.cjS dZddlmZmZddlmZmZmZmZddl m Z dZ dZ dZ dZd ed efd Zd ed ededed edzf dZGddeZdS)uAbstract base class for pluggable context engines. A context engine controls how conversation context is managed when approaching the model's token limit. The built-in ContextCompressor is the default implementation. Third-party engines (e.g. LCM) can replace it via the plugin system or by being placed in the ``plugins/context_engine//`` directory. Selection is config-driven: ``context.engine`` in config.yaml. Default is ``"compressor"`` (the built-in). Only one engine is active. The engine is responsible for: - Deciding when compaction should fire - Performing compaction (summarization, DAG construction, etc.) - Optionally exposing tools the agent can call (e.g. lcm_grep) - Tracking token usage from API responses Lifecycle: 1. Engine is instantiated and registered (plugin register() or default) 2. on_session_start() called when a conversation begins 3. update_from_response() called after each API response with usage data 4. should_compress() checked after each turn 5. compress() called when should_compress() returns True 6. on_session_end() called at real session boundaries (CLI exit, /reset, gateway session expiry) — NOT per-turn )ABCabstractmethod)AnyDictListOptional)redact_sensitive_textipiiz+ ...[memory provider context truncated]... memory_contextreturnct|dd}t|tkr|S|dtt z|t dzS)zBPrepare provider context for a context-engine/LLM egress boundary.T)forceredact_url_credentialsN)r striplenMEMORY_CONTEXT_MAX_CHARS_MEMORY_CONTEXT_HEAD_CHARS!_MEMORY_CONTEXT_TRUNCATION_MARKER_MEMORY_CONTEXT_TAIL_CHARS)r sanitizeds 6/home/ice/.hermes/hermes-agent/agent/context_engine.pysanitize_memory_contextr(sx%#I  9~~111---. + , //00 1 2enginephasedefault_messagecontextNc t|ddsdSt|dd}t|r |d||d|}n|}|dSt|}|pdS)aDResolve host-visible status for an automatic compaction event. Engines can suppress routine automatic status with ``emit_automatic_compaction_status = False`` or customize it by defining ``get_automatic_compaction_status_message(...)``. Empty strings and ``None`` mean "do not emit a lifecycle status". emit_automatic_compaction_statusTN'get_automatic_compaction_status_message)rr)getattrcallablestrr)rrrr formattermessages r#automatic_compaction_status_messager&8s 6=t D Dt I4PPI ") +     "t'll  ""G ?drcVeZdZUdZeedefdZdZe e d<dZ e e d<dZ e e d<dZ e e d<dZe e d <dZe e d <d Zee d <d Ze e d<dZe e d<dZee d<edeeefddfdZedBde defdZdBde ddfdZe dCdeeeefdee deeded edeeeeff d!Z dBdeeeefde dzdeeeeefe ffd"Zdddd#d$eeeefd%eeeefd&eeefd'e deeeeff d(Z dBdeeeefdeeefd)eddfd*Z!deeeefdefd+Z"d,e defd-Z#d.ed/ed0ededzfd1Z$deeeefdefd2Z%d3eddfd4Z&d3edeeeefddfd5Z'dDd6Z(deeeeffd7Z)d8ed9eeefdefd:Z*deeeffd;Z+ dEded?ed@eddfdAZ,dS)F ContextEnginez.Base class all context engines must implement.r cdS)z,Short identifier (e.g. 'compressor', 'lcm').Nr selfs rnamezContextEngine.name^rrlast_prompt_tokenslast_completion_tokenslast_total_tokensthreshold_tokenscontext_lengthcompression_countg?threshold_percentprotect_first_nprotect_last_nTrusageNcdS)aUpdate tracked token usage from an API response. Called after every LLM call with a normalized usage dict. The legacy keys ``prompt_tokens``, ``completion_tokens``, and ``total_tokens`` are always present. Newer hosts also include canonical buckets: ``input_tokens``, ``output_tokens``, ``cache_read_tokens``, ``cache_write_tokens``, and ``reasoning_tokens``. Engines should treat those fields as optional for compatibility with older hosts. Nr )r+r9s rupdate_from_responsez"ContextEngine.update_from_responser-r prompt_tokenscdS)z0Return True if compaction should fire this turn.Nr r+r<s rshould_compresszContextEngine.should_compressr-rztuple[bool, str | None]c0||dfS)a4Return ``(should_compress, reason)``. The base implementation is backward-compatible: engines that only implement ``should_compress`` get ``(should_compress(prompt_tokens), None)``. Concrete