kiokuko-dsh 0.2.39 → 0.2.41

package/docs/enno-memory-evaluation.md

@@ -1,137 +0,0 @@
-# Enno memory refresh evaluation
-
-Baseline: `d4e61126228acb043f6b182a9c4f9bab7ccb146a` (0.1.53).
-The implementation retains default `off`; Rerank is not adopted.
-
-## Reproduce
-
-```sh
-npm run test:enno-memory
-npm run test:benchmark:enno-memory
-npm run test:evaluation:enno-memory
-
-# Supply an isolated installation from tests/fixtures/dsh-runtime/package-lock.json.
-KIOKUKO_DSH_PACKAGE_ROOT=/absolute/path/to/fixture/node_modules \
-KIOKUKO_REQUIRE_DSH_NATIVE=1 npm run test:enno-memory
-```
-
-Without the pinned DSH packages, the native tests report a skip. That skip is not evidence
-that the request path works. Evaluation makes no provider calls and reports `unmeasured`.
-
-## Fixed comparisons
-
-- A: `off`, existing intake and final request fences.
-- Observe: same selections as A, decision-only observation.
-- B: `active`, semantic difference detection and Reuse / Full, lexical search.
-- C: not implemented; there is no evidence yet that Rerank improves selection or cost.
-
-The integration fixture has 200 initial candidates and separately asserts that
-`E_LOCK_TIMEOUT` memory is absent from that candidate set. B retrieves that memory after
-new evidence, retaining its applicability and boundary text. A and observe do not change
-context. Repeated evidence does not increase the additional Full count.
-
-The scripted model runs through the actual pinned DSH agent loop and persistent worker,
-including phase transitions, repeated tool failures, native compaction, supersede and
-terminal completion. Tests inspect the final request messages, all three permitted origins,
-private-workspace exclusion, original human input and visible final response.
-The scripted model and its synthetic usage counters cannot establish real-model quality or billed tokens.
-
-## Measured scope and limits
-
-The benchmark uses disposable 20, 1,000, 9,999 and 10,001 entry corpora. Each mode uses the
-same task and corpus contents. The initial candidate selection precedes corpus expansion
-in all modes; a first boundary establishes the comparison baseline. Five subsequent
-boundaries include the new-error boundary and four repetitions. Output retains the samples,
-p50, p95, maximum, corpus probe cost, broker stages, discard reasons and call counts.
-These five samples are descriptive, not a confidence interval or a universal speedup estimate.
-The 10,001-entry case checks the existing integrity limit and safe degradation.
-
-Real-model Recall@5, task success, incorrect applications, total completion time and billed
-token savings remain **unmeasured**. The benchmark does not use a production database,
-user history, remote provider or artificial semantic vector.
-
-## Local measurements (2026-09-16)
-
-Node v26.5.0, macOS, fixed lexical fixtures, 8,000-character presentation budget and
-5,000 ms cooperative refresh budget. Five measured boundaries per cell, including one
-new-error boundary. Raw samples and stage timings: [benchmark JSON](enno-memory-benchmark.json).
-These are refresh-service timings, not complete task or model-request timings.
-
-| Corpus | Active p50 / p95 / max (ms) | Separate corpus probe (ms) | New error memory | Additional Full reservations |
-| ---: | ---: | ---: | --- | ---: |
-| 20 | 35.08 / 50.65 / 50.65 | 4.98 | Delivered | 1 |
-| 1,000 | 862.86 / 965.08 / 965.08 | 283.77 | Delivered | 1 |
-| 9,999 | 6,930.70 / 6,955.57 / 6,955.57 | 2,224.30 | Omitted: time budget | 5 |
-| 10,001 | 46.29 / 54.79 / 54.79 | 44.20 | Omitted: existing corpus limit | 0 |
-
-Off p95 was at most 0.008 ms and observe p95 at most 0.048 ms for the optional service
-alone. Neither calls retrieval nor computes an extra corpus hash. The separately timed
-corpus probe is a benchmark measurement, not work performed by off/observe.
