Borrowing it
Nothing to install: this file belongs to l-nmch/hailo-10h-llm-compiler. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.
curl -O https://raw.githubusercontent.com/l-nmch/hailo-10h-llm-compiler/main/.claude/skills/device-debug/SKILL.mdgit clone --depth 1 https://github.com/l-nmch/hailo-10h-llm-compilerWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/l-nmch/hailo-10h-llm-compiler/device-debug)<a href="https://agentmods.dev/skills/l-nmch/hailo-10h-llm-compiler/device-debug"><img src="https://agentmods.dev/badge/skills/l-nmch/hailo-10h-llm-compiler/device-debug/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/l-nmch/hailo-10h-llm-compiler/device-debug"><img src="https://agentmods.dev/badge/skills/l-nmch/hailo-10h-llm-compiler/device-debug.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00046 | $0.01069 |
| Opus 5.5 | $0.00018 | $0.00428 |
| Sonnet 5.5 | $0.00009 | $0.00214 |
| Haiku 4.5 | $0.00005 | $0.00107 |
Grade A, and why
device-debug scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 1mo ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 86 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Device-side debugging playbook
The most common trap in this project: a generic client-side error
(HAILO_TIMEOUT, a hang, a vague failure code) hides a specific,
immediately-legible error that only exists in the device's own logs.
Chasing client-visible symptoms (file size, PCIe throughput, retry logic)
before checking device-side logs has cost real time more than once — check
this first, not last.
Why the client alone can lie to you
The Hailo-10H runs its own firmware with an embedded mini-Linux (visible
in dmesg as u-boot-spl.bin / fitImage / image-fs loaded by the PCIe
driver at boot). The "server" side of any client/server exchange
(hailort.log talks about "sending N chunks to server") is that firmware,
not a host process. It can fail immediately and specifically, then have
its own generic timeout/error path fire seconds later — which is the
only thing the client ever reports. If you only read the client log, you
are debugging the echo, not the cause.
The technique
hailortcli logs runtime # LLM/inference server logs — start here
hailortcli logs system_control # control-plane / configuration logs
hailortcli logs nnc # neural network core logs
Run this immediately after a failure, before forming hypotheses from client-side symptoms alone. Look for the first CHECK/error line, not the last — the device's own error-reporting path can itself time out and add a second, misleading failure a few seconds later.
General elimination order for opaque failures
- Device-side logs first (above) — often gives the exact failing check and file:line in one shot.
- Structural comparison against a known-good official artifact
(
hef_audit.py,parse-hef) — rules in/out the artifact itself before blaming the runtime. - Client-side logs (
HAILORT_CONSOLE_LOGGER_LEVEL=debug,hailort.log) — timing and chunking, useful once you know what to look for. - Version alignment — driver, HailoRT, and server packages must come from the same software-suite drop; mixed versions fail in confusing ways. Check all three before deep-diving logic.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 1mo ago Changed · +13 lines ed29ea80c96c
- 1mo ago First seen · 73 lines · 46 tokens per session scan A 95c48dafa2b4
device-debug is a skill published in the GitHub repository l-nmch/hailo-10h-llm-compiler (23 stars, last pushed 26d ago), licensed MIT. It adds 46 tokens to every session and 1,069 once invoked, about $0.0002 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
trulens-instrumentation
Instrument LLM apps with TruLens OTEL-based tracing - from setup to debugging and optimization.
trulens-diagnosis
Diagnose low evaluation scores and generate actionable improvement recommendations.
bug-reproducer
Find likely software bugs in a codebase, rank concrete bug candidates, and prove or reject them with focused regression tests before proposing a fix. Also turn bug reports, stack traces, screenshots, failing behavior, support tickets, and regressions into minimal reproducible cases with red-to-green evidence. Use when…
root-cause-delivery
Use for a reported bug, regression, crash, failed workflow, or substantive code change that must reach a verified candidate rather than a local patch. Map the affected layer, freeze a bounded plan, require incident reproduction red-to-green or planned-change contract proof, force re-triage after two failed attempts…
screenpipe-incident-timeline
Reconstruct a bounded incident timeline and separate observed events from suspected causes.
screenpipe-bug-report
Turn a recorded software failure into a reproducible bug report with timestamps and observed behavior.