)jO UdZddlZddlZddlZddlZddlZddlmZddlm Z m Z m Z m Z m Z ddlmZddlZddlZddlmZmZddlmZejeZdeezfd Zehd Zeeed <ej d ej!Z"ej#d Z$dedefdZ%ia&e ee ee ffed<da'e(ed<iZ)e ee ee ffed<dZ*e(ed<dZ+iZ,e ee ee ee fffed<iZ-e ee(fed<dZ.gdZ/e/dZ0dZ1iddddddddddd dd!dd"dd#dd$dd%d&d'd(d)d*d+d*d,d(d-d.d/d.id0d*d1d2d3d.d4d5d6d7d8d7d9d7d:d;ddd?dd@ddAddBd.dCd;dDd5dEdidFdGdHd;dIddJdKdLdMdNd7dOd7dPd=dQdRdSdRdTddUd7dVd;dWd;dXd;dYdGdZdGid[d;d\dGd]dGd^d;d_d;d`dadbdGdcdGdddGdedKdfdGdgd5dhd5did5djdGdkdGdldMZ2dmZ3dedefdnZ4doZ5dpZ6dqZ7drZ8dsedefdtZ9ddvede eeffdwZ:dsedefdxZ;dsedefdyZm?Z@e@D]&ZAeABZCeCreCe=vr eAjDe=eC<'n #eE$rYnwxYwdsede efdZFdsedefdZGdsedefdZHddsedvede efdZIde fdZJdde deKdeKde eKfdZLde ee fdeMedfde eKfdZNde ee fde eKfdZOde ee fde eKfdZPde ee fde ee ffdZQde ee ee ffdede ee fddfdZRddede ee ee fffdZS ddsedvedede ee ee fffdZT ddedsedvede eKfdZUdefdZVde eeKffdZWdedsedeKddfdZXdedsede eKfd„ZYdedseddfdÄZZdeKde eKfdńZ[dede eKfdDŽZ\dedeKde eKfdɄZ]dede eKfdʄZ^dededefd̈́Z_ddedsedvede eKfd΄Z`ddedsedvede eKfdτZadedefdЄZbdedefdфZcdedefd҄Zdddedsedvede eKfdӄZededefdԄZfdedsedvede eKfdՄZgdddd.dddddddל Zhe eeKfed<iaie eeKfed<daje(ed<dZkdede eeKffd݄Zl ddedede eKfdބZm ddedsedvede e eKeffd߄Zn ddedsedvedeKdzdedeodzdeKfdZpdedeKfdZqde e ee fdeKfdZrde ee fdeKdeKfdZsde ee fdeKfdZtduddde e ee fdede e e ee fdeKfdZudS)zModel metadata, context lengths, and token estimation utilities. Pure utility functions with no AIAgent dependency. Used by ContextCompressor and run_agent.py for pre-flight context checks. N)Path)AnyDictListOptionalTuple)urlparse)base_url_host_matchesbase_url_hostname)OPENROUTER_MODELS_URLreturncdD];}tj|}|r#tj|r|cS>#&& JJJ 4>Ix-aix.aiz-aiz.aikimi-cnarcee-ai deep-seek gmi-cloud novita-ai minimax-cn nvidia-nim moonshot-cn qwen-portal xiaomi-mimo opencode-zen github-models google-gemini minimax-oauth tencent-cloudgoogle-ai-studiogoglmgminimxaizaizengrokkilokimimimonousqwenarceelocalzhipualiyunclaudecustomgeminigithubgooglenovitanvidiaollamaxiaomialibabaarceeaicopilotminimaxstepfuntencentdeepseekgmicloudkilocodemoonshotnemotronnovitaaiopencodetokenhub anthropic dashscope openrouter tencentmaas qwen-oauth copilot-acp kimi-coding opencode-go ollama-cloud openai-codexgithub-copilotkimi-coding-cntencent-tokenhub_PROVIDER_PREFIXESzE^(\d+\.?\d*b|latest|stable|q\d|fp?\d|instruct|chat|coder|vision|text)z 100.64.0.0/10modelc,d|vs|dr|S|dd\}}|}|tvr0t |r|S|S|S)uaStrip a recognised provider prefix from a model string. ``"local:my-model"`` → ``"my-model"`` ``"qwen3.5:27b"`` → ``"qwen3.5:27b"`` (unchanged — not a provider prefix) ``"qwen:0.5b"`` → ``"qwen:0.5b"`` (unchanged — Ollama model:tag) ``"deepseek:latest"``→ ``"deepseek:latest"``(unchanged — Ollama model:tag) :http) startswithsplitstriplowerrd_OLLAMA_TAG_PATTERNmatch)reprefixsuffix prefix_lowers r_strip_provider_prefixrsVs %5++F33 [[a((NFF<<>>''))L)))  $ $V\\^^ 4 4 L Lr_model_metadata_cache_model_metadata_cache_time_novita_metadata_cache_novita_metadata_cache_timei_endpoint_model_metadata_cache#_endpoint_model_metadata_cache_timei,)i}i>i@r|zclaude-fable-5@Bz claude-fablezclaude-opus-4-8zclaude-opus-4.8zclaude-opus-4-7zclaude-opus-4.7zclaude-opus-4-6zclaude-sonnet-4-6zclaude-opus-4.6zclaude-sonnet-4.6r@i@ gpt-5.5iz gpt-5.4-nano gpt-5.4-minigpt-5.4gpt-5.3-codex-sparkr{z gpt-5.1-chatgpt-5zgpt-4.1izgpt-4rBizgemma-4rzgemma4z gemma-4-31bzgemma-3igemmai zdeepseek-v4-prozdeepseek-v4-flashz deepseek-chatzdeepseek-reasonerrOllamaz qwen3.6-pluszqwen3-coder-plusz qwen3-coderir; minimax-m3rL r0iz grok-buildzgrok-code-fastz grok-2-visionz grok-4-fastiz grok-4.20grok-4.3zgrok-4zgrok-3zgrok-2r6r8z