߲jUdZddlmZddlZddlZddlZddlZddlZddlZddl m Z ddl m Z ddl mZmZmZddlmZdZd Zd Zd Zd ZiZd ed<ej4Zd d!dZd"dZd"dZd#dZd$dZ d%dZ!d&dZ"d'dZ#d'dZ$ejJdejLZ'e de(Z)ded<d(dZ*ejVd)dZ,d*dZ-y)+aHistory-reuse budgets for native vision embeds (config section ``vision``). A native ``vision_analyze`` result bakes the image into conversation history, where it is re-sent on every later API call. Two knobs bound that cost: ``vision.embed_target_bytes`` (how large one embed may be) and ``vision.max_calls_per_image`` (how often the same image may be embedded per session). See #112095: a delegated subagent re-loaded five screenshots 158 times in 15 minutes because nothing refused the repeat. ) annotationsN) ContextVar)Path)AnyIteratorOptional) tool_erroriii@izdict[tuple[str, str], int]_repeat_countscX ddlm}m}||d||S#t$r|cYSwxYw)zJ``vision.`` from config.yaml; ``default`` when config is unavailable.r)cfg_get load_configvisiondefault)hermes_cli.configr r Exception)keyrr rs F/home/ubuntu/.hermes/hermes-agent/tools/vision_tools_history_budget.py _cfg_visionr&s1:{}hWEE s  ))ctdt} t|tr t dt |}tt|ttS#t ttf$r tcYSwxYw)zZ``vision.embed_target_bytes`` clamped to 64 KiB..4 MiB; the 256 KB default on a bad value.embed_target_bytesrzboolean is not a byte budget) r_DEFAULT_EMBED_TARGET_BYTES isinstancebool ValueErrorint TypeError OverflowErrorminmax_MIN_EMBED_TARGET_BYTES_MAX_EMBED_TARGET_BYTES)rawtargets rresolve_embed_target_bytesr&/sj *4O PC+ c4 ;< <S s6235L MM z= 1+**+s&AA43A4ctd} |*|dk7r%t|tstt |dSddlm}|rtSdS#t t f$rY&wxYw)uPer-image embed cap for this session; 0 = unlimited. ``vision.max_calls_per_image`` unset (or unparseable) → ``_SUBAGENT_REPEAT_CAP`` inside a delegated subagent, unlimited for the main agent. An explicit value applies everywhere. max_calls_per_imager)"is_delegated_child_process_context) rrrr!rrragent.delegation_contextr*_SUBAGENT_REPEAT_CAP)r$r*s rresolve_repeat_capr-;si + ,C ?sbyC1Fs3x# #L#E#G NQN z "   s+AA A ct|jdr9dtj|jddj ddzS|j dddj d }d |vrd |zSd tttjj|jzS) uStable identity for an image source: local paths resolve (symlinks, ``~``, ``file://``), URLs drop their fragment, data URLs hash. Region crops are NOT part of the key — the incident loop alternated full loads and crops of the same files, and both re-embed the image.zdata:zutf-8ignoreN #rzfile://z://zurl:zfile:) startswithhashlibsha256encode hexdigestsplit removeprefixstrrospath expanduserresolve) image_urlstrippeds r _image_keyrAKsG$ (8(8((KLVVXY\Z\]]]sA&q)66yAH   Sbgg00:;CCEF FFc6ddlm}|ddt|fS)Nr)get_session_envHERMES_SESSION_IDr))gateway.session_contextrDrA)r?rDs r _count_keyrGWs7 . 3Z 5J JJrBct}|dkryt|}t5tj |d}||krt | dddy dddt dd|ddS#1swYxYw)a|Reserve one native embed of ``image_url`` for the current session; tool-error JSON when the per-session cap is already spent, else ``None``. Check and count are ONE lock section: a parallel tool batch on the same image (the incident's 4 concurrent calls) must not all pass a check taken before any of them recorded. Callers ``release_embed`` when the embed then fails.rNzHvision_analyze refused: this image has already been loaded into context z time(s) in this session (region crops of the same file count too), and every native load re-sends the full image on each later API call. Answer from what you can already see, or ask the user. (vision.max_calls_per_image = z; 0 = unlimited)F)success)r-rG _repeat_lockr get_record_embed_lockedr )r?caprcounts rrepeat_refusalrO\s  C ax Y C ""3* 3;  %    RSXRYZ/03e3C E   s (A00A9ctj|ddzt|<tttkDrCtj t t ttttkDrByy)Nrr2)r rKlen_REPEAT_COUNTS_MAX_KEYSpopnextiter)rs rrLrLssT(,,S!4q8N3 n  7 74^ 456 n  7 7rBcbt5tt|dddy#1swYyxYw)uCount one successful native embed of ``image_url`` for the current session (uncapped sessions only — capped ones are counted by the reservation in :func:`repeat_refusal`).N)rJrLrG)r?s r record_embedrWzs) 4Z 23444s%.ct|}t5tj|d}|dkDr |dz t|<ntj |ddddy#1swYyxYw)zRGive back a slot reserved by :func:`repeat_refusal` when the embed did not happen.rr2N)rGrJr rKrS)r?rrNs r release_embedrYs[ Y C *""3* 19"'!)N3    sD ) ***s ?AA#z&^\[Image attached(?: at)?: (.+?)\]\s*$vision_native_turn_imagesrzContextVar[frozenset[str]]_native_turn_imagesct|trtd|Ds tSdj d|D}tdt j |DS)Nc3hK|]*}t|txr|jddk(,yw)typer?NrdictrK.0ps r z$_native_turn_keys..s15AB 1d<f  <<5s02 c3K|]<}t|ts|jddk(s)|jdd>yw)r^textr)Nr_ras rrdz$_native_turn_keys..s9p1 1d@SXYX]X]^dXeioXoQUU62&psAAAc3pK|].}|jst|j0ywN)striprA)rbms rrdz$_native_turn_keys..s'aqWXW^W^W`Z *as66)rlistany frozensetjoin_NATIVE_HANDLE_REfindall) user_messagergs r_native_turn_keysrss[ lD )5FR52{ 99p p pD a4E4M4Md4Sa aarBc#Ktjt|} dtj|y#tj|wxYww)zKScope the images natively attached to ``user_message`` to the running turn.N)r[setrsreset)rrtokens rnative_turn_imagesrxsC # #$5l$C DE) !!%(!!%(sA<AAAcx|t|tjvrytjddddS)aText tool result when ``image_url`` already rides the active user turn natively, else ``None``. A native re-embed would put the identical pixels into the same request twice (the Telegram case in #76411); a ``region`` crop still embeds because it returns new detail.NTuThis image is already attached natively to the current user message — you can see it now. Answer with your built-in vision; pass a `region` to zoom into part of it.)rIalready_in_contextmessage)rAr[rKjsondumps)r?regions rnative_turn_duplicatersEZ 2:M:Q:Q:SS ::" ^  rBri)rr:)returnr)r?r:rr:)r?r:rtuple[str, str])r?r:r Optional[str])rrrNone)r?r:rr)rrrrzfrozenset[str])rrrrzIterator[None])r?r:r~zOptional[list]rr).__doc__ __future__r contextlibr4r|r;re threading contextvarsrpathlibrtypingrrrtools.registryr rr"r#r,rRr __annotations__LockrJrr&r-rArGrOrLrWrYcompile MULTILINErprnr[rscontextmanagerrxrrBrrs# "**%)#)-/*/y~~  N O GK .74* BJJH",,W2<=Xbkbm2n/nb )) rB