The RDNA4 offload campaign left three HIP probes and a driver sitting loose in $HOME, where the next agent cannot find them and a stray `rm -rf` ends them. They answer a recurring question on the R9700 box — is decode limited by the link or by the engine's read path — so they move into scripts/probes/ with a driver, tests, and a page. Two things changed in the code, not just the address: - The driver measures only after a guard that runs BEFORE anything is built: gpu_busy_percent over PCIE_PROBE_MAX_BUSY, or any process holding /dev/kfd, is a refusal (exit 3) rather than a measurement. A benchmark taken beside a live endpoint is both disruptive and worthless, and silently unloading a production service is not a tool's decision — --force is the human saying otherwise. The refusal-before-build ordering is pinned by a test, because the sysfs and kfd paths are overridable and a refusal is otherwise untestable off the card box. - hip_smoke.hip and pcie_probe.hip called hipGetDeviceProperties, which SEGFAULTS on this ROCm build even though hipGetDeviceCount works; only pcie_bw.hip avoided it, and that is the one that produced the campaign's numbers. Device identity now comes from the sysfs topology block the driver prints, so all three probes are runnable on the build we ship. Verified here: refusal on a real busy APU (exit 3), missing toolchain (exit 4), missing sources (exit 2). Compile-and-run still needs the box with hipcc. Refs #53
21 KiB
type, title, description, tags, timestamp
| type | title | description | tags | timestamp | |||
|---|---|---|---|---|---|---|---|
| changelog | Bundle Change Log | Chronological update history for the OKF knowledge bundle. |
|
2026-10-19 |
Bundle Change Log
Entry law (since 2026-10-19): one dated entry = what changed + the durable
lesson, a few lines. Full incident narratives are recoverable from git
(git log --follow omp-docs/log.md); knowledge that outlives an entry lives
on its concept page, linked here. Never re-expand an entry in place.
2026-10-20 — two instruments rescued from $HOME script sediment
$HOME had accumulated 123 loose scripts at its root (archived to
.AGENTTRASH/home-scripts-2026-10-20, ~1.8 GiB of stray audio consolidated into
~/audiosamples). Two were durable instruments and now live in the repo with
black-box tests and pages: tools/remote-job and
tools/pcie-probe. Refs #53. Laws learned:
- A readiness pattern in a log is not evidence of completion.
ssh host 'cmd | tail -f log | sed -n /ready/q'fails three independent ways — the broker aborts the long connection, the pattern can match a header printed before the work starts, andsed -n /x/qexits the moment it matches, tearing down the job it was watching. The only honest signal is a sentinel the job writes last, read over a short-lived connection. - Never benchmark beside a live service, and never let a tool decide to stop one. The guard refuses before anything is built; unloading a production endpoint is a human's call. A number taken as noise beside serving traffic is both disruptive and wrong.
- An instrument that segfaults is not an instrument. Two of the three shipped
.hipprobes calledhipGetDeviceProperties, which crashes on this ROCm build whilehipGetDeviceCountworks fine; device identity now comes from sysfs, and the trap is recorded in the sources so it is not reintroduced.
2026-10-20 — the secret gate was reading the wrong files, and one credential had already escaped
Found while deciding what else belongs in dotfiles/. (1) The gate whitelisted
extensions (*.md *.yml *.json *.toml *.sh *.ts *.js), so a .env, .container,
.ini or extension-less dotfile was never read — exactly the shapes the dotfiles
tier is made of. It now scans the whole staged diff for token shapes and, in
non-prose files, for credential assignments, and it fails closed when git diff
itself errors. That last clause earned its place: the rewrite's first attempt
passed -I to git diff, which is not valid there, got an empty file list, and
reported success on every commit — that is what a blind gate looks like, and it
looks identical to a clean tree. (2) A live credential had already escaped into
tracked content: a Caddy basicauth bcrypt hash, its plaintext password, and the
Pi-hole internal password, in an auto-learned memory file. Redacted — but both
credentials need rotation, because redaction does not un-leak history that
every machine has already pulled. (3) home/dank/.omp/dotfiles/** and
.pre-pull-backup/** were tracked fiction invisible to deploy.sh, and 23
wallpaper JPEGs (80 MiB) rode along in every clone. Laws: a gate must fail closed,
since a silently-passing gate is worse than none — it buys the belief of safety;
auto-learned memory is written by a process that has read your terminal, so it is
a leak surface and gets gated like any config; systemctl --user enable
snapshots an absolute path, so enabling from a worktree means the seven-day reap
silently kills the unit while is-enabled goes on saying enabled
(blankpad-desk-autoswitch.service was dangling that way). Details:
dotfiles-deploy,
home-config-census.
