c1j\SdZddlmZddlZej j ddddlmZ e dZ ddl m Z ddlmZdd lmZddlZddlZddlZddlZddlZddlZddlZddlZdd lmZmZdd lmZed Zd Zd"dZ d#dZ!d$d%dZ"d&dZ#d&dZ$d'dZ%d(dZ&d)dZ' d* d+dZ(dZ)d,d-dZ*d.d/dZ+d,d0dZ,d1dZ-d2dZ.d3d4dZ/d5dZ0e1dk(r7e jed ejfe0ejhd!dyy)6u Aria Knowledge Pipeline (AKP) — Stage 3: Deep Process KAR-74 · 2026-05-12 Liest alle triaged-promote Videos ohne deep_processed-Eintrag, schickt vollen Transcript an Opus 4.7 mit kombiniertem Prompt: - 7-Punkt Praxisanalyse (Standing Order feedback_video_analysis_protocol) - Cross-Reference zu Brain via aria-brain-search.py - Klassifikation Brain-Entry/Issue/Skill/Spike/Ignore + Priority P0-P2 - Inline Multi-Level-Challenge (Devils-Advocate -> Steel-Man -> Synthese) - Confidence-Score (Opus self-rated) Output: Brain-Note in 00-Inbox/Videos/--.md (Standing-Order Frontmatter) Auto-Issue gated auf P0 + confidence>=0.85 + Cross-Ref dedupe. Gated auf ANTHROPIC_API_KEY. ) annotationsNz/root/aria/lib) get_logger aria-akp-deep)audit)DeepProcessedItem)trace_llm_call)datetimetimezone)Pathz%/root/aria/brain/youtube/.config.yamluDu bist Aria im Deep-Process-Modus fuer KAR-74 AKP. Auftrag: dieses Video tief analysieren nach dem Standing Order "feedback_video_analysis_protocol" und das Ergebnis als JSON liefern. Video-Metadata: - Titel: {title} - Channel: {channel} - Dauer: {duration_min} min - Upload: {upload_date} - URL: {url} Cross-Reference (top 3 verwandte bestehende Brain-Notes, falls Match): {cross_ref_block} Transcript: {transcript} --- Liefere AUSSCHLIESSLICH ein JSON-Objekt mit dieser Struktur (keine Markdown-Codefences drumherum, kein Vor- oder Nachwort): {{ "kernaussage": "1-2 Saetze", "praxisanalyse": {{ "1_was_lernen": "...", "2_wie_einsetzen": "...", "3_vorteil_aria_kadi": "...", "4_besser_machen": "...", "5_lohnt_warum": "...", "6_naechste_schritte": "..." }}, "kritik": "Pflichtfeld — was ist schwach, redundant, schon-bekannt, ueberholt? Nicht schoenfaerben.", "klassifikation": "Brain-Entry|Issue|Skill|Spike|Ignore", "prioritaet": "P0|P1|P2", "confidence": 0.0, "thema": "Schlagwort fuer Frontmatter", "nutzen": "Ein-Satz-Nutzen fuer Aria/KADi/MRR", "umsetzungsidee": "konkrete Action wenn klassifikation=Issue/Skill/Spike, sonst null", "tags": ["array", "von", "tags"], "issue_proposal": {{ "title": "Wenn klassifikation=Issue: KAR-Issue-Vorschlag-Title, sonst null", "body_md": "Wenn klassifikation=Issue: kurzer Body inkl. why+how+effort, sonst null", "estimate": "S|M|L|XL" }}, "challenge": {{ "devils_advocate": "Warum sollte das NICHT umgesetzt werden? 1-2 Saetze.", "steel_man": "Bestes Argument FUER Umsetzung. 