+ iHjRt^RIHt^RIHtHtRRRRRR/tRtRtR R lt R R lt R Rlt RRlt RRlt RRltRRltRRltR#)aGemini long decode: one response spread over several generateContent hops. Gemini 4 caps each request (32,768 tokens unary, 65,536 streamed) but lets ``maxOutputTokens`` reach 1,048,576 minus the input. A hop that hits the per-request cap with budget left ends with ``finishReason: CONTINUATION`` plus a ``continuationToken``; the client resends the SAME request with that token (never the partial output) until any other finish reason. Each hop re-reads the prompt plus everything generated so far, so per-hop ``promptTokenCount`` grows: usage is summed for billing (what the google-genai SDK reports and Google bills) while the first hop's prompt is kept as the context size, because compaction fed the summed prompt would fire far too early. A dropped stream resumes from its last checkpoint token rather than restarting (Google bills a dropped stream up to that checkpoint). Docs: https://ai.google.dev/gemini-api/docs/generate-content/long-decode ) annotations)AnyOptionalxhighimaxiultraic$V^8dQhRRRRRR/#)modelstrreasoning_configrreturnz Optional[int])formats"=/home/ubuntu/.hermes/hermes-agent/agent/gemini_long_decode.py __annotate__rs&____ _cT;'gRP4P4PR4pVPR4'd,\ V\ 4'dVP R4RJdR#\P \VP R4;'gR4P4P44#)z`Total output budget a Gemini 4 request gets for its reasoning effort, or None (per-request cap).zgoogle/zgemini-4enabledFNeffort) striplower removeprefix startswith isinstancedictget_LONG_DECODE_BUDGETSr )r r names&& rlong_decode_max_output_tokensr s KKR   & & ( 5 5i @D ??: & &j9I4.P.PTdThThirTsw|T|  # #C(8(<("D"DZXikoMpMp vwF qVG{*+ ,,rc V^8dQhRRRR/#)r api_payloadrr r#r)rs"rrr-s^^c^c^rc\V\4'dVPR4MRp\V\4'd\VPR44#^#)aNon-streaming stale budgets are one timer around every unary hop of a long decode: scale it by the hop count so a healthy multi-hop answer isn't killed and re-billed from hop 1. The run-budget cap still applies afterwards; streaming re-arms on every chunk and needs no scaling. extra_bodyNr")rrrr()r*extras& rlong_decode_stale_multiplierr.-sJ.8 T-J-JKOOL )PTE?I%QU?V?V EII&9: ;]\]]rc V^8dQhRRRR/#)r payloaddict[str, Any]r r)rs"rrr5s444>4rc\V\4'dVPR4MRp\V\4'dV'd V^,MRp\V\4'dV#/#) candidatesN)rrrlist)r0r3firsts& r _candidater65sN.8$.G.G\*TJ' D99jJqMdEud++533rc V^8dQhRRRR/#)r r0r1r z Optional[str]r)rs"rrr;sAAA=Arc\V4pVPR4p\VPR4;'gR4P4R8Xd!\ V\4'd V'dV#R#)zWThe token to resend with when this hop ended on CONTINUATION; None on any other finish.continuationToken finishReasonr CONTINUATIONN)r6rr upperr)r0candtokens& rcontinuation_tokenr?;si g D HH( )E066B7==?>QV`afhkVlVlqv5A}AArc V^8dQhRRRR/#r hopslist[dict[str, Any]]r r1r)rs"rrrBs.>rca\Uau/uF[o\;QJdV3RlV4F 'gK RM RM!V3RlV44'gKES\V3RlV44bK] ppV'd V^,M/p\VP R4;'g^4VR&\VP R4;'g^4VR&V#uupi) zcSum every count across hops; ``initialPromptTokenCount`` (the google-genai name) is hop 1's prompt.c3.<"TF pSV9xK R#5i)Nr.0hkeys& r "merge_hop_usage..Ds^vqulm_bfg_gqusTFc3l<"TF)p\VPS4;'g^4xK+ R#5i)N)r#rrFs& rrJrKDs':Ts155:??++Ts!44promptTokenCountinitialPromptTokenCountcachedContentTokenCountinitialCachedContentTokenCount)_USAGE_COUNT_KEYSanysumr#r)rBrImergedr5s&` rmerge_hop_usagerVBsFW wFWs[^[^^vqu^v[^[^[^^vqu^v[v:c3:T:::FWF wDGE(+EII6H,I,N,NQ(OF $%/2599=V3W3\3\[\/]F +, M xsCCCCc V^8dQhRRRR/#rAr)rs"rrrKsKK2K~KrcVUUu.uFHp\V4PR4;'g/PR4;'g.Fq"NK KJ ppp\VR,4p/VCR/VPR4;'g/CRRRV/C/Cp\VUu.uFqPR4;'g/NK up4p/VR,CRV.RV/C#uuppiuupi)zbOne generateContent response whose model turn joins every hop's parts, ending as the last hop did.contentpartsroler usageMetadatar3)r6rrV)rBhoppartrZlast candidateusages& rmerge_hop_responsesrcKs! jTc 30C0CI0N0T0TRT/Y/YZa/b/h/hfh/hdT/hTTE j d2h De4e$d(;(A(Ar$dFGU\^c$deI tLtWW_5;;;tL ME Jd2h J yk?E JJ kMs%CCCC7Cc$V^8dQhRRRRRR/#)r eventr1 hop_usagerCr z$tuple[dict[str, Any], Optional[str]]r)rs"rrrTs(UU^U@TUY}Urc\V4pV'dTPVPR4;'g/4\V4P 4UUu/uFwr4VR9gKW4bK pppVP 4UUu/uFwr4VR8wgKW4bK uppRV./,V3#\V4pV'dHVPR4'd1/VCR\ .VOVPR4;'g/N4/CR3#VPR4'dRMVPR4pT\ V\4'd V'dV3#R3#uuppiuuppi)a6``(event to translate, token to resend with if the stream ends here)`` for one SSE event. A CONTINUATION finish is not the end: its finish reason and (per-stream cumulative) usage are held back in ``hop_usage`` and the token returned for the next hop. The real final event gets the usage summed over every hop. A token on an event with no finish is a mid-stream checkpoint (every ~5K tokens past 32K): it commits that event and everything before it, so the caller resumes a dropped stream by resending the request with it; a drop with no clean checkpoint restarts. r\r:r9r3N)r:r9)r?appendrr6itemsrVrr )rerfr>kvr= checkpoints&& rsplit_stream_continuationrmTs> u %E ?399r:!+E!2!8!8!:m!:aGl>l!:m!&G!2FGrusa #  %GYO  _-^4 AKUr