User decisions (2026-11-10): the game-mode minimum on a 16 GB card is the
2 GB expert placement, the normal placement leaves the STT its share by
formula instead of Strata's automatic placement, and STT partials slow
down while the LLM answers with finals handled at once. Also notes the
user's report of occasional, minor frame drops during the live run.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The gamemode_stt run showed Strata's automatic placement filling the card
next to the STT (15.9 of 16.3 GB, first token 1.6 -> 7.5 s) and the STT's
back-to-back partial passes cutting LLM output to a third. The new
gamemode_tune stage measures both fixes the user chose:
- stt_room: one start without experts (STT loaded) gives the LLM's base
VRAM; --expert-vram gets what is left after the STT's ten-speaker
growth (from this PC's stt stage) and a 768 MiB margin.
- stt_pace: a ten-speaker partial pass every 2 s, or none, with finals
every 0.5 s and finals first; records final wait and STT GPU share.
run_m0_gamemode_tune.bat runs it unattended.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The voice-start figures add medians from different runs and leave out
the TTS, the VRAM-full cause of the 7.5 s first token is a hypothesis,
and the 2 GB placement is 2.1x faster with the STT on the GPU but 1.5x
with it on the CPU.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
gamemode (third run), gamemode_stt and gamemode_live finished on the
game PC. The STT on the GPU takes 2.7 GB at its 10-speaker peak, as
estimated, and leaves the game 3.6 GB (experts 0 GB) or 2.6 GB (2 GB).
The STT stays on the GPU in games: on the CPU one utterance takes
3.0-3.8 s. In a real game the 2 GB expert placement is twice as fast as
0 GB. The GPU contention between the STT and the LLM, the in-game floor
placement and an STT margin for the normal placement wait for the user.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A worker that did not stop in time returned its earlier "loaded" note, was
killed, and the placement still counted as ok with no transcription times.
finish() now reports a timeout, and a worker that timed out, died, never
started or measured no ten-speaker or single-clip timing is an error, so
the stage is partial and the next run retries it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The user meant the chat channel's "answer every message", not the voice
policy "in order": the LLM takes two or more requests at once only while a
chat channel answers every message, otherwise one whatever the voice policy.
The roadmap and the open measurements point at the new gamemode_stt and
gamemode_live stages, which also decide whether the STT moves to the CPU
during a game.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
gamemode now also starts the 2 GB expert placement with a 16K context and
drops the --kv q4_0 and --mtp off rows the Strata 0.1.59 qwen36 engine
ignores; the normal placement answers one request at a time.
gamemode_stt loads the STT worker (stt.py mode=load) before the LLM picks its
placement and keeps it transcribing ten speakers at once while the LLM is
measured, so the VRAM left is the real game share. The game placements run
with the STT on the GPU and on the CPU. gamemode_live repeats the game rows
while the user runs a real game; it only runs when asked for
(run_m0_game_live.bat) and is skipped when nobody is at the PC.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The M0 stt stage measured 1300 MiB for the model and another 1440 MiB while
transcribing ten speakers at once. The game share now subtracts the 2.7 GB
peak: about 3.1 GB on the 16 GB game PC (not 4.6 GB), and the 2 GB expert
placement returns at a reserve of about 2.2 GB. Both are marked as sums of
separate LLM and STT runs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The user chose one request at a time by default (97 tok/s on a 16 GB card
against 70-75 with two) and two or more only under the "in order" policy,
which answers every request. The second gamemode run showed a 16 GB card
cannot free 8 GB for a game: the smallest placement still leaves 6.3 GB.
The Strata 0.1.59 qwen36 engine source confirms why --kv and --mtp off made
no difference: it takes no --kv flag and always loads the MTP block. The
game placement is now one request, CPU vision, 8 workers, 16K context and a
computed expert share, falling back to the minimum placement on this PC.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first game PC run showed that a 2 GB expert budget still left a 16 GB
card only 4.3 GB free: the weights, the 32K int8 conversation memory, the
MTP drafter and buffers take about 8.7 GB on their own. The game placements
now also shrink the memory format (q4_0), the context (16K, 8K) and the
drafter, with half the CPU workers (no speed cost measured). The engine log
is appended across starts, so each placement now reads only its own lines;
when no placement leaves the reserve the result names the most free one.