hy3-previewrStrinityelephantzQwen/Qwen3.5-397B-A17BzQwen/Qwen3.5-35B-A3Bzdeepseek-ai/DeepSeek-V3.2izmoonshotai/Kimi-K2.5zmoonshotai/Kimi-K2.6zmoonshotai/Kimi-K2-ThinkingzMiniMaxAI/MiniMax-M2.5zXiaomiMiMo/MiMo-V2-Flashz mimo-v2-proz mimo-v2.5-proz mimo-v2.5z mimo-v2-omniz mimo-v2-flashz zai-org/GLM-5)z grok-3-minizgrok-4.20-multi-agentrc|pdsdSdD]"}|vr|dd#tfdtDS)a$Return True when an xAI Grok model accepts ``reasoning.effort``. Allowlist by substring (matches both bare ``grok-3-mini`` and aggregator-prefixed ``x-ai/grok-3-mini``). Conservative by design: if a future Grok model isn't listed, we send no effort dial rather than 400. F)/ric3BK|]}|VdSN)rj).0rpnames r z1grok_supports_reasoning_effort..&s/SS6tv&&SSSSSSr)rlrmrsplitany_GROK_EFFORT_CAPABLE_PREFIXES)reseprs @rgrok_supports_reasoning_effortrs KR   & & ( (D u++ $;;;;sA&&r*D SSSS5RSSS S SSr) context_lengthcontext_window context_sizemax_context_lengthmax_position_embeddings max_model_lenmax_input_tokensmax_sequence_length max_seq_len n_ctx_trainn_ctxctx_size)max_completion_tokensmax_output_tokens max_tokens) localhostz 127.0.0.1z::1z0.0.0.0)z.docker.internalz.containers.internalz.lima.internalbase_urlcT|pddS)Nrr)rlrstriprs r_normalize_base_urlrHs& N ! ! # # * *3 / //rrapi_keyc^t|pd}|siSdd|iS)Nr AuthorizationBearer )strrl)rtokens r _auth_headersrLs@  2   $ $ & &E  .u.. //rc"t|dS)N openrouter.ai)r rs r_is_openrouter_base_urlrSs ? ; ;;rc^t|}t|ot| Sr)rboolr)r normalizeds r_is_custom_endpointrWs.$X..J    G$;J$G$G GGrzapi.openai.comopenaiz chatgpt.comapi.anthropic.comrWzapi.z.air4zopen.bigmodel.cnzapi.moonshot.air]zapi.moonshot.cnrbz api.kimi.comzapi.stepfun.airMzapi.stepfun.comz api.arcee.air<z api.minimaxzdashscope.aliyuncs.comrIzdashscope-intl.aliyuncs.comzportal.qwen.air[rrYz!generativelanguage.googleapis.comr:rK fireworksr^r3rFrHr1rErcr_)zinference-api.nousresearch.comzapi.deepseek.comzapi.githubcopilot.comzmodels.github.aizmodels.inference.ai.azure.comzapi.fireworks.aiz opencode.aizapi.x.aizintegrate.api.nvidia.comzapi.xiaomimimo.comzxiaomimimo.comzapi.gmi-serving.com api.novita.aiztokenhub.tencentmaas.comz ollama.com_URL_TO_PROVIDER)list_providersct|}|sdStd|vr|nd|}|jp|j}t D] \}}||vr|cSdS)aInfer the models.dev provider name from a base URL. This allows context length resolution via models.dev for custom endpoints like DashScope (Alibaba), Z.AI, Kimi, etc. without requiring the user to explicitly set the provider name in config. N://zhttps://)rr netlocrmrritems)rrparsedhosturl_partproviders r_infer_provider_from_urlrs%X..J t EZ$7$7jj=T =T=T U UF =   7FK$5$5$7$7D.4466( t  OOO  4rc$t|duSr)rrs r_is_known_provider_base_urlrs #H - -T 99rct|}|sdSd|vr|nd|} t|}|jpdn#t$rYdSwxYwtvrdSt fdt DrdSrdvrdS tj}|j s|j s|j rdSt|tj r |tvrdSn#t$rYnwxYwd}t#|dkr t%|d t%|d }}|d krdS|d krd |cxkrdkrnndS|dkr|dkrdS|dkrd|cxkrdkrnndSn#t$rYnwxYwdS)aReturn True if base_url points to a local machine. Recognises loopback (``localhost``, ``127.0.0.0/8``, ``::1``), container-internal DNS names (``host.docker.internal`` et al.), RFC-1918 private ranges (``10/8``, ``172.16/12``, ``192.168/16``), link-local, and Tailscale CGNAT (``100.64.0.0/10``). Tailscale CGNAT is included so remote-but-trusted Ollama boxes reached over a Tailscale mesh get the same timeout auto-bumps as localhost Ollama. Frzhttp://rTc3BK|]}|VdSr)endswith)rrqrs rrz$is_local_endpoint..s/ I IV4== I I I I I Ir.rri d@)rr hostname Exception _LOCAL_HOSTSr_CONTAINER_LOCAL_SUFFIXES ipaddress ip_address is_private is_loopback is_link_local isinstance IPv4Address_TAILSCALE_CGNAT ValueErrorrklenint) rrurlraddrpartsfirstsecondrs @ris_local_endpointrs5%X..J u++**1G:1G1GC#$" uu |t I I I I/H I I IIIt 4t #D)) ? d. $2D 4 dI1 2 2 t?O7O7O4       JJsOOE 5zzQ aMM3uQx==6E{{t||f 2 2 2 2 2 