This page records host-local operational pointers that are useful during driver, runtime, service, and benchmark work. It must never contain secret values.
The local sudo password file is outside this repo:
/home/steve/SUDOPASSWORD.txt
Use it only when privileged local operations are actually needed, such as driver/runtime package checks, systemd service changes, or recovery from a known device/runtime failure. Never print the file contents, paste them into a note, or commit a copy.
The repo .gitignore and user global Git ignore both exclude
SUDOPASSWORD.txt, *.password.txt, and common sudo-password variants. If a
new helper creates another credential filename, add that filename to both ignore
lists before using it.
When Claude or OpenCode is orchestrating work, prefer delegating concrete research, audit, patch, and validation tasks to Codex/GPT through the CLI. GPT token use is much less constrained on this host, so Claude/OpenCode should act as manager/reviewer and let Codex handle bulky searches, source reading, and iteration-heavy implementation where practical.
Useful Codex CLI forms:
codex --cd /home/steve/llm-optimizations
codex exec --cd /home/steve/llm-optimizations "audit the Qwen docs and propose focused cleanup"
codex review --cd /home/steve/llm-optimizations
codex resume --last
Keep tasks bounded: provide the target repo, files or lane, expected output, and what must not be touched. For active experiments, explicitly say whether Codex may stage, commit, push, or only report findings.
Codex should use subagents whenever reasonable and available, especially for parallel source audits, independent review of risky changes, log/result classification, and research synthesis. The main Codex agent remains responsible for final decisions, edits, verification, and not disturbing active experiment processes.
Primary hot-path model cache is still on the internal NVMe:
/mnt/fast-ai/llm-cache/hf
As of 2026-07-03, an external 4 TB USB drive is available for overflow model storage and archived benchmark artifacts:
/mnt/usb-models
Device identity observed at setup:
/dev/sda2;ntfs3;CorsairExternal;/mnt/usb-models;/mnt/usb-models/llm-cache/hf;/mnt/usb-models/bench-results;/mnt/usb-models/models.Use the internal NVMe cache for active benchmark hot paths unless disk pressure or model count makes that impractical. Use the USB drive for alternate model variants, overflow downloads, and archived large artifacts. Do not commit model weights, USB paths full of artifacts, or generated cache contents to Git; record only the model identity, local path, checksum if useful, and result summaries.
For Laguna S 2.1, the stricter policy established on 2026-07-23 overrides the
general overflow rule: the external Corsair ntfs3 volume is backup-only.
Laguna model reads, caches, temporary files, logs, run roots, and recovery
evidence must stay on the internal NVMe/ext4 filesystem and fail closed rather
than falling back to /media/steve/CorsairExternal. The active model root is
/mnt/fast-ai/llm-models/laguna-s-2.1; the active artifact root is
/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1. Frozen historical
evidence may retain its original USB path in notes and manifests.