+ ӃiTvRt^RIt^RIt^RIHtHtHtHt^RIH t ^RI H t H t ]P!]4tRtRtRtRtR tR t^tR t!R R 4tR#)aAutomatic context window compression for long conversations. Self-contained class with its own OpenAI client for summarization. Uses auxiliary model (cheap/fast) to summarize middle turns while protecting head and tail context. Improvements over v1: - Structured summary template (Goal, Progress, Decisions, Files, Next Steps) - Iterative summary updates (preserves info across multiple compactions) - Token-budget tail protection instead of fixed message count - Tool output pruning before LLM summarization (cheap pre-pass) - Scaled summary budget (proportional to compressed content) - Richer tool call/result detail in summarizer input N)AnyDictListOptional)call_llm)get_model_context_lengthestimate_messages_tokens_roughai[CONTEXT COMPACTION] Earlier turns in this conversation were compacted to save context space. The summary below describes work that was already completed, and the current session state may still reflect that work (for example, files may already be changed). Use the summary and the current state to continue from where things left off, and avoid repeating work:z[CONTEXT SUMMARY]:i皙?i.z/[Old tool output cleared to save context space]iXcha]tRt^5toRtR$V3RlRlltV3RlRltR%V3RlRlltV3R lR ltV3R lR lt V3R lRlt V3RlRlt V3RlRlt V3RlRlt ]V3RlRl4t]V3RlRl4tV3RlRltV3RlRltV3RlRltR%V3RlR lltR%V3R!lR"lltR#tVtR#)&ContextCompressoraCompresses conversation context when approaching the model's context limit. Algorithm: 1. Prune old tool results (cheap, no LLM call) 2. Protect head messages (system prompt + first exchange) 3. Protect tail messages by token budget (most recent ~20K tokens) 4. Summarize middle turns with structured LLM prompt 5. On subsequent compactions, iteratively update the previous summary Ncj<V^8dQhRS[RS[RS[RS[RS[RS[RS[RS[R S[R S[R ,R S[/ #) modelthreshold_percentprotect_first_nprotect_last_nsummary_target_ratio quiet_modesummary_model_overridebase_urlapi_keyconfig_context_lengthNprovider)strfloatintbool)format __classdict__s"5/home/ubuntu/hermes-agent/agent/context_compressor.py __annotate__ContextCompressor.__annotate__@s};:;:;:!;: ;:  ;: $ ;:;:!$;:;:;: #Tz;:;:c WnWnWnWnW nW0nW@n\R\VR44Vn W`n \WV V V R7Vn \VPV,4Vn^Vn\VPVP,4p Wn\\VPR,4\"4VnV'gg\&P)RYPVPV^d,VP^d,VP T ;'gRT;'gR4 RVn^Vn^Vn^VnT;'gRVnRVnR VnR#) g?g?)rrrrg?zContext compressor initialized: model=%s context_length=%d threshold=%d (%.0f%%) target_ratio=%.0f%% tail_budget=%d provider=%s base_url=%snoneFN)rrrrrrrmaxminrrrcontext_lengthrthreshold_tokenscompression_counttail_token_budget_SUMMARY_TOKENS_CEILINGmax_summary_tokensloggerinfo_context_probedlast_prompt_tokenslast_completion_tokenslast_total_tokens summary_model_previous_summary_summary_failure_cooldown_until) selfrrrrrrrrrrr target_tokenss &&&&&&&&&&&& r__init__ContextCompressor.__init__@sg    !2.,$'c2F.M$N!$6 g"7  !$D$7$7:K$K L!"D11D4M4MMN !."% ##d* +-D#  KK***D,A,A!C')B)BS)H&&""FH$6$6  %"#&'#!"399r1569,r"c6<V^8dQhRS[S[S[3,/#)r usagerrr)rrs"rr r!}s>>$sCx.>r"c VPR^4VnVPR^4VnVPR^4VnR#)z-Update tracked token usage from API response. prompt_tokenscompletion_tokens total_tokensN)getr2r3r4)r8r=s&&rupdate_from_response&ContextCompressor.update_from_response}s<"'))OQ"?&+ii0CQ&G#!&>1!