2 2 2 2t||# t||f 3 3 3 3 3 3 3 3 3t    D  5sK: AA)C,#C CC 0E<=E< E<#E<< F F cddl}t|}|}|dr |dd}t|} |d|5} ||d}|jdkr dddd Sn#t$rYnwxYw ||d }|jdkr8 |}d |vr dddd Sn#t$rYnwxYwn#t$rYnwxYw ||d }|jdkr||d}|jdkrd|j vr ddddSn#t$rYnwxYw ||d}|jdkr&|}d|vr ddddSn#t$rYnwxYwdddn #1swxYwYn#t$rYnwxYwdS)zDetect which local server is running at base_url by probing known endpoints. Returns one of: "ollama", "lm-studio", "vllm", "llamacpp", or None. rN/v1g@timeoutheaders/api/v1/models lm-studioz /api/tagsmodelsrG /v1/props/propsdefault_generation_settingsllamacppz/versionversionvllm) httpxrrrClientget status_coderjsontext) rrrr server_urlrclientrdatas rdetect_local_server_typer s` LLL$X..JJ5!!%_ G$$G* \\#w\ 7 7' 6 JJ*<<<===C''& ' ' ' ' ' ' ' ' (     JJ*77788=C'' vvxx#t++#+#' ' ' ' ' ' ' '  ,$     JJ*77788=C'' j#8#8#899A=C'',IQV,S,S%9' ' ' ' ' ' ' ' :     JJ*66677=C''6688D D((%K' ' ' ' ' ' ' ' L    M' ' ' ' ' ' ' ' ' ' ' ' ' ' ' P       4sG/G##B G/G# BG#BG#!#C=C, G/+C=, C96C=8C99C=<G#= D G# D  G#AE, G/+G#, E96G#8E99G#=;G9 G/G# GG#GG# G/#G''G/*G'+G// G<;G<valuec#Kt|tr2|V|D]}t|Ed{VdSt|tr|D]}t|Ed{VdSdSr)rdictvalues_iter_nested_dictslist)r nesteditems rrrs%0 llnn 2 2F)&11 1 1 1 1 1 1 1 1 2 2 E4 0 0 0D)$// / / / / / / / /00 0 0r逖minimummaximumc$ t|trdSt|tr(|dd}t |}n#t tf$rYdSwxYw||cxkr|krnn|SdS)N,r)rrrrlreplacer TypeErrorr)r rrresults r_coerce_reasonable_intr%s eT " " 4 eS ! ! 3KKMM))#r22EU z "tt&####G##### 4sA&A A&&A;:A;payloadkeys.cd|D}t|D]W}|D]@\}}t||vr)t |}||ccSAXdS)Nc6h|]}|Srm)rkeys r z%_extract_first_int..4s * * *cciikk * * *r)rrrrmr)rrkeysetmappingr$r coerceds r_extract_first_intr)3s * *T * * *F%g..!--//  JC3xx~~v--,U33G"#   4rc,t|tSr)r)_CONTEXT_LENGTH_KEYSrs r_extract_context_lengthr-? g'; < <z$_extract_pricing..Zs.PPP*#uc#hhnn&&PPPrc3NK|]}tfd|DV dS)c3 K|]}|vV dSrr")raliasrs rrz-_extract_pricing...[s(@@uu *@@@@@@rN)r)raliasesrs rrz#_extract_pricing..[s@ccW3@@@@@@@@@ccccccr>Nr)rrfloatrrrr) r novita_input novita_outputpricing alias_mapr'targetrLrKrs @r_extract_pricingrSGs;;899LKK :;;M=#<"$  # #E,$7$7&$@9$L M MGH   $$'m(<(2)F)F)J)JKbdh)i)i &(33 99Y33 E uh 6 6 6 "2B77I .sVJq$4G4Gj!!Z/rrrrrr model_aliasz1Failed to fetch model metadata from %s/models: %s)rrrxrryra_ENDPOINT_MODEL_CACHE_TTLrrappendrr rcrrdrrrrr1rSrZrrr-rrokrerf)#rrr[rcached cached_at candidates alternater last_errorrrhrrTrerUrVrinstcfgctxrrPalt_idexc candidater is_llamacppbase_verify props_respprops gen_settingsrrus# rfetch_endpoint_model_metadatars#%X..J 0<< /33J??7;;JJJ  49;;#:>W"W"WMJ5!!'ssO**3// & %Yj00)$$$8?G 3' 3 344RG&*J$$21 ' GDDD SS>>"YYt__F!&#..A6Af>P>P*5&%@@@=B.z:BF)++3J? ]T^   JJJJJJ  44 s##i/2 |C"MeMgMghhhH  % % ' ' 'mmooG/1E VR00 ; ;!%.. 99T??)/681L1L(M!8!?!?!-.Bikk / ;LLL   JJJJJJ b H*V`aaa13":.6:ikk' 3 IsW1I%M M."M))M.D/W?C#V#"W# V0-W/V00)W W3&W..W3ct||}||}|sht|dkr/tt |}n&|D]\}}||vs||vr|}n|r,|d}t|tr|SdS)zDResolve context length from an endpoint's live ``/models`` metadata.rmrirN) rrrnextiterrrrr)rerrendpoint_metadatamatchedr$rVrs r _resolve_endpoint_context_lengthr#s 6hPPP##E**G   ! !Q & &4 1 8 8 : :;;< context_length cache from disk.utf-8encodingNcontext_lengthsz'Failed to load context length cache: %s) rexistsopenyaml safe_loadrrrerf)rfr rjs r_load_context_cacher@s " $ $D ;;==  $ ) ) ) +Q>!$$*D + + + + + + + + + + + + + + +xx)2...  >BBB s@A;A A;AA;!A"A;; B,B'!B,'B,lengthc|d|}t}|||krdS|||<t} |jddt |dd5}t jd|i|d dddn #1swxYwYt d ||d dS#t$r&}t d |Yd}~dSd}~wwxYw)zPersist a discovered context length for a model+provider combo. Cache key is ``model@base_url`` so the same model name served from different providers can have different limits. @NTparentsexist_okwrrrFdefault_flow_stylez%Cached context length %s -> %s tokensrz'Failed to save context length cache: %s) rrrparentmkdirrrdumpreinforrf)rerrr$rTrrrjs rsave_context_lengthrNsp   X  C  ! !E yy~~E#J " $ $DC $666 $g . . . O! I(%0! N N N N O O O O O O O O O O O O O O O ;SV--PPPPP CCC >BBBBBBBBBCs<.C3B CBC B!!C C4C//C4cV|d|}t}||S)zBLook up a previously discovered context length for model+provider.r)rr)rerr$rTs rget_cached_context_lengthrcs0   X  C  ! !E 99S>>rc|d|}t}||vrdS||=t} |jddt |dd5}t jd|i|d ddddS#1swxYwYdS#t$r'}t d ||Yd}~dSd}~wwxYw) zCDrop a stale cache entry so it gets re-resolved on the next lookup.rNTrrrrrFrz6Failed to invalidate context length cache entry %s: %s) rrrrrrrrrerf)rerr$rTrrrjs r!_invalidate_cached_context_lengthrjsV   X  C  ! !E % c " $ $DW $666 $g . . . O! I(%0! N N N N O O O O O O O O O O O O O O O O O O WWW MsTUVVVVVVVVVWs;.BB6 BBB B B CB<<Ccurrent_lengthc.tD] }||kr|cS dS)z@Return the next lower probe tier, or None if already at minimum.N)CONTEXT_PROBE_TIERS)rtiers rget_next_probe_tierrzs/# . KKK ! 4r error_msgc|}gd}|D]O}tj||}|r6t|d}d|cxkrdkrnK|cSPdS)a?Try to extract the actual context limit from an API error message. Many providers include the limit in their error text, e.g.: - "maximum context length is 32768 tokens" - "context_length_exceeded: 131072" - "Maximum context size 32768 exceeded" - "model's max context length is 65536" )zY(?:max(?:imum)?|limit)\s*(?:context\s*)?(?:length|size|window)?\s*(?:is|of|:)?\s*(\d{4,})z:context\s*(?:length|size|window)\s*(?:is|of|:)?\s*(\d{4,})z)(\d{4,})\s*(?:token)?\s*(?:context|limit)z">\s*(\d{4,})\s*(?:max|limit|token)z(\d{4,})\s*(?:max(?:imum)?)\brirrNrmresearchrgroup)r error_lowerpatternspatternrolimits rparse_context_limit_from_errorrs//##KH ';//   A''Eu**** ***** 4rcurrent_context_lengthc<t|}|dS||kr|SdS)uReturn a provider-reported lower context limit, if one is present. Context-overflow recovery must not invent a new model window size. Some providers only say that the input exceeds the context window without reporting the actual maximum. In that case callers should keep the configured context length and try compression only, rather than stepping down through guessed probe tiers (1M → 256K → 128K → ...). N)r)rr parsed_limits r&get_context_length_from_provider_errorrs32)< window. Fix: reduce max_tokens (the output cap) for this call. Do NOT touch context_length — the window hasn't shrunk. Anthropic's API returns errors like: "max_tokens: 32768 > context_window: 200000 - input_tokens: 190000 = available_tokens: 10000" Returns the number of output tokens that would fit (e.g. 10000 above), or None if the error does not look like a max_tokens-too-large error. ravailable_tokenszavailable tokensz in the outputzmaximum context length requestedz output tokensN)zavailable_tokens[:\s]+(\d+)zavailable\s+tokens[:\s]+(\d+)z =\s*(\d+)\s*$rizmaximum context length is (\d+)zK\((\d+)\s+of text input,\s*(\d+)\s+of tool input,\s*(\d+)\s+in the output\)z'maximum context length is (\d+)\s*tokenz!prompt contains (\d+)\s*characterr)rris_output_cap_errorrrrotokens_m_ctx_m_parts _available _m_ctx_tok_m_chars_ctx _est_inputs r(parse_available_output_tokens_from_errorrs"//##K  # U ; . S2D 2S ;& 4 $ 3  !K/ + ; & + { *!$ tH  ';//  Q((F{{ Y9; G GFyVH(a))Cq0A0A,B,BBSXYIZIZE[E[[ ?? E{SSJy={KKHh:##A&&''(..++,,q0Q6 J& ??  