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.
|
||
|---|---|---|
| .. | ||
| src | ||
| Cargo.lock | ||
| Cargo.toml | ||
| README.md | ||
kei-model-router
Model selection (Haiku 4.5 / Sonnet 4.6 / Opus 4.7) for Claude Code Agent spawns. Empirical-posterior decision rule keyed on task-class DNA + Beta posterior + cost minimization, with kernel-smoothing for unseen task classes.
Math
Decision rule:
m*(d̂) = argmin_{m ∈ M} { c(d̂, m) | P[q(d̂, m) ≥ q*] ≥ 1 − δ }
Empty feasible set → fallback (top tier) per RULE -1 NO DOWNGRADE.
Posterior: q(d, m) ~ Beta(α₀ + n⁺, β₀ + n⁻) with uniform prior. n⁺ counts rows where outcome='functional' AND escalation_depth=0; n⁻ everything else.
Kernel-smoothed transfer for unseen task classes:
K(d, d') = α_role · 1[role=role'] +
α_caps · |caps ∩ caps'| / |caps ∪ caps'| +
α_scope · 1[scope=scope'] +
α_body · jaccard_bigram(body, body')
Verified pricing
[VERIFIED: https://platform.claude.com/docs/en/docs/about-claude/pricing 2026-04-30]
| Model | Input/MTok | Output/MTok |
|---|---|---|
| Haiku 4.5 | $1.00 | $5.00 |
| Sonnet 4.6 | $3.00 | $15.00 |
| Opus 4.7 | $5.00 | $25.00 |
Opus 4.7 uses a new tokenizer that may produce up to 35% more tokens on identical text — multiply quote accordingly when comparing against Haiku/Sonnet on the same input.
CLI
kei-model-router pricing # print pricing table
kei-model-router select <agent> [--prompt P]
kei-model-router calibrate # re-fit kernel weights
kei-model-router --help
Orchestrator integration (Path B — runtime)
Per RULE 0.13, the orchestrator owns Agent spawning. Before spawning a
non-trivial agent the orchestrator can consult the router and pass an
explicit model parameter:
kei-model-router select code-implementer-rust \
--prompt "Add multi-tool integration test for parser"
# → model: sonnet (if posterior built up); model: opus (initial fallback)
Then in the orchestrator's Agent invocation:
Agent({ subagent_type: "code-implementer-rust", model: "sonnet", ... })
Until posterior data accumulates the router conservatively returns
top-tier (Opus). As outcome column fills via agent-fork-done.sh STATUS-TRUTH
parsing, posterior diversifies and cheaper tiers begin to qualify.
Assembler integration (Path A — compose-time, deferred)
Rebaking the model into generated .md files at assemble time is
deferred. Current default model: opus in 55/55 manifests is safe;
adopt Path B (orchestrator discipline) until ledger has ≥100
outcome-tagged rows per common task class.
Cubes
| File | LOC | Concern |
|---|---|---|
| pricing.rs | 167 | Verified per-MTok constants (microcents) |
| dna_class.rs | 113 | DNA component extraction (role/caps/scope) |
| complexity.rs | 178 | τ-estimator (heuristic regex+role+length) |
| posterior.rs | 197 | Beta posterior from ledger + Wilson lower b. |
| kernel.rs | 134 | Substrate similarity kernel for unseen DNAs |
| escalate.rs | 73 | Retry-ladder bookkeeping |
| select.rs | 197 | Decision rule (argmin cost s.t. q_lb ≥ q*) |
| calibrate.rs | 193 | Offline LOO weight refit (grid search) |
| main.rs | 142 | CLI dispatch |
| lib.rs | 35 | Module barrel + re-exports |
All cubes within Constructor Pattern budgets (≤200 LOC, ≤30 LOC/fn).
Schema dependency
Requires kei-ledger schema v9+ which adds:
tokens_in INTEGER
tokens_out INTEGER
stubs_count INTEGER DEFAULT 0
outcome TEXT CHECK (outcome IN ('functional','partial','scaffolding','fail',NULL))
escalation_depth INTEGER DEFAULT 0
task_class_dna TEXT GENERATED ALWAYS AS (...) VIRTUAL
INDEX idx_agents_task_class ON agents(task_class_dna)
Hooks
agent-fork-logger.sh(PreToolUse:Agent, advisory) — writes 'running' row with DNA at spawn.agent-fork-done.sh(PostToolUse:Agent) — closes row + parses STATUS-TRUTH MARKER from agent's tool_response → updates outcome, stubs_count, tokens_in, tokens_out.
Lock
2026-04-30. Phases 1-9 of kei-model-router rollout complete (Phase 9 orchestrator-discipline; assembler refactor deferred until ≥100 outcome-tagged rows accumulate).