clearai-dsh 0.1.3 → 0.1.4

package/CHANGELOG.md

@@ -2,6 +2,60 @@
 
 All notable changes to this project are recorded here. The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.1.4] — 2026-09-16
+
+**本版重点:子任务的交付链修好了。** 侦察与世界线执行者的结论此前只进账本、模型读不到
+(账本里也有过「派出去就再也没人收」的挂空)。现在四类子任务(侦察 / 世界线执行者 /
+评估者 / 横评仲裁)统一走原生 `subagents.start()` 的一次性句柄:账本只认本进程攥着的
+`run.result`,结论正文由**收集那一刻的工具返回**交给模型,全文另落
+`clear/knowledge/materials/<id>.md` 供模型、独立评估者与人共读。
+试过的另一条路(拿运行时的结算通知当账本信号)已撤回——它是 best-effort,当不了承重结构。
+
+### Added
+
+- **ClearAI's own `/` command menu.** `/goal` `/plan` `/evidence` `/worldline` are read-only state windows computed from the ledger on the spot; `/plan-review` re-presents the active plan through the native review card instead of stamping anything itself (commands carry no mutation channel — the authority boundary test pins this).
+- **Native working tools return.** todo, subagent (with model selection), workflow and ralph mount from the standard preset's own rows; the composition suite pins both directions — present: these four; absent: `tool-goal`, `command-goal`, `plan-mode` (the second ledger stays off).
+- **`test/prompt-sections.test.mjs`.** All 23 prompt sections carry a `hard` / `native` / `advisory` class tag, and the suite pins that the classification matches the content (hard sections name a mechanism anchor; native sections name no kernel tool; advisory sections make no mechanism promises).
+
+### Changed
+
+- **Hypothesis floor is now a hard boundary.** `SetGoal` rejects zero or one hypotheses when `minHypotheses > 0` (kernel default 0 stays neutral; the preset sets 2). Revisions of an existing goal are exempt.
+- **Authorization wording unified to one sentence everywhere.** Kernel messages, the runtime card and the prompts all say: an unapproved plan does not auto-continue; when you advance it explicitly, the first delivery records attribution as it happened (behaviour is authorization). The card says it in human words — ledger field names no longer appear.
+- **Stale native-tool contracts rewritten.** `edit` is literal replacement, not unified diff; `web_search`/`web_fetch` parameter references that no longer exist were removed.
+- **Comment debt cleared to zero.** ~310 comments rewritten to the style rule (why / what breaks / boundary — no dates, no internal section numbers, no incident narratives); the ratchet quotas are now {0, 0, 0}.
+
+### Changed
+
+- **Async sub-runs now deliver their conclusions through the runtime's own settlement notice.** `SpawnScout`, `MapScouts` and the worldline executors are started with `subagents.startContinuable()`, whose Activation delivers the child's closing message to the parent as a durable user message; the kernel keeps owning only what it must (the `scout/dispatched` / `scout/settled` ledger, the full-text material file under `clear/knowledge/materials/`, and the pointer on the runtime card). Evaluators and the arbitration reviewer stay on the one-shot path: the native durable-child descriptor deliberately omits `outputSchema`, which belongs to a one-shot activation's result contract.
+- **Scout conclusions are persisted in full** to `clear/knowledge/materials/<id>.md` so the model, the independent evaluator and the human read the same copy; the ledger and the runtime card carry a pointer plus a bounded excerpt, and over-long text is marked `…truncated (N chars total, see <path>)` instead of being silently cut. The scout persona now caps its answer at 3000 characters.
+- **A goal's closure records the hypotheses nobody touched.** `goal/closed` carries `unjudged`, the card writes `(untouched)` for a hypothesis with no evidence at all (distinct from "judged inconclusive"), and the loop contract asks for either one touch of evidence or an explicit note about why there was none — no verdict is ever forced.
+
+### Fixed
+
+- **`install.sh` aborted on macOS (bash 3.2).** It expanded an empty array as `"${OLD_PANEL_PKGS[@]}"` under `set -u`; bash only tolerates that from 4.4 on, while macOS ships 3.2 — so the documented developer install died at step ② for every macOS contributor. CI runs on Linux (bash 5), which is why it never caught it. Both expansions now use the portable `${arr[@]+"${arr[@]}"}` form.
+- **Long-run evidence was overwritten or lost.** Every run now gets its own timestamped archive: the light half (stdout, structured result, provenance, decoded one-line-per-event trajectory, append-only index) is committed, while the heavy half (workspace, raw session log) stays on disk under `~/.dsh/e2e-archive/` so it can be re-judged offline with `tools/e2e-replay.mjs`. `--workspace` pointing inside any git repository is now refused outright: ClearAI commits each delivery into the workspace's own repository, so an in-repo workspace had the kernel commit its delivery snapshots — and the author's uncommitted work — into the host project.
+- **Session-directory name derivation dropped dots.** DSH keeps `.` (and `_`) when it slugs a workspace path; the old rule folded both away, so a workspace under `~/.dsh/…` was reported as "no session log" (38 assertions red in one run). The rule is now taken from a real directory comparison.