=r"c&<V^8dQhRS[RS[/#)r r@return)rr)rrs"rr r!s//S/D/r"c BVeTM VPpW P8#)z3Check if context exceeds the compression threshold.)r2r*)r8r@tokenss&& rshould_compress!ContextCompressor.should_compresss#"/";AXAX....r"cL<V^8dQhRS[S[S[S[3,,RS[/#r messagesrG)rrrrr)rrs"rr r!s(77$tCH~2F747r"c 4\V4pW P8#)z>Quick pre-flight check using rough estimate (before API call).)rr*)r8rNrough_estimates&& rshould_compress_preflight+ContextCompressor.should_compress_preflights7A!6!666r"c6<V^8dQhRS[S[S[3,/#r rGr>)rrs"rr r!s  DcN r"c  RVPRVPRVPRVP'd/\^dVPVP, ^d,4M^RVP/#)z3Get current compression status for display/logging.r2r*r) usage_percentr+)r2r*r)r(r+)r8s&r get_statusContextCompressor.get_statussl !$"9"9  5 5 d11 `d`s`s`sSt'>'>ATAT'TWZ'Z\yz !7!7   r"c <V^8dQhRS[S[S[S[3,,RS[RS[S[S[S[S[3,,S[3,/#)r rNprotect_tail_countrG)rrrrrtuple)rrs"rr r!sIT#s(^,BE tDcN#S( )r"c V'gV^3#VUu.uFq3P4NK pp^p\V4V, p\V4FspWG,pVPR4R8wdK#VPRR4p V 'd V \8XdKJ\V 4^8gK\/VCR\/CWG&V^, pKu WE3#uupi)aReplace old tool result contents with a short placeholder. Walks backward from the end, protecting the most recent ``protect_tail_count`` messages. Older tool results get their content replaced with a placeholder string. Returns (pruned_messages, pruned_count). roletoolcontentr%)copylenrangerC_PRUNED_TOOL_PLACEHOLDER) r8rNrZmresultprunedprune_boundaryimsgr_s &&& r_prune_old_tool_results)ContextCompressor._prune_old_tool_resultssQ; $,-Hq&&(H-V'99~&A)Cwwv&(ggi,Gg)AA7|c!HsHI/GH ! '~!.sCcL<V^8dQhRS[S[S[S[3,,RS[/#r turns_to_summarizerGrrrrr)rrs"rr r!s- N N$tCH~:N NSV Nr"c \V4p\V\,4p\\\ W0P 44#)a Scale summary token budget with the amount of content being compressed. The maximum scales with the model's context window (5% of context, capped at ``_SUMMARY_TOKENS_CEILING``) so large-context models get richer summaries instead of being hard-capped at 8K tokens. )rr_SUMMARY_RATIOr'_MIN_SUMMARY_TOKENSr(r.)r8rncontent_tokensbudgets&& r_compute_summary_budget)ContextCompressor._compute_summary_budgets788JK^n45&F4K4K(LMMr"cL<V^8dQhRS[S[S[S[3,,RS[/#)r turnsrGrrrr)rrs"rr r!s(1"1"Dc3h,@1"S1"r"c  .pVEF/pVPRR4pVPR4;'gRpVR8XdVVPRR4p\V4R8dVR,R ,VR R ,pVPR V R V 24KVR 8XEdL\V4R8dVR,R ,VR R ,pVPR.4pV'd.pVFp \V \4'dqV PR/4p V PRR4p V PRR4p \V 4R8dV R,R,p VPRV RV R24K\ V RR 4p V 'd\ V RR4MRp VPRV R24K VRRP V4,R,, pVPRV 24EK\V4R8dVR,R ,VR R ,pVPRVP4 R V 24EK2 RP V4#)!zSerialize conversation turns into labeled text for the summarizer. Includes tool call arguments and result content (up to 3000 chars per message) so the summarizer can preserve specific details like file paths, commands, and outputs. r]unknownr_r%r^ tool_call_idi :NiNz ...[truncated]... Nz [TOOL RESULT z]: assistant tool_callsfunctionname? argumentsi:NiNz...z ()z(...)z [Tool calls:  z ]z [ASSISTANT]: [ i)rCraappend isinstancedictgetattrjoinupper) r8rxpartsrir]r_tool_idr~tc_partstcfnrargss && r_serialize_for_summary(ContextCompressor._serialize_for_summarys C7769-Dggi(..BGv~''."5w<$&%en/FFQUQVWG }WIS BC{"w<$&%en/FFQUQVWG WW\26 !H(%b$//!# B!7B#%66&##6D#%66+r#:D"4y3'+DzE'9$OObavQ,?@!(Z!>B?A72vs#;sD$OObe,<=)1DIIh4GG%OOG }WI677|d"!%.