The gamemode runner skips the desktop stage, which is already measured.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The gamemode stage was "ok" when only the normal placement started; a
placement that does not start, or a measurement error inside one, now marks
the stage partial. The design no longer raises the expert budget above the
computed one: it shrinks the game reserve to what the smallest placement
leaves, shows that, and pauses the LLM during games when nothing is left.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Drop the recommended model (Qwen3.8-Flash-Next) and the experimental one:
Qwen3.6-35B-A3B passed every tool call and answers first. Replace game-mode
actions (a)(b)(c) with automatic tuning: a normal and a game placement of the
same model, computed per PC from VRAM, a first-run self-check that lowers
defaults on slow PCs, and latency watching that lowers restart-free settings
first. Record the agent-desktop results and choose the per-call CLI.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add a gamemode stage that restarts Qwen3.6-35B-A3B with a smaller routed
expert VRAM budget (6, 4, 2 GiB, and 2 GiB with half the CPU workers), then
the normal placement again, and records switch time, free VRAM, output speed,
first spoken sentence and 8K-context first token for each. It names the
fastest placement that leaves the game its 8 GB reserve. Strata 0.1.59's
qwen36 engine cannot shrink VRAM while running, so each placement is a restart.
The desktop stage now matches running apps by the process image name Windows
lists (Discord.exe) and closes the calculator it launched by image name when
close-app cannot find a UWP app.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
check_mcp timed five CLI and five MCP list-windows calls but ignored their
answers, so a server returning isError/ok=false still scored ok. Read the
envelope of every call, record the error code (never the message), and
fail the check on any failed call.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The gaming-PC rerun scored the fast model's honest refusals ('...할 수 있는
도구는 없습니다', '...준비되어 있지 않습니다') as failures, so the kit
downloaded two more fast-model candidates for nothing. Rescored with the
fixed judge, UD-Q4_K_XL, UD-Q6_K_XL and Q8_0 all pass 27/27; the
recommended model's Q2_0/IQ3_XXS/IQ3_S still fail 3/27 by calling an
unrelated tool when no tool fits. Record both results in the v2 docs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add a 'desktop' stage to the M0 kit that installs the pinned agent-desktop
1.2.0 Windows build (SHA-256 checked) and checks, on the gaming PC:
calculator buttons by automation id (12+3=15), set-value and read-back on
a test page in a throwaway browser profile, Discord/Explorer read time,
and a warm MCP server against one CLI process per call. Results keep only
counts, timings, error codes and the calculator digits.
Record agent-desktop (UI Automation) as the planned M7 control path in the
v2 design docs, with raw Windows input only for screens without
accessibility.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The temperature and minute checks dropped the sign and the decimal part,
so "13.9도" and "-13도" passed as 13 degrees. Numbers are now read whole
("-", "영하" and decimals included) and must equal the tool result.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The tools stage decides whether to keep a model or measure the next one,
so a made-up answer must never pass. Each case now has a judge for the
spoken answer: numbers must match exactly with their unit (15 or 50
minutes fail a 5-minute timer), weather answers must agree with the rain
in the result and give no other temperature, and a refusal fails when it
also says the thing was done. A new case asks for a timer with no timer
tool offered, the shape Jarvis has today. Unit tests cover the reviewer's
counterexamples and that a lying model makes the stage move on to the
next candidate.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When the skipped-tool guard could not run the planned call (not a literal
call, or arguments the tool does not take) it delivered the answer it had
held back, which was given without the tool and could be a made-up value.
The planned call is now recorded as a failed tool result and the model
answers from that, so the user hears that it could not check or do it.
M0 kit: a case passes only when the final answer uses the tool result
(13 degrees, not a made-up 24) or, for the no-tool request, says it
cannot do it; the no-tool request now counts. The kit mirrors the
engine's failed-result path, and a failing model moves on to the next
candidate on these stricter checks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
User rule: if the tool-calling fix does not work, change the model. The
tools_fast / tools_recommended stages now measure the next, less
compressed model of the same family (fast: UD-Q6_K_XL then Q8_0;
recommended: IQ3_XXS then IQ3_S) when the default still gets a scored
case wrong, stop at the first one that passes every case, and record it
as tools_verdict.chosen. Candidates without enough free disk are skipped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
M0 on Strata caught both v2 models skipping a tool they had been given:
the recommended model made up "24 degrees, sunny" instead of calling the
weather tool, and the fast model said "timer set" without calling the
timer tool (3 of 3 each). qwen2.5:7b does the same through the engine.
