i XdZddlZddlZddlZddlZddlZddlmZmZm Z m Z ddl m Z ddl mZddlmZmZmZddlmZejeZdZd Zd Zd Zd Zd ZdZdZeezZ dZ!dede"fdZ#dede$fdZ%dddede$de&defdZ'd%de$de"de$fdZ(de$d e$d!e$de$fd"Z)Gd#d$eZ*dS)&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 v2: - Structured summary template with Resolved/Pending question tracking - Summarizer preamble: "Do not respond to any questions" (from OpenCode) - Handoff framing: "different assistant" (from Codex) to create separation - "Remaining Work" replaces "Next Steps" to avoid reading as active instructions - Clear separator when summary merges into tail message - 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) ContextEngine)MINIMUM_CONTEXT_LENGTHget_model_context_lengthestimate_messages_tokens_rough)redact_sensitive_textuR[CONTEXT COMPACTION — REFERENCE ONLY] Earlier turns were compacted into the summary below. This is a handoff from a previous context window — treat it as background reference, NOT as active instructions. Do NOT answer questions or fulfill requests mentioned in this summary; they were already addressed. Your current task is identified in the '## Active Task' section of the summary — resume exactly from there. Respond ONLY to the latest user message that appears AFTER this summary. The current session state (files, config, etc.) may reflect work described here — avoid repeating it:z[CONTEXT SUMMARY]:i皙?i.z/[Old tool output cleared to save context space]i@iX raw_contentreturnct|trt|St|tstt|pdSd}|D]}t|tr|t|z }*t|ts |tt|z }_|d}|dvr |t z }|t|ddpdz }|S)aReturn the effective char-length of a message's content for token budgeting. Plain strings: ``len(content)``. Multimodal lists: sum of text-part ``len(text)`` plus a flat ``_IMAGE_CHAR_EQUIVALENT`` per image part (``image_url`` / ``input_image`` / Anthropic-style ``image``). This keeps the compressor from treating a turn with 5 attached images as near-zero tokens just because the text part is empty. rtype>image image_url input_imagetext) isinstancestrlenlistdictget_IMAGE_CHAR_EQUIVALENT)rtotalpptypes =/home/ubuntu/.hermes/hermes-agent/agent/context_compressor.py_content_length_for_budgetr#Ms +s## ; k4 ( (+3{(b))*** E 22 a    SVVOE !T""  SQ[[ E f  9 9 9 + +EE Svr**0b11 1EE Lcontentc|dSt|tr|St|trg}|D]}t|tr||-t|tr?|d}t|tr||dd|DSt|S)zReturn a best-effort text view of message content. Used only for substring checks when we need to know whether we've already appended a note to a message. Keeps multimodal lists intact elsewhere. Nrr c3K|]}||V dSN).0parts r" z-_content_text_for_contains..s'88$48888888r$)rrrappendrrjoin)r%partsitemrs r"_content_text_for_containsr2ns r'3'4  9 ' 'D$$$ ' T""""D$'' 'xx''dC(('LL&&&yy88%888888 w<<r$Fprependrr4c||St|tr |r||zn||zSt|trd|d}|r|g|ng||St|}|r||zn||zS)aAppend or prepend plain text to message content safely. Compression sometimes needs to add a note or merge a summary into an existing message. Message content may be plain text or a multimodal list of blocks, so direct string concatenation is not always safe. Nr)rr)rrr)r%rr4 text_blockrendereds r"_append_text_to_contentr8s '3=!(zB_truncate_tool_call_args_json.._shrink..s':::daAwwqzz:::r$c&g|] }|Sr*r*)r+rArBs r" zB_truncate_tool_call_args_json.._shrink..s!,,,1GGAJJ,,,r$)rrrritemsr)r=rBr;s r"rBz._truncate_tool_call_args_json.._shrinks c3   3xx*$$;J;'*:::J c4  ;::::ciikk::: : c4  -,,,,,,, , r$F) ensure_ascii)jsonloads ValueError TypeErrorrdumps)r:r;parsedshrunkenrBs ` @r"_truncate_tool_call_args_jsonrOs2D!!  "  S S       wvH :hU 3 3 33s // tool_name tool_args tool_contentc  |rtj|ni}n#tjtf$ri}YnwxYw|pd}t |}|r|ddznd}|dkrp|dd}t |dkr |dd d z}tj d |}|r| dnd } d |d| d|dS|dkr9|dd } |dd} d| d| d|ddS|dkrb|dd } |dr,|ddddznd } d| d| dS|dkr|dd } |dd} |d d}tj d!|}|r| dnd }d"|d#| d$| d%|d& S|d'kr9|dd } |d(d)}d*|d+| d|dd,S|d-vrI|d.d}|d/d}|rd0|n|rd1|nd}d2|d3|d|ddS|d4kr |d5d }d6|d7|dd,S|d8kr|d9g}t|tr |r|dnd }t|tr,t |dkr|d:t |dz d;z }d<|d|ddS|d=kr@|d>d}t |d?kr |dd@d z}dA|d7|dd,S|dBkrh|dCpddd? dd0}t |dCdd?kr|d z }dD|dE|dFS|dGvr#|dHd }d2|dI|d|ddS|dJkr(|dKdddL}dM|d7|ddS|dNkr4|dOd }|d d }dP|dQ|S|dRkrdSS|dTkrdUS|dVkrdW|ddS|dXkr|dOd }dY|S|dZkr4|dOd }|d[d }d\|d]|Sd}t|dd^D]'\}} t| dd_}!|d0|d`|!z }(d2|d3|d|dd,S)aaCreate an informative 1-line summary of a tool call + result. Used during the pre-compression pruning pass to replace large tool outputs with a short but useful description of what the tool did, rather than a generic placeholder that carries zero information. Returns strings like:: [terminal] ran `npm test` -> exit 0, 47 lines output [read_file] read config.py from line 1 (1,200 chars) [search_files] content search for 'compress' in agent/ -> 12 matches rr'rterminalcommandPNM...z"exit_code"\s*:\s*(-?\d+)?z[terminal] ran `z ` -> exit z, z lines output read_filepathoffsetz[read_file] read z from line z (,z chars) write_filer%z[write_file] wrote to z lines) search_filespattern.targetz"total_count"\s*:\s*(\d+)z[search_files] z search for 'z' in z -> z matchespatchmodereplacez[patch] z in z chars result))browser_navigate browser_clickbrowser_snapshot browser_typebrowser_scrollbrowser_visionurlref z ref=[] web_searchqueryz[web_search] query='z' ( web_extracturlsz (+z more)z[web_extract] delegate_taskgoal<9z[delegate_task] ' execute_codecodez[execute_code] `z` (z lines output)) skill_view skills_list skill_managenamez] name=vision_analyzequestion2z[vision_analyze] 'memoryactionz [memory] z on todoz[todo] updated task listclarifyz[clarify] asked user a questiontext_to_speechz"[text_to_speech] generated audio (cronjobz [cronjob] process session_idz [process] z session=(=)rHrIJSONDecodeErrorrKrstripcountrresearchgrouprrrfrFr)"rPrQrRr:r% content_len line_countcmd exit_match exit_coder\r] written_linesrarc match_countrrermrndetailrsruurl_descrw code_previewrrrsid first_argr@rAsvs" r"_summarize_tool_resultrs (19tz)$$$r  ) , bGg,,K,3MMOOBt$$q((JJhhy"%% s88b==crc(U"CY;WEE +5>J$$Q'''3 W#WWWWjWWWWKxx$$(A&&T4TTFTTkTTTTTL  xx$$CG88ICVCV_B//55d;;a??\_ FFF FFFFN""((9c**xx$$(I..i >9u$DI4IIKIIIIIN""((.B4<deded ed ed ed eddfd Z d?dede dedede de ded ed ededzd ed efdZ de eeffdZd@dede fdZ d@dee eefd ed!edzdeee eefeffd"Zd#ee eefdefd$Zd%Zd&Zd'Zd'Zd(Zd)ee eefdefd*Zd@d#ee eefd+edeefd,Zed-edefd.Zedefd/Zdee eefdee eeffd0Z dee eefd1edefd2Z!dee eefd1edefd3Z"dee eefd4edefd5Z#dee eefd6ed4edefd7Z$ d@dee eefd4ed8edzdefd9Z%dee eefde fd:Z&dAdee eefd;ed+edee eeffd<Z'xZ(S)BContextCompressoruDefault context engine — compresses conversation context via lossy summarization. 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 rcdS)N compressorr*)selfs r"rzContextCompressor.nameJs|r$Nctd|_d|_d|_d|_d|_d|_d|_d|_ d|_ d|_ dS)z/Reset all per-session state for /new or /reset.FNrY@) superon_session_reset_context_probed_context_probe_persistable_previous_summary_last_summary_error_last_summary_dropped_count_last_summary_fallback_used_last_aux_model_failure_error_last_aux_model_failure_model_last_compression_savings_pct_ineffective_compression_count)r __class__s r"rz"ContextCompressor.on_session_resetNso   """$*/'!