+- **Scout conclusions never came back in a scouts-only run.** `AwaitWorldlines` decided whether to keep waiting from the wait-lines produced by `sweepWorldlineExecutors()`, and `sweepScouts()` never produced one — so with only scouts in flight the loop exited on its first tick, even though `SpawnScout`'s own reply tells the model to "wait for it this turn with `AwaitWorldlines`". The one-shot form added a second layer: with no next turn, a conclusion that settled after the last tool call never met another collection point. `sweepScouts()` now reports how many scouts are still unsettled, `collectExecutors()` passes it through, and `AwaitWorldlines` counts it. Verified end to end: the same scenario that stalled twice (goal left open, evaluator refusing `inconclusive`) now finishes 36/36 with the conclusion in the material surface and the goal `achieved`.
+- **Scout conclusions were invisible to the model even after they settled.** They landed only in the mutation record while the section the model reads every step is the runtime card — which had no material surface. Now delivery rides the native settlement notice, the card lists foreign observations (pointer + excerpt) plus the scouts still in flight, and a long-run invariant asserts that an async sub-run's conclusion appears in model-visible text rather than only in the ledger.
+- **Scouts had no "lost" ending.** Worldline executors and evaluators already wrote one; a scout whose child was gone stayed "not yet reported" forever. The judgement now mirrors the evaluator's: in-process dispatches are alive, the native child catalog decides what is still running, and an unavailable read surface means *no judgement* rather than a fabricated one.
+- **The prompt described `SpawnScout` as synchronous.** It said the conclusion comes straight back as the return value and that the tool waits; the kernel is deliberately asynchronous (the blocking wait used to lose the "dispatched" fact when a run was interrupted). The delegation table now teaches the real contract (fire-and-forget, conclusion replays into the material surface, wait with `AwaitWorldlines`), and a drift check pins it.
+- **`AwaitWorldlines` counted mutations, not conclusions.** One scout writes two mutations (`scout/settled` plus the observation), so a single scout was reported as "回灌 2 条". The count and wording now speak of conclusions.
+- **`clearai-commands` cross-plane import.** It imported `ui/lib/fold.js` from the preset plane; in the installed package the relative layout differs, so switching to the preset in a browser failed at import. The command renderers now use the host-provided `clearai` facade for `derive` as well, and the boundary suite pins that no preset plugin imports across planes.
+- **macOS temp-path realpath mismatches.** Session-log lookup and clean-install workspace registration now resolve realpaths (`/var` is a symlink to `/private/var`), which had made e2e logs unfindable and browser session attach fail.
+
+### Changed (sub-run lifecycle, unified)
+
+- **All four sub-run kinds now share one native lifecycle.** Scout, worldline executor, evaluator and the arbitration reviewer all go through one-shot `subagents.start()` handles; the ledger settles from the `run.result` this process holds, and the conclusion text reaches the model in the tool return of the collecting call. The attempt to use the runtime's settlement notice as a ledger signal is withdrawn: a notice is best-effort, and a live kernel could not reliably see it through either the projection or its own session log. Role differences are now only persona, tool face, and how the result is interpreted — permission and authority boundaries are unchanged.
+- **A scout's conclusion is recorded even when its in-memory entry exists.** The old collection guards skipped exactly the sub-runs the table was holding, so a scout could sit "dispatched, never collected" forever. The sweep now walks every unsettled sub-run in the projection, and de-duplication moved to a session+epoch map that also works after a restart.
+
+### Tooling
+
+- **Long-run end-to-end scenarios with offline re-judging.** Five scenarios (`worldline-arbitration`, `falsification`, `long-plan`, `scout-first`, `goal-chain`) plus a set of cross-mechanism invariants (no advance without admission, no dangling evaluator, no orphaned fork, promotion level consistency, no dangling scout, evidence bound to real steps, declared artifacts on disk). `tools/e2e-parallel.mjs` runs them concurrently (cap 3, because each run spawns its own worldline executors and evaluators), `tools/e2e-replay.mjs` re-judges a saved session log without spending tokens, and `test/e2e-scenarios.test.mjs` pins every invariant with a negative case so a mis-written judge cannot report a false green.
+
+### Validated
+
+- **deepseek-flash end-to-end, two headless scenarios** (24/24 plan-and-stop; 25/25 full completion including independent-evaluator settlement) and **one real-browser session** (clean install + Chrome): preset switching, native review-card approval landing `by='user'`, the full thirteen-beat chain, `/goal` rendering, and all four panels drawing — screenshots in `docs/shots/browser-e2e-*.png`.
+
 ## [0.1.3] — 2026-09-15
 