- Prompt: TOOL_HONESTY_RULES in the persona (live facts only from a tool
result this turn, actions only claimed after a tool reports success, no
answer text next to a tool call), restated at the end of the system
message by tool_honesty_reminder() with the turn's real tool names.
- Engine: skipped-tool guard. When a native-tool model answers before any
tool has run and the plan names a callable tool as a concrete call, the
answer is dropped and the plan steps are direct-executed, then the model
answers from the results. Only literal, schema-valid steps run (regex
fast path, no LLM repair); otherwise the held-back answer is delivered.
- Planner: the AVAILABLE TOOLS lines now carry argument names
(planner_tool_catalog), so concrete steps use the tool's real keys.
- Eval evals/test_tool_skip_confabulation.py; on qwen2.5:7b, 3 runs each,
available-tool cases went from 0/12 to 9/12 calling the tool.
- M0 kit: tools_fast / tools_recommended stages (run_m0_tools.bat) rerun
tool calling before and after the fix on the real models, loading the
production system_prompt.py and planner.py.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Keep the fast model as the default: the recommended model decodes faster
(131 vs 82 tok/s) but starts every voice answer, long-context reply and
image answer later. Note its working VRAM shrink for game mode and its
fabricated weather answer for the M3 evals.
A named share question now captures the popout straight away when that
person is sharing, even with several sharers (user decision).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Only capture a named person's share when they are the sole streamer or
the user has confirmed the selection in the popout. A timeout no longer
captures whatever is showing, a non-streaming target gets an honest
answer, and clicking the popout after confirmation asks again.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Discord allows only one popped-out share window and picks the stream
inside it, so there is never a choice between windows. Use the voice
channel's streaming state instead: one streamer means capture directly,
two or more means asking the user to pick that person in the popout
when the requested person changes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first run left three results out of the plan: the fast model skipped
the timer tool and claimed it was set, the popped-out share window carries
the voice channel name rather than the streamer's, and a covered popout
keeps drawing. The recommended model failed to start (DFlash2 needs a
native pack), so M0 stays open for that one stage with the fixed kit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The user wants the app to work on any Windows PC from its own installer,
so the Parsec driver is a test aid for M0, not a product dependency.
Scope the target to x64 Windows 10/11 outside S mode (ARM64 needs WHQL)
and re-verify the Discord share, cursor and unplug behaviour with the
driver that ships.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The user wants the bot window on no real screen at all. Add a vdisplay
stage that drives the Parsec Virtual Display Adapter already on the PC
(no install, the display dies with the script), moves the clock window
onto it as a normal and as a tool window, and asks about real monitors,
Alt+Tab, Discord's Screens tab and whether the cursor can wander there.
Record the M0 defaults the user confirmed in docs/v2.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Strata 0.1.59 gives Qwen3.8-Flash-Next Q2_0 the DFlash2 drafter alone
(--draft-fallback none), but DFlash2 needs a native pack and the canonical
Q2_0 pack is not one, so the engine stops at start ("DFlash2 is off here:
it needs a native pack ... the start ends here"). Ask for the MTP drafter
instead; Strata falls back to no drafter when the MTP layer is missing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first gaming-PC run failed the recommended model with only Strata's
one-line reason ("no CUDA-capable device"), which is the picture
reader's side message, not the cause. The engine log stayed on the PC.
A launch that does not become ready now copies the newest
*.engine.log (scrubbed, last 200 KB) into logs/, which the results zip
already packs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The context test re-estimated characters per token from the server's
prompt_tokens after every request. A server that reports a tiny count
made each next prompt grow by orders of magnitude, so the kit tests
reached 8 GB and were killed by the host memory guard. Clamp the
estimate to 0.5-4 characters per token.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- A Strata install or start that never reached a server is saved as
"failed" (not "partial"), so the next run retries the stage.
- A launch that exits non-zero is not ready, even if something answers
on the port afterwards.
- Before the cold start, the stage stops if another program already
listens on port 8080 and tells the user to close it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
One double-click kit (scripts/m0/run_m0.bat) that measures the gaming PC
and packs the results into a single zip:
- PC info (GPU, driver, RAM, disks, monitors)
- Windows checks with the user: bot browser window hidden from Alt+Tab
but listed in Discord's share picker and still drawing while covered;
Discord pop-out stream window identification and occlusion; installed
games, anti-cheat services and the borderless game detector
- MeloTTS, CosyVoice2 and Fun-CosyVoice3 Korean synthesis speed
- faster-whisper large-v3-turbo: final/partial latency, CER, wake-word
hint, batched decoding for several speakers
- Strata LLM (fast and recommended presets): TTFT by context length,
decode speed, tool calls, picture reading per quality tier, Korean
answers, parallel requests, VRAM give-back and restart/switch timings
uv and Python, per-stage venvs, models and Strata all live inside the
work folder. Results are scrubbed of user/computer names and home paths.
