Every speak-aloud path in the harness gates `--voice` against the bundled
Kokoro enum, so the frank-tts cloned voices could only ever produce files.
`agent/extensions/` auto-loads, so one tracked extension adds the missing
transport: `/say <text>`, `/say on|off` (narrate the final answer at
`agent_end`, silent when the harness schedules a continuation),
`/say voice|voices|stop|status` — OpenAI speech POST plus playback through
pw-play/paplay/ffplay. The endpoint and served id come from the
`api: openai-speech` row of `agent/models.yml`, so the file carries no host
and no secret; markdown is stripped, over-long utterances are cut at a
sentence bound, and the temp WAV dir is always removed.
Verified live in a real TUI session on the workstation (roster
narrator1/narrator2/narrator3; 2.08 s and 1.44 s WAVs played; auto-speak
spoke each finished answer, 1.20 s WAV; unknown voice answers with the
roster). Black-box seam `tests/dankvoice.bats` (12 tests) drives the
extension against a fixture speech endpoint and a stub player.
cwd: /home/blank/.omp/wt/feat-21-dankvoice
session: wt-20261017-181502-6d1fc1f
'Too Claudish': drop the aside ('the gotchas no doc confesses') and the
simile ('watches it fall'); lead with the constraint itself. Facts
unchanged: exclusive dialog, scouts before ask, Esc ends the run, the
ledger survives.
The ask tool's recommended flag only labels option 0; the reply always
names the selected options, so 'the user's single Enter accepts it'
overclaimed a default-accept mechanic that does not exist. Stated
truthfully: the label invites one-glance acceptance, the reply is
self-documenting.
- opening restores 'shared understanding' alongside 'nothing is
silently assumed', and 'Map this as a design tree'
- cluster rationale restated: tight clusters keep each scout aware of
possible adjacencies
- web_search: explicit 'do not delegate to a subagent'
- 'Done' heading -> 'Definition of Done'
Read all 36 upstream skills, then rewrote to their house style per
writing-for-agents' own levers: leading words as tokens (frontier,
brief, ledger, artefact, scout), vocabulary-first and used exactly,
positive phrasing with prohibitions kept only as hard guardrails,
completion criteria sharpened to checkable+exhaustive ('counted before
sending', 'frontier empty and nothing sits blocked'), zero history,
gotchas the environment cannot confess kept inline. All settled
behaviour from the grilling-on-grilling session survives intact.
- drop provenance blockquote and the absorbed-law attribution: the skill
is instructions, not a history
- 'concurrency is capped' replaces the 3-scout figure of one machine
- completion distillate: project's knowledge base or a log markdown file,
not a command-center-specific path (skill travels outside this repo)
Rewritten per grilling-on-grilling (2026-10-18, all branches settled via
ask rounds): ask tool is the default round vehicle (one artefact per
question, recommended:0 = bare-yes semantics, hybrid boundary ask-default
/ prose-exception); todo as visible decision tree (phase=branch, blocked
reasons) + durable local://grilling-state.md ledger surviving the
Esc-aborts-the-run path; scout=environment facts (spawn-before-ask,
exclusive concurrency), web_search=world facts, inline=cheap lookups;
completion gate = summary + one closing ask, then distilled decisions
land in omp-docs/log.md via branch -> PR.
Absorbs managed skill one-decision-per-ask-question (Frank's choice;
reworded: ideas heavily encouraged via own atomic question, ban is on
smuggling into option text only). Managed skill deleted here; heartbeat
drops it from the live tree.
Observed failure mode: tracking flips land in one batch at the end of a
task bill, so the operator's cockpit view stays frozen for the whole run.
Rule pins the mechanics: init before first substantive step, done/start
batched with the step they describe, block-with-reason on stalls, no
lone todo calls. Global tier (APPEND_SYSTEM.md), every session affected.
omp's skill scanner ignores agents/*.yaml (Eve/Codex binding metadata
carried along by the skills CLI store). The file kept vanishing from the
live tree to a background sweep; an authored omp skill does not need it.