4r candidate_id lookup_modelcb||krdSd|vr"|ddd|krdSdS)aReturn True if *candidate_id* (from server) matches *lookup_model* (configured). Supports two forms: - Exact match: "nvidia-nemotron-super-49b-v1" == "nvidia-nemotron-super-49b-v1" - Slug match: "nvidia/nvidia-nemotron-super-49b-v1" matches "nvidia-nemotron-super-49b-v1" (the part after the last "/" equals lookup_model) This covers LM Studio's native API which stores models as "publisher/slug" while users typically configure only the slug after the "local:" prefix. TrriF)r)rrs r_model_id_matchesr sH|##t l|223::1=MMt 5rc@ddl}t|}|d}|dr |dd} t ||}n#t $rYdSwxYw|dkrdSt |} |d| 5}||d d |i } | j d kr ddddS| } | dd} d| vr| dD]s} d| vrm| } t| dkr4 t| dccdddS#t $rYowxYwt| di}|D]B\}}d|vr9t%|tt&frt|ccdddSC dddn #1swxYwYn#t $rYnwxYwdS)ayQuery an Ollama server for the model's context length. Returns the model's maximum context from GGUF metadata via ``/api/show``, or the explicit ``num_ctx`` from the Modelfile if set. Returns None if the server is unreachable or not Ollama. This is the value that should be passed as ``num_ctx`` in Ollama chat requests to override the default 2048. rNrrrrmrG@r /api/showrrr parametersrnum_ctx rr model_infor)rrsrrr rrrpostrrrrkrlrrrrrrM)rerrrrYr server_typerrrespr paramslinerrr$r s rquery_ollama_num_ctxrsLLL'..J%%J5!!%_ .xIII tthtG$$G \\#w\ 7 7 &6;;*777vz>R;SSD3&& & & & & & & & &99;;DXXlB//FF"""LL..%%D D(( $ 2 2 4 4u::??%'*59~~ 5 5 & & & & & & & &$.%%% $%,33J(..00 & & U#s**z%#u/N/N*u::%%- & & & & & & & &( &) & & & & & & & & & & & & & & &.      4s A A)(A)H)H HBHE7(H* H7 FHFA#H' H4H6 HHH H H HHcddl}|d}|dr |dd}t|} |d|5}||dd |i }|jd kr ddddS|}|d i} | D]J\} } d | vrAt| ttfr%t| } | dkr| ccdddSK|dd} d| vr| dD]|}d|vrv| }t|dkr= t|d} | dkr| ccdddSl#t $rYxwxYw}dddn #1swxYwYn#t"$rYnwxYwdS)uQuery an Ollama server's native ``/api/show`` for context length. Provider-agnostic: works against ANY Ollama-compatible server regardless of hostname — local Ollama, Ollama Cloud (``ollama.com``), custom Ollama hosting behind a reverse proxy, etc. For non-Ollama servers the POST returns 404/405 quickly; the function handles errors gracefully. For hosted servers the GGUF ``model_info.*.context_length`` is the authoritative source: the user can't set their own ``num_ctx``, and the OpenAI-compat ``/v1/models`` endpoint correctly omits ``context_length`` per the OpenAI schema. Resolution order for hosted Ollama: 1. ``model_info.*.context_length`` — GGUF training max (authoritative) 2. ``parameters`` → ``num_ctx`` — server-side Modelfile override The order is flipped vs ``query_ollama_num_ctx()`` because local users control ``num_ctx`` themselves; hosted users can't. rNrrrg@rrrrrrrrrrrrrr)rrrrrrrrrrrrrMrkrlrrr)rerrrrrrrr rr$r rrrrs r_query_ollama_api_showrSs&LLL%%J5!!%_ G$$G \\#w\ 7 7 %6;;*777vuo;NND3&& % % % % % % % %99;;D,33J(..00 # # U#s**z%#u/N/N*e**Cd{{"  % % % % % % % % XXlB//FF"""LL.. % %D D(( $ 2 2 4 4u::??%&)%)nn#&$;;+.JJ3 % % % % % % % %0$/#-%%% $%5 % % % % % % % % % % % % % % %8      4s G# )G G#A;G G#A/GF:*G, G#9G: GGGG G#GG#GG## G0/G0c\|}|dpd|vS)aBReturn True if the model name looks like a Kimi-family model. Catches ``kimi-k2.6``, ``kimi-k2.5``, ``kimi-k2-thinking``, ``moonshotai/Kimi-K2.6``, and similar variants. Used as a guard against stale OpenRouter metadata that underreports these models as 32K context when they actually support 262K+. r8rR)rmrj)rerms r_model_name_suggests_kimirs/ KKMME   F # # :zU'::rc.d|vS)aReturn True if the model name looks like MiniMax M3. Catches ``MiniMax-M3``, ``minimax/minimax-m3``, and similar variants across surfaces (native MiniMax-M3, OpenRouter/Nous minimax/minimax-m3). Used as a guard against stale cache entries seeded by pre-catalog builds that resolved M3 via the generic ``minimax`` catch-all (204,800) before the ``minimax-m3`` (1M) entry existed in DEFAULT_CONTEXT_LENGTHS. rr#res r_model_name_suggests_minimax_m3rs 5;;== ((rc.d|vS)azReturn True if the model name looks like a Grok 4.3 variant. Catches ``grok-4.3``, ``grok-4.3-latest``, and similar