2026-10-20 — engineering skills configured against this KB
setup-matt-pocock-skills ran, and its output moved out of the skill's default
docs/agents/ into the knowledge base: issue tracker
(Gitea on Dankclaw is the tracker — no gh, no tea, token via a 600 curl
config), triage labels (the five canonical roles,
created on the repo, which carried zero labels before), domain
docs (single-context; glossary withheld until a term has
no other owner), plus an empty decisions with its format
rule. AGENTS.md gained an ## Agent skills pointer block. Law: this repo has
one knowledge base, so per-tool configuration is a concept page in omp-docs/ —
a second docs/ tree means a second set of index rules, which is where
documentation starts disagreeing with itself. Known cost: re-running that setup
skill will not recognise omp-docs/skills/* and will offer to recreate
docs/agents/; the AGENTS.md pointer is the thing that stops it.
2026-10-19 — dankvoice auto-speak armed at startup
Sessions spoke only on explicit /say: the settings file carried no auto
key, and arming is session-scoped by design (/say on never persists). Set
"auto": true in ~/.omp/cache/dankvoice.json. Same report, second half:
auto-speak listened only to agent_end — a settle that fires once per prompt
and is skipped outright when the harness schedules a continuation — so nothing
was narrated while an agent worked; fixed in dank/dankvoice#3, which narrates
turn_end as well. Law: silence at session start is a config state — read
/say status for auto-speak on, then watch pgrep -a pw-play for one player
per segment, before blaming the endpoint. Law: whether the harness delivers an
event to extensions is settled in minutes by a throwaway probe extension that
timestamps turn_end/agent_end — never inferred from audio timing, which the
humanizer lags by seconds. Details: dankvoice.
2026-10-19 — LazyVim installed as an NVIM_APPNAME profile
lazyvim had never been installed here — shell history shows the command typed
twice and dropped; ~/.config/nvim held only an orphaned matugen.lua palette
that nothing loaded. Now: LazyVim starter in dotfiles/lazyvim/ symlinked to
~/.config/lazyvim, a lazyvim zsh function, and the palette applied through
RRethy/base16-nvim on ColorScheme. Law: the distro stays out of
~/.config/nvim so bare nvim keeps working as a vanilla quick-edit editor.
Details: neovim.
2026-10-19 — dankvoice cut over to the dank/dankvoice package (Refs #34)
Drop-in agent/extensions/dankvoice.ts and its 26-test suite retired; the
package carries the behaviour (41/41 black-box suite) and is installed with
pi install git:. Law: install and prove /say status speaks on a box
before pulling the cutover commit — the shared settings file keeps menu
state across both transports, so the swap is stateless once installed.
2026-10-19 — dankvoice: silence and timeouts, root-caused (Refs #32)
The humanizer's raw fetch to frank-zealot omitted chat_template_kwargs. enable_thinking: false, so the Qwen35 hybrid thought instead of answering:
6.9 s of reasoning_content, content: "" — timing out the rewrite twice
(12 s dead air) and, when it did answer inside the window, speaking silence.
Root-caused by replaying the exact wire body against the live endpoint.
Fix: pin the kwarg in the request body; treat empty-content-with-reasoning as
failure (mechanical voice), not "nothing worth saying"; the menu's test phrase
fires without blocking its handler. Lesson: the harness injects thinking-off
for managed roles, but raw HTTP callers must set it themselves — 26/26 green
with three red-captured regressions.
2026-10-19 — dankvoice: humanized narration and thinking aloud (Refs #27)
Spoken units are rewritten by the @tiny role (frank-zealot) under the
bundle's speech budgets before synthesis, with mechanical cleanup as fallback;
auto-speak narrates the last message's thinking blocks first; all toggles
persist. 7 new scenarios, 22 green; live smoke through real zealot + frank-tts.
2026-10-19 — dankvoice: segmented sequential speech
agent_end awaiting one whole-answer synthesis hit the harness's 30 s
extension-handler timeout — lesson: extensions must never await work that can
exceed 30 s. Ported the built-in vocalizer's utterance-splitting state machine
(boundary = [.!?…]+ + whitespace, abbreviation/code guards, per-answer
segment budget); segments synthesize strictly sequentially. DANKVOICE_MAX is
now a segment budget. 17/17 in tests/dankvoice.bats.