Refs #16
Retroactive provenance: the 10 pi-era skill snapshots were byte-identical
to upstream. Canonical installs now live in ~/.agents/skills tracked by
~/.agents/.skill-lock.json (mattpocock/skills x38 after grilling removal,
find-skills from vercel-labs/skills).
Repo tier is authored overrides only: clear all 10 tracked upstream
snapshots; omp loads the store directly through the agents provider
(skills.enableAgentsUser, default on). grilling returns as a real
directory — extracted from the store (CLI + lockfile pruned) so Frank can
customize it without skills update resurrecting a twin.
Docs: skill-store-wiring.md (+index). Pre-cutover snapshot of the old
tracked dirs at /tmp/agent-skills-snap-20261018.
Frank's outline restyle: focus-ring red #e01b24 -> #ffb000 -> #ff8c00
(orange), thickness 2 -> 4 in base layout.kdl. The visible ring is
painted by noctalia.kdl, included LAST in config.kdl — its focus-ring
active/urgent colors moved to #ff8c00 too; tab-indicator, insert-hint,
and recent-window highlights deliberately stay red.
noctalia.kdl was hand-written live-only (Oct 14, no repo copy — the
override that silently beat every layout.kdl edit); first durable copy
lands in the zz-blankpad mirror.
Frank veto: DirectoryNotEmpty re-fires on every file event (GUI writes,
wallpaper shuffle) un-debounced -> hundreds of heartbeat spawns/day.
The sweep is cheap (57 ms CPU) but the trigger design was event spam.
Exposure-window fix is an open design question; timer unchanged.
Timer-only heartbeat left GUI dotfiles edits uncommitted for up to 59
minutes; a trunk-sync session clearing the dirty tree on 2026-10-18
destroyed a month of uncommitted noctalia bar layout inside that gap.
live-sync.path (DirectoryNotEmpty on dotfiles/) fires the existing
flock-guarded oneshot within seconds of any repo-side dotfiles write,
shrinking the exposure window from ~1h to ~2s. Timer stays as the
hourly catch-all for the machine tiers.
Documents the 16:46 layout loss (uncommitted GUI state stashed, stash
entry destroyed during trunk sync) and the git-fsck recovery landed in
c87b378. Lesson codified in managed skills.
The rich GUI-built layout (cat/clock/clock-utc/weather start, taskbar
center, capsule groups g1-g4 + control-cheat-sheet end, thickness 36,
font_scale 1.3, opacity 0.9) existed only in the working tree; it was
stashed at 16:46 during the trunk merge and the stash entry was later
destroyed. Recovered from dangling stash commit 7fe3ed8 via git fsck.
This commit is the durable copy the layout never had before.
Uncommitted machine/session edits (live config, models, docs drift) were
blocking the merge of origin/main. Committed as-is to preserve them; the
merge that follows resolves overlapping prose files toward upstream
canonical, with local variants recoverable from this commit. Pre-merge
copies also in .AGENTTRASH/pre-merge-2026-10-18/.
First dogfooding round stamped the build worktree instead of the session's
own working directory (the value the paired session id resolves to). Rule
now says so explicitly. Refs #11
Every agent-opened issue/PR body ends with cwd + omp session id so machine
filed artifacts trace back to the box and session that produced them - the
live-sync daemon and filing agents make 'who/what wrote this' a real
question now. Mirrored in dank/serverAppConfigs by its own PR.
Refs #11
Operator decision at review: 15 min is aggressive for config churn; hourly
keeps the stale-window acceptable (<=1 h) with a fifth of the wakeups.
Timer unit, tier law, and failure-table wording aligned.
The checkout IS the deployment and symlinks make live state (memories,
auto-learned skills, agent config, desktop settings) dirty tracked files
continuously; no commit law covered machine churn, so it accumulated
unpushed for weeks — one disk held the only copy of ~61 managed skills,
a settings fix stayed local-only, and dirty trees refused merges.
scripts/live-sync.sh (systemd user timer, ~15 min) commits ONLY the
allowlisted machine/desktop-live tiers with hooks armed, merges
origin/main, pushes, and reaps wt/* worktrees older than 7 days. Task
dirt outside the allowlist is never staged; conflicts abort and alert
via ntfy; branch guard refuses off main; never --no-verify, never force.
