Pre-public Phase 1. Remove personal/IP traces that should not ship in a
general-purpose kit; keep only intended author attribution.
- no-github-push.sh + hooks-and-blocks.md + ci-scaffold: drop "KeiTech
unfiled patent IP / trade secrets / priority date" wording; reword as a
generic opt-in guard for keeping code on a private remote.
- check-error-patterns.sh: remove author-local absolute path from the
tombstone comment.
- graph-export-watcher.sh: default viz dir to ~/.local/share/kei/graph-viz
(was a personal project path).
- agent manifests (cost-guardian, modal-runner, infra/ml/code-implementer)
+ ci.yml: strip private memory references and dated personal incidents;
keep the generic cost/ops lessons. Snapshots regenerated; golden 3/3.
Kept intentionally: author attribution (NOTICE / README / Cargo / plugin).
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Thin skill over the kei-message jsonl mailbox: /msg reads inbox, /msg @name text sends (identity = cwd basename), /msg all broadcasts, /msg list|who. send now accepts leading @name. Mirror of keigit 7b453aac.
Mirror of keigit 3aff0029. lib-hooks.sh apostrophe-in-jq closed the bash single-quote and broke every install (fixed). cortex-ui removed from profiles+menu; forgejo/zoekt source lib-launchd; portable kei-message id (BSD date fallback); non-TTY default minimal; install stamps .kei-profile; README hook count 54; web-install warns before discarding local edits in the managed clone.
Any Claude Code session can message any other (not just Agent-Teams teammates),
no tmux. Append-only jsonl bus + UserPromptSubmit hook pulling unread per turn.
kei-message.sh (send/inbox/list, address by cwd-basename or "all"),
mailbox-inject.sh (cursor dedup, first-turn baseline, no self-echo), bin/kei
`message` dispatch, lib-scaffold copies all scripts/*.sh, snippet wires the hook.
Bypass KEI_MAILBOX_BYPASS=1. Verified by 2-session simulation.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
SessionStart first-run hook now injects an ordered post-install checklist Claude
walks the user through: (1) /onboard projects → per-project agents,
(2) /sleep-setup → nightly REM (recommend local-only), (3) /cortex-setup (only
if cortex daemon installed). Confirm + run each, skippable. Fires once.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1. SessionStart hook first-run-onboard.sh: on first Claude Code session after
install, nudges user/Claude to run `/onboard ~/Projects/*` (scan stack +
create a project-specialist agent per project; delegates to /new-agent).
Fires once (marker), then silent. Wired in settings-snippet under SessionStart.
2. lib-summary next-steps: lead with /onboard ~/Projects/*, then /new-agent.
3. lib-hooks merge: normalize null/absent matcher → "" (Claude Code /doctor
rejects null; pre-kit hooks often lack a matcher field).
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
KEISEI ASCII logo (голубой) at install start; [install] prefix тёмно-жёлтый
(38;5;178); primitive names + language codes голубой (38;5;39); language names
тёмно-жёлтый. lib-log COLOR now enables under curl|bash (detect terminal via
[ -t 1 ] || /dev/tty — web-install tees stdout so -t 1 was false → colors off).
Color detection, not an interactivity gate. Verified piped-under-pty.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PATH-wiring (~/.claude/bin, home of the `kei` entry-point) was gated on
`[ -t 0 ] && [ -t 1 ]`; curl|bash tees stdout so -t 1 is false → pathway skipped
→ `kei: command not found`. Same tee/-t1 trap as onboarding (this one in the
top-level install.sh). Eradicated across tree: install.sh, lib-menu.sh,
lib-wizard.sh all gate on [ -t 0 ]. Verified piped-under-pty.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Splash was cyan; rebrand to blue/yellow: sky-blue (38;5;39) logo + dim-blue
separators, gold (38;5;220) brand line + field values. Version v0.16→v0.38.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1. CRASH on word input (e.g. "claude") at any menu: $((ans-1)) under set -u →
"unbound variable" → install died. Added _onb_read_choice validation (all 4 menus).