2026-10-18 — dankvoice: /say options menu with output-device routing (Refs #23)
/say menu (pi-peon-style): auto-speak, voice, PipeWire output device
(wpctl → pw-play --target), player, char cap, test phrase. Precedence
env > ~/.omp/cache/dankvoice.json > defaults — born from the incident where
PipeWire's remembered Bluetooth route played answers into nowhere.
2026-10-18 — dankvoice: the cloned voices finally speak in the TUI (Refs #21)
The harness gates --voice against the bundled Kokoro enum, so frank-tts's
cloned voices were unreachable in-TUI. Added agent/extensions/dankvoice.ts
(/say …) synthesizing over the OpenAI speech route, endpoint read from the
api: openai-speech row of agent/models.yml (no host or secret in code).
2026-10-18 — Noctalia bar layout clobbered by trunk-sync stash, recovered
GUI-built bar customizations existed only in the working tree; a trunk-sync
stash-then-drop destroyed them. Recovery: git fsck --lost-found surfaced the
dropped stash as a dangling commit; validated in a sandboxed XDG_*_HOME
before restore. Lesson: GUI-written config is only durable once committed —
the motivation for the live-sync heartbeat below.
2026-10-18 — Fleet law: issue/PR attribution footers
Every agent-opened issue/PR body ends with cwd: + session: footer, in both
dank/command-center and dank/serverAppConfigs AGENTS.md: machine-made
artifacts must trace to the box and session that produced them.
2026-10-18 — live-sync heartbeat for machine-written paths (Refs #11)
A month of silent divergence in the live checkout proved the checkout-is-
deployment model needs a committer: scripts/live-sync.sh + systemd user
timer auto-commits allowlisted tiers (memories, managed-skills, live config,
dotfiles) with hooks armed, merges origin/main, pushes, reaps stale wt/*.
Non-allowlist dirt never staged; conflicts abort-and-alert. See
tools/live-sync.md; tier law in AGENTS.md.
2026-10-18 — Fleet law: filed PRs open in Frank's browser immediately
xdg-open <PR-URL> the moment a PR opens — codified in both repos' git rules.
2026-10-18 — Trunk law: extension route closed unmerged; direct repo alignment
Operator closed PR #9 (cwd-scoped before_agent_start extension) in favor of
editing each repo's AGENTS.md. Durable findings: every context-file provider
(AGENTS.md walk-up, .omp/ dirs, CLAUDE.md) resolves at or below the repo
root — ancestor-scope files can never load; user-level APPEND_SYSTEM.md
broadcasts to all sessions; serverAppConfigs never had git-flow.
2026-10-18 — SDLC switched to trunk-based development
develop retired (single-dev repo). Law in AGENTS.md § Trunk-based SDLC:
branch off freshest origin/main, PRs target main, --no-ff merges, tags
vX.Y.Z on main. Managed skill gitflow-back-merge-plumbing is historical.
2026-10-18 — Sibling-vault overlap census; system docs realigned to live ground truth
Census of omp-docs/ vs serverAppConfigs-docs (subagents + live-ssh probes)
found four contradicting pairs; ours fixed, theirs flagged for their PR flow.
Recorded then: RTX 3090 moved Dankserver→Dankclaw 2026-09; dankserver = 2×
Radeon AI PRO R9700 (the "board swap" story was false); served models per
live units. Current state always: /system/index.md
(27B-era zealot figures here are historical; live census 2026-10-19 below).
2026-10-18 — Status line rebuilt on custom segments; full-source audit
New page core/statusline.md with an audit log — four
first-pass claims were false and corrected against the minified bundle + a
live screenshot matrix (context meter is box/band only; composer.shape
routes segment arrays; mode states = Plan/Prewalk/Goal/Vibe/Loop). Bar
rebuilt on band + custom segments; frank-archon display name slimmed in
models.yml.