slugs. Used as a guard against stale cache entries seeded by pre-catalog builds that resolved grok-4.3 via the generic ``grok-4`` catch-all (256,000) before the ``grok-4.3`` (1M) entry was added to DEFAULT_CONTEXT_LENGTHS on 2026-05-15. rr#rs r_model_name_suggests_grok_4_3rs  &&rc ddl}t|}|d}|dr |dd}t |} t ||}n#t $rd}YnwxYw |d|5}|d krL||d d |i }|j d kr$| } | dd} d| vr| dD]s} d| vrm| } t| dkr4 t| dccdddS#t $rYowxYwt| di} | D]B\}}d|vr9t%|tt&frt|ccdddSC|dkr| |d}|j d kr| } | dgD]}t)| dd|s$t)| dd|r| dgD]j}| di}| d}|r;t%|tt&frt|cccdddSkn| |d|}|j d kr| } | dp)| dp| d}|r7t%|tt&frt|cdddS| |d }|j d kr| } | d!g}|D]}t)| dd|rz| dp)| dp| d}|r9t%|tt&frt|ccdddSdddn #1swxYwYn#t $rYnwxYwdS)"z4Query a local server for the model's context length.rNrrrrmrrrGrrrrrrrrrrrrrrrr$r^rorpz /v1/models/rrz /v1/modelsr )rrsrrrr rrrrrrrkrlrrrrrrMr)rerrrrrrrrr rrrrr$r rsr~rr models_lists r_query_local_context_lengthrs#LLL #5 ) )E%%J5!!%_ G$$G.xIII  E \\#w\ 7 7B ,6h&&{{j#;#;#;65/{RR#s**99;;D"XXlB77F F**$*LL$6$6--D(D00(, (:(:(<(<#&u::??%-/259~~(=(='B ,B ,B ,B ,B ,B ,B ,B ,(,6%-%-%-(,%-"&,!;!;J&0&6&6&8&8.. U+s22z%#u7V7V2#&u::--5B ,B ,B ,B ,B ,B ,B ,B ,Bk))zzZ"?"?"?@@#s**99;;D!XXh33"",QUU5"-=-=uEE"IZ[\[`[`aegi[j[jlqIrIr"().@"(E(E44&*hhx&<&<&)gg.>&?&?#&!4:cC<+H+H!4+.s88OOOOYB ,B ,B ,B ,B ,B ,B ,B ,Z"E"::????@@D3&&yy{{hh//g488N8O8OfSTSXSXYeSfSf,:cC<#@#@,#&s88OOEB ,B ,B ,B ,B ,B ,B ,B ,B ,B ,B ,B ,B ,B ,B ,B ,B ,B ,B ,B ,B ,B ,B ,F       4sA** A98A9=SB2R7E*R7 S* E74R76E77A#R7 S'DR79 SB(R7. S;C"R7 S*R7+ S7R;;S>R;?S SSc.|ddS)zNormalize version separators for matching. Nous uses dashes: claude-opus-4-6, claude-sonnet-4-5 OpenRouter uses dots: claude-opus-4.6, claude-sonnet-4.5 Normalize both to dashes for comparison. r-)rrs r_normalize_model_versionrs ==c " ""rcr|r|drdS |d}|dr |dd}|d}|dd}tj||d t }|jd krdS|}|d gD]O}|d |kr4|d} t| tr | dkr| cSPn2#t$r%} t d| Yd} ~ nd} ~ wwxYwdS)zQuery Anthropic's /v1/models endpoint for context length. Only works with regular ANTHROPIC_API_KEY (sk-ant-api*). OAuth tokens (sk-ant-oat*) from Claude Code return 401. z sk-ant-oatNrrrz/v1/models?limit=1000z 2023-06-01)z x-api-keyzanthropic-versionrrnrr r^rrz%Anthropic /v1/models query failed: %s) rjrrrcrrrrrrrrerf) rerrrrrrr rsrrjs r_query_anthropic_context_lengthrsz g((66tAs## ==   9D,,, !-  |C"E]E_E_```  s " "4yy{{&"%%  AuuT{{e##ee.//c3''C!GGJJJ   AAA  &>&>?+--      s " " LLZ    Iyy{{  5s;;; )34(>(>Fdhhx$$$BGF''$%%  xxhh'(( dC  'ZS%9%9 'cAgg#&F4::<<  .%+"*-' Ms$AB! B!! C+C C Cct|}|sdS|ret|}||vr||S|}|D]!\}}||kr|cS"|}t t ddD] \}}||vr|cSdS)zResolve a Codex OAuth model's real context window. Prefers a live probe of chatgpt.com/backend-api/codex/models (when we have a bearer token), then falls back to ``_CODEX_OAUTH_CONTEXT_FALLBACK``. Nc,t|dSNrrxs rz5_resolve_codex_oauth_context_length..sS1YYrTr$reverse)rsrlrrmrsortedr)rer model_barelive model_lowerrrs r#_resolve_codex_oauth_context_lengthrs(..4466J t1,??    # # &&((   ID#zz||{** +""$$K%++--3F3FPT c ;  JJJ  4rc8|rt|||}||dfSt}dtdtdtt fd}||vr||||}||dfSt |}|D]\}} d |vr| d d d n|} | |ks%t | |kr||| }||dfcS|} |D]\}} d |vr| d d d n|} | | ft | |ffD]g\} } | | rMt| t| ks| t| d vr||| }||dfccShd S) uMResolve Nous Portal model context length. Tries the live Nous inference endpoint first (authoritative), then falls back to OpenRouter metadata with suffix/version matching. Nous model IDs are bare after prefix-stripping (e.g. 'qwen3.6-plus', 'claude-opus-4-6') while OpenRouter uses prefixed IDs (e.g. 