1-2 Saetze.", "synthese": "Aria-Verdict: Promote / Refine / Drop — mit kurzer Begruendung." }} }} Klassifikations-Regeln: - Brain-Entry: Wissen aufschreiben, kein Action-Item - Issue: konkretes KAR-Backlog-Item das gebaut werden sollte - Skill: neuer .claude/skills/-Eintrag oder bestehender Skill-Update - Spike: Architektur-Untersuchung noetig, mehrere Optionen offen - Ignore: trotz Triage-Promote bei Deep-Inspect doch nicht relevant (Override) Confidence: ehrlich self-rated 0.0-1.0. Bei Klassifikation=Issue: nur >=0.85 erlaubt Auto-Issue. Prioritaet: P0=jetzt, P1=diese Woche, P2=Backlog. ctj5}tj|cdddS#1swYyxYwN) CONFIG_PATHopenyaml safe_load)fs #/root/aria/scripts/aria-akp-deep.py load_configras1    !q~~a !!!s4=cZtj|}tj|_|Sr )sqlite3connectRow row_factory)db_pathcons rdb_connrfs //' "CkkCO Jc:t|jd|fS)NaSELECT i.video_id, i.channel, i.title, i.duration_seconds, i.url, i.published_at AS upload_date, i.raw_path, t.avg_score, t.aria_relevance, t.one_line_reason FROM ingested i JOIN triaged t ON t.video_id = i.video_id LEFT JOIN deep_processed d ON d.video_id = i.video_id WHERE t.verdict = 'promote' AND d.video_id IS NULL ORDER BY t.avg_score DESC, t.triaged_at ASC LIMIT ?)listexecute)rlimits rget_pending_promoter"ls&       rc$|dz dz|dz dzzS)N@B.@gR@ input_tokens output_tokenss ropus_pricing_usdr*|s! 9 $ , 0IT/Q QQrc$|dz dz|dz dzzS)Nr$g@r%r&r's rsonnet_pricing_usdr,s! 9 $ +}y/HD.P PPrctjtjj j }|j d|d|fj}|rt|dSdS)NzSELECT z FROM daily_costs WHERE date=?r) r nowr utcdate isoformatr fetchonefloat)rkeytodayrows r todays_costr8s` LL & + + - 7 7 9E ++u$BCeX N W W YC5Q=(S(rctjtjj j }|j d|||||f|jy)NzINSERT INTO daily_costs (date, triage_usd, deep_usd, total_usd) VALUES (?,0,?,?) ON CONFLICT(date) DO UPDATE SET deep_usd=deep_usd+?, total_usd=total_usd+?)r r/r r0r1r2r commit)r deep_deltar6s radd_costr<sQ LL & + + - 7 7 9EKK U  J J? JJLrc t|jsgS tjd||dt |dgddd}|j dk7r tjd||dt |gddd}|j dk7rgSg}|jjD]}|j}|r|jdr'tjd |}|r$|j|jd dd c|jd su|j|dd |d |S t!j"|j}t%|t&r|d |S gS#tj $rgcYSwxYw#tj $rgcYSwxYw#t j($rYgSwxYw)zNCall aria-brain-search.py for top-N matches. Best-effort, no crash on failure.python3z--topz--jsonT)capture_outputtexttimeoutr)z==Top#---z^\s*[\d.]+\s+(.+\.md)\s*$)pathscore.mdN)r exists subprocessrunstrTimeoutExpired returncodestdout splitlinesstrip startswithrematchappendgroupendswithjsonloads isinstancerJSONDecodeError)querybrain_search_pathtop_nroutlinemdatas r cross_refres ! " ) ) +  NN )5'3u:x PdB   ||q -ugs5zJ#$A <<1 IHH'') 7D::=F>c tj} |jj|dd|dg}d jd |jD}|} d} d | vra| jd } | D]K} | j} | jd r| d dj} | jdsI| } n| jd} | jd}| dk7r |dk7rtj | | |dz} |j&j(|j&j*d}t |||||t tj|z dz|| |fS#t$rV}t |||t tj|z dz|t|dt|dddfcYd}~Sd}~wwxYw#tj"t$f$rd} YwxYw)Ni user)rolecontent)model max_tokensmessagesi)scriptrjprompt latency_msvideo_iderrorr)rqr(r)c3NK|]}t|ds|jyw)rAN)hasattrrA).0bs r zcall_opus..sJ!wq&7IqvvJs%%z```rY{}rFr')rmrjrn response_textusagerorp)time monotonicrlcreate ExceptionrintrMjoinrisplitrRrSfindrfindrYrZr\ ValueErrorr}r(r))clientrjrnrp trace_scriptt0resperaw_textrAparsedpartspfirstlastr}s r call_opusrs  B N%%%&9:& wwJt||JJH D F D=JJu%E GGI<<'!" A<<$D   #tzz# B;42:ZZU46 23F"ZZ44tzzG_G_ `E,eF!)!4>>#3b#8D"@AHV 5=7 Nl%"%t~~'7"'<&D"E (A 8s1vq1MMM N*  * -s8"E/AG?A G/ G8A G G GG/.G/uDu bist ein UNABHAENGIGER, adversarialer Critic. Ein anderes Modell hat aus einem Video-Transcript die folgende Analyse erzeugt. Pruefe sie streng — NICHT schoenfaerben, im Zweifel kritischer. Pruefe: 1. Halluzination: Steht alles in der Analyse wirklich im Transcript, oder erfunden/uebergeneralisiert? 