engines with richer block reasons (e.g. a summary-LLM cooldown or an anti-thrashing guard) override this to surface a human-readable reason so callers can warn the user instead of silently skipping compression. Added for the silent-overflow warning fix (#62625) so plugin engines don't raise AttributeError. N)r?r>s rshould_compress_infoz"ContextEngine.should_compress_infos##M22D88rFmessagescurrent_tokens focus_topicr r cdS)uCompact the message list and return the new message list. This is the main entry point. The engine receives the full message list and returns a (possibly shorter) list that fits within the context budget. The implementation is free to summarize, build a DAG, or do anything else — as long as the returned list is a valid OpenAI-format message sequence. Args: focus_topic: Optional topic string from manual ``/compress ``. Engines that support guided compression should prioritise preserving information related to this topic. Engines that don't support it may simply ignore this argument. force: Whether a user-requested compression should bypass an engine-owned cooldown. Engines without cooldowns may ignore it. memory_context: Text returned by memory providers immediately before compaction. Summarizing engines should include non-empty text in their handoff prompt. Older engines may omit this parameter; the host filters unsupported optional arguments by signature. Nr )r+rCrDrEr r s rcompresszContextEngine.compressr-rc |dfS)uDeterministically trim old tool-result payloads without an LLM call. Runs on a low, cost-oriented trigger independent of ``should_compress`` so large-window engines can reclaim re-sent tool output long before full compaction would fire. Returns ``(messages, n_pruned)``. Default is a safe no-op: the list is returned unchanged with ``0`` pruned. Engines that don't implement a cheap prune — and any engine that predates this hook — inherit this default, so the agent loop's post-tool-call prune path never raises ``AttributeError`` on them. The built-in ContextCompressor overrides this with the real implementation. rr )r+rCrDs rprune_tool_results_onlyz%ContextEngine.prune_tool_results_onlys"{r)conversation_messagesincoming_message budget_tokensrequest_messagesrJrKrLcdS)u Optionally choose/replace the context for THIS request, pre-generation. Called every turn after the request message list is assembled and before it is dispatched to the provider — independent of ``should_compress()``. This lets an engine *select* which context enters the prompt (retrieval, topic routing, role/branch switching) rather than *shrink* context that is already there. The two verbs are orthogonal: - ``compress()`` : context is too long -> make it shorter. - ``select_context()``: this turn belongs to a different context -> use that one instead. Without this hook, engines that need per-turn access to the message list have to force ``should_compress()`` to return ``True`` so that ``compress()`` is invoked every turn purely as a callback — which conflates selection with compression and degrades behaviour when the engine's backend is unavailable. ``select_context()`` removes the need for that workaround. The returned list is request-only: it replaces the messages sent to the provider for this single call and MUST NOT be treated as persisted transcript state. The conversation history in the session DB is left untouched, so nothing leaks across turns. Return ``None`` to leave the request unchanged. Unlike the ``pre_llm_call`` plugin hook (which appends to the user message and intentionally never rewrites the list, to preserve the cache prefix), ``select_context()`` may *replace* the message list. Ordering / cache contract: the host runs this hook **before** prompt cache-control and **before** every request sanitizer (orphaned-tool cleanup, thinking-only/role normalization, whitespace/JSON normalization). So (a) whatever the hook returns still passes through the same validation as any request — a malformed replacement cannot reach the provider — and (b) prompt-cache stability (an AGENTS.md invariant) is preserved: the default no-op leaves the request byte-identical, so cache behaviour is unchanged for the built-in compressor and any non-implementing engine. An engine that *does* replace the list changes its own cache prefix by definition; that is the engine's concern, and cache-control breakpoints are re-derived on the selected list. The hook is evaluated per provider request (so it re-runs on retries within a turn), consistent with "select the context for THIS request". Args: request_messages: The assembled request message list (system prompt + history + any ephemeral prefill), in OpenAI format. conversation_messages: The unmodified persisted conversation history, for reference only (do not mutate). incoming_message: The current turn's user message, if available. budget_tokens: The active model's context length, or 0 if unknown. Default returns ``None`` (no-op) — zero impact on the built-in compressor or any existing engine. Nr )r+rMrJrKrLs rselect_contextzContextEngine.select_contexts @trkwargsc dS)uS Observe a finished user turn (post-turn ingestion / observation). Called from the standard turn-finalization path once the assistant/tool loop completes, with the finalized in-memory transcript snapshot. This is the complement to ``select_context()``: selection happens *before* the request, while observation happens *after* the turn. It lets an engine ingest, index, summarize, or update routing / topic / session state from what actually happened — so the next ``select_context()`` can act on it. Coverage: this fires from the normal finalization seam. Some abnormal early-return paths in the loop (e.g. a content-policy block or a provider terminal failure) persist and return without routing through finalization, and therefore do not currently emit this hook. Treat it as a best-effort post-turn observation for completed turns, not a guaranteed callback for every possible early exit; unifying all terminal paths behind one finalization seam is a separate follow-up. Together the two hooks remove the need to abuse ``should_compress()`` / ``compress()`` as a generic per-turn callback just to observe history, and they cover the case where a turn finishes and there may be no next request from which to infer the previous turn. ``messages`` is a shallow copy and should be treated as read-only: return values are ignored and this hook must not rely on transcript mutation for persistence. ``kwargs`` may include ``turn_id``, ``task_id``, ``api_call_count``, ``interrupted``, ``failed``, and ``turn_exit_reason``. ``usage`` carries the completed turn's canonical token usage (the same dict shape passed to ``update_from_response`` — ``prompt_tokens`` / ``completion_tokens`` / ``total_tokens`` plus the canonical ``input_tokens`` / ``output_tokens`` / ``cache_read_tokens`` / ``cache_write_tokens`` / ``reasoning_tokens`` buckets) so an engine can weigh how large/expensive the selected context actually was when deciding the next ``select_context()``. It is ``None`` on finalized turns that never reached a provider response (e.g. interrupt); engines must treat it as optional. Default is a no-op. Nr )r+rCr9rPs ron_turn_completezContextEngine.on_turn_completes ^trcdS)zQuick rough check before the API call (no real token count yet). Default returns False (skip pre-flight). Override if your engine can do a cheap estimate. Fr r+rCs rshould_compress_preflightz'ContextEngine.should_compress_preflightLs ur rough_tokenscdS)aReturn True when preflight should trust recent real usage instead. Built-in compression uses this to avoid re-compacting from known-noisy rough estimates after a compressed request has already fit. Third-party engines can ignore it safely. Fr )r+rVs r$should_defer_preflight_to_real_usagez2ContextEngine.should_defer_preflight_to_real_usageTs urrrrc |jsdS|S)aReturn user-visible status for automatic host-triggered compaction. Return ``None`` to suppress successful automatic lifecycle status for this compaction event. ``phase`` identifies the host call site (for example ``"preflight"`` or ``"compress"``). ``context`` contains best-effort fields such as ``approx_tokens`` and ``threshold_tokens``. This hook does not control warning/error messages or explicit manual commands such as ``/compress``. N)r)r+rrrs rrz5ContextEngine.get_automatic_compaction_status_message]s"4 4rcdS)uQuick check: is there anything in ``messages`` that can be compacted? Used by the gateway ``/compress`` command as a preflight guard — returning False lets the gateway report "nothing to compress yet" without making an LLM call. Default returns True (always attempt). Engines with a cheap way to introspect their own head/tail boundaries should override this to return False when the transcript is still entirely protected. Tr rTs rhas_content_to_compressz%ContextEngine.has_content_to_compressts tr session_idc dS)zCalled when a new conversation session begins. Use this to load persisted state (DAG, store) for the session. kwargs may include hermes_home, platform, model, etc. Nr )r+r\rPs ron_session_startzContextEngine.on_session_startr-rcdS)uCalled at real session boundaries (CLI exit, /reset, gateway expiry). Use this to flush state, close DB connections, etc. NOT called per-turn — only when the session truly ends. Nr )r+r\rCs ron_session_endzContextEngine.on_session_endr-rc>d|_d|_d|_d|_dS)zyCalled on /new or /reset. Reset per-session state. Default resets compression_count and token tracking. rN)r.r/r0r3r*s ron_session_resetzContextEngine.on_session_resets* #$&'#!"!"rcgS)zReturn tool schemas this engine provides to the agent. Default returns empty list (no tools). LCM would return schemas for lcm_grep, lcm_describe, lcm_expand here. r r*s rget_tool_schemaszContextEngine.get_tool_schemass  rr,argsc >ddl}|dd|iS)zHandle a tool call from the agent. Only called for tool names returned by get_tool_schemas(). Must return a JSON string. kwargs may include: messages: the current in-memory message list (for live ingestion) rNerrorzUnknown context engine tool: )jsondumps)r+r,rerPrhs rhandle_tool_callzContextEngine.handle_tool_calls-  zz7$JD$J$JKLLLrc|jdkr|jnd}||j|j|jrtd||jz dznd|jdS)zsReturn status dict for display/logging. Default returns the standard fields run_agent.py expects. rd)r.r1r2 usage_percentr3)r.r1r2minr3)r+ last_prompts r get_statuszContextEngine.get_statusso261H11L1Ld--RS "- $ 5"1&.Ct'::S@AAA,-!%!7   rmodelbase_urlapi_keyproviderapi_modec||_ddlm}t|ds |j|_||t |di|j|_|j|_t||jz|_ dS)aCalled when the user switches models or on fallback activation. Default updates context_length and recalculates threshold_tokens from threshold_percent. Override if your engine needs more (e.g. recalculate DAG budgets, switch summary models). r)resolve_model_threshold_config_threshold_percentmodel_thresholdsN) r2agent.context_compressorrwhasattrr4rxr!_base_threshold_percentintr1)r+rqr2rrrsrtrurws r update_modelzContextEngine.update_models- EDDDDDt899 D.2-CD *'>'> 74!3R88  *( ( $"&!= #NT5K$K L Lr)N)NNFrB)r N)rBrBrBrB)-__name__ __module__ __qualname____doc__propertyrr#r,r.r}__annotations__r/r0r1r2r3r4floatr6r8rboolrrr;r?rArrrGtuplerIrOrRrUrXrr[r^r`rbrdrjrpr~r rrr(r(Ys88;c;;;^X; "#C###scNCs $u###OSNC .2$d111  $sCx.  T    ^  ??S?D???^? 9 9# 9AZ 9 9 9 9)-%)   tCH~& !  c]       d38n    ^ D&*tCH~&d  tDcN#S( ) 27;+/ @@@tCH~.@ $DcN3 @ sCx. @  @ d38n @@@@J!%//tCH~&/CH~/ /  ////f$tCH~2F4    t . T#s(^0D      3 T      T#s(^8L QU    ####$tCH~"6 MS MS#X MS M M M M DcN    6MMMM M  M  MM MMMMMMrr()rabcrrtypingrrrr agent.redactr rrrrr#rr&r(r rrrs,6$#######,,,,,,,,,,,,......!""$S! C C            4Z BPMPMPMPMPMCPMPMPMPMPMr