%#' +,(+0(-1*-1*-2*./+++r$rmodelcontext_lengthbase_urlapi_keyproviderapi_modecR||_||_||_||_||_||_t t||jzt|_ t|j |j z}||_ tt|dzt|_dS)z>Update model info after a model switch or fallback activation.皙?N)rrrrrrmaxintthreshold_percentr threshold_tokenssummary_target_ratiotail_token_budgetmin_SUMMARY_TOKENS_CEILINGmax_summary_tokens)rrrrrrr target_tokenss r" update_modelzContextCompressor.update_model\s        , # !77 8 8 "! !  D1D4MMNN !."% % & &(?# # r$?r Frprotect_first_nprotect_last_nr quiet_modesummary_model_overrideconfig_context_lengthc J||_||_| |_| |_| |_||_||_||_tdt|d|_ ||_ t||| | | |_ tt|j |zt|_d|_t|j|j z} | |_tt|j dzt&|_|s@t*d||j |j|dz|j dz|j| pd|pd d |_d|_d|_|pd |_d|_d |_d|_d |_d|_d|_ d |_!d|_"d|_#dS) Ng?g?)rrrrrrzContext compressor initialized: model=%s context_length=%d threshold=%d (%.0f%%) target_ratio=%.0f%% tail_budget=%d provider=%s base_url=%sdnoneFrr)$rrrrrrrrrrrrr rrr rcompression_countrrrloggerinforlast_prompt_tokenslast_completion_tokens summary_modelrrr_summary_failure_cooldown_untilrrrrr)rrrrrrrrrrrrrrs r"__init__zContextCompressor.__init__xs        !2.,$'c2F.M.M$N$N!$6 Hg"7   !$ #&77 8 8 "! ! "#D1D4MMNN !."% #d* + +-D# #   KK*t*D,A!C')BS)H&"FH$6    %"#&'#39r1549*34+69,26 12(16( =A*<@***r$usagecr|dd|_|dd|_dS)z-Update tracked token usage from API response. prompt_tokensrcompletion_tokensN)rrr)rrs r"update_from_responsez&ContextCompressor.update_from_responses5"'))OQ"?"?&+ii0CQ&G&G###r$rc||n|j}||jkrdS|jdkr)|js td|jdSdS)aCheck if context exceeds the compression threshold. Includes anti-thrashing protection: if the last two compressions each saved less than 10%, skip compression to avoid infinite loops where each pass removes only 1-2 messages. NFruCompression skipped — last %d compressions saved <10%% each. Consider /new to start a fresh session, or /compress for focused compression.T)rrrrrwarning)rrtokenss r"should_compressz!ContextCompressor.should_compressso#0";AX D) ) )5  .! 3 3? /7  5tr$messagesprotect_tail_countprotect_tail_tokensc |s|dfSd|D}d}i}|D]}|ddkr|dpgD]}t|tr\|dd} |di} | d d | d df|| <st|ddpd} t|dd } | rt| d d nd } | rt| d dnd} | | f|| <Ɍ|R|dkrKd} t |}t |t |}t t |d z ddD]}||}|dpd}t|}|tzdz}|dpgD][}t|trD|did d}|t |tzz }\| |z|krt ||z |kr|}n| |z } |}t|t ||z }nt ||z }i}t t |d z ddD]}||}|ddkr$|dpd}t|trQt |dkretj | ddd d}||vri|ddi||<|d z }||ddf||<t |D]}||}|ddkr$|dd}t|trP|r |tkr^|drtt |dkrP|dd}||d\}}t#|||}i|d|i||<|d z }t |D]}||}|ddks|ds9g}d}|dD]}t|trh|did d}t |dkr+t%|}||kri|di|dd |ii}d}|||r i|d|i||<||fS)azReplace old tool result contents with informative 1-line summaries. Instead of a generic placeholder, generates a summary like:: [terminal] ran `npm test` -> exit 0, 47 lines output [read_file] read config.py from line 1 (3,400 chars) Also deduplicates identical tool results (e.g. reading the same file 5x keeps only the newest full copy) and truncates large tool_call arguments in assistant messages outside the protected tail. Walks backward from the end, protecting the most recent messages that fall within ``protect_tail_tokens`` (when provided) OR the last ``protect_tail_count`` messages (backward-compatible default). When both are given, the token budget takes priority and the message count acts as a hard minimum floor. Returns (pruned_messages, pruned_count). rc6g|]}|Sr*)copy)r+ms r"rEz=ContextCompressor._prune_old_tool_results..s ---q!