2. Wert: Ist der behauptete Nutzen fuer Aria/KADi real oder generisches AI-Hype-Wiederkaeuen? 3. Klassifikation: Ist die klassifikation (Brain-Entry/Issue/Skill/Spike/Ignore) angemessen, oder wurde Hype als Issue/Skill ueberbewertet? 4. Confidence: Ist die self-rated confidence gerechtfertigt? Transcript (Auszug): {transcript} Analyse (JSON des Generators): {analysis} Liefere AUSSCHLIESSLICH ein JSON-Objekt (keine Codefences, kein Vor-/Nachwort): {{ "verdict": "keep|refine|drop", "confidence": 0.0, "hallucination_risk": "low|medium|high", "issues": ["konkrete Schwaeche 1", "..."], "corrected_klassifikation": "nur wenn die urspruengliche falsch ist, sonst null", "one_line": "1-Satz-Urteil" }} Verdict-Regeln: - keep: akkurat + wertvoll, Klassifikation passt. - refine: brauchbar, aber Schwaechen (ueberbewertete Klassifikation, schwache Belege). - drop: halluziniert, wertlos oder reines Hype — sollte Ignore sein. Sei im Zweifel strenger (Default refine statt keep).ctj|xsdddtj|ddd}t ||||d S) zKAR-747: separater, unabhaengiger Critic-Call (eigener Context, Cross-Model). Generator != Evaluator. Gibt (critic_dict_or_None, usage) zurueck. rrNi@F) ensure_asciiip) transcriptanalysiszaria-akp-deep-critic)rpr)CRITIC_PROMPT_TEMPLATEformatrYdumpsr)rrjrrrprns r call_criticrsS $ * *$"et,F7>+F VUFXLb ccrctjdd|j}tjdd|jd}|d|S)Nz[^\w\s-]rrz[\s_-]+-)rTsublowerrR)smaxlens rslugifyrsC {B *A z3"((-A Wf:rcT t|dd}|jddtjj d}t |dxs|dd}|d t |d xsd d d |d } || z } dd|dxsdj tdtddd|ddd|d d|dd|jddd|jddj tdtddd|jdxsdj tdtddd |jd!d"d#|jd$d%d&|jd'd(d)d*d+j|jd,ggd-zd.d/|d0ddg} gd1|ddd2|d d3t|dxsd4d5z d6|ddd7|jd8d9dd:d;|jd|jd.\s9Q"QC9sz - (keine)z<## Adversarial-Critic (separater Call, Cross-Model, KAR-747)z **Verdict:** verdictu · **Confidence:** u · **Halluzinations-Risiko:** hallucination_riskz **Urteil:** one_linecorrected_klassifikationz **Korrigierte Klassifikation:** z**Schwaechen:**z ## Cross-Reference (Brain-Match)rGz- [[z ]] (score rH)z- (keine Brain-Matches)z## Klassifikation: **u** · Prioritaet **u** · Confidence **z**issue_proposalz"## Issue-Proposal (auto-generated)z **Title:** z**Estimate:** estimateMbody_mdu-*Auto-generated by aria-akp-deep.py · cost $.4fu · *)r mkdirr r/strftimerreplacechrgetrrstemrVr r0r2 write_textrM)cfgrr cross_refscostcriticout_dirr6slugfnamerGfm_lines body_linesr issue_blockcrlinkips rwrite_brain_noters3v;123G MM$M. LLN # #J /E 5>6U:%6 ;DgQwuY/85"=>avS IE U?D  E'N(b11#b'#b'BC1E 5<.! E)$%& U#5678 &**WR()* FJJx+33CGCGDEQG VZZ(89?RHHRQTUWQXYZZ[\ vzz,t456 6::&6}EFG vzz,s345 $))FJJvb14UUVWWXY    %H(% U7^ % % i()eDV>W>\[\^`=`9a8bbstyzuAtB C% %  %  ="67 % % #% )%  ?B'++NC@% % /%  ?B'++,=sC% % Y%  ?B'++,A3G!%" #%$ 8%%&  ?B'++,=sC'%( )%* 5+%,  ?B'++OSA-%. /%0 #1%2  ?B'++,A3G3%4 5%6 "7%8  8Y'9%: ;%< 4=%>  ;r : > >?PQT UVW?