Black-box contract pinned by tests/live-sync.bats (9 tests, suite 36/36).
Refs #11
Trunk-based SDLC gains the delivery habit: xdg-open the PR URL the moment
it's filed (mirrors dank/serverAppConfigs step 6, riding their PR #6).
Changelog entry added.
Single-dev repo: develop bought nothing and cost a back-merge every
release. New law: branch off freshest origin/main per work item, PRs
target main directly, --no-ff merges kept, tags vX.Y.Z straight on main.
No release branches, no back-merges. Changelog entry added.
Census against serverAppConfigs-docs + read-only ssh probes on both boxes
(2026-10-18). Corrections to OURS, flags only for THEIRS:
- Hardware drift: RTX 3090 moved dankserver->dankclaw 2026-09; dankserver
now 2x Radeon AI PRO R9700 (Ryzen 7 7700X stays), 246 GiB RAM, no swap;
dankclaw 121 GiB RAM + 64 GiB swap. Served: frank-archon radiance TP=2
(:7770 dankserver), frank-dragoon halogen NATIVE v2 (:7770 dankclaw
iGPU), frank-zealot llama.cpp PrismML Ternary-Bonsai-2-27B PQ2_0
(:7771 dankclaw, 3090). Kernel 7.0.0-38 both; dankclaw :7770 no auth.
- Gitea: gitea-caddy quadlet retired (fleet Caddy fronts the domain);
data tree moved to /home/dank/Containers/gitea/ (2026-09-01 standard),
/opt/gitea is a do-not-use leftover; LAN IP + live domain.
- pi-era pages (core/settings, pi/models + role tables) demarcated as
ARCHIVE with banners; llama-cpp fleet note modernized.
- AGENTS.md machine-facts table corrected (it still routed prod warnings
at frank-default-large).
Their-side staleness flagged in the changelog entry for the operator,
not edited: OMP Harness thresholdPercent 70 (live is 95), Ubuntu 24.04
OS table, 2026-09-03 stack mapping, superseded pi-KB backlink.
Refs #5
Full-source re-verification of pi/statusline.md against dist/cli.js plus a
live screenshot matrix. False and corrected: claude's top rule is NOT the
context meter (meter is box/band only; rule elsewhere is a session-accent
filler); 'Tern-only' composer-bar claim (real knob: composer.tokenRate);
segmentOptions gated to preset custom (they merge under every preset - only
the arrays are gated); mode segment yolo (states are Plan/Prewalk/Goal/
Vibe/Loop). Expanded: 7-preset defaults table, per-shape statusAttachment/
bottomBar model incl. plugin shape registry, contextLine per-value behavior,
27-segment catalog with triggers and options, segmentOptions schema, full
statusLine.*/composer.* knob registry. Changelog entry added.
Refs #5
Left group keeps the where-am-I identity (host, path, git, pr); the right
cluster carries what's-running (model) plus cost and elapsed time, painted
in array order. context_pct verified to hoist into the inline meter spacer
regardless of array position. Docs + changelog updated.
Band merges the segment groups into one row above the editor and paints
the labeled inline context meter (60% - 200K); the session title rides
the working row. Cumulative token totals live in /session - off the bar.
Docs and changelog updated to the band final state.
stripWorkPrefix true: worktree sessions show projectName/worktreeName with
the worktree glyph instead of the on-disk cwd; matches the live bar tuned
and screenshot-approved 2026-10-18.
Named presets silently ignore leftSegments/rightSegments - the nerd
preset's baked arrays were what we saw, and the configured arrays were
dead. preset: custom with explicit arrays: chip (right) = token_in,
token_out, cost, context_pct, time_spent; strip (left) = session_name,
vim, mode, hostname, model, path, git (branch text off - worktree
branches are autogenerated; dirty badges kept), pr. Cut tok/s duplicate,
clock, cache trio, context_total, subagents, session hash. Composer
shape stays claude after live box/band/rule trials. models.yml display
name frank-archon drops the (dankserver) suffix; id untouched.