2. Claude Code default: added claude-code subscription provider (kei-registries
submodule c559065→b904993), subscription default transport, claude-code default
provider — Enter,Enter,Enter → Claude Code, no API key.
3. install died at line 178 for no-key providers (claude-code/codex/local):
onboarding_run ended on a false `&&` → returned 1 → set -e. Added `return 0`.
Plus en+ru explanations and auto-select single-option steps. Verified piped-under-pty.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
curl|bash reads this script byte-by-byte from the pipe; a bare `exec < /dev/tty`
then made bash read the NEXT line (the bootstrap exec) from the keyboard →
infinite hang after "delegating", bootstrap never started. Fix: merge the
redirect into the bootstrap exec (one command). Verified piped-under-pty:
buggy=HANG, fixed=full install completes.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
web-install.sh tees stdout to a logfile, so -t 1 is false even in an interactive
curl|bash; the /dev/tty fix reattached only stdin. Wizard gates required BOTH
-t 0 and -t 1, so onboarding (language select) and bootstrap's profile wizard
were silently skipped on the primary install path. Prompts go to stderr and read
from stdin — interactive stdin is the only real requirement.
Gates: bootstrap.sh, lib-onboarding.sh (×2), lib-preflight.sh, lib-hooks.sh.
Non-interactive (CI / </dev/null) still skips — verified.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
E2E (curl install.keisei.app | bash, no controlling terminal) died at
`exec < /dev/tty` — "No such device or address". `[ -r /dev/tty ]` stat's
the path readable but the device can't be opened without a controlling
terminal (CI/cron/ssh non-interactive) — the documented `| bash -s -- --yes`
CI path. Now probes a real open via `{ : < /dev/tty; }` and reattaches only
on success; else proceeds non-interactively. Verified RC=0 end-to-end.
Root causes found by reproducing a clean install from keigit:
1. PROFILE_PRIMS resolved only inside check_prereqs → unbound for
--no-execute (plan showed 0 prims for every profile) and silently
empty for --skip-prereqs. Now resolved unconditionally in install.sh
before any reader (SSoT).
2. Every profile (even minimal, advertised "no Rust compile") fell back
to a 5-15 min `cargo build --workspace` because no prebuilt release
binaries exist. Auto-set KEI_SKIP_RUST for profiles with no rust
primitives → minimal installs in ~18s (assembler only). cargo stays a
hard prereq because the agent assembler always compiles.
3. The assembler aborted the WHOLE install on any single bad manifest
(set -e). generate_agents is now tolerant: bad manifests print FAIL
but hooks/skills/settings still land. Commit-time validate stays strict.
4. Data bugs that broke the assembler:
- duplicate [taxonomy] table in _roles/{auditor,merger}.toml
- fal-ai-runner handoff → keimd-expert (not shipped in kit)
- infra-implementer-cicd forbidden_domain literal `${{ secrets.NAME }}`
collided with assembler ${{ }} placeholder detection
5. Metadata: KeiSei84 (nonexistent GitHub org) → KeiSeiLab/KeiSeiKit-1.0
across plugin manifests, bootstrap, README, docs, Cargo/npm metadata.
.claude-plugin/{plugin,marketplace}.json 0.16.0 → 0.38.0. SECURITY.md
supported version 0.14.x → 0.38.x.
feat: ship KeiSei tamagotchi statusline into the kit
- scripts/keisei-pet{,-update}.sh (portable, state under ~/.claude/pet/)
- install copies them to ~/.claude/scripts/
- settings-snippet adds statusLine (set-if-absent, never clobbers an
existing one) + 4 pet-update hooks (prompt/rust_write/github_block/sleep)
Verified: clean minimal install RC=0, zero FAIL, 38 agents + 52 hooks +
68 skills, settings valid, statusLine wired, pet renders, idempotent re-run.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
gx10 fail при первой установке (фото юзера 2026-05-18):
jq: error: Could not open file /home/keisei/.claude/settings.json
[error] jq-merge produced invalid output; ... unchanged
[error] install failed at line 126; rolled back
Root cause:
backup_file() (install/lib-backup.sh:56) делал `mv target backup`,
затем activate_hooks() → _jq_merge_hooks() пыталась читать оригинал
target'а через jq — но оригинал уже унесён mv'ом, открыть нечего.
jq fail → rollback всей установки.