+BBWTU^S LL1TZZ\N#gY7 8OR{{5!!r"c\<V^8dQhRS[S[S[S[3,,RS[S[,/#rm)rrrrr)rrs"rr r!s1HHDc3h4HHXVY]Hr"c \P!4pW P8d*\P RVPV, 4R#VP V4pVP V4pVP'dRVP RV RV R2pM RV RV R 2pR R R R RRV/.RV^,/pVP'dVPVR&\R/VBpVP^,PPp\V\4'gV'd \V4MRpVP4p WnRVnVP!V 4# \"dE\P!4\$,Tn\&P(!R\$4R#\*dMp \P!4\$,Tn\&P(!RT \$4Rp ? R#Rp ? ii;i)uGenerate a structured summary of conversation turns. Uses a structured template (Goal, Progress, Decisions, Files, Next Steps) inspired by Pi-mono and OpenCode. When a previous summary exists, generates an iterative update instead of summarizing from scratch. Returns None if all attempts fail — the caller should drop the middle turns without a summary rather than inject a useless placeholder. z:Skipping context summary during cooldown (%.0fs remaining)NzYou are updating a context compaction summary. A previous compaction produced the summary below. New conversation turns have occurred since then and need to be incorporated. PREVIOUS SUMMARY: z NEW TURNS TO INCORPORATE: u) Update the summary using this exact structure. PRESERVE all existing information that is still relevant. ADD new progress. Move items from "In Progress" to "Done" when completed. Remove information only if it is clearly obsolete. ## Goal [What the user is trying to accomplish — preserve from previous summary, update if goal evolved] ## Constraints & Preferences [User preferences, coding style, constraints, important decisions — accumulate across compactions] ## Progress ### Done [Completed work — include specific file paths, commands run, results obtained] ### In Progress [Work currently underway] ### Blocked [Any blockers or issues encountered] ## Key Decisions [Important technical decisions and why they were made] ## Relevant Files [Files read, modified, or created — with brief note on each. Accumulate across compactions.] ## Next Steps [What needs to happen next to continue the work] ## Critical Context [Any specific values, error messages, configuration details, or data that would be lost without explicit preservation] Target ~u tokens. Be specific — include file paths, command outputs, error messages, and concrete values rather than vague descriptions. Write only the summary body. Do not include any preamble or prefix.zCreate a structured handoff summary for a later assistant that will continue this conversation after earlier turns are compacted. TURNS TO SUMMARIZE: u Use this exact structure: ## Goal [What the user is trying to accomplish] ## Constraints & Preferences [User preferences, coding style, constraints, important decisions] ## Progress ### Done [Completed work — include specific file paths, commands run, results obtained] ### In Progress [Work currently underway] ### Blocked [Any blockers or issues encountered] ## Key Decisions [Important technical decisions and why they were made] ## Relevant Files [Files read, modified, or created — with brief note on each] ## Next Steps [What needs to happen next to continue the work] ## Critical Context [Any specific values, error messages, configuration details, or data that would be lost without explicit preservation] Target ~u! tokens. Be specific — include file paths, command outputs, error messages, and concrete values rather than vague descriptions. The goal is to prevent the next assistant from repeating work or losing important details. Write only the summary body. Do not include any preamble or prefix.task compressionrNr]userr_ max_tokensrr%r&ztContext compression: no provider available for summary. Middle turns will be dropped without summary for %d seconds.zWFailed to generate context summary: %s. Further summary attempts paused for %d seconds.)time monotonicr7r/debugrurr6r5rchoicesmessager_rrstrip_with_summary_prefix RuntimeError!