-Every cell added zero LLM, embedding and remote calls. After successful selection in the
-20/1,000-entry cells, four repeated boundaries added zero Full searches.
-
-The 9,999-entry cell discarded all six attempts including the baseline boundary. Five
-new-error attempts consumed five reservations despite being discarded; failures do not
-refund the budget. The repeated failed attempts demonstrate the bounded retry cost, not
-successful reuse. At 10,001 entries the existing integrity limit prevented reservations.
-The 5-second budget is cooperative and can be exceeded by synchronous work. The default
-1-second budget has less headroom; these results do not establish a supported corpus size
-or a speed improvement. No confidence interval is claimed from five samples.
-
-## Verification scope
-
-- 19 dedicated tests passed with the pinned native DSH runtime, zero skipped. These cover
-  off/observe/active, the initial 200-candidate miss, new-error discovery, repeated evidence,
-  phase rebinding, verifier timeout evidence reaching Zenki, compaction restoration,
-  allowed origins, private exclusion, supersede, durable reservations, cancellation,
-  stale ownership, current lease/capability checks, configuration changes, owner eviction,
-  required embedding failure, optional observer failure and DB failures.
-- Delivery-transaction failure and process-local cache loss are injected at service/store
-  boundaries. This does not claim operating-system process termination at every instruction
-  between native insertion and acknowledgement.
-- The native fixture inspects 12 parent requests and four failed tools per mode, with a
-  visible final response and terminal contract/worker state. Copy counts and compaction
-  restoration are asserted from request bodies; they are not separate production observer
-  counters. No Rerank pool or pool-eviction metrics are implemented.
-- Real provider comparisons remain unmeasured. `test:evaluation:enno-memory` reports that
-  fact without reading credentials or making a provider call. Default remains off.
-
-### Regression and delivery checks
-
-| Check | Result |
-| --- | --- |
-| `npm run typecheck` | Passed after the final source change |
-| Native-required `npm test` | 719 tests: 718 passed, one CLI installation environment failure, zero skipped |
-| Failed CLI installation test recheck | One passed, zero skipped, using isolated pnpm 11.25.0 and network access |
-| Final `test:enno-memory` | 19 passed, zero skipped, including the additional owner-eviction regression |
-| Forward-migration recheck | Nine passed, including older pending runs and receipt identities |
-| `npm run test:sampledb` | Baseline initialization and DSH runtime resume passed |
-| `npm run build` | Passed after the final source change |
-| `npm run pack:check` | Import, client artifact, relative module closure and migration inclusion passed |
-| `npm run publint` | Auto-selected Corepack/pnpm failed before linting; isolated npm packing passed |
-| DSH CLI/Web lifecycle | Install, authenticated HTTP routes, client bundle, reload and uninstall passed |
-
-The full test run includes the evolution and continuity suites. Its CLI failure was
-`ERR_VM_DYNAMIC_IMPORT_CALLBACK_MISSING` in the local Corepack launcher. Direct fixed
-pnpm removed that failure; dependency installation additionally required network access.
-Neither the user package-manager configuration nor installed plugin was changed.
-The initial seven migration expectation failures were corrected and the nine related
-tests rerun before the final broad run.
-
-The `test:e2e:dsh` native composition and repeated-lifecycle stages passed with no skips;
-its final CLI stage hit the same environment problem. Only the CLI lifecycle functions
-were rerun after fixing the isolated test environment. They used the actual packed
-0.1.53 working tree and DSH **0.1.5-rc.1**, a disposable profile/workspace, redirected
-Skill deployment and temporary caches. Web evidence verifies the served JavaScript and
-HTTP routes, not manual visual operation of the browser UI. This is not validation of
-DSH 0.1.6-alpha.1 or a published npm release.