Fix: cp -a вместо mv (как уже сделано в backup_path() для директорий).
Атомарный безопасный бэкап; оригинал остаётся для jq, потом
_jq_merge_hooks атомарно переписывает его mv tmp → target.
CI run 26014515768 — `x86_64-apple-darwin` упал на webauthn-attestation-ca
→ openssl-sys 0.9.115: pkg-config не настроен для cross-compile
(macos-latest = arm64, target = x86_64).
Фикс: добавлен прямой dep `openssl-sys = { version = "0.9", features =
["vendored"] }` в kei-auth-webauthn — компилит openssl из исходников,
не требует системного. Транзитивно достаточно одного объявления feature
во всём workspace (Cargo unifies features).
CI v0.38.0 mcp-server билды все 5 упали на `bun install --frozen-lockfile`
с «lockfile had changes, but lockfile is frozen». В lockfile отставали:
- @keisei/mcp-server 0.14.0 → 0.14.6
- ajv 8.18.0 → 8.20.0
- express-rate-limit 8.3.2 → 8.4.1
- hono 4.12.14 → 4.12.16
Пересоздан `rm bun.lock && bun install` под bun 1.3.14 (тот же что в CI).
audit-2026-05-18 (соседняя сессия на gx10, 3 бага из 5 подтверждены):
#1 [HIGH] cortex profile сломан — kei-cortex depends on kei-router
path-dep'ом, а kei-router → kei-model. kei-model не было в
profile list, cargo build падал с «failed to read kei-model/
Cargo.toml». Добавил kei-model в cortex profile.
#3 [HIGH] cortex-ui отсутствует в репо — был в profile list, но
исходника нет ни в _primitives/_rust/, ни в _ts_packages/.
Удалил из profile.
#4 [MEDIUM] web-install.sh теряет /dev/tty — exec | tee + curl|bash
делают stdin pipe, wizard'у read нечего читать, onboarding падает
на первом prompt. Переподключаю stdin к /dev/tty перед exec
bootstrap.sh, если /dev/tty доступен.
#2 (linux-arm64 prebuild) — tag v0.38.0 не запушен на github, CI
release.yml не сработал. Требует RULE 0.1 triple-confirm.
#5 (non-TTY default = minimal) — уже корректно, install.sh:141
`PROFILE="${PROFILE:-minimal}"` + lib-menu.sh skip on non-TTY.
Авто-сгенерирован из _blocks/*.md по категориям префикса:
API / AUTH / CI / DB / DEPLOY / DOCS / DOMAIN / MODE / OBS /
PATH / RULE / SCRAPER / SECURITY / STACK / TEST + прочие.
Каждая запись: `name` — первая H1-строка файла.
Использование в _manifests/<agent>.toml:
blocks = ["baseline", "rule-pre-dev-gate", "api-anthropic", ...]
Assembler читает блоки из _blocks/, склеивает в финальный
_generated/<agent>.md. Новый блок = просто .md в _blocks/.
Регенерация INDEX.md — TODO build-index.sh скрипт (сейчас
сгенерирован вручную bash циклом).
Новые словари:
uk Українська
de Deutsch
fr Français
es Español
pt Português
it Italiano
tr Türkçe
ar العربية
hi हिन्दी
zh 简体中文
ja 日本語
ko 한국어
id Bahasa Indonesia
vi Tiếng Việt
Каждый файл — 17 STR_* ключей (тот же контракт что en.sh + ru.sh).
lib-i18n.sh::i18n_available_languages — единый список (en + 15)
для меню выбора в мастере. i18n_load_lang упрощён: всегда грузит
английский как fallback, потом поверх — словарь языка (отсутствующий
ключ остаётся английским).
onboarding_pick_language теперь генерирует whiptail/bash select из
i18n_available_languages динамически — добавление нового языка =
один файл `install/i18n/<код>.sh` + одна строка в available_languages,
дальше всё подхватится автоматически.