2026-10-09 — Blankpad DE redesign: desk profiles, keybinds, Handy wiring
Per-desk niri profiles keyed by EDID long name (connector names reshuffle
between docks; identity travels) + scripts/desk.sh; niri 26.04 has no
hotplug event (OutputConfigChanged is only an ACK per niri-ipc) — the
autoswitch daemon diffs output sets on WorkspacesChanged. Keybind cheat
sheet now generated (scripts/gen-keybinds.sh); noctalia docs corrected to
the v5 noctalia msg API. Handy: niri owns autostart; open breaker —
frank-secondary answers in reasoning_content with empty content, and
Handy cannot send chat_template_kwargs.enable_thinking=false. 27/27 green.
2026-10-09 — Incident: agent sudo-loop tripped pam_faillock, locked out the operator
A subagent retried sudo yay (no TTY) for ~15 min; pam_faillock deny=3
locked user blank — correct-password sudo failed, mimicking a password
change. Recovery: faillock --user blank --reset. Law: agents never run
sudo/yay/pacman/pkexec; root steps go to the operator as exact
commands. Skill faillock-sudo-lockout-recovery.
2026-10-09 — Full secret audit of the command-center repo: clean
Five layers (live-value scan across 77 revs/765 blobs, whole object store
incl. dangling, shape triage, entropy ≥4.3, LFS payloads) — no secret in any
reachable or dangling object; quarantine list holds. Fixed hf_token_write
0644→0600. Open gap: the pre-commit gate lacks hf_/npm_/pypi-/AIza/
Stripe/Telegram shapes and basic-auth URLs — widening it is an open decision.
2026-10-09 — Gitea token moved out of every remote URL
Remotes bare; credential in ~/.config/secrets/gitea (0600) behind a global
credential.helper store. Trap (git 2.56.0, measured): git credential-store rejects the protocol=/host= block shape the man page
shows — the store file must be one URL per line, port included. Verified with
GIT_TERMINAL_PROMPT=0. Details: tools/gitea/client-access.md.
2026-10-09 — Working tree reconciled: callsign renames, compaction move, gitignore quarantine
200-entry dirty tree committed as seven atomic units. Harness model ids
renamed frank-{medium,large,small} → frank-{archon,dragoon,zealot} —
services keep old names on the wire; the engines accept any request model
id, so renames are harness-side only. Server-side compaction moved to
zealot's 3090 (threshold 70→85 %). .gitignore quarantine for worktree-keyed
memory buckets and runtime dirs. Skill renames are not losses: git renders
directory renames as mass delete+add; pre-cutover content intact at 12c5e88.
2026-09-28 — Retired workstation purged from the knowledge base
Sold box: hostname, both LAN addresses, Wi-Fi SSID gone bundle-wide; the
knowledge distilled into system/network.md (multi-link
resolver law: a static resolver must be pinned on each connection) and
tools/niri/displays.md (portrait transforms, the
three ways a multi-monitor rig breaks). Hostname stays only where it is a
lookup key (install.sh --host, packages/<hostname>.pkgs).
2026-09-01 — Retry backoff and thinking default (pi era)
Pi settings.json: 10 s base retry backoff (10/20/40 s), defaultThinkingLevel: medium (no per-model thinking field exists in models.json).
2026-08-28 — KB rebuilt to OKF v0.1; live-verified fleet facts
omp-docs/ copied from legacy ~/.pi/agent/docs, rebuilt as git-tracked OKF
v0.1 (checkout is the deployment); unverifiable claims tagged unverified.
Conformance pass promoted concept-in-disguise indexes, stripped index
frontmatter, resolved all links; scripts/kb-lint.sh exits 0. Hardware facts
were corrected then (3090-on-dankserver since moved — see 2026-10-18).
Inventory census recorded under inventory/ for the pruning session.
2026-08-23 — Homepage deployed to production; Caddy KB doc
Homepage live on Dankclaw via scripts/deploy.sh (status checker timer,
deployed.js, rendered basicauth hashes, health check). Hard-won Caddy lesson:
top-level directives are sorted, so rewrite runs before basicauth and
auth matchers saw rewritten paths — serve each page from its own handle
block to preserve order; a handle inline matcher is a single token; a
leading @name <directive> parses as a matcher definition. Now documented on
caddy/caddyfile.md. A compose mount change needs a
container recreate, not a reload. 116 tests green.
2026-08-23 — Homepage design settled (grilling complete)
All 38 open questions answered; full record lives in serverAppConfigs
(docs/Infrastructure/Homepage Design Grilling.md). Key deltas: per-service
tcp/http probes, 5 s timeout, TUI-style panel grid, no client auto-refresh,
git-driven deploy, full TDD. Correction recorded: the Dankserver media stack
runs as bare processes, not containers — never probe it as a compose stack.
