- dashboard/handlers/files.py: handle_files_download() servește mp3/wav/zip
din WORKSPACE_DIR cu Content-Disposition attachment, resolve dedicat
(nu _resolve_sandboxed, care nu ajunge niciodată la WORKSPACE_DIR)
- dashboard/api.py: rutează /api/files/download
- tools/epub_to_audio.py: tool nou de conversie EPUB → audio
- cron/jobs.json, memory/kb/index.json: stare auto-generată (job runs,
regenerare index KB)
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Fallback-ul local (Qwen3.5-2B, LXC 104) raspundea ca nu are unelte sau
internet. Acum conversează, are unelte read-only si istoric per canal.
Unelte (allowlist strict read-only, src/local_fallback_tools.py):
- cauta_web — DuckDuckGo Lite, fara API key
- masini — status Proxmox/LXC prin SSH paralel (src/net_status.py)
- vremea, sold, facturi, trezorerie, doctor, logs, email, kb,
cauta_memorie, citeste_pagina
Protocol: function calling OpenAI nativ. Protocolul text anterior
("TOOL: nume arg") pierdea argumentul in 100% din cazuri.
Uneltele cu output deja formatat se intorc verbatim — modelul de 2B
transforma un `doctor` cu 5 linii OK in "Sistemul este in 5/5 state.".
Doua treceri: prima decide uneltele (cu tools, temp 0), a doua reface
turnul fara ele — prezenta definitiilor strica raspunsurile simple
("cat fac 128/4?" -> "Nu stiu ce inseamna 128/4"). Few-shot doar pe
cereri creative: pe turnuri factuale strica aritmetica (17*23 -> 471).
Ocoliri deterministe pentru ce modelul greseste reproductibil: vremea si
preturile live (raspundea din memorie), plus sarcini "instructiune: text"
(payload-ul dicta unealta — "scrie mai politicos: da-mi raportul acum"
chema `sold`). llama.cpp nu e determinist nici la temperatura 0 si
ignora tool_choice, deci promptul singur nu ajunge.
Masurat: 36/36 pe set combinat (17 alegere unealta + 19 comprehensiune),
de la 14/19 pe comprehensiune inainte de regulile negative din prompt.
Comenzi: /f (inlocuieste /testfallback, ramas alias), /f reset, /masini.
Prefixul "Claude e la limita" apare doar pe calea de rate limit.
roa2web: /otp <cod> pentru 2FA din chat — mesajul de eroare trimitea la
verify_2fa(code, email), imposibil de rulat din Discord.
Corectat memory/kb/tools/infrastructure.md: LXC 101 si 110 sunt pe pve1.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B88u3KkzvUNnzU56y7GBn2
Markerul opac [tts-lang:en] depindea de regula din VOICE_MODE.md, aflată
la ~30k caractere distanță în system prompt — modelul o rata pe ~1 din 5
turnuri. Răspunsul în română era apoi respins integral de pocket-tts
(English-only) și turnul rămânea fără audio.
- src/router.py: markerul poartă instrucțiunea inline (un singur token,
acoperit în continuare de strip-ul anti-jailbreak); 5/5 EN la repro
- src/voice/tts_stream.py: dacă modelul tot scapă română pe o voce
pockettts, blocul cade pe Supertonic (M2, ro) în loc de tăcere;
coada TTS transportă (clause, voice, lang) per clauză
- tools/tts.py: looks_romanian public (folosit de tts_stream)
- personality/VOICE_MODE.md: regula gated pe prefixul [tts-lang:en
- tests: 4 teste noi fallback + expectații marker actualizate
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Două cauze pentru audio mut pe pocket-tts după fix-ul [tts-lang:en]:
1. normalize_for_tts expanda numerele/orele în cuvinte românești chiar și
în text englezesc ("It's 9:46 PM" → "nouă și patruzeci și șase de
minute") → diacritice → pocket-tts respingea clauzele. push_text rezolvă
acum lang din engine-ul vocii curente (engine_for_voice) și sare
expansiunile RO când lang != 'ro'. [scris de Echo în sesiunea Discord]
2. sanitize_for_supertonic (cap 400 chars — limită ONNX per request) rula
pe textul întreg în normalize, înainte de segmentarea în clauze —
răspunsurile lungi erau retezate mid-word și trunchierea la 200 cuvinte
cu sufixul "Restul l-am scris în chat" nu se mai aplica. Separat
map_tts_punctuation (fără cap, folosit de normalize) de
sanitize_for_supertonic (cap, aplicat în _synthesize_supertonic).
