OKFS: HINDSIGHT P0 entry (persist-under-B4 fix measured + kairos genesis)

Author: nihilistau

memory-okf/LUT.md

@@ -196,3 +196,4 @@ Lookup before you build. `python tools/okf_mem.py lookup --root <root> <kw>`
 | 1264a8623b7c6c0e | agent | audit-2026-07-10,memory persistence,cross-session,SP_NIGHTSHIFT_PERSIST,batch set no-op,registry rot,dead rows,orphan episodes,memory_doctor,SP_SPINE leak,run_console_reason,launcher fix,silent reload | Memory-not-persisting = broken launcher flags (set VAR with no =1 is a batch NO-OP; no registry; L5 off) + 13/17 registry rows pointing at deleted dirs (silently skipped, no log) + 43 orphaned episode dirs (no disk rescan). Fixed launchers + new tools/memory_doctor.py; live-verified grow->restart->recall 4/5. NEW: SP_SPINE=1 leaks persona record on foreign questions (A/B proven) - keep SPINE off in serve until gated. | GREEN | sum/1264a8623b7c6c0e.md |
 | dd122e50e2a82fea | agent | audit-2026-07-10,personality,PF-B4 wired,SP_PERSONALITY,gateway-only personality,daemon personality-blind,interceptor dead code,persona.md,okf conformance,sp_status normalize,slug addr defect | Personality gates 5/5 PASS. Daemon :3000 personality-blind by design (gateway :8800 only). FIXED: PF-B4 @personality pack now in all_tools() when SP_PERSONALITY=1 (was orphaned). Open: interceptor registered-but-never-invoked (spine run_post_turn does the real write-back, gated on typed not SP_PERSONALITY); PF-B5 curation rarely armed. OKFS: 50 conformance errors -> 18 (enum case normalized, address-safe); 18 legacy no-frontmatter notes + 6 slug-addr entries need owner decision. | GREEN | sum/dd122e50e2a82fea.md |
 | 6baa9fa3efef7b2a | agent | audit-2026-07-10-pt2,agent stack,gateway 8800,tool calling dead,fence drift,tool_code space,python fence,run_gateway_system,run_console_agent,fastmcp,mcp server,mcp bridge,SP_MCP_TOOLS,gateway autodetect,persona editor,ADR-012 committed,numeric transcription residual | Agentic abilities dead = gateway :8800 not in the path (console posts :3000; run_gateway.bat armed nothing) PLUS reason-model fence drift ('` tool_code' space / '`python' wrappers) the parser rejected. Fixed both; console now autodetects the gateway + has a persona editor; FastMCP server+bridge shipped (G-MCP-SERVER 3/3); live-verified {tool}/{recall} events + real execution. ADR-012 committed ordered (36c2da3->4d307fd). Residual: model garbles numeric tool results when paraphrasing. | GREEN | sum/6baa9fa3efef7b2a.md |
+| 8ed4d23c31cab75d | agent | hindsight,kairos,SP_PERSIST_B4,persist under B4,agent turn latency,minutes to seconds,prewarm,prefix snapshot,numeric fidelity,rebuild,gemma4_tail_cpu link,build-cpu first,production repo,migration map,G-KAIROS-P0 | SP_B4_NIGHTSHIFT force-disabled persist-KV = the 2-5min agent turns. SP_PERSIST_B4=1 fix measured: 6.7-9.2s extensions, cache survives capture turns, gateway prewarm now functional. Post-tool low-temp verbatim rule fixes number garbling. Daemon clean-rebuilt (build-cpu FIRST for submodule changes). kairos repo born: HINDSIGHT charter + migration map, G-KAIROS-P0 GREEN. | GREEN | sum/8ed4d23c31cab75d.md |

