "connection timeout" to any *.franksdank.cloud service from a LAN client is usually a CLIENT split-DNS problem, not a server outage. The public DNS record for franksdank.cloud points at the WAN IP (71.227.227.172) which has NO 443 port-forward (filtered). Pi-hole (192.168.0.143) returns the correct LAN answer 192.168.0.143 via addn-hosts. If a client's systemd-resolved is using 1.1.1.1 (or any external resolver) as its current server, franksdank.cloud resolves to the WAN IP → TCP SYN to a filtered port → timeout (curl rc=28). LAN-only names (dankclaw/grafana/marreta) return NXDOMAIN instead. Diagnose: dig +short www.franksdank.cloud @192.168.0.143 (→143) vs @1.1.1.1 (→WAN); resolvectl status shows the current server. Fix on the client: a systemd-resolved drop-in routing the domain route-only to Pi-hole. Verify the server itself is fine by pinning the LAN IP: curl -sk --resolve www.franksdank.cloud:443:192.168.0.143 https://www.franksdank.cloud/ → 200. (context: 2026-09-12 homepage "links gone / not loading / connection timeout" on dankware. Server was fully healthy (Caddy @home 200, Anubis challenge valid, Authentik outpost healthy, files render). Root cause was dankware's resolved using 1.1.1.1. Fixed with /etc/systemd/resolved.conf.d/franksdank-lan.conf [Resolve] DNS=192.168.0.143 Domains=~franksdank.cloud ~franksdank.casa + restart systemd-resolved.)
• Dankware fleet — "service unreachable / connection timeout" triage for Anubis+Authentik+Caddy guard-chain vhosts. Two durable gotchas: (1) Rootless-podman published-port SNAT makes Anubis log x-real-ip: 10.89.1.22 (the podman gateway) for EVERY client — the real client IP is lost at the published-port layer, so never use Anubis x-real-ip for per-client identity/rate-limit reasoning. (2) The WAN IP 71.227.227.172 has NO 443 port-forward; any client resolving *.franksdank.cloud to the WAN IP (i.e. NOT using Pi-hole 192.168.0.143 as DNS, or off-LAN) gets a TCP connection timeout — the SYN hits the filtered public address. LAN clients must resolve via Pi-hole to 192.168.0.143. To PROVE a guard-chain service is server-side healthy (and thus the fault is the complaining client's DNS/network, not the server): test from a LAN peer — ssh dankserver then dig +short www.franksdank.cloud (expect 192.168.0.143), /dev/tcp/192.168.0.143/443 (expect OPEN), curl -sk https://www.franksdank.cloud/ (expect 200). Also check Pi-hole addn-hosts has the 192.168.0.143 host line and iptables -L INPUT policy is ACCEPT. If the LAN peer gets 200 but one device times out, the server is fine — fix that device's DNS (point at 192.168.0.143) or its LAN connectivity. Note: a curl to the bare IP without matching SNI returns code 35 (SSL name mismatch) — that is NOT a real failure; always use the hostname URL with --resolve or real DNS. (context: 2026-09-12: homepage reported "links gone" then "connection timeout" from danklivingroom. Exhaustive server-side trace (Anubis challenge page valid, Authentik outpost healthy, Caddy @home/@data 200, files render with visible links, no CSP/redirect-loop, no fail2ban) proved the stack healthy; dankserver LAN peer got https 200 in 7ms. Root cause isolated to the complaining client resolving to the fi)
• On dankclaw's Anubis→Authentik→Caddy guard chain, the rootless-podman published-port SNAT masks the real client IP: Anubis logs x-real-ip: 10.89.1.22 (the Caddy container's podman-network address) for EVERY client — verified with both host-local curl and a browser from a different machine (blankpad). So Anubis cannot distinguish clients by IP; clearance is cookie-based, not IP-based. This is pre-existing fleet behavior, NOT caused by config edits, and does not break clearance. When a user reports a guarded page "broken" but the server is healthy, the break is client-side/session — you need the authenticated view (Authentik creds) or the exact symptom, not more server probing. (context: 2026-09-12 homepage "links gone / not loading" investigation; server-side fully verified healthy, root cause not reached (awaiting user symptom).)
• TRANSPORT UPDATE (2026-09-12, supersedes the old "blankpad has no fleet SSH" lesson): blankpad (~/.ssh on user blank) NOW has passwordless SSH to the whole fleet. An ed25519 key was generated (id_ed25519, comment blankpad-deploy-20260912) and its pubkey installed on dank@192.168.0.143 (dankclaw) and dank@192.168.0.235 (dankserver). ~/.ssh/config defines Host aliases dankclaw and dankserver (User dank, IdentityFile ~/.ssh/id_ed25519, IdentitiesOnly yes). So ssh dankclaw / ssh dankserver and ./scripts/deploy.sh dankclaw ... work from this box now — no more "ask Frank for a transport." Passwordless sudo on the targets is separate (FRANK_SUDO_PW / NOPASSWD). MARRETA/LARAVEL APP_KEY GOTCHA: a Laravel APP_KEY must be the base64:-prefixed form — base64:$(openssl rand -base64 32). A bare base64 string (no prefix) is parsed as a raw key of the wrong length and every app route 500s with Unsupported cipher or incorrect key length at Encrypter.php:61, while the health route /up (which skips the encrypter) still returns 200 — so a green healthcheck masks a fully-broken app. Fix before first boot (no encrypted data persisted yet). (context: Deployed marreta on dankclaw 2026-09-12; had to set up SSH keys from blankpad first, then fix a bare-base64 APP_KEY that 500'd all app routes.)
• On the Dankware/Dankclaw fleet, LLM services that are rootless Quadlet units (e.g. frank-large on Dankclaw, user dank) must be read WITHOUT sudo: sudo podman logs name fails with "no such container" because rootless storage is per-user. And because Quadlet runs --replace --rm, podman logs only covers the CURRENT container instance — history across recreations lives only in journalctl --user -u name.service. So: current instance → podman logs; anything before the last recreate → journalctl --user. Separately: the dev box /home/blank/Projects/serverAppConfig (user blank) has NO ssh key or host alias for the fleet hosts — ~/.ssh holds only known_hosts, and dank@192.168.0.143 / frank@192.168.0.143 both return Permission denied (publickey,password). Do not burn calls probing fleet SSH from this box; ask Frank for a working transport before claiming live verification, and label repo-grounded answers as not live-verified. (context: Answering "how do I retrieve logs for frank-large?" (2026-09-12); grounded in dankclaw/frank-large/quadlet/frank-large.container and serverAppConfigs-docs/LLM/halogen-flash-server.md:191-219. Live verification attempt failed on missing SSH credentials.)