%@ A%B &**[488SIJKC%D E%F K377 3GHIG%H I%JLH%+=Cdii9&99  JFJJy566J6::VbcfKgJhiHIOISISThilImHn o 6::j56 7 \b\f\fhB]C/ ;U0V/W XIK     *J LAfb!A7<KK t >8ZTZZXbMceijJJLrc\|jd}|j|jS)Nz>INSERT INTO run_log (stage, status) VALUES ('deep', 'running'))r r: lastrowid)rcurs r log_run_startrs$ ++V WCJJL ==rcR|jd|||||f|jy)NzUPDATE run_log SET finished_at=datetime('now'), status='completed', items_processed=?, items_skipped=?, items_failed=?, notes=? WHERE id=?)r r:)rrun_id processedskippedfailednotess rlog_run_finishrs-KK U GVUF3JJLrc t}t|dd}tjj d}|s%t |d}t dt|dy ddlm }|| }|d d }|d d}t|d} t|d} | |k\s| |k\rt d| dd|d| dd|yt |}|s t dyt dt|d|| z dd|| z dt|} |dd} |dd} |dd}|dd}t|dj dd }|dj d!d"}i}d}d}d}|D]i}t|d|k\st|d|k\rt d#n= tjt!|d$j#}|j d&xsd'd| }|d(d)|d*dd+}t)||| }|sd,nd-j+d.|D}t,j/|d(xsd'|d*xsd't1|d/xsdd0z |d1xsd2|d3xsd'||4}t3||||d56\}}t5|d7|d8}t7||9|rd:|vrt9||d5d;d} |j d?d<}!t;|j d@d=}"d}#|rt=|||||d56\}#}$|$rO|$j dAs>t?|$j d7d|$j d8d}%t7||%9||%z }|#r*|#j dBdC}&|j |&dd%z||&<tA||||||#D}'t9||d5| |!|"d|'|| |d%z }tCjDdEl|rdF|nd'}(tG|| |||dGt|dd|(t dH|dI|dJ|dKt|dd|( |jIy#t$rt d Yy wxYw#t$tj&f$r |d%z }YwxYw)LNpathsstate_dbANTHROPIC_API_KEYi'z&[akp-deep] ANTHROPIC_API_KEY not set. z promoted videos queued.r) Anthropicz![akp-deep] anthropic SDK missing.)api_key cost_caps deep_max_usd total_max_usd total_usddeep_usdz#[akp-deep] cost-cap reached. deep=$rz/$z , total=$z&[akp-deep] no pending promoted videos.z [akp-deep] z pending. budget left: deep=$z total=$rtranscript_max_charscross_ref_top_nmodels deep_model brain_searchcritic_enabledT critic_modelzclaude-sonnet-4-6z$[akp-deep] cost-cap reached mid-run.raw_pathrFrrrr rz(keine Matches)rc3nK|]-}d|jddd|jddd/yw)rrG?z (score rHrrN)r)rurs rrwzmain..s>C FGbvc"#8AEE'#,>+?q AC s35rr upload_dateunknownr)rr duration_minrrcross_ref_blockrrp)rpr(r))r;rIgnorerr.rrrrqrr)rg?z critic_verdicts=z deep_today=$z[akp-deep] processed=z skipped=z failed=z deep_today=$)%rrosenvironrr"printlen anthropicr ImportErrorr8rboolrYrZr read_textOSErrorr\rerDEEP_PROMPT_TEMPLATErrrr*r<rr4rr,rr~sleeprclose))argvrrrpendingrrcost_cap total_cap todays_total todays_deepr excerpt_maxcross_nrjrrrverdict_countsrrrr7rawr cross_querycrefs cref_blockrnrr}rrrrrc_usagec_costvr critic_notes) rmainr;sR -C #g,z* +Cjjnn01G %c51 6s7|nD\]^'w 'F;/HK 1IsK0Lc:.Kh,)"; 3K3DBxjPYZfgjYkkmnwmxyz!#&G  67 KG ~%B8KCWX[B\\deno{e{|eA BC 3 Ff+45K&k+,G M, 'Ew</L#f+//*:DABNx=$$^5HIL%'NIG F7 sJ '8 3{3 7TXa7a 8 9  **T#j/2<<>?Cggl+1r +|W=.3&C KPC : &,,g,$" N(bc"45:b@AM*7iE  b&!- "&%#j/R n 5u_7MN&)7 sC OXtS$ESW X aKF $4mD::lD16::lC89  )&, F]`ak]lmOFGw{{73+GKK,JGKKXgijLkl0JJy#.$2$6$6q!$sqlite3.Connection)d)rr@r!rr>zlist[sqlite3.Row])r(rr)rr>r4)rr@r5rMr>r4)rr@r;r4r>None)r]rMr^rMr_rr> list[dict])Nr) rjrMrnrMrp str | NonerrMr>tuple[dict | None, dict]r ) rjrMrrMrr?rprDr>rE)2)rrMrrr>rM)rr?rz sqlite3.Rowrr?rrCrr4rz dict | Noner>rM)rr@rprMrrMrrMrr4rrDrrMrrCrr4r>rB)rr@r>r)rr)rr@rrrrrrrrrrMr>rB)r*z list[str]r>r)5__doc__ __future__rsys_sysrGinsert aria_loggingr _get_logger_log aria_auditr_audit aria_schemasr _DeepItemaria_llm_tracerrYrrTrrKr~rr r pathlibr rr'rrr"r*r,r8r<rerrrrrrrrr;__name__eventexitr*r&rrrXs"# $%2?#&7)9999':; <~!   R Q ) 'TGK"1$$6N$N8> d nb k \ zJJ~ CHHT#((12,  r