Answers spoken with /say were verbatim markdown-cleaned text while the
built-in vocalizer speaks prose rewritten by a small model. Port that
enhancer (bundle Qyr/T6o: same prompt, 1536 tokens, 6 s, 4000-char
window, null-fallback/empty-skip contract) and resolve its model from
modelRoles.tiny exactly as the harness does, so auto-speak sounds like
the enhanced TUI voice instead of a pager. agent_end now also speaks the
last assistant message's thinking blocks as their own unit before the
answer; narration-flagged thinking counts as answer text. Both knobs are
menu-toggleable, persisted, and env-overridable; the fixture gains a
chat route (echo/prose) and seven scenarios cover the round-trip, the
404 fallback, ordering, the kill switches and role resolution (22 green,
58 suite-wide; live smoke through real frank-zealot + frank-tts).
Refs #27
agent_end awaited one whole-answer synthesis; long answers hit the
harness 30s extension-handler timeout (handler timed out after 30000ms)
and the single frank-tts request was the stall. Port the built-in
speak-aloud vocalizer's utterance segmenter (fHe in dist/cli.js) so
segmentation matches the known-working stock path: [.!?…]-run
boundaries, abbreviation/decimal/inline-code guards, fenced code and
tables swallowed, markup cleaned. Play segment N to player exit before
synthesizing N+1 (audio never stacks); fire-and-forget the chain so
utterances outlive the handler timeout; a run token retires superseded
chains. DANKVOICE_MAX becomes a per-answer budget of whole segments;
stripForSpeech is gone. Refs #23 follow-up.
Bare /say and /say menu open a ctx.ui.select settings surface: auto-speak
toggle, voice from the frank-tts roster, PipeWire output device
(pw-play --target / paplay --device=), player binary, spoken cap, and a
test phrase. Choices persist to ~/.omp/cache/dankvoice.json (env wins
over file); detectPlayer becomes availablePlayers + playerFor. Fixes
the silent-route failure mode where synthesized audio played into an
offline Bluetooth sink. 15/15 dankvoice bats, 51/51 suite, live pty
TUI smoke armed the footer and wrote the settings file.
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.
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/.
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.
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.
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.
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).
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.
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.
- 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.
- 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.
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.
Memory-file flush for the dotfiles-deploy session with the blankpad admin
credential scrubbed to a placeholder; sessions transcripts untracked
(gitignored) since harness transcripts can carry raw secrets.
Managed skills: command-center-dotfiles-import, ssh-fish-host-remote-exec,
stash-supersession-audit-before-drop.
Frank
dank