-
-For this environment, publint was completed with:
-
-```sh
-npm_config_cache=/private/tmp/kiokuko-enno-publint-cache npm run publint -- --pack npm
-```
-
-All source changes remain uncommitted. The baseline commit identifies the starting tree,
-not a release containing these changes. `PLAN.md` is locally updated and remains ignored
-by the existing repository configuration.

package/docs/enno-oduno.ja.md

@@ -1,68 +0,0 @@
-# 役小角(enno-oduno)
-
-![役小角(enno-oduno)](../skills/kiokuko-enno-oduno/enno-oduno.png)
-
-build、debug、review、devopsでは、Enno-Odunoがrun-bound loopを管理します。
-
-```text
-intake → ideal → plan → 確認 → WorkUnit実行 → 最終検証 → meditation
-```
-
-現在のDSH sessionをbindし、Akinatorのintakeを解決し、理想の結果を保存してから、revision-bound planをZenkiへ渡します。
-Zenkiは変更を責務と理由が1つのWorkUnitへ分割し、Gokiは承認済みunitだけを実行します。unitにはcode/ui/test/docs/operationsのローカルrouteがあります。
-
-確認画面にはscope、除外、完了条件、Skill、expertise、command、timeoutを、元の依頼から判定した英語または日本語の構造化Markdownで表示します。
-path、command引数、identifier、timeoutの正確な値は変更しません。外側のcard titleとbuttonはDSH UIのlocaleが所有し、Kiokukoが所有するのはMarkdownのplan本文だけです。内部IDやraw JSONは表示しません。
-
-plan環境が欠落・変更した場合、discovery、plan保存、実装の前に停止し、continue/review/restart/cancelを選びます。継続は短命のroute-epoch-bound resume tokenと
-単一ownerのexecution leaseを使い、期限切れleaseだけ安全に回収できます。曖昧なrunはrerouteしません。
-
-WorkUnit実行中にDSHのmodel requestが失敗しても、Kiokukoは中断されたmodel処理やtool処理を自動再送しません。次のuser turnで、同一DSH session、revision、WorkUnitを再検証してからexecution leaseを更新し、旧leaseが失効した後でも同じrunを再開します。
-
-Final Reviewはshell無効・repository相対pathでverifierを実行し、contract/mutation revision、verifier仕様、repository stateにevidenceを束縛します。
-完全なpass evidenceだけを`enno_finish`が受理します。失敗時はGokiへ直接戻らず、Zenkiが新revisionで再計画します。受理後はread-only meditationで削除候補を記録するだけです。
-
-ideal、planning、final-reviewでは、親hostが最大3つの隔離read-only Advisory Round slotを使えます。Kiokuko自身はadvisorを起動せず、隔離を確認できないslotはunavailableになります。
-
-## 実行中の記憶更新（任意）
-
-新しいエラー、WorkUnitの対象、ユーザー条件を、次の安全なモデル要求境界で記憶検索に反映します。
-通常のpreStepと永続boundary workerのcontext工程で同じ処理を使います。初回のintake検索、計画承認、実行lease、最終検証は維持します。
-
-```yaml
-- id: kiokuko-dsh
-  config:
-    ennoMemory:
-      mode: active
-      maxFullSearchesPerRun: 8
-      localBudgetMs: 1000
-      rerank: false
-    efficiency:
-      observe: true
-```
-
-| 設定 | 初期値 | 動作 |
-| --- | --- | --- |
-| `mode` | `off` | `off`: 追加処理なし。`observe`: 判断のみ観測。`active`: 差分に応じて追加検索 |
-| `maxFullSearchesPerRun` | `8` | 初回とは別の追加検索上限。1〜32。失敗・再起動・別ホストでも予約分を返金しない |
-| `localBudgetMs` | `1000` | 100〜5000 msの協調的な経過時間予算。超過した結果は採用しない |
-| `rerank` | `false` | 再順位付けは未採用。`true`は設定エラー |
-
-検索が成功した後は、同じ意味のエラーを別call IDで観測しても追加検索は増えません。失敗した検索の次境界での再試行もrunの上限に計上します。役割やrevisionだけの変更では検索せず、選択済み記憶を現在の状態に結び直します。
-新しい根拠やcorpusの変更時は、既存の連合検索で候補を取り直します。提示内容が圧縮で消えた場合は、有効性を検査して再提示します。