&&((---r$role assistant tool_callsidrfunctionrunknown argumentsNrTr% toolr9zutf-8rf)errors u>[Duplicate tool output — same content as a more recent call] tool_call_idrZz[Duplicate tool output)rrFiT)rrrgetattrrrranger#_CHARS_PER_TOKENrrhashlibmd5encode hexdigest_PRUNED_TOOL_PLACEHOLDER startswithrrOr.) rrrrresultprunedcall_id_to_toolmsgtccidfnrargs_str accumulatedboundary min_protectirr msg_tokensr:prune_boundarycontent_hashesr%hcall_idrPrQsummarynew_tcsmodifiednew_argss r"_prune_old_tool_resultsz)ContextCompressor._prune_old_tool_resultss. Q; --H----/ @ @Cwwv+--'',//52 @ @B!"d++ @ ffT2..VVJ3302vy0I0I266R]_aKbKb/c,,%b$339r$RT::ACRwr69===CE#M72{B#?#?#?204h/?,,  */BQ/F/FK6{{H0#f++>>K3v;;?B33  Qi!ggi006B 8EE (,<,>,D"EEGv~~''."55w<<$"333%&9t'9&9:=TTW^`d`r_r_s_sWttG BWBBBBCCC{""w<<$"333%&9t'9&9:=TTW^`d`r_r_s_sWttG WW\266 P!H( > >%b$// >!# B!7!7B#%66&##6#6D#8 R9P9P#Q#QD"4yy4+>>>'+,AT-A,A'BU'J$OO,?,?,?,?,?,?@@@@!(Z!>!>B?A#J72vs#;#;#;sD$OO,<,<,<,<====1DIIh4G4GG%OOG 6W667777||d///!"54#5"569PPSZ\`\n[n[o[oSpp LL7TZZ\\77g77 8 8 8 8{{5!!!r$ focus_topicc tj}||jkr%td|j|z dS||}||}d}d|d}|jr|d|jd|d|}n |d |d |}|r |d |d |d z } d|j|j |j |j |j dd|dgt|dzd} |jr |j| d<td+i| } | jdjj} t'| t(s| rt)| nd} t+| } | |_d|_d|_d|_|| S#t4$rEtjt6z|_d|_t9jdt6YdSt<$r} t?| ddpt?t?| dddd}t)|  }|dvp d|vpd|vpd |v}|r|jr|j|jkrt?|d!dsd"|_t9jd#|j| |jt)| p | j!j"}tG|d$kr|dd%$d&z}||_%|j|_&d|_d|_|'||'cYd} ~ S|jr|j|jkrt?|d!dsd"|_t9jd(|j| |jt)| p | j!j"}tG|d$kr|dd%$d&z}||_%|j|_&d|_d|_|'||'cYd} ~ Sd)}tj|z|_t)| p | j!j"}tG|d$kr|dd%$d&z}||_t9jd*| |Yd} ~ dSd} ~ wwxYw),uWGenerate a structured summary of conversation turns. Uses a structured template (Goal, Progress, Decisions, Resolved/Pending Questions, Files, Remaining Work) with explicit preamble telling the summarizer not to answer questions. When a previous summary exists, generates an iterative update instead of summarizing from scratch. Args: focus_topic: Optional focus string for guided compression. When provided, the summariser prioritises preserving information related to this topic and is more aggressive about compressing everything else. Inspired by Claude Code's ``/compact``. 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)NuYou are a summarization agent creating a context checkpoint. Your output will be injected as reference material for a DIFFERENT assistant that continues the conversation. Do NOT respond to any questions or requests in the conversation — only output the structured summary. Do NOT include any preamble, greeting, or prefix. Write the summary in the same language the user was using in the conversation — do not translate or switch to English. NEVER include API keys, tokens, passwords, secrets, credentials, or connection strings in the summary — replace any that appear with [REDACTED]. Note that the user had credentials present, but do not preserve their values.u## Active Task [THE SINGLE MOST IMPORTANT FIELD. Copy the user's most recent request or task assignment verbatim — the exact words they used. If multiple tasks were requested and only some are done, list only the ones NOT yet completed. The next assistant