+ iHjURt^RIHt^RIt^RIHt^RIHtHtH t H t ^RI H t ^RI HtHtHtHtHtHtHt^RIHt^RIHtHtHtHtHtHt]P<!R 4t]!R R 44t R R lt!]!RR44t"RRlt#R#)a/Pre-API preflight compression gate and post-tool compression for the turn loop. Runs once per API call after the request pressure is measured: grow a managed llama.cpp window (last resort), or compress when over threshold (deferring on noisy estimates, in failure cooldown, or while the review fork's first request is pending), handle the provider-proven overflow re-check fail-closed, and emit the deduped blocked/uncompressed overflow warnings. Nothing here imports ``agent.conversation_loop`` at module level (cycle); loop-internal helpers resolve lazily.) annotationsN) dataclass)AnyDictListOptional)#automatic_compaction_status_message)#PRE_API_COMPRESSION_STATUS_TEMPLATE_reset_read_dedup_cachescompression_blocked_transientlycompression_skipped_due_to_lockcontext_compression_timed_out&conversation_history_after_compression&ensure_compression_feasibility_checked)"_review_fork_first_request_pending)_apply_grown_window_blocked_compress_reason_clear_overflow_warn_refund_api_call _reanchor#_reset_retry_state_after_compactionzagent.conversation_loopc]tRt^t$RtR]R&R]R&R]R&R]R&R]R&R]R &R]R &R]R &R]R &R]R &R]R&R]R&R]R&R]R&RtR]R&RtR#)PreflightGateVerdictu``action``: ``"fallthrough"`` (make the API call), ``"continue"`` (window grown or history compacted — the call/budget was refunded, re-enter the turn loop and re-measure), ``"break"`` (turn ends: compression timeout or non-actionable compaction handoff — ``final_response``/``failed``/``_turn_exit_reason`` set) or ``"return"`` (``result`` is the turn's result dict). The other fields are the loop locals the gate rebinds; ``_last_preflight_pressure`` is re-armed only by a compression pass.stractionrpending_moa_prepared_requestmessagesactive_system_promptconversation_historyapi_call_countcompression_attemptsfinal_responsefailed_turn_exit_reason_compression_timeout_exhausted_preflight_compression_blocked#_provider_overflow_recovery_pending_last_preflight_pressureNOptional[dict[str, Any]]result)__name__ __module__ __qualname____firstlineno____doc____annotations__r)__static_attributes__r*9/home/ubuntu/.hermes/hermes-agent/agent/turn_preflight.pyrrsfU K"%%M K$''$''),,!!'+F $+r2rcHV^8dQhRRRRRRRRRR R RR RR RR RRRRRRR/ #)agentrvr compressorrequest_pressure_tokensintprovider_overflow_preflightbooldefer_preflightmoa_prepared_requestsystem_message user_messagemax_compression_attemptseffective_task_idreturnr*)formats"r3 __annotate__rE9sh h h 'h 8;h VYh !%h 8;h SVh h (+h GJh _bh  h r2c  aaaa a^RIHp Hp Hp HpHoHpRRV3RlllpRVVV VV3Rllp\VRR4!4pSP;'d0\SP4^8;'dSPS 8pV'd \SS4V'EdQ\S4'Eg?SP'd V'Ed$V!S4'd V'Ed V'EgVPS4'EdV!SVS4pV'd1\!SVV4\#SSP$4SnV!R 4#VeVSnS;P^, un \)S4\+\VR ^4;'g^4p\+\VR ^4;'g^4p\,P/R SR VR \VR ^4'dVR MRSPS 4\1VR\2P4!SR7SVVSP6SPS R7 pV'dSP9V4SSnSPpSP=SPVSV RR7wSn Sn\AS4'd@\#SSP$4SnV Sn!RSn"RSn#RSn$V!R4#SPVJd\\KS4'g\MS4'd:S;P^,un RSnSP&VJdRSnEM!\OS4\QSSPSPR4Sn)\#SSP$4SnV!SPV4'd>\,P/R4SPB'gV Sn!RSn$V!R4#V!R 4#V'dVV'dNSPUSPSPR4V!RV !SSPSP$RR74#V'd SPS 8dV!RV!44#V'd]V!S4'gOV'dG\WVS4pV'd2SPYTS\+\VR ^4;'g^44MSP'g~\SP4^8dd\\SRR4R R4p\[V\*4'd6V^8d/SV8d(\SRR4p\]V4'd V!SV4V'dV!RV!44#V!R4#)uMirror of the turn-prologue guard chain (defer on noisy estimate → skip in failure cooldown → ``should_compress``), rebinding the loop locals on ``v`` and setting ``v.action``. A compression pass that never reaches the provider refunds the call/budget in every branch (skip, re-run, timeout) so ``api_call_count`` never over-reports; a lock/transient skip refunds the attempt and leaves the progress blocker unarmed. A forced provider-overflow preflight that any gate blocks fails closed (llama.cpp may silently truncate).)