v0.16.15

What's Changed

• docs: register the held-out probe sets by hash (ADR-0065 §3) by @NickFlach in #1092
• remember events: publish on a handshake-only connection, without the JetStream probes by @NickFlach in #1093
• serve prompt arms: no-state, no-reference, no-state-reference (wrapper study E2) by @NickFlach in #1094
• ADR-0065: the serve figure is the graded 18–19 of 20, not the mechanical 11 by @NickFlach in #1095
• ci: allow clippy::approx_constant on the two PHI literals by @NickFlach in #1097
• docs(adr): ADR-0066 mailboxes for outside research contributors by @NickFlach in #1096
• train_lora: --init-adapter, --eval-every, --max-holdout-rise (kannaka-loop-c1) by @NickFlach in #1098
• snapshots: list-snapshots shows the newest 500, and the stream caps messages at 1 MiB by @NickFlach in #1099
• train_lora: opt-in memory measures and a VRAM guard for kannaka-loop-c1 on an 8 GiB card by @flaukowski in #1100
• nats: a refused subscription is remembered for the process and never re-asked; swarm tail drops a refused subject (#1101) by @NickFlach in #1102
• release: v0.16.15 by @NickFlach in #1103

New Contributors

• @flaukowski made their first contribution in #1100

Full Changelog: v0.16.14...v0.16.15

v0.16.14

What's Changed

• test(config): make the HRM path tests absolute on Windows too by @NickFlach in #1073
• fix(nats): let radio publish prime's remember events, scoped to that leaf by @NickFlach in #1074
• fix(nats): confirm memory events once per subject, on a probe connection (#1071) by @NickFlach in #1072
• docs(swarm): say who actually recalls through the serve daemon by @NickFlach in #1075
• ADR-0065: the dogfood loop — improvement compounds only if use is graded and the ruler is frozen by @NickFlach in #1083
• fix(nats): an observatory user, scoped to recall requests on kannaka-prime by @NickFlach in #1079
• nats-accounts: the mail-deny on every scoped user, matching the hub by @NickFlach in #1082
• recall --remote: connect for request/reply only, skip the JetStream probes (#1080) by @NickFlach in #1081
• test: bound the second write alone in nats_down_does_not_fail_the_write (#1085); recall --remote help by @NickFlach in #1086
• ask log: KANNAKA_ASK_LOG records what each served ask showed the model (ADR-0065 item 1) by @NickFlach in #1084
• ADR-0065: record the blind-regraded probe numbers (bare model 4/20, not 1/20) by @NickFlach in #1088
• serve: KANNAKA_SERVE_PROMPT_ARM — switchable prompt arms for the wrapper study by @NickFlach in #1089
• docs(agent): the prompt arm is a serve-path switch; the work queue always runs baseline by @NickFlach in #1090
• release: v0.16.14 by @NickFlach in #1091

Full Changelog: v0.16.13...v0.16.14

v0.16.13

What's Changed

• mail: one NATS seat per agent (mail-seat.py); record the anonymous-read fix by @NickFlach in #1048
• mail: outbound through Resend; Stalwart admin helper; citizens' seats recorded by @NickFlach in #1049
• mail: new-mailbox.py; 0xscada-qe gets a mailbox by @NickFlach in #1050
• p2: kannaka-brain-7b-v2 — Qwen3-8B on citizen task data, pairwise task judge, pre-registered gate by @NickFlach in #1051
• publish_hf: cards say only what is declared; size not guessed from packed 4-bit weights by @NickFlach in #1052
• serve_debain2: bake the chat template, promote family aliases, stop filling the disk by @NickFlach in #1053
• publish_hf: a declared corpus must also declare its provenance by @NickFlach in #1054
• ADR-0064: native mail — the mailbox is the record, memory holds the meaning by @NickFlach in #1055
• fix(nats-acl): let serve publish the served-recall event by @NickFlach in #1056
• tools/mail: agent-mail.py, so SpaceChild and 0xSCADA-QE can use their own mailboxes by @NickFlach in #1058
• fix(recall): carry the chiral intuition flag into RecallResult (#1005) by @NickFlach in #1059
• ADR-0064 P0: kannaka mail — read-only accounts/sync/status/thread/close (gate 96.1%) by @NickFlach in #1060
• fix(hrm): store a content_id beside the file digest (#984) by @NickFlach in #1062
• fix(status-cache): count memories without the O(n²) assessment after writes by @NickFlach in #1064
• fix(inbox): report whether inbox send reached a listener (NATS no-responders) by @NickFlach in #1065
• feat(events): every write path publishes MemoryRemember (ids by default) (#1057) by @NickFlach in #1066
• fix(config): KANNAKA_DATA_DIR wins over an absolute hrm.path outside it (#1067) by @NickFlach in #1068
• release: v0.16.13 by @NickFlach in #1069