Docs: pi/statusline.md records the preset-vs-custom trap, shape->segment
routing, and the context-meter truth.
Mod+R now cycles 1/4 <-> 1/2 only, matching the quarter-based
preset-window-heights; verified live (840px quarter / 1696px half on the
3440px panel). Mod+F still maximizes a column.
- set-column-width / set-window-height nudges 10% -> 5% (Mod+-/= and
Mod+Shift+-/=).
- New preset-window-heights { 1/4, 1/2 } in layout.kdl; Mod+Ctrl+Shift+R
no longer falls back to the width-preset ladder.
- Mod+Ctrl+Left/Right (and H/L) retargeted swap-window-* ->
consume-or-expel-window-*, matching Mod+brackets; swap-window now unbound.
Validated with niri validate; live via hot-reload.
Live census 2026-10-14: every flameshot mode on this box captures via
xdg-desktop-portal-gnome → niri's single-output API (niri#4313) — full
AND screen hang ~15 s and error docked (verified: exit 124, no file).
Replace all six flameshot binds with niri-native equivalents:
all-monitors script (Print/Ctrl+Shift+1/Mod+Shift+S), screenshot-screen
(Alt+Print/Ctrl+Shift+2), interactive screenshot UI (Ctrl+Print/
Ctrl+Shift+3). Absolute script paths everywhere: niri's spawn-sh PATH
is system-only, ~/.local/bin absent — the bare screenshot-desk.sh and
desk.sh binds were silently dead. bats 27/27.
flameshot full dies docked: its portal path hits niri's
screenshot_all_outputs ensure!(outputs.len()==1) (upstream niri#4313,
niri 26.04), so every Print press on the home desk produces error
bubbles. niri's built-in screenshot UI selects only one output, so add
scripts/screenshot-desk.sh: per-monitor focus + screenshot-screen with
explicit --path, focus restored, one PNG per monitor in
~/Pictures/Screenshots. Bind Mod+Shift+S (Frank's call); noctalia
settings moves to Mod+Shift+O. Restore screenshot-path (was null, now
root-level as config.screenshot_path requires) for UI saves. Live-
verified docked: valid 1920x1080 PNGs for eDP-1 + DP-2, focus restored;
bats 27/27 incl. niri validate on both desk profiles.
learned.md: noctalia v5 bar layering + hot-reload loop; secret-gate self-match convention (quoted literals redacted per the convention the line itself states); niri 26.04 event-stream has no hotplug event. New managed skills: git-repo-secret-audit, omp-system-prompt-retrieval.
Imports the Git Workflow / Testing law from dank/serverAppConfigs: pull+clean tree, atomic early commits, bats+secret gate as commit precondition, push on green, imperative messages with Refs #N.
Bar start = launcher, wallpaper, workspaces, taskbar; center keeps only the cat plugin widget. settings.toml is the GUI-override layer, symlinked from this repo.
live store wired 2026-10-09: provider custom -> frank-secondary @
192.168.0.143:7771, prompt default_improve_transcriptions,
post_process on, self-autostart off (niri owns it)
- desk.sh: list/current/switch/cycle/match/capture over cfg/desk-*.kdl
profiles; apply = relink display.kdl, niri validate (rollback on fail),
load-config-file, noctalia notification; match scores // requires: lines
(EDID long names) against live niri msg outputs
- desk-autoswitch.sh: watch niri msg --json event-stream for
OutputsChanged, debounced desk.sh match --apply, reconnect loop
- blankpad-desk-autoswitch.service: ConditionHost=blankpad user unit
- gen-keybinds.sh: keybinds.kdl -> markdown cheat sheet (sections from
// --- headers, hotkey-overlay-title preferred over action)
- tests/: 27 bats cases, all green (desk 16, gen-keybinds 6,
niri-config 5 incl. real niri validate on the assembled tree with
each desk profile active)
Five layers: live-value match across all 77 revs / 765 reachable blobs; the whole
object store including the 2 dangling blobs (one synthetic sk-shaped fixture,
never committed on any ref); 21 provider shapes with every hit triaged; an
entropy scan of HEAD; all 23 git-lfs payloads (81 MiB) scanned as bytes.