Перевод формальный, без излишеств. Welcome баннер всегда EN (юзер
ещё не выбрал на момент показа).
Проверено: bash -n всех 16 словарей чисто, roundtrip всех языков
работает (i18n_load_lang en/ru/uk/de/fr/es/pt/it/tr/ar/hi/zh/ja/ko/id/vi
выдают локализованные STR_DONE_TITLE + STR_TR_DIRECT_API), non-TTY
smoke install --no-execute проходит.
- preflight failure handling: вместо `|| true` (молчаливое продолжение
при упавшем preflight) — явный prompt «продолжить? [y/N]» с return 1
при отказе. Без TTY печатает warning и продолжает. Это закрывает
HIGH bug-9: «.onboarded флаг выставляется при нерабочей конфигурации».
- lib-preflight.sh::preflight_check_cli — общий helper (command -v +
offer-install + version echo). Убирает 6-file boilerplate (хотя сами
per-provider файлы пока не переписаны под него — это отдельный шаг).
- onboarding_fallback_providers: расширен с 3 до 14 провайдеров,
покрывает все 7 транспортов. Был дрейф vs providers.toml (14 vs 3),
юзер без submodule видел только anthropic+openai+ollama.
- STR_PICK_PROVIDER plural mismatch: whiptail и plain ветки теперь
используют один fallback "Provider within" (раньше plain имел
"Providers within", whiptail — "Provider within").
- STR_DONE_NEXT удалён из en.sh + ru.sh (мёртвый ключ).
- Новые ключи: STR_MENU_* (для lib-menu.sh) + STR_PREFLIGHT_FAILED +
STR_PREFLIGHT_CONTINUE. lib-menu.sh начал использовать
STR_MENU_TITLE / STR_MENU_SUBSTRATE (частичная локализация, остальное
меню — отдельной задачей).
Тесты: bash -n чисто, i18n round-trip EN/RU работает, non-TTY smoke
install --no-execute проходит.
1. HIGH-1: onboarding ↔ kei-model-router связка
До: onboarding мастер писал ~/.claude/config/onboarding.toml,
но router его не читал — выбор провайдера декоративный.
После: lib-onboarding.sh::onboarding_write_config доп. пишет
~/.claude/config/user-model-override.toml; registry.rs::Registry
получил load_user_override() возвращающий UserModelOverride.
Приоритет: --pinned > user-override > agent-profiles default_model_ref.
2 новых теста (round-trip TOML, optional transport).
2. HIGH-2: eval "$install_cmd" → bash -c "$install_cmd"
До: lib-preflight.sh::preflight_offer_install делал eval.
После: bash -c с явным subshell + печать команды юзеру до запуска.
3. HIGH-3: codex.sh regex false-pass
До: grep -qiE "logged.in|active" пропускал "not logged in" как pass.
После: сначала negative-pattern (not logged|signed out|please log in),
потом positive (\blogged in\b|status: active|auth: yes).
4. HIGH-4: path traversal в source preflight
До: lib-preflight.sh::preflight_run делал source без валидации
provider id — `../../../evil` сработал бы.
После: whitelist regex ^[a-z0-9][a-z0-9_-]{0,63}$ + realpath
проверка что resolved путь не вышел за PREFLIGHT_DIR.
5. HIGH-5: curl|sh без verification
ollama-local.sh + google-vertex.sh теперь печатают предупреждение
что Linux-установка тянет shell-скрипт с внешнего сервера без
проверки хэша/подписи, и предлагают альтернативу.
MEDIUM попутно:
- anthropic-bedrock.sh: один вызов aws sts get-caller-identity
вместо двух (экономит 1-3с), различает cred-error от network
по тексту stderr, маскирует account ID в ARN перед печатью.
- mlx-local.sh: pip install --user mlx-lm вместо global pip install
(не требует sudo, не загрязняет system Python).