must pick up exactly here. Example: "User asked: 'Now refactor the auth module to use JWT instead of sessions'" If no outstanding task exists, write "None."] ## Goal [What the user is trying to accomplish overall] ## Constraints & Preferences [User preferences, coding style, constraints, important decisions] ## Completed Actions [Numbered list of concrete actions taken — include tool used, target, and outcome. Format each as: N. ACTION target — outcome [tool: name] Example: 1. READ config.py:45 — found `==` should be `!=` [tool: read_file] 2. PATCH config.py:45 — changed `==` to `!=` [tool: patch] 3. TEST `pytest tests/` — 3/50 failed: test_parse, test_validate, test_edge [tool: terminal] Be specific with file paths, commands, line numbers, and results.] ## Active State [Current working state — include: - Working directory and branch (if applicable) - Modified/created files with brief note on each - Test status (X/Y passing) - Any running processes or servers - Environment details that matter] ## In Progress [Work currently underway — what was being done when compaction fired] ## Blocked [Any blockers, errors, or issues not yet resolved. Include exact error messages.] ## Key Decisions [Important technical decisions and WHY they were made] ## Resolved Questions [Questions the user asked that were ALREADY answered — include the answer so the next assistant does not re-answer them] ## Pending User Asks [Questions or requests from the user that have NOT yet been answered or fulfilled. If none, write "None."] ## Relevant Files [Files read, modified, or created — with brief note on each] ## Remaining Work [What remains to be done — framed as context, not instructions] ## Critical Context [Any specific values, error messages, configuration details, or data that would be lost without explicit preservation. NEVER include API keys, tokens, passwords, or credentials — write [REDACTED] instead.] Target ~u tokens. Be CONCRETE — include file paths, command outputs, error messages, line numbers, and specific values. Avoid vague descriptions like "made some changes" — say exactly what changed. Write only the summary body. Do not include any preamble or prefix.z You 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 completed actions to the numbered list (continue numbering). Move items from "In Progress" to "Completed Actions" when done. Move answered questions to "Resolved Questions". Update "Active State" to reflect current state. Remove information only if it is clearly obsolete. CRITICAL: Update "## Active Task" to reflect the user's most recent unfulfilled request — this is the most important field for task continuity. a Create a structured handoff summary for a different assistant that will continue this conversation after earlier turns are compacted. The next assistant should be able to understand what happened without re-reading the original turns. TURNS TO SUMMARIZE: z Use this exact structure: z FOCUS TOPIC: "z" The user has requested that this compaction PRIORITISE preserving all information related to the focus topic above. For content related to "u", include full detail — exact values, file paths, command outputs, error messages, and decisions. For content NOT related to the focus topic, summarise more aggressively (brief one-liners or omit if truly irrelevant). The focus topic