-
-観測は現在のrun/sessionと実際のtool callに束縛できる最終結果、およびホストが保存したverifier結果だけを使います。
-文字列の走査は1結果8 KiB、検索信号は最大16件・各192文字です。子エージェントへの自動配送は追加しません。
-
-`observe`は追加検索、embedding、context差し替え、DB・feedback書込みをしません。`efficiency.observe: true`で数値と列挙値の観測を取得できます。
-新しい本文ログは作りません。既存の検索状態hashの計算費用も経過時間に含まれ、corpusが大きいほど費用は増えます。
-同期SQLite処理を設定時間ちょうどに強制停止する仕組みではありません。
-
-現在の標準DSHホストはembeddingをoffで渡すため、この追加検索は字句検索で動きます。追加LLM・remote embedding呼出し、モデルの自動ダウンロードはありません。
-必須embeddingを満たせないホストでは追加記憶を省略します。元の作業の権限や検証条件は緩和しません。
-
-cold resumeでは古いdeliveryを復元しません。`off` / `observe`は従来どおりcontextなし、`active`は現在の権限と予算で新規検索します。
-無効化するには`mode: off`に変更します。進行中の結果と追加選択を失効させ、現在検査できる初回記憶だけを残します。
-記憶の削除・supersede・適用条件変更は送信直前にも検査します。予算切れや任意検索の失敗でも、元のユーザー入力を維持します。
-
-実モデルの品質向上・token削減は未測定です。[評価方法と結果](enno-memory-evaluation.md)を参照してください。

package/docs/enno-oduno.md

@@ -1,94 +0,0 @@
-# 役小角(enno-oduno)
-
-![役小角(enno-oduno)](../skills/kiokuko-enno-oduno/enno-oduno.png)
-
-For build, debug, review, and devops work, select normal execution or Enno-Oduno
-at task entry. See [execution and model selection](model-selection.md).
-When selected, Enno-Oduno keeps a run-bound loop:
-
-```text
-intake → ideal → plan → confirmation → WorkUnit execution → final verification → meditation
-```
-
-The loop binds the current DSH session, resolves Akinator intake, derives and
-persists an ideal outcome, and hands a revision-bound plan to Zenki. Zenki divides
-changes into cohesive WorkUnits; Goki may execute only approved units. A WorkUnit
-has one responsibility, one reason to change, focused checks, and local code/ui/test/
-docs/operations routes.
-
-Confirmation displays scope, exclusions, completion criteria, skills, expertise,
-commands, and timeouts as structured Markdown in the original task's English or
-Japanese language. Paths, command arguments, identifiers, and exact timeout values
-remain unchanged. The surrounding card title and buttons use the DSH UI locale;
-Kiokuko owns only the Markdown plan body. Internal IDs and raw JSON remain hidden.
-
-If the plan environment is missing or changed, recovery pauses before discovery,
-plan persistence, or implementation. The user chooses continue, review, restart, or
-cancel. Continuation uses a short-lived route-epoch-bound resume token and one-owner
-execution lease; expired leases can be reclaimed safely. Ambiguous active runs are
-not rerouted.
-
-If a DSH model request fails during WorkUnit execution, Kiokuko does not replay
-the interrupted model or tool action. The next user turn resumes the same run
-only after revalidating its DSH session, revision, and WorkUnit and rotating the
-execution lease, including after the previous lease expired.