Тесты: cargo test --lib 80/80, bash -n всех изменённых файлов чисто.
Структура локализации:
install/i18n/en.sh — английский словарь (дефолт, fallback)
install/i18n/ru.sh — русский словарь
install/lib-i18n.sh — лоадер + welcome banner
Поток:
1. install.sh source'ит lib-i18n.sh и зовёт i18n_load_default →
все строки на английском.
2. Если onboarding нужен — печатается welcome banner ASCII-рамка
на английском (язык ещё не выбран).
3. onboarding_pick_language — единственный двуязычный шаг
("Choose language / Выберите язык"). По выбору вызывает
i18n_load_lang ru|en — перегружает словарь.
4. Все последующие шаги (transport / provider / model / auth /
completion) идут на выбранном языке.
Fallback: если ru-словарь не имеет ключа — используется английское
значение (load_default вызывается до загрузки ru.sh, переменные
перезаписываются поверх).
lib-onboarding.sh переведён со смешанных hardcoded строк на
${STR_*} placeholders.
Тесты: bash -n всех 5 файлов чисто, i18n loader unit-тест показывает
EN/RU перегрузку, non-TTY smoke install --no-execute проходит.
1. OID-check в parse_x25519_pkcs8_pem
До: брался последний 32-байтный slice любого PKCS#8 DER, OID не
проверялся. RSA/EC/Ed25519 ключ молча давал 32 неправильных байта
→ decrypt падал с generic "wrong key" без объяснения.
После: строгая проверка длины (48 байт) + OID 1.3.101.110 (X25519,
byte slice 9..12 = 0x2b,0x65,0x6e). Внешний openssl ключ другого
алгоритма теперь даёт явную ошибку с указанием реального OID.
Константы X25519_OID + X25519_PKCS8_DER_LEN.
RFC 8410 §3 + §7 ссылка в doc-комментарии.
2. x25519-dalek feature `zeroize`
До: features=["static_secrets"] — StaticSecret хранил priv-ключ
в куче без затирания при Drop. Локальный priv_raw.zeroize() стирал
только стек-копию, оригинал в куче оставался до GC.
После: features=["static_secrets","zeroize"] — StaticSecret сам
реализует ZeroizeOnDrop, ключ затирается при выходе из scope.
3. Два новых теста:
- parse_rejects_wrong_length_der — 32-байтный DER (вместо 48)
отклоняется с сообщением про "48 bytes"
- parse_rejects_wrong_oid — DER с OID Ed25519 (0x2b,0x65,0x70)
отклоняется с сообщением про "X25519"
8/8 тестов модуля проходят, cargo check workspace чисто.
Старая 0.14.5 mcp-server (с source maps содержавшими /Users/
denisparfionovich/...) удалена с keigit.com отдельной операцией
через Forgejo DELETE API.
Новый интерактивный мастер при первой установке:
1. Язык интерфейса (RU/EN)
2. Транспорт (direct-api / aws-bedrock / azure-openai / google-vertex
/ local / proxy / subscription)
3. Провайдер внутри транспорта (14 вариантов суммарно)
4. Модель из выбранного провайдера (3 моделей Anthropic, и т.д.)
5. Ключи/креды (silent read, пишет в ~/.claude/secrets/.env chmod 600)
Skip-логика:
- флаг ~/.claude/.onboarded
- env KEISEI_SKIP_ONBOARD=1
- не-TTY запуск
Запись:
~/.claude/config/onboarding.toml — выбор lang/transport/provider/model
~/.claude/secrets/.env — ключи провайдера
~/.claude/.onboarded — флаг прохождения
Парсер toml — pure awk (без зависимостей). Реестры из submodule
_blocks/registries. Submodule bumped до afe0c6f с новым полем transport.
Fallback если submodule не подтянут: anthropic + sonnet.
- _ts_packages/tsconfig.base.json: sourceMap=false, declarationMap=false
(source maps leaked absolute dev paths in published tarballs).
- All 6 @keisei/* packages: publishConfig.registry = keigit.com.
mcp-server bumped 0.14.5 -> 0.14.6 (republished without maps).
