pi-plans 0.8.0 → 0.8.2

package/tests/exec.test.ts

@@ -337,9 +337,13 @@ describe("task-tree execution core", () => {
 				persistTaskProgress(ctx);
 			}
 			// Simulate an exhausted budget persisted from earlier attempts.
+			// v0.9.3: pin the (now non-default) 5-round budget so this test keeps
+			// exercising the same numeric-exhaustion path.
+			getExecution()!.reviewBudget = 5;
+			getExecution()!.reviewRoundsTotal = REVIEW_MAX_ROUNDS;
 			getExecution()!.audit.rounds = REVIEW_MAX_ROUNDS;
 			getExecution()!.audit.failed = ["VC-002"];
-			mutateCheckpoint(workdir, runId!, (cp) => applyExecutionProgress(cp, { audit: { rounds: REVIEW_MAX_ROUNDS, lastResult: "VC-002" } }));
+			mutateCheckpoint(workdir, runId!, (cp) => applyExecutionProgress(cp, { reviewBudget: 5, reviewRoundsTotal: REVIEW_MAX_ROUNDS, audit: { rounds: REVIEW_MAX_ROUNDS, lastResult: "VC-002" } }));
 			const snapshot = getExecution();
 			const ctxUi = { ...makeCtx(workdir), mode: "tui" as const };
 			await restoreFromSession(ctxUi, [{ type: "custom", customType: "pi-plans-exec", data: snapshot }]);
@@ -363,6 +367,8 @@ describe("task-tree execution core", () => {
 			applyTaskUpdate(getExecution()!.tasks, id, "complete", `${id} evidence`);
 			persistTaskProgress(ctx3);
 		}
+		getExecution()!.reviewBudget = 5;
+		getExecution()!.reviewRoundsTotal = REVIEW_MAX_ROUNDS;
 		getExecution()!.audit.rounds = REVIEW_MAX_ROUNDS;
 		getExecution()!.audit.failed = ["VC-002"];
 		const headlessCtx = { ...makeCtx(wd2), hasUI: false } as never;
@@ -415,9 +421,20 @@ describe("task-tree execution core", () => {
 			persistTaskProgress(ctx);
 		}
 		// Exhaust the budget, then pause at the cap exactly like the loop does.
+		// v0.9.3: an explicit 5-round budget keeps the old numeric-cap shape.
+		getExecution()!.reviewBudget = 5;
+		getExecution()!.reviewRoundsTotal = REVIEW_MAX_ROUNDS;
 		getExecution()!.audit.rounds = REVIEW_MAX_ROUNDS;
 		getExecution()!.audit.failed = ["VC-001", "VC-002"];
-		const ctxTui = { ...makeCtx(workdir), mode: "tui" as const };
+		// The grant re-opens the picker (v0.9.3): this TUI host answers through
+		// its menu with the same 5-round budget; a bare panel resolution would
+		// mean "closed without a choice" and keep the run paused.
+		const baseCtx = makeCtx(workdir);
+		const ctxTui = {
+			...baseCtx,
+			mode: "tui" as const,
+			ui: { ...(baseCtx as { ui: Record<string, unknown> }).ui, select: async (_title: string, options: string[]) => options.find((o) => o.startsWith("5 round")) },
+		};
 		const snapshot = getExecution();
 		await restoreFromSession(ctxTui, [{ type: "custom", customType: "pi-plans-exec", data: snapshot }]);
 		await __awaitReviewRoundForTests();
@@ -433,14 +450,16 @@ describe("task-tree execution core", () => {
 		assert.equal(getExecution()!.audit.rounds, beforeRounds, "budget survives input");
 		// The explicit surface grants the fresh budget and completes the run.
 		const { resumeActiveExecution } = await import("../src/exec.ts");
-		const resumed = resumeActiveExecution(ctxTui);
-		assert.equal(resumed, true);
-		assert.equal(getExecution()!.audit.rounds, 0, "resume grants a fresh review budget");
+		const resumed = await resumeActiveExecution(ctxTui);
+		assert.equal(resumed.resumed, true);
+		assert.equal(resumed.grantedBudget, 5, "the grant reports the re-picked budget");
 		await __awaitReviewRoundForTests(); // detached grant chain completes the run
 		assert.equal(runnerCalls, 1, "review re-ran after the resume (fresh budget)");
 		const final = loadCheckpoint(workdir, runId!);
 		assert.ok(final.status === "ok");
 		assert.equal(final.checkpoint.phase, "completed");
+		// The granted round committed as budget round 1 — proof the counter reset.
+		assert.equal(final.checkpoint.execution?.audit?.rounds ?? 1, 1, "the granted budget restarted at round 1");
 		__setAuditRunnerForTests(null);
 	});
 