Include vocea clonată "Marius 4" (tts_voices.json + default în config).
Fixează test_truncate_exactly_200_words_unchanged și
test_truncate_over_200_words_appends_suffix.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Instrucțiunea „engleză pentru audio" nu ajungea la model: VOICE_MODE.md nu
avea nicio regulă de limbă, iar condiția din TOOLS.md („când generez pentru
o voce pockettts") era neevaluabilă — modelul nu știa ce voce e activă.
Rezultat: mirroring pe română → filtrul de diacritice bloca sinteza → tăcere.
- tools/tts.py: helper public engine_for_voice() (catalog → tts.default_engine)
- src/router.py: injectează [tts-lang:en] în prefixul voice când vocea activă
e pockettts (config citit fresh — vede swap-ul live /voice setvoice);
anti-jailbreak strip extins la [tts-lang:...]
- personality/VOICE_MODE.md: regulă de limbă gated pe marker
- tests/test_router.py: 5 teste noi pentru marker + strip
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Kyutai pocket-tts nu suportă română (vezi eval anterior). Codul rula
oricum sinteza pe orice text primit; acum _synthesize_pockettts
respinge diacriticele RO înainte de request, fără fallback silențios
pe altă voce.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- synthesize() rezolvă engine din tts_voices.json (fallback la tts.default_engine din config)
- _synthesize_pockettts nou: POST multipart la :7789/tts cu voice_wav/voice_url, ignoră lang
- fallback tehnic automat pe _synthesize_supertonic(M2, ro) via _PocketTTSUnavailable; erori de
conținut (voce lipsă, text gol) nu declanșează fallback; rezultatul include engine_used
- gates rulate: tests PASS (1043 passed, 22 eșecuri preexistente neschimbate), /workflow:simplify
(4 agenți) cu 1 fix aplicat, /review (backend, manual) PASS
Pre-existing work committed before starting Ralph self-improvement run on
ralph/echo-improve branch, so that branch's diff stays isolated to the
pocket-tts integration.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Jobul intern rula pe zile greșite (vineri/sâmbătă/marți în loc de
joi/vineri/luni) din cauza convenției diferite de day-of-week dintre
APScheduler și crontab standard. Scriptul standalone via OS crontab
e mecanismul corect și funcțional — rămâne singura sursă.
chore: sync stare working-tree — cron, KB notes, anaf-monitor snapshots
Pre-existing uncommitted changes swept in with the STT work:
anaf-monitor snapshots/versions, cron job + newsletter state, 9 youtube KB
notes, tools/ocr_bon.py, and tools/tts.py.
Note: the tts.py change breaks 2 truncation tests in test_voice_normalize.py
(sanitize word-count) — flagged for a separate follow-up.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Gemma 4 cloud audio was infeasible (31b-cloud has no audio; E4B broken
upstream, no deploy host), so improve faster-whisper instead.
- Pin temperature=0.0 to disable the fallback ladder that re-decoded unclear
audio up to 6x (source of the 16-24s latency outliers); reject hallucinated
segments via avg_logprob/compression_ratio in the new pure _filter_segments.
- Adopt mikr/whisper-small-ro-cv11 (CT2 int8) via configurable voice.stt_model:
spike showed WER 24%->10%, numbers fixed at source, +0.33s p50 (in budget).
- Add tools/voice_stt_mine.py (log mining) + tools/voice_stt_spike.py (model
eval with diacritic scoring) + tests for the gate and miner.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Expose a navigation layer to the agent and harden RAG, after analyzing the
OKF note and testing on the real KB.
- memory_search.search(): dedupe best-chunk-per-file (a relevant note can no
longer be buried by another file's chunks) + keyword fallback tagged
degraded:True when Ollama is unreachable (no more hard crash).
- update_notes_index.py: emit per-folder index.md + root router; prune empty
folders; fix latent subcategory->project bug.
- Exclude generated index.md from RAG rglob (reindex/incremental) + indexer
scans + heartbeat freshness check (prevents self-pollution / reindex thrash).