- .github/workflows/release.yml split into two jobs:
npm-publish-keigit: primary. Activates on KEIGIT_NPM_TOKEN +
KEIGIT_NPM_USER secrets. Publishes via direct curl PUT
(Forgejo requires Basic auth; npm CLI sends Bearer).
npm-publish-npmjs: reserved for future. Activates on NPM_TOKEN
secret. Currently no token -> job skipped gracefully.
End-to-end verified: clean dir + scope @keisei -> keigit + npm install
pulls 145 deps, no leaked paths, no .map files in any of 6 packages.
Thin wrapper (88 LOC) that lets a fresh machine install KeiSeiKit with
one line, no prior clone:
curl -fsSL https://install.keisei.app | bash
curl -fsSL https://install.keisei.app | bash -s -- --profile=dev --yes
## Why a third install entry point
Existing install path required `git clone` + `cd` + `./bootstrap.sh` —
three commands the user has to type, plus they must already have the
repo URL handy. For private repos with SSH auth this is real friction.
`web-install.sh` is meant to be served as a static file (Cloudflare
Pages / nginx) at install.keisei.app. It does ONE thing: prereq → clone
→ delegate to ./bootstrap.sh. Single source of truth — no duplicated
install logic.
## What it does
1. Splash + log to ~/.keisei-install.log
2. Hard prereq: git (the one thing bootstrap.sh cannot self-install)
3. SSH auth probe for git@github.com — clear error message if key missing
4. Clone/pull KeiSeiLab/KeiSeiKit-1.0 to $KEISEI_ROOT (default ~/.local/share/keisei)
5. git submodule update --init (pulls kei-registries)
6. exec ./bootstrap.sh "$@" — pass-through all flags
## Env overrides
- KEISEI_ROOT — install location
- KEISEI_REPO — git URL
- KEISEI_REF — branch/tag/sha
## Hosting
Push this file to install.keisei.app (Cloudflare Pages, S3+CF, nginx
static — anything that serves a single .sh over HTTPS).
## README updated
Quick start section now shows the curl one-liner as the recommended
path. Repo URL corrected from KeiSei84 → KeiSeiLab org.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Closes audit HIGH-1 (SSoT drift between KSK and marketplace).
_blocks/registries/ now tracks the single canonical kei-registries repo;
marketplace consumes the same submodule.
closes MEDIUM/LOW from feat/agent-dna-three-layer audit:
- models.toml: cache_write_1h_per_mtok_micro added to all 11 entries
(Anthropic: 200M/600M/1000M micro = $2/$6/$10 per MTok per pricing page;
other providers: 0 placeholder)
- main.rs (kei-model-router): WAL pragma + busy_timeout errors now logged
to stderr instead of silently dropped (previously .ok() swallowed both)
- models.toml: Haiku id pin TODO documented (router still hardcodes alias)
Adds `agent_shell_dna` cube parsing the new agent-shell::<p>:<m>:<c>::<scope>::<body>-<nonce>
format emitted by keisei-marketplace/src/lib/cryptoid.ts::agentDna.
Companion to legacy 4-segment `dna_class` (untouched per RULE Don't-Rewrite).
Accepts both 8-hex (legacy) and 16-hex (current) lengths for forward-compat.
- new file: src/agent_shell_dna.rs (235 LOC, 13 tests all pass)
- lib.rs: pub mod agent_shell_dna + module doc
Closes HIGH-2 (dna-three-layer audit).
Codex CRITICAL + 4 HIGH + 5 MEDIUM/LOW from RULE 0.23 dual-review and
RULE 0.25 multi-critic swarm — all closed.
CRITICAL fix
- Model::slug() ledger compatibility: posterior.rs + select_kernel.rs
query `WHERE model = ?2 OR model = ?3`, binding canonical + legacy
slug pair via new `Model::legacy_slug()`. Production ledger rows
written under "haiku"/"sonnet"/"opus" remain visible to posterior
aggregation. Regression test ledger_legacy_slug_counted.