@@ -455,6 +474,9 @@ describe("task-tree execution core", () => {
 			applyTaskUpdate(getExecution()!.tasks, id, "complete", `${id} evidence`);
 			persistTaskProgress(ctx);
 		}
+		// v0.9.3: pin the 5-round budget (the no-UI default is 3) so the retry
+		// chain still spends exactly REVIEW_MAX_ROUNDS rounds.
+		getExecution()!.reviewBudget = 5;
 		const snapshot = getExecution();
 		await restoreFromSession(ctx, [{ type: "custom", customType: "pi-plans-exec", data: snapshot }]);
 		const ex = getExecution()!;
@@ -464,6 +486,7 @@ describe("task-tree execution core", () => {
 			assert.equal(ex.tasks.find((t) => t.id === id)?.status, "complete", `${id} not rolled back`);
 		}
 		assert.equal(ex.audit.rounds, REVIEW_MAX_ROUNDS, "the retry chain spent the whole budget");
+		assert.equal(ex.reviewRoundsTotal, REVIEW_MAX_ROUNDS, "the run-cumulative counter tracks every committed round");
 		assert.equal(ex.stall.paused, true, "the loop pauses at the cap");
 		assert.match(ex.stall.pausedReason ?? "", /execution review exhausted 5 rounds/);
 		assert.equal(
@@ -549,6 +572,8 @@ describe("task-tree execution core", () => {
 			applyTaskUpdate(getExecution()!.tasks, id, "complete", `${id} evidence`);
 			persistTaskProgress(ctx);
 		}
+		// v0.9.3: pin the 5-round budget so the retry chain keeps its old length.
+		getExecution()!.reviewBudget = 5;
 		const snapshot = getExecution();
 		await restoreFromSession(ctx, [{ type: "custom", customType: "pi-plans-exec", data: snapshot }]);
 		const ex = getExecution()!;
@@ -617,6 +642,8 @@ describe("task-tree execution core", () => {
 			applyTaskUpdate(getExecution()!.tasks, id, "complete", `${id} evidence`);
 			persistTaskProgress(ctx);
 		}
+		getExecution()!.reviewBudget = 5;
+		getExecution()!.reviewRoundsTotal = REVIEW_MAX_ROUNDS;
 		getExecution()!.audit.rounds = REVIEW_MAX_ROUNDS;
 		getExecution()!.audit.failed = ["VC-002"];
 		const ctxUi = { ...makeCtx(workdir), mode: "tui" as const };