Gate gap noted: hf_/npm_/pypi-/AIza/basic-auth-URL shapes are absent from
.githooks/pre-commit. Quoting the fixture literal in this commit message was
rejected by the gate, so it is described instead.
Also chmod 600 ~/.config/secrets/hf_token_write (was world-readable).
git 2.56 credential-store parses one URL per line (port included), not the
key=value block shape the man page shows for ~/.git-credentials; a block-format
file yields a silent empty get and git dies at askpass.
Captured from this session's cutover: reference-grep list, the YAML
cross-check regex (the provider label contains a space and parentheses, so a
whitespace split yields false unresolved hits), the per-endpoint 4-token probe
that proves a harness-side id rename is wire-safe, the fit check against the
served window, and the commit shape.
New tools/gitea/client-access.md: the credential file (~/.config/secrets/gitea,
0600) behind a global 'store --file' helper, the git 2.56 credential-store trap
(the key=value block shape parses to an empty get; the file must be one
http://user:token@host:port URL per line), the GIT_TERMINAL_PROMPT=0
verification loop, the lfs.locksverify key that silences the URL echo, and the
rotation procedure. Cross-linked from the gitea index and management pages.
Both workstation remotes are now bare URLs; the token lives in one file.
Callsign renames verified against the live endpoints (all three accept the
new request ids and return reasoning_content), compaction endpoint moved and
22 model references checked for integrity, runtime scratch quarantined in
.gitignore, skill libraries shown to be renames with history intact at
12c5e88.
The managed-skill library was regenerated: 94 narrow per-incident skills
replaced by 25 broader ones (radiance-*, llamacpp-*, homepage-*, fleet/
endpoint ops). Git renders this as mass delete + mass add because whole
directories changed name; old content stays in history at 12c5e88.
Surface check: the 25 on disk are exactly the skills the harness advertises
this session.
DP-2: transform 90 -> normal with mode 3440x1440@239.999 pinned, moved
(0,0) -> (1581,-600). HDMI-A-1: transform 270 -> 90 (portrait kept, top bezel
now on the left). eDP-1 re-pinned (2400,840) -> (5021,-240) to sit rightmost.
All three positions stay pinned so niri's auto-layout cannot reflow them.
frank-medium -> frank-archon (dankserver), frank-large -> frank-dragoon
(dankclaw-gpu), frank-small -> frank-zealot (dankclaw-3090). The Dankclaw
service/unit names keep the old frank-large spelling; only the OMP model ids
move, per the comments at each entry.
Windows refit to the served context minus the compaction headroom:
archon 196608 -> 200000 ctx, 32768 -> 16384 maxTokens
dragoon 200192 -> 217088 ctx, 57344 -> 45056 maxTokens, input +image
zealot 262144 -> 200704 ctx, 32768 -> 16384 maxTokens
All three now carry compactionModel frank-zealot, reasoningContentField
reasoning_content, and Qwen-recommended sampling in compat.extraBody
(merged last, so it beats the global -1 sampler settings).
autoresearch/ runs, agent/predict/ ngram models, local/ artifact dumps, wt/
worktrees and memories/--home-blank-.omp-wt-* buckets are runtime state keyed
on ephemeral paths; none were tracked, so the rules are purely additive.
Finishes the purge. These two surfaces were left out of the earlier commit
because the memory store is auto-generated, but both are read back into every
future session, so the name was still reaching the agent.
58 sites across 16 files: hostname, both LAN addresses, the SSID. Special-cased
the compound names a blind token swap breaks ('dank@the workstation/dankclaw',
'on the workstation/Dankclaw/Dankserver') and reworded the four that came out
ungrammatical. Facts kept: the hardware inventory of the box, the noctalia
architecture history, and today's own purge lessons all survive intact — only
the identifiers are gone.
Non-memory: the tracked .pre-pull-backup snapshot of a managed skill, whose
live counterpart was already clean.