HIGH fixes
- cmd_select(): no longer early-returns on profile match. Profile's
default_model_ref now becomes DecisionInput.fallback; select() always
runs, posterior/kernel evidence wins if present. RULE 0.20 cost
optimisation restored for all 18 registered agents.
- Registry pricing SSoT: DecisionInput now carries Option<Arc<Registry>>.
estimated_cost() tries registry first; hardcoded match is documented
fallback only. select_posterior.rs no longer duplicates models.toml
constants.
- registry.rs portability: include_str!() embeds the three TOMLs at
compile time. load_embedded() new; disk path tried first via
KEI_REGISTRIES_DIR, embedded as fallback. `cargo install`d binaries
now find registries unconditionally. embedded_registry_matches_disk
test ensures embedded ≡ disk source.
- next_model() ambiguity: replaced Option<&Model> with EscalationResult
enum (Next(&Model) / AtTop / NotFound). Callers can distinguish typo
from ceiling. 5 new tests.
MEDIUM fixes
- posterior.rs u32 overflow: `(n_plus + n_minus) as u32` →
`u32::try_from(n_plus.saturating_add(n_minus)).unwrap_or(u32::MAX)`.
overflow_guard_on_huge_n test with i64::MAX.
- pick() unknown-model: now returns None when default_model_ref's model
is absent from registry. Inverted the deprecation guard.
- HOME unset: disk_registries_dir() returns None on empty HOME and
falls through to embedded registries. open_ledger() logs warning
and returns None instead of opening at malformed path.
- SQLite WAL + busy_timeout: applied to ledger connection in
open_ledger() — concurrent CLI invocations no longer SQLITE_BUSY.
LOW fixes
- impl Model consolidation: next_tier() moved to pricing.rs.
escalate.rs uses current.next_tier() instead of duplicating logic.
- complexity.rs: removed duplicate "ml-implementer" in HEAVY_ROLES.
- dna_class.rs: role("") now returns None instead of Some("").
Verification (orchestrator-side, RULE 0.13 §Verify-before-commit):
- cargo check → clean
- cargo test --release → 63 passed / 0 failed (was 58 → +5 new tests
cover legacy-slug, EscalationResult, overflow, unknown-model, embedded)
- Constructor Pattern → all files ≤ 200 LOC (max registry.rs 196)
- Largest fn from_ledger 28 LOC / limit 30
DNA-INDEX.md regenerated by kei-registry hook (cosmetic).
=== STATUS-TRUTH MARKER ===
shipped: functional
stubs: 0
cargo-check: PASS
behaviour-verified: yes
follow-up-required:
- (none from this commit; next audit pass before merge to main)
Splits agent definition into stable provider + swappable model + role-bound
profile. Adding a new LLM API is one row in providers.toml; new model is
one row in models.toml; agent invocation picks any (provider, model) pair
through agent-profiles.toml default_model_ref.
- providers.toml: 10 providers — anthropic, openai, codex (OAuth), xai,
deepseek, google, ollama-local, mlx-local, lmstudio-local, litellm-proxy
- models.toml: 11 models with cost_*_per_mtok_micro + context_window +
verified_at + deprecated_at
- agent-profiles.toml: 18 representative profiles; manifest_path points
to the canonical .md in ~/.claude/agents/
Three-layer DNA per the new architecture:
agent-shell::<provider>:<model>:<caps>::<scope8>::<body8>-<nonce8>
This commit only adds registries — kei-model-router still hardcodes
the Claude-only Model enum. Wave 4 will rewire it to read TOML.
Profile bundles the crates kei-buddy needs at runtime:
kei-buddy, kei-telegram-webhook, kei-shared, kei-chat-store,
kei-social-store, kei-memory-sqlite, kei-router, kei-llm-bridge-mlx.
Used by keisei-marketplace cloud-init when user clicks 'хочу своего
KeiBuddy' on /keibuddy/setup — Hetzner VPS spawn pulls KeiSeiKit and
runs install.sh --profile=buddy --yes.