Full Changelog: v0.16.12...v0.16.13

v0.16.12

What's Changed

• qubo: Simulated Bifurcation as a ConsolidationSolver (Wave 4) by @NickFlach in #1043
• fix(recall): drop memories expired before the as-of instant (#1044) by @NickFlach in #1046
• release: v0.16.12 by @NickFlach in #1047

Full Changelog: v0.16.11...v0.16.12

v0.16.11

What's Changed

• fix(nats): a stream-create refusal is not a server error (#969) by @NickFlach in #1027
• chore(deps): rustls-webpki 0.103.13, rand 0.8.6 — closes 5 dependabot alerts by @NickFlach in #1028
• fix(ooda): three of seven rotation slots could never move the fitness by @NickFlach in #1029
• docs(recall): KANNAKA_RECALL_OBSERVE=0 does not stop the write — correcting #1010 by @NickFlach in #1030
• fix(swarm): self-origin is every identity this node published under (#890) by @NickFlach in #1032
• fix(fs): parent-dir fsync, temp-litter sweep, bounded rename retry (#934 P2-P4) by @NickFlach in #1033
• fix(store): a ghost survives the rebuild that follows the dream that made it (#917) by @NickFlach in #1035
• ops: host-metrics must notice when its own POST fails by @NickFlach in #1034
• docs(changelog): the seven changes since v0.16.10, and a warning before the roll by @NickFlach in #1036
• docs(changelog): #1034 was missing — the release range moved while CI ran by @NickFlach in #1037
• events: publish recall events from the daemon recall responder by @NickFlach in #1038
• changelog: #1038, recall events on the bus by @NickFlach in #1040
• fix(store): a fresh store is chiral from birth, so --batch builds the same store as remember (#917, #1031) by @NickFlach in #1039
• fix(recall): a recall hit is counted in the sidecar, not by rewriting the store (#977) by @NickFlach in #1041
• release: v0.16.11 by @NickFlach in #1042

Full Changelog: v0.16.10...v0.16.11

v0.16.10

What's Changed

• fix(nats): remember the stream-create refusal per PROCESS by @NickFlach in #1018
• fix(ops): an empty unit set is not a successful roll by @NickFlach in #1019
• fix(nats): a reply may only go to the inbox that asked (#943) by @NickFlach in #1020
• docs(changelog): the two fixes since v0.16.9 by @NickFlach in #1021
• fix(test): close the KANNAKA_FACET_DECOMPOSE race by scoping it per thread by @NickFlach in #1022
• fix(ooda): run the prebuilt research binary and stop discarding its stderr by @NickFlach in #1023
• fix(ooda): autoresearch runs as its owner, and finds cargo — the real #939 cause by @NickFlach in #1024
• fix(ooda): report cargo's exit status, not the if's by @NickFlach in #1025
• release: v0.16.10 by @NickFlach in #1026