- CLAUDE.md: reframe memory as hybrid (navigation first, RAG for fuzzy recall).
- Delete stale orphan kb/youtube/index.json; correct the OKF source note.
- Tests: dedup, keyword fallback, index.md exclusion. Plan + review in docs/.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- tools/youtube_subs.py: get_subtitles() returneaza acum (title, desc, transcript).
Functii noi is_description_about_video() si extract_relevant_description()
detecteaza daca descrierea contine capitole/timestamps (nu doar promotie autori)
si curata trailing-urile promotionale inainte sa includa descrierea in output.
- dashboard/handlers/youtube.py: aceleasi functii adaugate; nota KB generata
include acum un bloc "Descriere / Index" daca descrierea e relevanta pentru video.
- memory/kb/youtube: nota Jeremy Grantham (AI bubble, investitii, toxicitate)
cu descrierea ca index de capitole.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Gmail preserves the original sender when forwarding — whitelist check
was blocking all Fwd: emails not from mmarius28@gmail.com.
echo@romfast.ro is private, so any Fwd: arriving there is from Marius.
Also strip ***SPAM*** prefix from slugs for cleaner filenames.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Squashed branch: voice/dave-recv → master. Closes Pas 12 (DAVE E2E) and lands
voice-mode UX polish + verbal voice control on top of the Pas 1-10 scaffolding
already on master.
## DAVE E2E receive-side decrypt (e4f3177)
Vendored fork: discord-ext-voice-recv 0.5.3a+echo.dave1. Patches the receive
pipeline to handle Discord's mandatory DAVE encryption on voice gateway v=8.
- `_maybe_dave_decrypt`: uses davey.can_passthrough(user_id) as primary gate,
falls through to dave.decrypt for DAVE-epoch peers, drops on decrypt failure
without killing the reader thread.
- VAD fix: silero-vad v5+ requires exactly 512 samples; our 100ms window
(1600 samples) was silently raising ValueError → STT never fired. Now slice
into 512-sample chunks.
- Whisper: bumped beam_size 1→5 and added RO initial_prompt.
- Tests: 11 DAVE unit tests + 2 callback integration tests + contract test
with fork-version guard.
## Voice UX polish (d1bc77e)
- Killed the 3s "mă gândesc" filler (always collided with Claude p50 4-7s).
- Barge-in via `ttsq.clear()` at top of `on_segment_done`.
- DTX silence-flush poller (200ms tick) — Discord stops sending RTP packets
when silent, so the inline silence-check in sink.write() never fired for
trailing audio; background thread handles it.
- `EchoStreamingAudioSource.read()` non-blocking — old `get_frame(timeout=0.1)`
wrecked Discord's 20ms cadence and the client interpreted bursts as
stuttering (Marius heard "4 de minute" instead of full sentence).
- RO time expansion: 23:09 → "douăzeci și trei și nouă minute".
- Supertonic Unicode sanitize centralized in tools/tts.py.
- Whisper local_files_only=True — no HF metadata GET on each startup.
- Diagnostic logging through sink → VAD → Claude stream → TTS chain.
## Voice mode iteration (e589e48)
- `personality/VOICE_MODE.md` — voice-tailored system prompt (short, no
markdown, no abbreviations, time without seconds, distances in
"mii"/"milioane"); plumbed via build_system_prompt(voice_mode=True).
- Isolated voice session key `voice:<channel_id>` — voice doesn't share
context with text adapter on the same channel; auto-applied without
/clear ceremony. /clear drops both keys.
- Metric units + Romanian thousands (normalize.py): "384.000 km" →
"trei sute optzeci și patru de mii de kilometri" with feminine-correct
pluralization and "de" particle for ≥20.
- `/voice setvoice <M1-F5>` slash command with native autocomplete; swaps
live + persists voice.default_voice to config.json.
- Verbal voice change (src/voice/voice_commands.py + 29 tests) — "schimbă
vocea pe M5", "voce em cinci", with permissive substring fallback for
Whisper-mangled forms like "Mâcinci"=M5 and "unul cinci"=M5. Whisper
initial_prompt now lists voice vocabulary to bias STT toward clean
outputs.