sections should receive roughly 60-70% of the summary token budget. Even for the focus topic, NEVER preserve API keys, tokens, passwords, or credentials — use [REDACTED]. compression)rrrrruserrr%g?)task main_runtimer max_tokensrrrrFz$no auxiliary LLM provider configuredztContext compression: no provider available for summary. Middle turns will be dropped without summary for %d seconds. status_coderesponse)iimodel_not_foundzdoes not existzno available channel_summary_model_fallen_backTzWSummary model '%s' not available (%s). Falling back to main model '%s' for compression.rYr.zMSummary model '%s' failed (%s). Retrying on main model '%s' before giving up.rxzWFailed to generate context summary: %s. Further summary attempts paused for %d seconds.r*)(time monotonicrrdebugrr-rrrrrrrrrchoicesmessager%rrr rr9r_with_summary_prefix RuntimeError!_SUMMARY_FAILURE_COOLDOWN_SECONDSloggingr Exceptionrlowerr__name__rrstriprr_generate_summary)rrr.nowsummary_budgetcontent_to_summarize_summarizer_preamble_template_sectionsprompt call_kwargsr7r%re_status_err_str_is_model_not_found _err_text_transient_cooldownerr_texts r"rJz#ContextCompressor._generate_summarys$n 5 5 5 LLL4s:   4556HII#::;MNN ,  9Gn o9G9G9Gv  ! .         FF/       F  v vvvNYvvv vF y %!Z $ $ #| $ !!'-@@A!.3"677  K! :'+'9 G$..+..H&q)19Ggs++ :*19#g,,,r,GMMOO<3C3CFg3gD 0'MD $ O.> ? ? ?44R R R  a55sJX\A]A]_lnr9s9sG1vv||~~H:%6$06#x/6*X5 $ [& [&$*44&BEJJ537/G&4: FFLLNNBak.B y>>C'' )$3$ 6 6 8 85 @I5>2595G2%'"7:4--.@k-ZZZZZZZZ" [&$*44&BEJJ537/D&4: FFLLNNBak.B y>>C'' )$3$ 6 6 8 85 @I5>2595G2%'"7:4--.@k-ZZZZZZZZ#% 37>3C3CFY3YD 01vv||~~=)=H8}}s""#DSD>0022U:'/D $ OB#     44444eR s;)C'FA S  S )E S3S 9C&SS %BSS rc|pd}ttfD]B}||r+|t |d}nC|r td|ntS)z@Normalize summary text to the current compaction handoff format.rNr')rLEGACY_SUMMARY_PREFIXSUMMARY_PREFIXrrlstrip)rrprefixs r"rBz&ContextCompressor._with_summary_prefixs 2$$&&,n=  Fv&& CKKLL)0022 /3F.**D***Fr$ct|tr.|ddp|ddpdSt|ddpt|ddpdS)zEExtract the call ID from a tool_call entry (dict or SimpleNamespace).rrr)rrrr)rs r"_get_tool_call_idz#ContextCompressor._get_tool_call_idsk b$   C66)R((BBFF4,<,<B Br9b))HWRr-B-BHbHr$c t}|D]a}|ddkrF|dpgD].}||}|r||/bt}|D]G}|ddkr,|d}|r||H||z r= fd|D}|js(t dt ||z }|rg}|D]|}|||ddkrL|dpgD]4}||}||vr|dd|d 5}|}|js(t d t||S) aFix orphaned tool_call / tool_result pairs after compression. Two failure modes: 1. A tool *result* references a call_id whose assistant tool_call was removed (summarized/truncated). The API rejects this with "No tool call found for function call output with call_id ...". 2. An assistant message has tool_calls whose results were dropped. The API rejects this because every tool_call must be followed by a tool result with the matching call_id. This method removes orphaned results and inserts stub results for orphaned calls so the message list is always well-formed. rrrrrctg|]4}|ddkr|dv2|5S)rrr)r)r+rorphaned_resultss r"rEz:ContextCompressor._sanitize_tool_pairs..sOf //AEE.4I4IM]4]4]4]4]4]r$z9Compression sanitizer: removed %d orphaned tool result(s)u@[Result from earlier conversation — see context summary above])rr%rz3Compression sanitizer: added %d stub tool result(s)) setrr_addrrrrr.) rrsurviving_call_idsrrrresult_call_idsmissing_resultspatchedrbs @r"_sanitize_tool_pairsz&ContextCompressor._sanitize_tool_pairss9#&%% 4 4Cwwv+--'',//5244B0044C4*..s333"uu - -Cwwv&((ggn---#'',,,+-??  