memory-okf/full/8ed4d23c31cab75d.md

@@ -0,0 +1,23 @@
+---
+type: memory
+title: HINDSIGHT P0: persist-KV alive under B4 (minutes->seconds, measured) + kairos production repo genesis
+description: SP_B4_NIGHTSHIFT force-disabled persist-KV = the 2-5min agent turns. SP_PERSIST_B4=1 fix measured: 6.7-9.2s extensions, cache survives capture turns, gateway prewarm now functional. Post-tool low-temp verbatim rule fixes number garbling. Daemon clean-rebuilt (build-cpu FIRST for submodule changes). kairos repo born: HINDSIGHT charter + migration map, G-KAIROS-P0 GREEN.
+timestamp: 2026-07-10T04:17:28Z
+resource: engine f92a024; harness 1865ef6; kairos genesis
+sp_status: GREEN
+sp_gate: persist timing probe 5/5 + PERSIST-KV reuse logs + G-KAIROS-P0 4/4
+sp_commit: engine f92a024; harness 1865ef6; kairos genesis
+sp_repro: python persist_timing.py against run_console_reason.bat serve; read kairos/HINDSIGHT.md
+mem_kind: agent
+mem_addr: 8ed4d23c31cab75d
+mem_verified: unverified
+mem_lifecycle: active
+tags: [hindsight, kairos, SP_PERSIST_B4, persist under B4, agent turn latency, minutes to seconds, prewarm, prefix snapshot, numeric fidelity, rebuild, gemma4_tail_cpu link, build-cpu first, production repo, migration map, G-KAIROS-P0, agent, tier-2]
+mem_tier: full
+---
+
+PROJECT HINDSIGHT P0 (2026-07-10, operator ask: overhaul the tool-calling/agentic/prefill/memory surface; clean-slate view of the 8-month build; merge to one production repo).
+LATENCY ROOT CAUSE (the big one): routes.rs force-disabled persist-KV whenever SP_B4_NIGHTSHIFT=1 -> EVERY memory-enabled serve re-prefilled the full prompt每 turn (~75-233ms/tok byte-exact = the 2-5 min agent turns). The B4 capture runs through the BATCHED gemma4_decode_cuda forward (model ptr, not the session) and the pos==cl guard + recalled-turn committed-clear already fail closed on real cache mutations -> the exclusion was over-broad. FIX: SP_PERSIST_B4=1 opt-in (engine f92a024; unset = old exclusion = byte-identical null floor). MEASURED live: T1 cold 7.4s, T2 extension 6.7s, T3 B4-capture turn 6.9s, T4 post-capture extension 9.2s (cache SURVIVES capture), T5 recall 13.8s; PERSIST-KV reuse logs confirm (e.g. reuse 240/240, suffix 16 vs full 256). The gateway _prewarm (persona+tools prefix hydration, app.py:388) was DORMANT under the flag conflict and now works: first turn after boot LCP-reuses the preamble; REWIND_BOUND=32 was designed exactly for the prewarm dummy-tail drop. Residual fresh-chat case (divergence deep in a long committed convo) = kairos P1 prefix-snapshot (engine has snapshot/rewind verbs from KAI-1b).
+NUMERIC FIDELITY: post-tool rounds now answer at temp 0.15/rep 1.05 + verbatim-numbers rule (harness 1865ef6) - the 0.6/1.3 chat sampling garbled tool-printed numbers (3304 -> "3334" live).
+REBUILD: clean daemon rebuilt (BUILD_EXIT=0, 14:09): first link FAILED on gemma4_tail_cpu unresolved - build-cpu math-core had NEVER been rebuilt with the ADR-012 submodule code (the old exe predated it differently); cmake --build build-cpu (ninja, works without env bats - cached compiler paths; the env-cpu.bat call KILLS a calling batch script, run steps separately) then cargo relink. LESSON: ADR-012-style submodule changes need build-cpu rebuilt BEFORE the daemon link.
+KAIROS: new production repo D:\F\shannon-prime-repos\shannon-prime-kairos (github nihilistau/shannon-prime-kairos, private). HINDSIGHT.md charter: 10 lessons (policy out of the kernel; config.toml not env ladders; one recall authority; the conversation cache is sacred; coherence gates; composition gates; live-play; native git; OKFS canonical; honest negatives ship default-off). MIGRATION-MAP.md: per-subsystem KEEP/REWRITE/RESEARCH/DEAD verdicts w/ receipts. Phases P0(genesis, G-KAIROS-P0 GREEN 4/4)..P5(llama.cpp parity ladder; targets: agent tool turn <=15s, decode >=40 tok/s via MTP re-gate, cold chat <=20s via prefix snapshot). PPT-LAT/L1 verdict: lattice takes ZERO code into the serve path; L1 ABI kept for dual-model co-residency verbs only. Staging repos = research lab, never rewritten.

memory-okf/sum/8ed4d23c31cab75d.md

@@ -0,0 +1,24 @@
+---
+type: memory
+title: HINDSIGHT P0: persist-KV alive under B4 (minutes->seconds, measured) + kairos production repo genesis
+description: SP_B4_NIGHTSHIFT force-disabled persist-KV = the 2-5min agent turns. SP_PERSIST_B4=1 fix measured: 6.7-9.2s extensions, cache survives capture turns, gateway prewarm now functional. Post-tool low-temp verbatim rule fixes number garbling. Daemon clean-rebuilt (build-cpu FIRST for submodule changes). kairos repo born: HINDSIGHT charter + migration map, G-KAIROS-P0 GREEN.
+timestamp: 2026-07-10T04:17:28Z
+resource: engine f92a024; harness 1865ef6; kairos genesis
+sp_status: GREEN
+sp_gate: persist timing probe 5/5 + PERSIST-KV reuse logs + G-KAIROS-P0 4/4
+sp_commit: engine f92a024; harness 1865ef6; kairos genesis
+sp_repro: python persist_timing.py against run_console_reason.bat serve; read kairos/HINDSIGHT.md
+mem_kind: agent
+mem_addr: 8ed4d23c31cab75d
+mem_verified: unverified
+mem_lifecycle: active
+tags: [hindsight, kairos, SP_PERSIST_B4, persist under B4, agent turn latency, minutes to seconds, prewarm, prefix snapshot, numeric fidelity, rebuild, gemma4_tail_cpu link, build-cpu first, production repo, migration map, G-KAIROS-P0, agent, tier-1]
+mem_tier: summary
+mem_full: 8ed4d23c31cab75d
+---
+
+# HINDSIGHT P0: persist-KV alive under B4 (minutes->seconds, measured) + kairos production repo genesis
+
+SP_B4_NIGHTSHIFT force-disabled persist-KV = the 2-5min agent turns. SP_PERSIST_B4=1 fix measured: 6.7-9.2s extensions, cache survives capture turns, gateway prewarm now functional. Post-tool low-temp verbatim rule fixes number garbling. Daemon clean-rebuilt (build-cpu FIRST for submodule changes). kairos repo born: HINDSIGHT charter + migration map, G-KAIROS-P0 GREEN.
+
+Full context: [full/8ed4d23c31cab75d.md](../full/8ed4d23c31cab75d.md)