#_COMPRESSION_TIMEOUT_FINAL_RESPONSE_HANDOFF_SKIP_FINAL_RESPONSE_compression_deferred_result_maybe_grow_local_window#_provider_overflow_exhausted_result-_should_skip_model_call_for_reference_handoffNc$V^8dQhRRRRRR/#)r5rrr)r(rCrr*)rDs"r3rE/run_preflight_compression..__annotate__Ks#c#;G[r2c$<WuSnSnS#N)rr))rr)r7s&&r3_done(run_preflight_compression.._doneKs#!(r2cV^8dQhRR/#)r5rCzdict[str, Any]r*)rDs"r3rErNOs  ~ r2cZ<S!SSPSPSPSS4#rP)rrr)rKr6rAr9r7sr3_exhausted_result4run_preflight_compression.._exhausted_resultOs/2 1::q55q7G7GI` $  r2'get_active_compression_failure_cooldowncR#rPr*r*r2r3+run_preflight_compression..Vstr2continuethreshold_tokenscontext_lengthzSPre-API compression: ~%s request tokens >= %s threshold (context=%s, attempt=%s/%s),unknownpre_api)tokens)phasedefault_message approx_tokensr\r]modelattempt max_attemptsrdtask_idtriggerTcontext_compression_timeoutbreakzgSkipping post-compaction model call: reference-only handoff would be the sole active user turn (#80622)!compaction_handoff_not_actionablerCtransient_block)reasoncontext_compressor#_warn_uncompressed_context_overflow fallthroughrP)/agent.conversation_looprGrHrIrJrKrLgetattrcompression_enabledlenrr rrr%should_compressrrrrrr:loggerinforr rDre _emit_statusr'_compress_contextrr r!r"r$r#r r rrr_persist_sessionr_warn_context_overflow_blocked isinstancecallable)r6r7r8r9r;r=r>r?r@rArBrGrHrIrJrLrQrU_compression_cooldown _eligible _grown_window _threshold_context_length_pre_api_status_pre_api_input _block_reason_ctx_len_warn_fnrKsff$d$$$$$d$ @r3run_preflight_compressionr9s   $=| !! > >  Oa  > > " "%= = .u6MN 25991115P5P !899=X=X%%  & &'> ? ?1 D[\  z= A/q7G7GHA $ $  +-AA * !# U#-?CHHqI gj2BAFKK!L  *&q)!n&-j:JA&N&Nq!T] " " $ > ?FF.2'*++**1       /%<"-2-D-D JJ6M%y.E. * A* ) / / 0q7G7GHA BA AH/3A ,"?A > ! :: ' +E 2 26UV[6\6\  " "a ' ")-A &--1EE15. / 6&Lqzz1#9#9&A " 0q7G7GHA  qzz1+A+ABX; 1::q//8I    %)?)?C[)[X0233 ?+BCCH]1=TU   0 06GJ(:A>CC!D  & & &3qzz?Q+>75*>EGWY]^ h $ $A:QT\:\u&KTRH!!0(;"X0233  r2cl]tRt^t$RtR]R&R]R&R]R&R]R&R ]R &R]R &R]R &R ]R &RtR#)PostToolCompressionVerdictu``end_turn`` True → a reference-only compaction handoff would be the sole active user turn: stop without another model call (``final_response`` / ``turn_exit_reason`` set).r<end_turnlist[dict[str, Any]]rrrrr:r r!turn_exit_reasoncurrent_turn_user_idxr*N)r+r,r-r.r/r0r1r*r2r3rrs8"N""r2rcLV^8dQhRRRRRRRRRRRRR R R R R RR RRRRR RR/ #)r5r6rrrr?r@rrr r:rArBr!rrrCrr*)rDs"r3rErEsXX X1XCFXVYXX58XPSX"X7:XLOX X36 X  Xr2c aaaaa a a ^RIHp Hp Hp^RIHpRVVVV V VV 3RllpVP p\\VRR44'd V!R4#^RI H pV!S\VRR 44pVeTpMcVP^8dVPpMEVPR8Xd^pM1T !TSS;'gR T!SVP;'gR R 74pVP'dSV8d \VV4VP'dSV8d\\VRR44'gVPV4'dS^, o\!V4VP#R 4SpVP%SVVVR R7wooSVJd\'V4'd S^,oEM\)VSS4oV!SV4'd*\*P-R4S 'gV o Ro V!R4#\/VSV4o MVP'd\1VVSR7pV'd2TP3TT\5\VR^4;'g^44\VRR 4p\7V4'dPV!SVR7wppV'd:VSJd4^RIHpV!VSVV4Vo\AYPB;'gR R7V!R4# \8d\*P;RRR7S^ppLqi;i)uPost-tool-call compression decision. Pressure comes from API-reported ``prompt_tokens`` (a tight lower bound; thinking models inflate completion tokens), ``0`` right after compression (no real count yet), else the route-aware overhead-inclusive estimate. Over threshold but blocked → deduped warning plus the deterministic tool-result-only prune, committed only when the engine returns a NEW list (never rebuild ``conversation_history`` for it).)rH _midturn_request_pressure_tokensrL)estimate_request_tokens_roughc V^8dQhRRRR/#)r5rr<rCrr*)rDs"r3rE1compress_after_tool_results..__annotate__s  4 $> r2c *<\VSSSSSSSR7#))rrrrr r!rr)r)rrr rrr!rrs&r3_verdict-compress_after_tool_results.._verdicts')G[!5L`)rs2#!,,D B   4 5 ,, ,2h V     Xr2