p#H? p WY\]mYnYnooo->  i,.G  s###776??k11!ggl339r"44R88/11#NN(.+m03,, H? i QSVWfSgSghhhr$idxc|t|krV||ddkr7|dz }|t|kr||ddk7|S)zPush a compress-start boundary forward past any orphan tool results. If ``messages[idx]`` is a tool result, slide forward until we hit a non-tool message so we don't start the summarised region mid-group. rrrTrr)rrrjs r"_align_boundary_forwardz)ContextCompressor._align_boundary_forward"sn CMM!!hsm&7&7&?&?6&I&I 1HCCMM!!hsm&7&7&?&?6&I&I r$ch|dks|t|kr|S|dz }|dkrI||ddkr*|dz}|dkr||ddk*|dkr<||ddkr||dr|}|S)aVPull 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). rrTrrrrrl)rrrjchecks r"_align_boundary_backwardz*ContextCompressor._align_boundary_backward,s !88sc(mm++JaqjjXe_0088FBB QJEqjjXe_0088FBB A::(5/--f55DDRWI\I\]iIjIjDC r$head_endctt|dz |dz dD]%}||ddkr|cS&dS)zMReturn the index of the last user-role message at or after *head_end*, or -1.rTrrr1)rrr)rrrqrs r"_find_last_user_message_idxz-ContextCompressor._find_last_user_message_idxHs\s8}}q((Q,;;  A{v&&&001rr$cut_idxc|||}|dkr|S||kr|S|jstd||t ||dzS)uGuarantee the most recent user message is in the protected tail. Context compressor bug (#10896): ``_align_boundary_backward`` can pull ``cut_idx`` past a user message when it tries to keep tool_call/result groups together. If the last user message ends up in the *compressed* middle region the LLM summariser writes it into "Pending User Asks", but ``SUMMARY_PREFIX`` tells the next model to respond only to user messages *after* the summary — so the task effectively disappears from the active context, causing the agent to stall, repeat completed work, or silently drop the user's latest request. Fix: if the last user-role message is not already in the tail (``messages[cut_idx:]``), walk ``cut_idx`` back to include it. We then re-align backward one more time to avoid splitting any tool_call/result group that immediately precedes the user message. rzjAnchoring tail cut to last user message at index %d (was %d) to prevent active-task loss after compressionrT)rsrrr?r)rrrtrq last_user_idxs r"!_ensure_last_user_message_in_tailz3ContextCompressor._ensure_last_user_message_in_tailQs},888LL 1  N G # #N  LLI    =(Q,///r$ token_budgetc2||j}t|}||z dkrtd||z dz nd}t|dz}d}|}t |dz |dz dD]} || } | dpd} t | } | tzd z} | d pgD][}t|trD|d id d}| t|tzz } \|| z|kr || z |krn|| z }| }||z }||kr|}||krt||dz}| ||}| |||}t||dzS) 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. Token budget is the primary criterion. A hard minimum of 3 messages is always protected, but the budget is allowed to exceed by up to 1.5x to avoid cutting inside an oversized message (tool output, file read, etc.). If even the minimum 3 messages exceed 1.5x the budget the cut is placed right after the head so compression still runs. Never cuts inside a tool_call/result group. Always ensures the most recent user message is in the tail (see ``_ensure_last_user_message_in_tail``). NrTrrg?rr%rrrrr) rrrrrrr#rrrrrprw)rrrqrxnmin_tail soft_ceilingr rtrrrrrrr: fallback_cuts r"_find_tail_cut_by_tokensz*ContextCompressor._find_tail_cut_by_tokenss(  1L MM/08|a/?/?3q!h,*+++Q<#-..  q1uhlB//  A1+C''),,2K4[AAK$(882=Jggl++1r @ @b$''@66*b1155k2FFD#d))/?"