Second half of the purge (the first covered the persona, AGENTS.md, README.md,
and the distillation of the per-host pages). Every remaining identifier is
gone: hostname, both LAN addresses, and the Wi-Fi SSID.
34 count-checked substitutions across 22 files — indexes, tool/research pages,
the inventory census, the changelog, one generated HTML report, and five
managed skills. Facts kept, names dropped: a history entry keeps its four-page
listing as a nameless summary, the syslog excerpts keep every timestamp and
line, and the census keeps the .zcompdump rows (the zsh version stamps are the
useful part, the hostname was not).
Deliberately unchanged: install.sh's --host case list and the
packages/<hostname>.pkgs filename, where the hostname is the lookup key.
Left out of scope by Frank's call: agent/memories/** — past-session records the
learn pipeline rewrites.
kb-lint: 0 violations. Residual grep for the hostname, its two IPs, and the
SSID across omp-docs/ and agent/managed-skills/: no hits.
Frank's call: no per-host graveyard. The four /system/dankware/ pages carried
two pieces of knowledge that outlive the box, and one that does not.
Distilled:
- omp-docs/system/network.md (new) — LAN topology plus the multi-link resolver
law: a client with two +DefaultRoute links flaps its DNS, and the fix is a
static per-connection resolver, not resolved.conf. Keeps the
flush-caches-loop verification, because one successful dig proves nothing.
- omp-docs/tools/niri/displays.md (new) — the multi-monitor rig: portrait
transforms, absolute positions, auto-placed third panel, the iGPU
render-node pin, and the three failure modes (duplicate output blocks,
positions that do not reflow, a mode the panel cannot drive).
Dropped: the machine's own spec sheet (CPU/GPU/PCI/NVMe inventory for a box
Frank sold) and its quick-facts table. The OS and userspace facts it repeated
already live in tools/cachyos.md.
omp-docs/system/dankware/ moved to .AGENTTRASH/20260928-retired-workstation/
(not deleted; git history holds it too). system/index.md and the niri index
repoint; niri's [[displays]] wikilink now resolves to the new page instead of
the trashed one. kb-lint: 0 violations.
AGENTS.md: the reader-preamble and the machine-facts table row no longer name
the sold client box; the row survives as the 'Dev workstation' class with its
two load-bearing facts (mainline llama.cpp build path, repo at ~/.omp).
README.md: same for the fleet table row and the 'primary host' sentence.
Deliberately untouched, because the name is load-bearing in both:
- scripts/install.sh:97 — the '--host' case list; --host <name> must still
accept this hostname.
- packages/dankware.pkgs:1 — install.sh resolves packages/$HOST.pkgs from
$(hostname) (line 25/44), so the manifest FILENAME is the lookup key.
Renaming it breaks package provisioning on this host.
APPEND_SYSTEM.md line 16 named the sold client box twice (as a fleet server
and as the host I operate from) and, while doing it, mis-listed the fleet as
'Dankware and Dankclaw' -- Dankserver was already missing.
Now: servers Dankclaw and Dankserver, plus several client workstations, and
the identity no longer claims a fixed host. Committed on main rather than a
feature branch because this checkout IS the live deployment (branch switches
swap live config under the running agent) and the tree already carries
in-flight WIP; only this file is staged.
Takes effect in NEW sessions -- the running one keeps its process-start
prompt snapshot.
consume-or-expel-window-left/right on Mod+[ and Mod+],
consume-window-into-column on Mod+, expel-window-from-column on
Mod+.. Moving a window into the next column joins (splits) that
column deterministically instead of relying on the hardcoded
drag-insert heuristic, which spawns a new column first.
- task.agentModelOverrides.scout: "@smol" (frank-small, xhigh thinking)
- task.maxConcurrency: 2 (semaphore on ACTIVE subagents; total
instantiations uncapped)
- compaction.remoteEndpoint -> frank-medium /v1/chat/completions
(192.168.0.235:7770), methodOrder remote-first; compactionModel:
frank-medium on every model so the wire model id is always valid
- models.yml header: single source of truth (pi-era import retired)
Verified: omp config get on all keys; omp -p startup probe passed;
live POST in the remote-compaction wire format answered.