_SUMMARY_FAILURE_COOLDOWN_SECONDSloggingwarning Exception) r8rnnowsummary_budgetcontent_to_summarizeprompt call_kwargsresponser_summaryes && r_generate_summary#ContextCompressor._generate_summarysnn 55 5 LLL44s: 556HI#::;MN  ! ! ! < DI&GFR < DC#GFJ"  ffi@Anq0K !!!'+'9'9 G$.+.H&&q)1199Ggs++*1#g,rmmoG%, "36D 0,,W5 5 37>>3CFg3gD 0 OO.> ? 37>>3CFg3gD 0 OOB1    s3$%E$ AE$):E$$A H2H;H/3.!D6*FFr"c <V^8dQhRS[/#rTr)rrs"rr r!s+++r"c \V\4'dVPRR4#\VRR4;'gR#)zEExtract the call ID from a tool_call entry (dict or SimpleNamespace).idr%)rrrCr)rs&r_get_tool_call_id#ContextCompressor._get_tool_call_ids9 b$  66$# #r4$***r"cr<V^8dQhRS[S[S[S[3,,RS[S[S[S[3,,/#rMry)rrs"rr r!s8::T$sCx.-A:d4PSUXPX>FZ:r"c  B\4pVFipVPR4R8XgKVPR4;'g.F/pVPV4pV'gKVPV4K1 Kk \4pVFGpVPR4R8XgKVPR4pV'gK6VPV4KI Wb, pV'dqVUu.uF3pVPR4R8XdVPR4V9dK1VNK5 ppVP'g \ P R\V44W&, p V 'd.p VF~pV PV4VPR4R8XgK,VPR4;'g.F3pVPV4pWY9gKV PRRRRRV/4K5 K T pVP'g \ P R \V 44V#uupi) aIFix orphaned tool_call / tool_result pairs after compression. Two failure modes: 1. A tool *result* references a call_id whose assistant tool_call was removed (summarized/truncated). The API rejects this with "No tool call found for function call output with call_id ...". 2. An assistant message has tool_calls whose results were dropped. The API rejects this because every tool_call must be followed by a tool result with the matching call_id. This method removes orphaned results and inserts stub results for orphaned calls so the message list is always well-formed. r]r}r~r^r|z9Compression sanitizer: removed %d orphaned tool result(s)r_u@[Result from earlier conversation — see context summary above]z3Compression sanitizer: added %d stub tool result(s)) setrCraddrr/r0rar) r8rNsurviving_call_idsrircidresult_call_idsorphaned_resultsrdmissing_resultspatcheds && r_sanitize_tool_pairs&ContextCompressor._sanitize_tool_pairss#&%Cwwv+-'',/5525B004Cs*..s36 #uCwwv&(ggn-3#'', +? ##af /AEE.4IM]4]8 ??? WY\]mYno-> ,.Gs#776?k1!ggl399r9"44R81#NN & )+m .,: H??? QSVWfSgh5s %.HHcR<V^8dQhRS[S[S[S[3,,RS[RS[/#r rNidxrGro)rrs"rr r!s0T#s(^0D3SVr"c tV\V48d(W,PR4R8Xd V^, pK7V#)zPush a compress-start boundary forward past any orphan tool results. If ``messages[idx]`` is a tool result, slide forward until we hit a non-tool message so we don't start the summarised region mid-group. r]r^rarC)r8rNrs&&&r_align_boundary_forward)ContextCompressor._align_boundary_forwards2 CM!hm&7&7&?6&I 1HC r"cR<V^8dQhRS[S[S[S[3,,RS[RS[/#rro)rrs"rr r!s0d38n1ECTWr"c *V^8:gV\V48dV#V^, pV^8d(W,PR4R8Xd V^,pK.V^8d<W,PR4R8Xd W,PR4'dTpV#)aPull a compress-end boundary backward to avoid splitting a tool_call / result group. If the boundary falls in the middle of a tool-result group (i.e. there are consecutive tool messages before ``idx``), walk backward past all of them to find the parent assistant message. If found, move the boundary before the assistant so the entire assistant + tool_results group is included in the summarised region rather than being split (which causes silent data loss when ``_sanitize_tool_pairs`` removes the orphaned tail results). r]r^r}r~r)r8rNrchecks&&& r_align_boundary_backward*ContextCompressor._align_boundary_backwards !8sc(m+JaqjX_008FB QJE A:(/--f5DI\I\]iIjIjC r"cf<V^8dQhRS[S[S[S[3,,RS[RS[R,RS[/#)r rNhead_end token_budgetNrGro)rrs"rr r!s@1*1*T#s(^,1*8;1*Dj1* 1*r"c Vf VPp\V4pVPp^pTp\V^, V^, R4FpW,p V P R4;'gRp \V 4\ ,^ ,p V P R4;'g.FZp \ V \4'gKV P R/4P RR4p V \V 4\ ,, p K\ Wk,V8dWH, V8dMWk, pTpK WE, pW~8dTpWr8:dTpVPW4p\Wr^,4#)aWalk