+ jHj&DRt^RIHt^RIt^RIt^RIHt^RIHt^RI H t H t H t H t HtHtRtRt^t^tR#t]!0R$m4t]!RR 7!R R 44tR R lt]3RRlltRRltRRltRRltRRltRRltRRlt RRlt!RRlt"R ^/R!R"llt#R#)%u-Memory-pressure bounds for the gateway's per-session AIAgent cache. Each cached ``AIAgent`` pins its full live transcript (tens of MB on a tool-heavy session); the LRU cap counts entries, not bytes, and the idle TTL defers eviction for busy sessions, so neither sees actual memory use. This module supplies that signal — own anonymous RSS against a budget derived from the cgroup limit — and ``GatewayRunner`` sheds LRU transcripts via soft eviction (rebuilt from the persisted session next turn). Pure/read-only; config under ``agent.agent_cache``. ) annotationsN) dataclass)Path)AnyCallableIterableListOptionalTupleg?iT)frozencb]tRt^ t$RtRtR]R&RtR]R&RtR]R&] t R]R &] t R]R &R t R#) AgentCacheBoundszOperator-facing bounds. ``max_size``/``idle_ttl_secs`` are ``None`` when unset so ``gateway/run.py`` keeps its defaults; ``memory_high_mb`` ``None`` = pressure eviction off.N Optional[int]max_sizezOptional[float] idle_ttl_secsmemory_high_mbintmax_evictions_per_passprotect_recent)__name__ __module__ __qualname____firstlineno____doc__r__annotations__rr_DEFAULT_MAX_EVICTIONS_PER_PASSr_DEFAULT_PROTECT_RECENTr__static_attributes__rA/home/ubuntu/.hermes/hermes-agent/gateway/agent_cache_pressure.pyr r s=f#Hm"%)M?)$(NM("ACA1NC1rr c V^8dQhRRRR/#)valuerreturnboolr)formats"r __annotate__r',sBB3B4Brc\\V\4;'d\V\4'*#)N) isinstancerr%)r#s&r _is_intr*,s! eS ! A A*UD*A&AArc$V^8dQhRRRRRR/#)r"r#rcastzCallable[[Any], Any]r$r)r&s"r r'r'0s'AASA 4AsArc\V\4'gVfRMV!V4pTe T^8dT#R# \\3dR#i;i)zG``cast(value)`` if it is a positive number (bools rejected), else None.N)r)r% TypeError ValueError)r#r,parseds&& r _positiver10sS#E400EMtE{'FQJ6@D@ z "s$5A  A c V^8dQhRRRR/#)r"pathrr$rr)r&s"r r'r'9s4444rc\VPRR7P44p^Tu;8d R8dT#R# \\3dR#i;i)zA cgroup memory limit file's value when it is a real cap; None for unreadable, empty, ``max``, or the v1 near-2^63 sentinel (all mean unlimited). utf-8-sigencodingNl)r read_textstripOSErrorr/)r3limits& r _finite_limitr<9sZDNNKN8>>@A))53)t3 Z s)?AAcV^8dQhRR/#r"r$rr)r&s"r r'r'Cs]rcX\PR8wdR#^RIHpV!4pV'dVR8wdRV 2.M.R.,pVUUu.uFpRF qC RV 2NK K uppR.,Fp\ \ V44pVfKVu# R# \dRpL|i;iuuppi) uMemory limit this process runs under, if cgroup-capped. Prefers v2 ``memory.high`` (the throttling point) over ``memory.max``, then v1. Own cgroup first (where a systemd unit's ``MemoryHigh=``/``MemoryMax=`` lands — root reads ``max`` there), then root for container-style limits. linuxN_own_cgroup_path//sys/fs/cgroupz/memory.z+/sys/fs/cgroup/memory/memory.limit_in_byteshighmax)sysplatformgateway.cgroup_cleanuprB Exceptionr<r)rBownrootsrf candidater;s r _cgroup_limit_bytesrQCs ||w; *-se$ %HXGY YE16PA18A3''PUBTCCC d9o.  LC  Qs B B& B#"B#cV^8dQhRR/#r>r)r&s"r r'r'Zs  ] rc6\\P!R44\\P!R44,# \\\ 3dMi;i^RIp\TP4P4# \dR#i;i) SC_PAGE_SIZE SC_PHYS_PAGESN) rossysconfr:r/AttributeErrorpsutilvirtual_memorytotalrK)rYs r _total_memory_bytesr\Zs{ 2::n-.RZZ5P1QQQ Z 0   6((*0011 s$AAAA"&B BBc V^8dQhRRRR/#)r"settingrr$rr)r&s"r r'r'gs??C?M?rc\V\4'd?VP4P4pVR8wdV\9dR#\ V4#MVRJdR#VRJd \ V4#\ 4;'g \4pV'gR#\V\,\, 4pV\8dV#R#)zAbsolute MB budget: ``"auto"`` derives from the cgroup limit (or total RAM when uncapped); a positive number is literal; anything falsy/off disables the pass.autoNFT) r)strr9lower _OFF_WORDSr1rQr\r_AUTO_BUDGET_FRACTION _BYTES_PER_MB_AUTO_BUDGET_FLOOR_MB)r^ normalizedr;budgets& r resolve_memory_high_mbrigs'3]]_**,  %34 N:9N N E   !!  ! : :%8%:E  ..