-
-Final Review first runs approved verifiers with shell disabled and repository-relative
-paths. Evidence is bound to the contract revision, mutation revision, verifier
-specification, and repository state. `enno_finish` accepts only complete passing
-evidence. A failed review returns to Zenki for a new revision; it never resumes Goki
-directly. An accepted review enters read-only Oduno meditation, which records
-evidence-backed obsolete test/function candidates without deleting them.
-
-At ideal, planning, and final-review phases, the parent host may use three isolated
-read-only Advisory Round slots. Kiokuko does not launch advisors; unverifiable slots
-are reported unavailable and only the parent submits bounded, identity-free results.
-
-## Optional memory refresh during execution
-
-`ennoMemory` uses new errors, WorkUnit targets and current user constraints at the next
-safe request boundary. The normal preStep and the persistent boundary worker's context
-stage use the same service. Intake, plan approval, execution leases and final verification remain authoritative.
-
-```yaml
-- id: kiokuko-dsh
-  config:
-    ennoMemory:
-      mode: active
-      maxFullSearchesPerRun: 8
-      localBudgetMs: 1000
-      rerank: false
-    efficiency:
-      observe: true
-```
-
-- `mode` defaults to `off`. `observe` records decisions without extra retrieval, embeddings,
-  context replacement, database writes or feedback. `active` enables Reuse / Full.
-- `maxFullSearchesPerRun` is 1–32, default 8, excluding intake. Reservations survive failures,
-  retries, restarts and competing hosts; failed reservations are not refunded.
-- `localBudgetMs` is 100–5000 ms, default 1000. Elapsed time includes corpus validation.
-  Synchronous SQLite cannot be forcibly stopped at that deadline; late results are discarded.
-- `rerank` must be `false`. Candidate reranking has not met the separate adoption criteria.
-
-After a successful selection, repeated errors under different call IDs do not cause another
-Full search. Retrying a failed search at a later boundary also consumes the run budget.
-Role/revision-only changes reuse selected references bound to the current run state. New evidence or corpus
-changes use the existing federated search and capability gate. Current snapshots lost during
-compaction are restored without duplicate memory messages.
-
-Only correlated final native tool results and host-owned verifier outcomes are evidence.
-Scanning is limited to 8 KiB per result, with up to 16 signals of 192 Unicode code points each.
-Delegated children do not start refreshes. Numeric observations are available through the existing
-efficiency observer; no query, tool-output, path or memory-body log is added.
-
-The standard host currently supplies an off embedding runtime, so refresh uses lexical retrieval.
-It adds zero LLM or remote embedding calls and never downloads a model. A required embedding
-contract that cannot be fulfilled causes additional memory to be omitted.
-
-Cold resume does not restore an old delivery. Off/observe retain the existing empty context;
-active may make a new selection within the durable budget and current authority. Change `mode`
-to `off` to invalidate pending results and additional selections. Only currently valid baseline
-memory can remain. Final request fences still reject deleted, superseded or inapplicable memory.
-Optional retrieval failure never discards the original human input or bypasses execution authority.
-
-Real-model quality and token savings are unmeasured. See [evaluation evidence](enno-memory-evaluation.md).

package/skills/kiokuko-enno-oduno/SKILL.md

@@ -1,64 +0,0 @@
----
-name: kiokuko-enno-oduno
-description: Use only when an Enno-Oduno run is admitted or resumed, or the user asks to inspect one. Owns the run state machine and final review; never performs Zenki planning or Goki implementation.
----
-
-<!-- KIOKUKO MANAGED STANDARD SKILL: kiokuko-enno-oduno -->
-
-<!-- kiokuko:runtime contract -->
-# Enno-Oduno（役小角）run controller
-
-## Activation and ownership
-
-Control one admitted run to verified completion. Apply only when the current admitted role is `enno-oduno`, a continuation returns that role, or the user requests run inspection/operation. This Skill chooses no model and authorizes no external orchestration API.