backward from the end of messages, accumulating tokens until the budget is reached. Returns the index where the tail starts. ``token_budget`` defaults to ``self.tail_token_budget`` which is derived from ``summary_target_ratio * context_length``, so it scales automatically with the model's context window. Never cuts inside a tool_call/result group. Falls back to the old ``protect_last_n`` if the budget would protect fewer messages. r_r%r~rr) r,rarrbrC_CHARS_PER_TOKENrrrr')r8rNrrnmin_tail accumulatedcut_idxrhrir_ msg_tokensrr fallback_cuts&&&& r_find_tail_cut_by_tokens*ContextCompressor._find_tail_cut_by_tokenssC  11L M&& q1uhlB/A+Cggi(..BGW)99B>Jggl+11r1b$''66*b155k2FD#d)/?"??J2',6AEh;N  %KG0|  !"G  "G//B7qL))r"c x<V^8dQhRS[S[S[S[3,,RS[RS[S[S[S[3,,/#)r rNcurrent_tokensrGro)rrs"rr r!5sGCCd38n!5CsCVZ[_`ceh`h[iVjCr"c \V4pW0PVP,^,8:dNVP'g:\P RVVPVP,^,4V#V'dTMVP ;'g \V4pVPWP^,R7wrV'd)VP'g\PRV4VPpVPW4pVPW4pWg8dV#WVpVP'g\PRVVP4\PRVPVP^d,VP4W7, p \PRV^,V\V4VV 4VPV4p .p \!V4F}p W,P#4p V ^8XdMV P%R4R8Xd7VP&^8Xd&V P%R 4;'gR R ,V R &V P)V 4K R pV 'dV^8d W^, ,P%RR 4MR pWs8dW,P%RR 4MR pVR9dR pMRpVV8XdVR 8XdRMR pVV8wdTpMRpV'gV P)RVR V /4M'VP'g\P+R4\!Ws4Fhp W,P#4p V'd7W8Xd1V P%R 4;'gR pV R,V,V R &R pV P)V 4Kj V;P&^, unVP-V 4p VP'gV\V 4pVV, p\PRV\V 4V4\PRVP&4V #)aCompress conversation messages by summarizing middle turns. Algorithm: 1. Prune old tool results (cheap pre-pass, no LLM call) 2. Protect head messages (system prompt + first exchange) 3. Find tail boundary by token budget (~20K tokens of recent context) 4. Summarize middle turns with structured LLM prompt 5. On re-compression, iteratively update the previous summary After compression, orphaned tool_call / tool_result pairs are cleaned up so the API never receives mismatched IDs. z-Cannot compress: only %d messages (need > %d))rZz-Pre-compression: pruned %d old tool result(s)z9Context compression triggered (%d tokens >= %d threshold)z,Model context limit: %d tokens (%.0f%% = %d)zISummarizing turns %d-%d (%d turns), protecting %d head + %d tail messagesr]systemr_r%z [Note: Some earlier conversation turns have been compacted into a handoff summary to preserve context space. The current session state may still reflect earlier work, so build on that summary and state rather than re-doing work.]Frr}TuCNo summary model available — middle turns dropped without summaryrz0Compressed: %d -> %d messages (~%d tokens saved)zCompression #%d complete)r}r^)rarrrr/rr2rrjr0rrr*r)rrrbr`rCr+rrr)r8rNr n_messagesdisplay_tokens pruned_countcompress_start compress_endrn tail_msgsr compressedrhri_merge_summary_into_taillast_head_rolefirst_tail_role summary_roleflippedoriginal new_estimatesaved_estimates&&& rcompressContextCompressor.compress5s] --0C0CCaG G???C((4+>+>>B O+9t?V?V@C@CZxzB[C"&!=!= )<)"   KKG V--55hO44XN  )O%\B KKK%%  KK>##&&,%%   #1I KK["&'  (();< ~&A+""$CAv#''&/X5$:P:PTU:UWWY'--2BBI   c "'$) Q_bcQcXq&89==ffMioNLXLeh488HkqO!66% * .)5)?+Vn,#*L 04,+!!6<G"LM??? bc|0A+""$C'A,=779-33!(6!1H!<I+0(   c " 1 !#..z: 9*EL+l:N KKBJ   KK2D4J4J Kr")r1r6r7rrr+r)r3r2r4r.rrrrrr5rr,rr*) g?r FNr%r%Nr%)N)__name__ __module__ __qualname____firstlineno____doc__r:rDrJrQrWrjrurr staticmethodrrrrrrr__static_attributes____classdictcell__)rs@rr r 5s;:;:z>> // 77   H N N1"1"fHHTGG++ ::x81*1*nCCCr"r )rrrtypingrrrragent.auxiliary_clientragent.model_metadatarr getLoggerrr/rrrrrqr-rcrrr rr"rr sz  ,,+   8 $- M$'!C C r"