Remap palette indices 1 and 9 to green so error output renders in
terminal green. Explicit palette entries override the generated
noctalia theme (palette-generate = true).
- DP-2 at 1536,0 and HDMI-A-1 at 2736,0, both 1200x1920 portrait, scale 1
- eDP-1 pinned at 0,0 with explicit scale 1.25 (output blocks reset unspecified props)
- niri transform is counter-clockwise: 90 = top bezel left, 270 = top bezel right
- remove 'include noctalia.kdl' from shared config.kdl: the file is a
per-host noctalia template render (colors from local palette), was
never committed, and made every config load fail hard on hosts where
it was absent — niri 26.04 falls back to DEFAULT config on parse
error, silently killing keybinds/rules/layout/autostart
- theme integration now only via keybinds (noctalia msg verbs) + the
host-local rendered files (see managed skill noctalia-template-offline-render)
- sync agent memories: delete retired memory-store skills, add new ones
config.yml: vision -> llamacpp (dankclaw)/frank-large:auto,
tiny -> llamacpp (dankserver)/frank-small:off,
task -> llamacpp (dankclaw)/frank-large:medium (was flm dankclaw-npu),
matching the flm (dankclaw-npu) provider removal in models.yml.
Memories/managed skills: rollouts and learned notes from the frank-large
KV-pool + sampler work and the noctalia recolor; merged the duplicate
noctalia-palette-recolor skill into noctalia-custom-palette-recolor (adds
the dotfiles-tracking step and the +show-config delta gotcha) and dropped
the superseded legacy memory skills whose ghostty USR1 reload recipe is
wrong for this fleet.
Adds custom palette "Gruber Red" (live at
~/.config/noctalia/palettes/Gruber Red.json, symlinked into this repo's
noctalia package). Derived from community "Gruber Darker" with jq:
dark mPrimary/terminal yellow/terminal foreground -> #f43841 (bright
yellow -> #ff4f58), light-mode yellows -> the light error reds. Custom
source, so the community-palette md5 sync cannot revert it.
gtk-3/gtk-4 noctalia.css are regenerated artifacts of templates-apply
(accent_* now red) committed so every host renders the same recolor.
Also carries previously-uncommitted edits already in the working file:
maxTokens 262144 -> 65536 (HALOGEN_MAX_TOKENS_CAP) with the KV-pool
admission note, and removal of the flm (dankclaw-npu) provider block.
Live check: halogen /health reports context=262144, so 200000 is inside
the served window; 200000 + 65536 = 265536 against a 401408-position pool.
Tracked the harness artifacts accumulated on blankpad (blank@), matching
the existing convention for the --home-dank-* memory stores and the 36
already-tracked managed-skills.
managed-skills (9): authentik-proxy-provider-app-link,
dankclaw-guard-chain-layer-diagnosis, e2e-arr-rawip-reachability-triage,
halogen-kv-pool-fit, lan-client-split-dns-fix, noctalia-reference,
rootless-quadlet-log-retrieval, ssh-key-bootstrap-no-sshpass,
swapfile-resize-safe-ordering.
memories (2 stores): --home-blank-.omp-- (dotfiles_git, handy_lifecycle,
input_permissions) and --home-blank-Projects-serverAppConfig--
(frank_large_quadlet).
Secret sweep before staging: no PEM blocks, no credential literals.
Content is fleet topology/IPs only, same class as the tracked dank stores.
halogen-flash-server has no server-side sampler defaults (temperature absent
or 0 = greedy), so the endpoint decoded greedy for OMP, whose global
sampling settings default to -1 (not sent).
compat.extraBody merges these top-level into every request body after the
compat policy, so they win:
temperature=1.0 top_p=0.95 top_k=20 min_p=0.0
presence_penalty=0.0 repetition_penalty=1.0 preserve_thinking=true
Verified against the live dankclaw endpoint: preserve_thinking is honoured
(350 vs 75 prompt tokens for a replayed reasoning turn), and the exact
body shape returns 200.
Frank