Verified on Linux GPU (LLM client against Ollama, STT, MeloTTS,
CosyVoice2/3) and on the Windows 11 CI VM (window checks, game scan,
MeloTTS/CosyVoice2/STT installs, Strata install and plan).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Chat channel name/description edits also rename the real Discord
channel and topic by default (toggle kept)
- Memory: separate toggles for registered chat channels and other
speakers in the same voice channel, both on by default
- Screen-share popout is requested by voice, never automated
- v1 head bb60c4d preserved as the remote legacy branch
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Replace the root README with the v2 plan (game PC fully local, Windows
management app) and split the plan into docs/v2/*.md: architecture,
voice-and-llm, conversation, memory, desktop-app, pc-features, privacy,
roadmap, decisions. The v1 README moves to docs/v1/README.md with its
relative links fixed; .env.example points to the new location.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Switch the default TTS engine from online Edge TTS to offline GPU
MeloTTS (privacy-first: reply text never leaves the host). Adds a
supervisord melo-worker program that runs in full/bot/brain roles and
synthesises Korean on the GPU (MELO_DEVICE=cuda, MELO_SPEED=1.5), a
baked /opt/melo py3.11 venv layer in the Dockerfile, and compose/env
defaults. Edge TTS stays available as an opt-in online fallback via
TTS_ENGINE=edge.
Verified on the RTX 5050 (Blackwell sm_120): melo-worker logs
"ready (lang=KR speed=1.5 device=cuda)", /health returns ok, and a
/synth round-trip returns a 177KB WAV.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Run the GPU-heavy compute (Whisper STT, Ollama LLM, TTS) on a dedicated
brain host and the bot + browser on a separate no-GPU app host, talking
over the LAN.
- supervisord: new `brain` role (bridge only) and `app` role (bot +
desktop/Chrome, no local bridge); `full`/`browser`/`bot` unchanged.
- compose: make BRIDGE_URL overridable (was hardcoded to loopback, which
prevented the bot from reaching a remote bridge).
- docs: DEPLOY.md "GPU split" layout with per-host .env blocks; document
the new roles in run-if-role.sh and .env.example.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
For the split where the LLM runs on a separate LAN host (e.g. a GPU box at
192.168.10.9) and the app stack runs elsewhere (192.168.10.5), the brain must
reach Ollama over the network. The javis service hardcoded
OLLAMA_BASE_URL=http://ollama:11434, ignoring any .env value, so the app could
only use the in-stack Ollama. Make it ${OLLAMA_BASE_URL:-http://ollama:11434}:
default unchanged (all-in-one), overridable to point at an external Ollama.
Document the external-Ollama setup in .env.example and DEPLOY.md, refresh the
browser-host IP examples (.9→.5), and de-hardcode the .9 example in novnc.ts.
moveMouse returned ok:true even when humanHover did nothing (no on-screen box)
or the selector never matched — recreating the "claims it moved but didn't"
bug. Now: humanHover returns a boolean (and brings the element into view first);
moveMouse returns ok:false when the target isn't found or has no on-screen box,
and when site=naver/... whose box isn't on the current page it navigates to the
site home first before hovering. search now reports input=human|api-fallback|api
so a silent fallback to cursor-less DOM input is visible, and the tool surfaces
that note in the reply instead of implying a human-like search happened.
The tool had no cursor-move/hover action, so "move the mouse to the search box"
had nothing to call and a weak model just claimed it had moved it. Add a
moveMouse action wired to human.humanHover (real xdotool cursor), targetable by
CSS selector or site= (that site's search box). Also clarify in the tool
description that 'navigate' types into the address bar (no mouse) while
'search'/'type' move the real cursor to the on-page box and click before
typing, so the model picks the visible-cursor path when the user wants it.
detect_model_size only matched 1b/3b/7b, so a genuinely small model such as
qwen*:4b fell through to LARGE and got the terse, less-guided prompt set it
can't follow — contributing to off-tone, rambling replies. Extend the small
patterns to cover all sub-8B sizes (1b-7b) across :/-/_ separators and sync the
spec table.