Full Changelog: v0.16.9...v0.16.10

v0.16.9

What's Changed

• ops/windows: a real scheduled task for the ask responder by @NickFlach in #995
• fix(nats): never re-issue a stream create the broker already refused by @NickFlach in #996
• fix(medium): three energy writes that did not respect ENERGY_CAP by @NickFlach in #999
• feat(medium): a .hrm digest that ignores when the file was saved by @NickFlach in #1000
• test: three assertions that were written down and never made by @NickFlach in #1001
• test(consolidation): assert the pair, and stop naming tests for a feature that was deleted by @NickFlach in #1002
• test: an assertion that could not fail, and an identity never checked by @NickFlach in #1003
• test: enable two git tests ignored for a reason that was not true by @NickFlach in #1004
• docs(recall): the intuition flag's doc described a refactor that never landed by @NickFlach in #1006
• docs(changelog): write the Unreleased section while the work is fresh by @NickFlach in #1007
• fix(hrm): clamp persisted energy on load, and make the test able to fail by @NickFlach in #1009
• recall: --at scores temporal recency as of a given time by @NickFlach in #994
• ADR-0051: the Phase 3 gate is satisfied, and the flag measured negative by @NickFlach in #998
• ADR-0063: referential memory — a fact with an authority must not be held as a wave by @NickFlach in #962
• feat(recall): KANNAKA_RECALL_OBSERVE=0 separates ranking cost from the write it provokes by @NickFlach in #1010
• feat(recall): KANNAKA_RECALL_INCLUDE_DREAMS=0 asks for evidence only by @NickFlach in #1011
• docs(changelog): cover three merges I had missed, and the two new knobs by @NickFlach in #1012
• test: probes that locate the O(n)-per-insert in ingest (#978) by @NickFlach in #1013
• test: make the chiral ingest probe actually mint facets (#978) by @NickFlach in #1014
• test: parametrise the chiral ingest probe, and close #978's range gap by @NickFlach in #1015
• fix(config): one data directory, one agent identity by @NickFlach in #1016
• release: v0.16.9 by @NickFlach in #1017

Full Changelog: v0.16.8...v0.16.9

v0.16.8

What's Changed

• fix(facet): flag guard restores KANNAKA_FACET_DECOMPOSE on drop (#942) by @NickFlach in #986
• feat(llm): configurable request timeout via [llm] timeout_secs (#904 part 1) by @NickFlach in #987
• docs(changelog): record #988 under 0.16.7, and #986/#987 as unreleased by @NickFlach in #990
• fix(swarm): sync's presence keeps the agent's display name by @NickFlach in #991
• ADR-0051 Phase 3 preconditions: M9, M8, M3 by @NickFlach in #992
• release: v0.16.8 by @NickFlach in #993

Full Changelog: v0.16.7...v0.16.8

v0.16.7

What's Changed

• ops: commit the fleet roll scripts (roll-node.sh, roll-o1.sh) by @NickFlach in #968
• corpus: attribute KAX replies to the brain that wrote them by @NickFlach in #971
• tools/mail: the mail membrane, first slice (ADR-0062) by @NickFlach in #972
• cli: remember --batch and recall --batch (NDJSON, one process) by @NickFlach in #973
• remember --batch: persist once and rebuild the cache once, not per item by @NickFlach in #974
• recall: KANNAKA_RECALL_XI_BOOST=off skips the xi-diversity reranker by @NickFlach in #975
• encoder: a new store defaults to the semantic encoder (closes #976) by @NickFlach in #980
• fix(swarm): sync publishes presence; serve reload exits 0 by @NickFlach in #970
• fix(medium): key-sort the two maps ChiralMedium::save writes (#952) by @NickFlach in #981
• fix(facets): backfill_all_facets rebuilds the memory cache (#947) by @NickFlach in #982
• fix(import): preserve the id on both paths that used to discard it (#949) by @NickFlach in #983
• feat(recall): attention beam on the default path — NEGATIVE RESULT, ships off by @NickFlach in #988
• release: v0.16.7 by @NickFlach in #989

Full Changelog: v0.16.6...v0.16.7

v0.16.6

What's Changed

• ADR-0060: one address per agent, as the root of agent identity (Proposed) by @NickFlach in #957
• adr-0060: decide retirement — an address is never reused by @NickFlach in #958
• ADR-0061: Declared faculties — agents are a cast, not a fleet (Proposed) by @NickFlach in #959
• Re-encode with an encoder that can fail, and let it reach a remote one by @NickFlach in #961
• ADR-0062: the mail membrane — inbound owned, outbound relayed, never a relay by @NickFlach in #964
• Cap energy at every write, and rank by similarity alone by default (#965) by @NickFlach in #966
• release: v0.16.6 by @NickFlach in #967

Full Changelog: v0.16.5...v0.16.6