'qwen/qwen3.6-plus', 'anthropic/claude-opus-4.6'). Version normalization (dot↔dash) is applied to handle name drifts. Returns ``(context_length, source)`` where ``source`` is one of: - ``"portal"`` — live /v1/models response (authoritative) - ``"openrouter"`` — OpenRouter cache fallback (non-authoritative; callers must NOT persist this to the on-disk cache or a single portal blip will freeze the wrong value in forever) - ``""`` — could not resolve rmNportalor_idrVr c|d}|dS|dkr-t|rtd||dS|S)Nrz{Rejecting OpenRouter metadata context=%s for %r (Kimi-family underreport, Nous path); falling through to hardcoded defaults)rrrer)rrVrs r _safe_ctxz/_resolve_nous_context_length.._safe_ctxsaii()) ;4 %<<5e<<< KK^U    4 rrYrriz-:.)Nr) rrkrrrrrrmrrkrjr)rerr portal_ctxmetadatarrrrrVbarerrquerys r_resolve_nous_context_lengthrsf4(5eXwWWW  !x' '#%%H  T hsm     ix// ? $ $)%006688J (()) u),u{{3""1%%5 ::<<5;;== ( (,DT,J,J,P,P,R,RV`,`,`)E5))CL((((++--K ((-- u),u{{3""1%%5"&**,, !r s - - Iu##E** -I#e**,, #e**0E0N0Niu--? ,,,,,,  - 8rconfig_context_lengthrcustom_providerscP |t|tr|dkr|S|r.|r,|r* ddlm}||||}|r|Sn#t$rYnwxYwt |}|rX|dkrQt ||}|>|dkr7|dkr1td|||d t||n|d kr?t|r0td |||d t||n|d kr?t|r0td |||d t||nw|dkr?t|r0td|||d t||n2t|dkrtd||n|S|dks4|rTt|dr2t#|dr" ddlm} | |S#t($rYnwxYw|dks|r;t#|dr+t+||pd|} | |rt-||| | St/|r4t1|s$t+|||} | | St1|st3|||} | t-||| | St5|r3t7|||} | r| dkr|dkrt-||| | Std||t8d |} t=t> ddD],\}}|| vr#td|d |||cS-t8S|dks|r*t|d krtC||pd!|} | r| S|}|r|d"vr|rt|}|r|}|d#vr) dd$l"m#}|||} | r| Sn#t$rYnwxYw|dkr6tI||pd%|pd%&\} }| r|r|d'krt-||| | S|dkr*tK||pd%(}|r|rt-||||S|d)kr|rt+|||} | | S|r't3|||} | t-||| | S|d*krltM}|'|}|rG|'d+}t|tr|dkr|d krt|s|S|rhdd,l(m)}|||} | rTt|rCt>'d-}|r'| |kr!td.| ||d |} | S|sgtM}||vrU||'d+t8}|d kr,t|rtd/||n|S|} t=t> d0dD] \}}|| vr|cS|rBt5|r3t7|||} | r| dkr|dkrt-||| | St8S)1aGet the context length for a model. Resolution order: 0. Explicit config override (model.context_length or custom_providers per-model) 1. Persistent cache (previously discovered via probing). Nous URLs bypass the cache here so step 5b can always reconcile against the authoritative portal /v1/models response. 1b. AWS Bedrock static table (must precede custom-endpoint probe) 2. Active endpoint metadata (/models for explicit custom endpoints) 3. Local server query (for local endpoints) 4. Anthropic /v1/models API (API-key users only, not OAuth) 5. Provider-aware lookups (before generic OpenRouter cache): a. Copilot live /models API b. Nous: live /v1/models probe first (authoritative), then OR cache fallback with suffix/version normalisation. Only portal-derived values are persisted to disk. c. Codex OAuth /models probe d. GMI /models endpoint e. Ollama native /api/show probe (any base_url, provider-agnostic) f. models.dev registry lookup (with :cloud/-cloud suffix fallback) 6. OpenRouter live API metadata (Kimi-family 32k guard) 7. Hardcoded defaults (broad family patterns, longest-key-first) 8. Local server query (last resort) 9. Default fallback (256K)Nr)"get_custom_provider_context_length)rerrlmstudior`rzaDropping stale Codex cache entry %s@%s -> %s (pre-fix value); re-resolving via live /models proberrziDropping stale Kimi cache entry %s@%s -> %s (OpenRouter underreport); re-resolving via hardcoded defaultsrzjDropping stale MiniMax-M3 cache entry %s@%s -> %s (pre-catalog value); re-resolving via hardcoded defaultsrzzhDropping stale Grok-4.3 cache entry %s@%s -> %s (pre-catalog value); re-resolving via hardcoded defaultsr:z@Bypassing persistent cache for %s@%s (Nous portal authoritative)bedrockzbedrock-runtime.z amazonaws.com)get_bedrock_context_lengthrErzhttps://api.novita.ai/openai/v1rmuCould not detect context length for model %r at %s — defaulting to %s tokens (probe-down). Set model.context_length in config.yaml to override.c,t|dSrrrs rr z*get_model_context_length..sc!A$iirTr zUUsing hardcoded