A 4b model replied to "하이" with "테니스 연습을 Trenches Gym에서..." — it copied
the literal "box at Trenches Gym" few-shot example embedded in the persona
prompt and mangled boxing into tennis, presenting a prompt example as if it
were a real user fact. Remove the copyable proper-noun example and add an
explicit guard: use ONLY names/places/activities that literally appear in the
memory section, never borrow them from the instructions or example wording.
A deployment had OLLAMA_INTENT_MODEL=qwen2.5:7b while the chat model was a 4b,
so every auxiliary call (intent judge, tool router, place extraction, query
decomposition) ran on the bigger, slower model and added latency to each
command. Make the .env.example comment state the invariant explicitly.
Replies stayed long because the prompt stack gave conflicting length signals:
the persona said "one sentence (two at the very most)" AND told the model to
"state the answer in a sentence, then add a dry observation" (a 2nd sentence),
while voice_style said "two to three sentences maximum". The model followed the
longest. Make all three sources agree on exactly one sentence: the persona's
aside must now fold into the same sentence as a trailing clause (never a 2nd
sentence), voice_style caps at one sentence, and agents/llm.md says 한 문장.
Shorter replies also cut Edge-TTS latency, since synth time scales with text
length. Specs (prompts.spec.md) and docs/llm_contexts.md updated; deterministic
prompt-contract tests added.
Whisper was decoding with beam_size=1 (greedy), the least accurate setting,
which hurt recognition on short/accented/noisy Discord-mic speech. Switch the
default to beam search (5, Whisper's own default) and stop conditioning on the
previous clip's transcript (which causes repetition/drift on isolated short
utterances rather than helping). Both are env-tunable (STT_BEAM_SIZE,
STT_CONDITION_ON_PREV) so accuracy/latency can be traded without a code change;
wired into docker-compose and documented in .env.example.
Follow-up to the OLLAMA_INTENT_MODEL split: document that the Docker stack runs
intent judging / tool routing / extraction on a small qwen2.5:3b (pulled by
ollama-init) kept separate from the big chat answer model, and that setting
OLLAMA_INTENT_MODEL = OLLAMA_CHAT_MODEL folds them back onto one resident model.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Intent judging, tool routing and arg extraction are classification/JSON calls,
not the spoken answer, yet the stack aliased OLLAMA_INTENT_MODEL back to the big
chat model — so each command paid the big model's cost 2-3 times for routing
before the reply even ran. With the GPU on, that round-trip stacking is the main
remaining per-turn latency. Default OLLAMA_INTENT_MODEL to qwen2.5:3b (the
project's reference small model, clean Korean on classification) and have
ollama-init pull it. The reply engine already routes these calls through
intent_judge_model, so answer quality is untouched; set OLLAMA_INTENT_MODEL =
OLLAMA_CHAT_MODEL to fold back onto one resident model.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
agents/llm.md promises "play the Nth video from the top", but browseAndPlay
only ever clicked the first result. Add an optional 1-based index argument
(default 1, backward-compatible) threaded to the Node helper, which now clicks
the Nth a#video-title and clamps to the number of results returned so asking
beyond the list plays the last available video instead of failing.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Spoken (TTS) replies are 1-2 sentences, so an unbounded num_predict only
exposes the worst case where the chat model rambles or loops. Add an
ollama_num_predict config (default 512, 0 disables) wired into the reply
loop's chat call on both the native- and text-tool paths. The 512-token
headroom stays well above this app's short tool-call JSON, so capping never
truncates a tool call. This keeps the user's quality model instead of
downgrading it. Configurable in the container via OLLAMA_NUM_PREDICT.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- run-chrome.sh: render web content (YouTube/Google/Naver) in Korean. --lang
only sets Chrome's UI; the Accept-Language sent to sites comes from the
profile's intl.accept_languages, which a persisted user-data-dir kept at
en-US. Seed the profile to ko-KR and add --accept-lang=ko-KR,ko.
- agents/llm.md: runtime instructions for the reply LLM (loaded by the agents
feature) — name "자비스", concise 1-2 sentence TTS replies (no emojis/markdown),
nuance-based intent, and YouTube voice controls (open/search/play Nth/pause/
back) via the on-screen browser tool.
- CLAUDE.md: drop the "use emojis in CLI output" rule — this assistant replies by
Discord voice, not CLI, so output should be plain.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>