2026-08-22 — vllmScripts merged into serverAppConfigs
Full-history two-parent import; one repo, one checkout per host, one
deploy.sh. The plaintext sudo password from old docs was NOT carried over
(a bats leak-check guards it); it remains in both repos' history on the LAN
Gitea — rotation recommended.
2026-08-21 — Fleet ops: passwordless sudo, Hermes SSH keys
/etc/sudoers.d/dank-nopasswd on both servers (per-command drop-ins remain);
dedicated ed25519 key for the Hermes container, verified into all three
machines.
2026-08-19 — pi-web-access: SearXNG endpoint re-pointed to dankclaw (pi era)
searxngBaseUrl moved off dead .235:8081 to .143:8081; pi caches
web-search.json per process — new endpoint needs a new session. yt-dlp
installed (2026.07.04).
2026-08-19 — SearXNG + Gluetun bring-up on dankclaw
Durable field notes (now on concept pages — see
agentic-web-research/vpn-routing.md,
searxng-tuning.md,
engineering/infra/containers.md):
manual-WireGuard Proton needs VPN_SERVICE_PROVIDER=custom + the single-peer
WIREGUARD_* env names (not protonvpn, not WIREGUARD_PEERS_*); gluetun
ships no curl and :8000 is the control server, not health; the limiter's
JSON cap (API_MAX=4/hour) is hardcoded Python — pass_ip is the only
bypass; rootless podman strips client IPs on published ports, so SearXNG must
sit behind Caddy for XFF; quadlet generator rejects CapAdd=/Device= and
generated units enable via [Install] WantedBy=default.target. ComfyUI on
gfx1151: build on AMD's ROCm base and filter the torch family out of
ComfyUI's requirements or you clobber the gfx1151 kernels.
2026-08-17 — Fleet redo + vLLM tool calling (pi era)
Fleet moved off the pipeline-parallel + NPU topology; Quark W8A8 quant chosen over the Avesed INT8 for clean Qwen3 XML tool calls.
2026-08-16 — Gitea boot-persistence fix (dankclaw)
unattended-upgrades reboot exposed that the compose-era Gitea had no
systemd units (~1 h down). Stack rewritten as quadlet (names, systemd-gitea
network, linger + WantedBy=default.target, app.ini env-override gotcha, UID
100999 ownership) across tools/gitea/*; podman generate systemd marked
deprecated in engineering/infra/containers.md.
2026-08-13 — Pi subagent limits recorded (pi era)
Effective caps: globalConcurrencyLimit 2, parallel.maxTasks 8 / concurrency 4, maxSubagentDepth 2, turnBudget 30; no per-model limit. See pi/packages/pi-subagents.md.
2026-08-11 — Engineering research area; skills leaderboard verified
Created engineering/ (SDLC, agent instruction files, stacks, infra,
plugins, skills ecosystem, Signal remote) via parallel research fleet.
Skills leaderboard replaced with the ranked list from the skills.sh API
(zero-based pagination, 200-skill pages).
2026-08-11 — Agentic web research section; exhaustive bundle audit
Created agentic-web-research/ (8 pages) and rewrote them on 2025/2026
evidence (SearXNG valkey deployment, real engine defaults, limiter budgets,
Gluetun pattern, open-model claim triage, MCP transports). Audit: 154+
wikilinks resolved, dead external links fixed, MemorySleepMode=deep
verified valid (systemd 256+), ASD-STE100 style pass. AGENTS.md created as
the system behavior contract.
2026-08-06 — Link fixes; ComfyUI docs; major documentation overhaul
Fixed 15 broken wikilinks (missing frontmatter) + title mismatches → zero
broken. Added ComfyUI pages, per-host system docs, pi config tree, tools
entries; orphaned root duplicates moved to .AGENTTRASH/.
2026-07-18 — Maintenance tooling
Expanded references/okf-maintenance.md (cross-links, health metrics,
troubleshooting); created core/log.md and .okf/config.yaml.
2026-01-15 — OKF conversion
Docs bundle converted to OKF v0.1; files migrated into core/, tools/,
packages/, references/ with frontmatter on all files.
2025-08-05 — Gitea documentation
Added tools/gitea/index.md (topology, management, data persistence) to the tools index.