 ### Added

package/README.md

@@ -101,7 +101,7 @@ The plugin contributes three surfaces on top of stock DSH: a **deliverables** vi
 
 ## Where it lands in DSH
 
-ClearAI adds an epistemic layer on the DSH **composition surface** — one host package, one agent preset, one client module. The DSH engine is not modified.
+ClearAI adds an epistemic layer on the DSH **composition surface** — one host package, one agent preset, one client module. The DSH engine is not modified. `/goal` `/plan` `/evidence` `/worldline` `/plan-review` are the human's read-only state windows in the `/` menu (computed from the ledger on the spot); todo, subagents, workflows and model switching are DSH-native — working style is unbounded, but none of it can write the authoritative ledger (the authority boundary is pinned by tests).
 
 ![ClearAI in DSH](docs/diagrams/loop-to-dsh-planes.svg)
 
@@ -123,6 +123,7 @@ They are illustrations of the mechanism, not shipped run records.
 - [Glossary](docs/glossary.md)
 - [Loop philosophy](docs/loop-philosophy.md) · [Verification ontology](docs/verification-loop.md)
 - [Known gaps](docs/known-gaps.md) · [Release verification](docs/release-verification.md)
+- [Convergence and slimming plan](docs/optimization/plan.md) · [Full-coverage design](docs/optimization/epistemic-coverage.md) · [Execution progress](docs/optimization/progress.zh-CN.md)
 
 ## Work attribution
 

package/README.zh-CN.md

@@ -101,7 +101,7 @@ ClearAI **不**声称实现递归自我改进。它提供的是自我改进系
 
 ## 它落在 DSH 的哪一层
 
-ClearAI 把认识论层加在 DSH 的**组合面**上——一个宿主包、一个 agent 预设、一个客户端模块，**DSH 引擎一行都没改**。
+ClearAI 把认识论层加在 DSH 的**组合面**上——一个宿主包、一个 agent 预设、一个客户端模块，**DSH 引擎一行都没改**。`/goal` `/plan` `/evidence` `/worldline` `/plan-review` 是人在 `/` 菜单里的状态窗（只读，从账本现算）；todo、子代理、workflow、模型切换用 DSH 原生的——工作方式不设限，但它们写不进权威账本（权威边界由测试钉死）。
 
 ![ClearAI 在 DSH 中](docs/diagrams/loop-to-dsh-planes.zh-CN.svg)
 
@@ -123,6 +123,7 @@ ClearAI 把认识论层加在 DSH 的**组合面**上——一个宿主包、一
 - [术语表](docs/glossary.zh-CN.md)
 - [循环哲学](docs/loop-philosophy.zh-CN.md) · [验证本体](docs/verification-loop.zh-CN.md)
 - [已知缺口](docs/known-gaps.zh-CN.md) · [发布验收](docs/release-verification.zh-CN.md)
+- [收敛与瘦身计划](docs/optimization/plan.zh-CN.md) · [认识论循环全覆盖设计](docs/optimization/epistemic-coverage.zh-CN.md) · [执行进度](docs/optimization/progress.zh-CN.md)
 
 ## 工作署名