- Fast barge-in: VAD ≥2 consecutive windows (~200ms) on Marius's user
while Echo has pending TTS frames → cut him off mid-sentence so user
doesn't wait the full silence + STT cycle. Acoustic echo bleed-through
still requires headphones (no AEC).
## Test suite
130 voice + router tests pass (test_voice_recv_dave, test_voice_session_cleanup,
test_voice_adapter_contract, test_voice_normalize, test_voice_commands,
test_router).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Parametrul `lang` era definit (DEFAULT_LANG = "ro") dar nu era inclus
in request-ul HTTP catre /v1/audio/speech. Adaugat "lang": lang in
body-ul JSON si lang="ro" explicit in _tts_synthesize().
OpenAPI-ul Supertonic confirma ca /v1/audio/speech accepta `lang`
ca parametru optional (OpenAISpeechRequest schema).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
email_digest.py folosea save_unread_emails() care salva în memory/kb/emails/.
Notițele KB trebuie create DOAR de heartbeat. Acum digest-ul face fetch
direct din IMAP (ca email_forward.py), fără side effects pe KB.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Instrucțiunea era prea restrictivă (doar formulare/documente "acționabile").
Acum include orice URL relevant: articole, linkuri de citit, resurse.
Același comportament adăugat și în HEARTBEAT pentru TL;DR din KB.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
La salvarea unui email forwardat, se extrage acum expeditorul original
din body și se elimină prefixul Fwd: din titlu — în loc de adresa lui Marius.
Corectat și fișierul deja salvat din 07 mai.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
approved-tasks.json, dashboard/status.json, anaf-monitor/monitor.log
are auto-modified by background processes (heartbeat, cron jobs, ANAF
monitor). Untracking them stops the noisy "auto-commit from dashboard"
churn. Files stay on disk; readers (router._load_approved_tasks etc.)
already handle missing files by returning empty defaults.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
State files updated by dashboard/heartbeat/cron jobs, plus new KB
captures (samsung firmware todo, scout song reel, weekly youtube notes).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Restructurare Ralph QC loop pe smart gate dispatcher tag-driven (în loc de
5 faze fixe), DAG dependsOn cu propagare blocked, retry guard 3-strike, rate
limit detection, plus dashboard live cu polling 5s.
Changes:
- tools/ralph_prd_generator.py: parametru optional final_plan_path; când e
furnizat, invocă Claude Opus pe final-plan.md pentru extragere user stories
cu schema extinsă (tags, dependsOn, acceptanceCriteria 3-5). Backward compat
păstrat — fără final_plan_path, fallback la heuristic-ul vechi.
- tools/ralph/prd-template.json: schema W3 (tags[], dependsOn[], retries,
failed, blocked, failureReason, requiresDesignReview).
- tools/ralph/prompt.md: 4 faze (impl, base quality, smart gates, commit) +
dispatcher pe story.tags. Tags vide → run-all-gates fallback (safe default).
- tools/ralph_dag.py (nou): tag validation heuristic anti-silent-regression
(force ui dacă diff atinge .vue/.tsx/.html/.css/.scss; force db pentru
migrations sau .sql; force vercel dacă există vercel.json) + topological
sort cu blocked propagation + atomic prd.json updates.
- tools/ralph/ralph.sh: --max-turns 30, DAG-aware story selection, retry
counter cu auto-fail la 3, rate limit detection (sleep 30min + 1 retry),
CLI subcommands prin tools/ralph_dag.py helper.
- dashboard/handlers/ralph.py (nou): /api/ralph/status + /<slug>/log + /prd
+ /stop. Defensive vs corrupt prd.json. Sandbox-ed PID kill.
- dashboard/ralph.html (nou): live cards 3/2/1 col responsive, polling 5s,
drawer pentru log/PRD viewer, status colors (--status-running/blocked/
failed/complete declarate inline), Lucide icons cu aria-labels.
- dashboard/api.py: mount /api/ralph/* (GET status/log/prd, POST stop).
- tests/: 72 teste noi (smart gates, DAG, retry, dashboard endpoint).
Note arhitecturale:
- Polling 5s ales peste SSE/WebSocket (suficient pentru iter Ralph 8-15min)
- Tag validation rulează POST-iter pe diff git pentru anti-silent-regression
- Rate limit retry: 1 dată per rulare, apoi mark failed=rate_limited
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