> ?F446>$>rc V^8dQhRRRR/#)r"configrr$r r)r&s"r r'r'yss/?rc P\V\4'd+VPR4;'g/PR4MRp\V\4'g/pVPR4p\V4pVf\ V4'd V^8Xd^p\ \VPR44\VPR4\ 4\VPRR44\VPR 44;'g\Vf \R 7#TR 7#) zRead ``agent.agent_cache`` from the *raw* config: the gateway loader does not deep-merge ``DEFAULT_CONFIG``, so callers can tell "operator chose 128" from "unset".agent agent_cacheNrrrrr`r)rrrrr) r)dictgetr1r*r floatrirr)rksectionrprotect_parseds& r resolve_agent_cache_boundsrtysAK6SW@X@Xvzz'"((b--m<^bG gt $ $[[!12N~.N'."9"9nPQ>Q 7;;z23 O r)r&s"r r'r'sMrc^a\PR8wdR#^RIHpV!4pV'dVR8XdR#\ RV 24o\ ;QJdV3RlR4F 'gK RM RM!V3RlR44pV'd"SP R 4PR R 7MR pTP4FQpTPR 4wrVpTR8XgK TP4P4'gKF\T4u# R# \\\3dR#i;i)ukAnonymous memory charged to this process's own cgroup v2 (``memory.stat`` ``anon``), or None. The budget is derived from the same cgroup's ``memory.high``/``memory.max``, and the kernel charges every process in the unit against it — execute_code kernels, terminal children — so a self-only reading under-counts by exactly the children's share (#110549). Anon, not ``memory.current``: the module's signal is heap, and reclaimable page cache is noise. Only for a *capped* cgroup: an uncapped one (a plain login session) is the whole user slice, and its budget is total RAM — self RSS stays the right scope there. r@NrArCrDc3N<"TFp\SRV 2, 4xK R#5i)zmemory.N)r<).0rOroots& r %_cgroup_anon_bytes..s$R/Q]4GA3-#788/s"%TFz memory.statr5r6 anonrE)rHrIrJrBranyjoinpathr8r:r/ ImportError splitlines partitionr9isdigitr) rBrLcappedtextlinekey_r#rys @r _cgroup_anon_bytesrs ||w ; cSjnSE*+R/RR/RROUt}}]+55{5K[]!s+  &=U[[]2244u: "  Z -s'DD-D(D#DD,+D,cV^8dQhRR/#r>r)r&s"r r'r's-rc\4pV'dV\,#^RIHpV!4pRF>pVP V4p\ V\ 4'gK,V^8gK5VR,u# ^RIp\ VP\P!44P4P\, 4# \dLbi;i \dR#i;i)u#Anonymous memory in MB (where cached transcripts live; file-backed pages are noise), or None. Own cgroup's ``memory.stat`` anon first — the scope the budget is charged against, so same-unit child processes count; then ``/proc/self/status``; psutil covers other platforms (total RSS only).)collect_memory_snapshotiN) rss_anon_kibrss_kib)rrehermes_cli.mem_trimrrpr)rrKrYProcessrVgetpid memory_inforss)chargedrsnapshotrkibrYs r read_anon_rss_mbrs !"G-''  ?*,.C,,s#C#s##ad{"/ 6>>"))+.::<@@=PQQ    s66CC% C0C3AC CC C'&C'c V^8dQhRRRR/#)r"rmrr$r%r)r&s"r r'r's X XC XD Xrc\VRR4\VRR4r!\V\4;'d"\V4;'dV\ V48#)a:True when the live transcript is fully on disk. Soft eviction rebuilds from the persisted session, so it is only safe once ``_last_flushed_db_idx`` (advanced only on a fully successful write) has caught up. Unknown shapes are *not* caught up: a skipped eviction costs memory, a wrong one costs the conversation. _session_messagesN_last_flushed_db_idx)getattrr)listr*len)rmmessagesflusheds& r transcript_persistence_caught_uprsK ':DA75RhjnCog h % W W''*: W Ww#h-?WWrrc ,V^8dQhRRRRRRRRRR /#) r"ordered_entrieszIterable[tuple[str, Any]] is_evictablezCallable[[str, Any], bool] max_evictionsrrr$zlist[tuple[str, Any]]r)r&s"r r'r's6.A[(+rc*\V4pV^8:g V'g.#\\V^4\V4^,4pV'dVRV)p.pVF:wrx\V4V8dV#V!Wx4'gK(VP Wx34K< V#)uoChoose which cached sessions to shed, least-recently-used first. ``ordered_entries`` must be LRU→MRU (the cache OrderedDict ``move_to_end``s on every hit). The batch is capped so one pass cannot stall the gateway. ``protect_recent`` is clamped to half the cache: a few huge transcripts can exhaust the budget alone, and a fixed guard would leave nothing to shed. N)rminrGrappend) rrrrentriesprotectplanrrms &$$$ r plan_pressure_evictionsrs?#G #na(#g,!*;r|offnonefalsedisabled)$r __future__rrVrH dataclassesrpathlibrtypingrrrrr r rdrfrrre frozensetrcr r*rr1r<rQr\rirtrrrrrrr rs# !AA"$ ? @  $222B8;A4. ?$*<4 X./r