context length %s for model %r (custom endpoint, catalog match on %r)rWrzhttps://api.anthropic.com>rArY>rKr\ra)get_copilot_model_contextr)rrr)rr1rYr)lookup_models_dev_contextrz[Rejecting models.dev context=%s for %r (MiniMax-M3 underreport); using hardcoded default %szpRejecting OpenRouter metadata context=%s for %r (Kimi-family underreport); falling through to hardcoded defaultsc,t|dSrrrs rr z*get_model_context_length..Xss1Q4yyr)*rrhermes_cli.configrrrsrrerrrrrrrfr rjr agent.bedrock_adapterr" ImportErrorrrrrrrrDEFAULT_FALLBACK_CONTEXTrmr DEFAULT_CONTEXT_LENGTHSrrhermes_cli.modelsr$rrrkragent.models_devr%)rerrrrrrcp_ctxryr"rr local_ctxr default_modelreffective_providerinferredr$source codex_ctxrrVor_ctxr%catalogs rget_model_context_lengthr7s B(Z8Ms-S-S(XmpqXqXq$$  H    L L L L L L77!!1F        D  #5 ) )E AH ***5(;;  >))f.?.? :8]] 2%BBBB5%>u%E%E :8]] 2%BBBB7""'Fu'M'M" :8]] 2%BBBB7""'DU'K'K" :8]] 2%BBBB*(33v== V8  9 h ' ' 2 23E F F "(O < <   H H H H H H--e44 4    D 8-B8_-]-].uh6cBcmtuuu ? :#E8S999J8$$-,-H-R-R-,9%SZ[[[  %! !*844) ,)'JJJC#E8S999  ** %7xQXYYY %Q:--+E8YGGG$$ KK.x$>>>>>''(:EBB  /u55 "155lCC"sW}}KKOUwNN "CJ  ')) H  e_(()9;STTF#image image_url input_imageri_anthropic_content_blocksrG _multimodal>rGrH)rrrr)rArCcountrEpartptypestashedinners rr=r=s E$.sD$9$9Ccggi   tG'4    DdD)) HHV$$E=== 6@d6K6KUcgg1222QUG'4    D$%% $((6*:*:g*E*E '4  W[[%?%? I&& eT " "   dD))dhhv.>.>BX.X.XQJE > !!rct|tstt|Si}|D]\}}|dkr |dkrt|t rg}|D]}t|trY|ddvr,||dddZ||p|||||<t|tr/|dr|dd ||<|||<|||< tt|S) zChar count for token estimation, excluding base64 image data. Base64 images are counted via `_count_image_tokens` instead; including their raw chars here would massively overestimate token usage. rJrErF>rGrHrIz [stripped])rFrGrK text_summaryr)rrrrrrrrw)rAshadowkvcleanedrMs rr<r<sv c4 3s88}}F 1 + + +  >>!T"" --D!$---88F++/TTT#NNDHHV4D4D|+\+\]]]]#NN40000t,,,,#q At$$ })=)= EE."55q q F1II s6{{  r) system_prompttoolsrWrXcd}|r|t|dzdzz }|r|t|z }|r%|tt|dzdzz }|S)uRough token estimate for a full chat-completions request. Includes the major payload buckets Hermes sends to providers: system prompt, conversation messages, and tool schemas. With 50+ tools enabled, schemas alone can add 20-30K tokens — a significant blind spot when only counting messages. Image content is counted at a flat per-image cost (see estimate_messages_tokens_rough). rrr)rrBr)r:rWrXtotals restimate_request_tokens_roughr[su E/ #m$$q(Q..: /999 , #c%jj//A%!++ Lr)r)rr)F)rF)rr)rrNrN)v__doc__rloggingrrrapathlibrtypingrrrrr urllib.parser rcrutilsr r rr getLogger__name__rerrr frozensetrd__annotations__compile IGNORECASErn IPv4NetworkrrsrtrurMrvrwrbrxryrvrr*MINIMUM_CONTEXT_LENGTHr+rrr+r0rrrrrrr providersr_list_providers_pp get_hostname_hostrrrrrr rrrtupler)r-r1rSrZrkrrrrrrrrrrrrrrrrrrrrrrrrrrrrr7r9rBr=r<r[r"rrrps  33333333333333!!!!!! ::::::::222222  8 $ $$**&/Y000&&IcN4!bjLM)9(99##&46tCc3h/0555$%E%%%46S$sCx.01666%&U&&&GIS$sDcN/B*C%C DIII8:#T#u*%5::: /q1   v g v G vwvwvwvwvwvvwvv  f!v.w/v0F1v2F3v4w5vB6CvDFEvvF VGvHwIvJ VKvN gOvRvSvT fUvV6WvXvYvZ T[vjykvlmvnYovpqvrsvv Vwv|G}v~vvv@6AvB FCvN'OvPvQvT 6Uvb&cvdfevfTgvh7ivjkvlmvn fovp fqvr fsvt Fuvx Fyv@6AvvvDEvHvIvLMvPfQvRFSvT UvVFWvXFYvZ"6[v\f]v^_v`7avbWcvdevfFgvhVivjVkvvJ!T#T$TTTT$ < 0#0#00000030S#X0000>Q9R!+!+!+!+LMMM MM #tCH~  MMMMf c] .;;;;; T#s(^    CsCcC3C4CCCC*SCHSM WS WC WD W W W W  chsm8c](TT TTTTnCst&444s4S4(SV-4444n::#::s:HUXM::::z ;S ;T ; ; ; ; )3 )4 ) ) ) ) ' ' ' ' ' '[[s[c[C[QYZ]Q^[[[[|#C#C####3#PXY\P]P!! # 11tCH~&.0DcN///),,,,!%0S0T#s(^0000h%' ! c]FGG GGG 8C=#  GGGGX(,$( m$m$ m$m$m$: m$  m$ Tk m$ m$m$m$m$`         3T$sCx.-A3c3333""T#s(^"S"S""""4c3hCH,0 4S>" Dc3h( )   s7J??KK