??JZ',66AEh;N;N : %KGG8| \ ! !"G h  ,1 55G//'BB887HUU7HqL)))r$cp|||j}|||}||kS)zReturn True if there is a non-empty middle region to compact. Overrides the ABC default so the gateway ``/compress`` guard can skip the LLM call when the transcript is still entirely inside the protected head/tail. )rmrr~)rrcompress_start compress_ends r"has_content_to_compressz)ContextCompressor.has_content_to_compresss<55h@TUU44X~NN  ,,r$current_tokenscd|_d|_d|_d|_d|_t |}|jdzdz}||kr%|jst d|||S|r|n|j pt|}| ||j |j\}}|r"|jstd||j}|||}|||} || kr|S||| } |jstd ||jtd |j|jd z|j|| z } td |dz| t | || || | } g} t-|D]}||}|dkrp|ddkrW|d}d}|t3|vr/t5|t7|t8r|rd|zn||d<| || sA|jst d| |z }||_d|_t<d|d} d}|dkr||dz ddnd}| |kr|| ddnd}|dvrd}nd}||kr|dkrdnd}||kr|}nd}|s| || dt-| |D]h}||}|r5|| kr/| dz}t5|d|d|d<d}| |i|xjdz c_| | } t| }||z }|dkr||z d znd}||_!|dkr|xj"dz c_"nd|_"|jsKtd|t | ||td|j| S) a~Compress 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. Args: focus_topic: Optional focus string for guided compression. When provided, the summariser will prioritise preserving information related to this topic and be more aggressive about compressing everything else. Inspired by Claude Code's ``/compact``. rFNrrTz-Cannot compress: only %d messages (need > %d))rrz-Pre-compression: pruned %d old tool result(s)z9Context compression triggered (%d tokens >= %d threshold)z,Model context limit: %d tokens (%.0f%% = %d)rzISummarizing turns %d-%d (%d turns), protecting %d head + %d tail messagesr<rsystemr%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.]r$uFSummary generation failed — inserting static fallback context markerTz% Summary generation was unavailable. z message(s) were removed to free context space but could not be summarized. The removed messages contained earlier work in this session. Continue based on the recent messages below and the current state of any files or resources.r1)rrrr2uZ --- END OF CONTEXT SUMMARY — respond to the message below, not the summary above --- r3rz8Compressed: %d -> %d messages (~%d tokens saved, %.0f%%)zCompression #%d complete)#rrrrrrrrrrrr rrrrrmr~rrrrJrrrr2r8rrr.r[rrirr)rrrr. n_messages_min_for_compressdisplay_tokens pruned_countrrr tail_msgsr compressedrrexisting_compression_note n_dropped_merge_summary_into_taillast_head_rolefirst_tail_role summary_roleflipped merged_prefix new_estimatesaved_estimate savings_pcts r"compresszContextCompressor.compresss*,-(+0(#' -1*-1*]]  014q8 * * *? C 1O+9Ct?V@CZxzB[C[C"&!=!= )< $ 6">" " ,  W W KKG V V V-55hOO44X~NN \ ) )O%n\&AB  KKK%    KK>#&,%     #\1I KK["&''    (();(UU ~&& # #A1+""$$CAvv#''&//X55779--%L!$,Fx,P,PPP%< 6@36O6OsT\s!222bs&&C N   c " " " " ? ighhh$~5I/8D ,/3D ,!ZZ7@ZZZ $) M[^_M_M_.1"4599&&IIIekHTWaHaHa(<044VVDDDgm 2 2 2!LL&L ? * *%1V%;%;kkG.((& ,0(' J   |HH I I I|Z00 # #A1+""$$C' 1A,=,=RR "9GGI&&! """I ,1(   c " " " " !#..z:: 5jAA ',6BPRSASAS~6<$(  &* &*,0RARARA!RA RA  RA $ RARA!$RARARA #TzRARARARARARAhH$sCx.HHHH SD8+/KKT#s(^,KBEK 4ZK tDcN#S( )KKKKb N$tCH~:N NSV N N N NLMMNO5"Dc3h,@5"S5"5"5"5"nHHDc3h4HHWZHfnorfsHHHHTGcGcGGG\GIIII\I :T$sCx.-A:d4PSUXPX>FZ::::xT#s(^0D3SVd38n1ECTW8T#s(^,8; -0tCH~&-0-0 -0 -0-0-0-0b$(>*>*T#s(^,>*8;>*Dj>* >*>*>*>*H -T#s(^0D - - - - -ssd38n!5sss`csostxy|BzBuCpDssssssssr$r)r9)+rrrHrErr=typingrrrragent.auxiliary_clientragent.context_engineragent.model_metadatar r r agent.redactr getLoggerrHrr[rZrrrrr_IMAGE_TOKEN_ESTIMATErrDrr#rr2rr8rOrrr*r$r"rsg&  ,,,,,,,,,,,,++++++...... /.....  8 $ $ L- M /1AA$'!CCB.IN;;;S;;;RU;;;;$+4+4+4+4s+4+4+4+4\wFcwFcwFwFQTwFwFwFwFtGGGGG GGGGGr$