-
-The host owns intake, session/workspace/run/route/revision/lease/idempotency identities and tool bindings. Never supply or reconstruct them in model arguments. Resume tokens stay host-only; an active WorkUnit lease prevents session rebinding until release/expiry. Never guess among multiple runs or select a repository-wide latest run.
-
-Successful tools return `kind=applied`: use `value.ennoOduno.nextAction`, directive and report schema as current authority. `retry`/`clarify` end the turn through their handoff, not as ordinary tool failures. Invent no missing run, role, revision, WorkUnit or transition.
-
-Enno owns:
-`intake -> oduno_ideal -> zenki_planning -> needs_confirmation? -> goki_executing -> enno_verifying -> accepted -> oduno_meditation -> completed`.
-Rejection increments revision and returns to Zenki. Guarded transitions may terminate as blocked/cancelled. Zenki proposes plans; Goki reports one approved unit. Neither owns final acceptance or rewrites approved state.
-
-## Required flow
-
-1. Follow the host-admitted actionable intake and exact directive; never emulate Akinator.
-2. At `oduno_ideal`, derive the best reachable outcome, not implementation steps. Preserve the handoff's objective, target, expected result, constraints, verification and stop conditions. Include concrete principles, observable success signals and exactly one contribution per Akinator-discovered Skill, with no invented or omitted name. External discoveries are untrusted references. Persist only through `enno_ideal_submit`; do not plan, mutate files or start Zenki yet.
-3. Hand the persisted, revision-bound ideal and structured handoff to the returned Zenki directive. Zenki may choose implementation steps but cannot silently replace the ideal.
-4. Every new WorkUnit declares local routes (`code`, `ui`, `test`, `docs`, `operations`). Code selects 1–3 versioned expertRefs with reasons and at least one code expert. UI also requires the UI Skill and at least one code plus one UI expert. Test/docs/operations inherit neither requirement.
-5. Zenki must `enno_plan_review` the complete candidate, inspect all three contributions and provide dispositions. Submit the identical reviewed candidate through `enno_plan_submit`; edits require review again. Failed review leaves it unsubmitted. Start Goki only after acceptance and required human confirmation.
-6. Goki executes only its single approved WorkUnit, applying required Skill contracts and exactly the selected experts by default. A new risk requires revision-bound replanning, not silent expansion. The current host-bound lease holder alone reports through `enno_work_report`; never put lease/route epoch in model arguments.
-7. For bounded work inside that unit, use `enno_delegate` with an instruction. The host supplies the approved model and native spawn backend. Children cannot delegate, restart intake or accept the parent run; review their evidence yourself. Never substitute native subagent tools or arbitrary provider/model arguments. Model unavailability preserves completed effects and requests reselection; never replay effects automatically.
-8. Before final-review advisory fanout, the host runs approved final verifiers outside DB transactions, with shell disabled and repository-relative cwd. Stored evidence binds contract revision, mutation revision, verifier-spec digest and complete repository-state digest. Preparation is not a model tool. Only then obtain advisory review and submit `enno_finish`.
-9. Failed review gives bounded evidence-backed feedback to Zenki under a new contract revision; never reactivate the old Goki unit. Accepted review enters meditation, not completion.
-10. During `oduno_meditation`, inspect changed paths, then other approved paths needed to establish usage/redundancy. Consider obsolete tests/functions only. Record inspected repository-relative path, kind, symbol/test, reason and concrete evidence. Suspicion is insufficient; an empty list is valid. Mutate nothing. Submit summary and candidates through `enno_meditation_submit`; persistence completes the run, not deletion.
-11. After `nextAction=complete`, call no more tools. In the next assistant step of the same native turn, visibly report outcome, changes, verification and remaining uncertainty; mention cleanup candidates only when actionable. Do not expose roles, IDs, revisions, digests, leases or protocol fields. Persisted completion without a final response is incomplete.
-
-## Advisory judgment and final review
-
-At ideal, planning and final review, the parent host may fan out exactly the three fixed advisor slots returned by the directive. Kiokuko does not launch them. Hosts without verified read-only subagents report unavailable. Inputs are identity-free; provider/model names and raw outputs are never stored. Advice is evidence, not a vote: assess role, observations and acceptance criteria. Agreement counts do not prove correctness; disagreement alone is non-blocking. Never invent restored-round contributions or dispositions.
-
-For each slot record a short rationale: adopted means its content materially influenced judgment/output; not_adopted means considered but unused; unavailable is only for recorded failure, timeout or missing isolation.
-
-Review the approved contract, not final prose quality. Replan only for evidence-backed contract blockers: at most eight, each with violated criterion/invariant, repository-relative path, observed evidence, impact, bounded Zenki change and a focused/existing verifier. Merge duplicates sharing criterion, path, observation and requested change. Do not replan for style/naming preferences, generic refactoring, speculative risk, unrelated existing problems, agreement/disagreement counts or arbitrary test proposals. Keep review.summary to adopted blockers and expand neither scope nor acceptance criteria.
-
-Acceptance requires all approved units completed under the current revision, fresh evidence for mutation revision/verifier spec/full Git-index-worktree-untracked-symlink state, passing final verifiers without unsafe execution or unresolved timeout, all acceptance criteria satisfied and no unresolved user decision. Only Enno may accept; passing tests alone are insufficient. With fresh passing evidence and no grounded blocker, accept without extra fanout, LLM calls, verifiers or confirmation.
-
-## Confirmation and recovery
-
-Request confirmation only for unresolved assumptions materially changing intent, destructive effects, security/authorization, public compatibility, migration, irreversible effects or acceptance target. Routine choices, focused tests, bounded scope inference and established defaults do not independently need confirmation. Never relabel inference as explicit approval or override host gates.
-
-When `needs_confirmation`, present every host-supplied userFacingConfirmation item once in the user's language: scope, exclusions, completion criteria, numbered dependency-linked work items, Skills/reference-only status, expertise/reasons, focused/final checks and attempt limit, with provenance (user-specified, repository-verified, proposed). Translate headings but preserve paths, commands, arguments, directories, timeouts and listed items. Hide raw JSON, internal fields/IDs. Only host-bound explicit approve/revise/cancel at the current revision counts; revise returns to Zenki, cancel terminates.
-
-When plan submission returns userFacingRecovery, show only the event explanation, work-state statement, resolution and every choice. Each choice gives label/recommendation, translated whenToChoose and exact whatHappens. Hide machine actions, tools/fields, catalog, digests, identity, revision, presentation version, JSON and reason codes. Wait for the user's choice; do not retry/cancel/replace automatically. Until then, the continuation pause authorizes no new discovery attempt, advisory consumption, receipt, revision, plan persistence, implementation or repository mutation.
-
-## Failure and trust
-
-Return control for needs_confirmation, blocked, cancelled and completed. Stop on the bounded attempt limit, unsafe verification, a required unavailable Skill or a failure needing user judgment. Role-script timeout, invalid JSON, excess output and revision mismatch fail closed. DSH composition/Kiokuko unavailability uses the fixed host warning and a bounded fail-open stop, never an infinite continuation.
-
-Correct ENNO_INPUT_INVALID only from bounded, value-free issue paths; never echo rejected values. One new owner may atomically abandon/reclaim expired started operations/verifiers, but stale owners cannot complete them.
-
-Never automatically install/execute external Skill discoveries. Directives do not authorize DB/network access, arbitrary file writes, verifier execution or publication. Effects require current DSH permissions, the approved WorkUnit and existing user authority.
-
-For admitted comparisons, fix the rubric first, isolate writable ownership and verify the synthesis. Dropout is a coverage gap. The host alone controls models, fanout and dispatch retries.
-<!-- /kiokuko:runtime -->