package/tests/fixtures/lattice-code-blocked/plan-v2-trimmed.md

@@ -0,0 +1,57 @@
+# PLAN_v2 - lattice-code v0.4 host/daemon migration (trimmed fixture)
+
+This file is a TRIM of the real plan at
+`lattice-code/.git/pi-plans/plans/2026-10-02-v0-4-ts-host-python-daemon/PLAN_v2.md`:
+only the tasks covered by execution-review round 1's failed checks, and those
+check lines, are kept verbatim. The run's checkpoint slices live in `state.json`.
+
+## Tasks
+
+- `Task-2`: S37 host 脚手架、隔离层与双栈门 — deps: Task-1; files: package.json, bun.lock, packages/host/package.json, packages/host/tsconfig.json, packages/host/src/cli.ts, packages/host/src/env.ts, packages/host/src/providers.ts, packages/host/src/gate.ts, packages/host/test/env.test.ts, Makefile, catalog/catalog_v1.json, tools/catalog/build.py, src/lattice/llm/catalog/resolver.py, tests/isolation/; wave: 2
+  - `Task-2.1`: 根 Bun workspace + `packages/host` 脚手架；三库锁版本安装；tsc `--noEmit` strict；`Makefile` 新增 `lint-ts`/`test-ts`/`gate`（pytest + `bun test` + `bun install --frozen-lockfile`）——自本步起每个提交双栈门禁
+  - `Task-2.2`: 隔离层：`$LATTICE_HOME` 布局（`pi-compat/`、`sessions/`、`daemon-venv/`、`bin/`、`logs/`）；启动即设六个 `PI_*` 环境变量（定位为纵深防御）；自管 `auth.json`（0600+原子写）显式 `apiKey` 注入 pi-ai；spawn 子进程环境清除一切 `*_API_KEY` 形态变量（杜绝 pi-ai 的 env 回退）；v0.4 不注册 OAuth 型 provider；pi-tui 构造显式传 `logDirectory: $LATTICE_HOME/logs`（该参数属 app 装配，本步先立约定与测试钩子）
+  - `Task-2.3`: catalog 中立迁移为基座：`catalog_v1.json`（context_window + effort 矩阵 + disabled_levels/disabled_reason）迁至仓库根 `catalog/`；`make catalog` 与 `tools/catalog/build.py` 重指向；冻结期 `llm/catalog/resolver.py` 改读新路径（保持 S46 删除前可用）；TS `providers.ts` overlay 合并以 `catalog/` 为基座（字段级合并/同 id 替换/自定 provider 透传），pi-ai `builtinModels()` 为未覆盖模型的次级补充
+  - `Task-2.4`: `tests/isolation/`：全 HOME 递归 manifest diff（path+size+mtime+sha256，排除 `$LATTICE_HOME`，复用 `util/hash.py` 的 sha256 语义）为唯一硬证明；wave-2 范围 = `bun run packages/host/src/cli.ts` 引导链路 + 六变量注入断言 + 预置非空 pi 后复跑（Task-10 后对编译二进制复跑同一套件）
+- `Task-3`: S38 pi-tui 界面重建 — deps: Task-2; files: packages/host/src/ui/app.ts, packages/host/src/ui/theme.ts, packages/host/src/ui/banner.ts, packages/host/src/ui/rail.ts, packages/host/src/ui/statusbar.ts, packages/host/src/ui/input.ts, packages/host/src/ui/cards.ts, packages/host/src/ui/panels.ts, packages/host/test/ui/; wave: 3
+  - `Task-3.1`: `app.ts` 四区装配（Planning box / Execution tasks / 卡片流 / 输入框）+ `theme.ts` 语义 token → pi-tui theme（四色彩模式；宽度计算复用 pi-tui 导出的 `visibleWidth`，不自建第二宽度引擎）
+  - `Task-3.2`: banner 三档梯子阈值即数据（full ≥60 / wordmark ≥14 / single <14；58 列对齐）+ 双 `⌗` 差异化（rail 与输入框前缀 token 不同、单色下字形不同）
+  - `Task-3.3`: rail 六态（色+单色字形）+ 扇出段（`N/M · K leases`）+ 路由段（`→ role`）；statusbar 弹性三段（plan title 吸收剩余宽度、model 段贴右误差 ≤1、定宽防抖；Ctx 段数据源 `catalog/`）
+  - `Task-3.4`: 输入框 + 卡片三态 + >24 行折叠标注；纯函数 + `render(width)` 行宽不变量（Lattice 侧义务：8–130 列扫描断言 `visibleWidth(line) <= width`）+ 多行样式逐行重应用；bun test 快照用例
+- `Task-4`: S39 daemon 与 JSONL 协议（含幂等对账） — deps: Task-2; files: schema/protocol.json, tools/gen_protocol_ts.mjs, tools/gen_protocol_py.py, packages/host/src/protocol.ts, packages/host/src/daemon.ts, packages/host/src/gen/protocol.gen.ts, packages/host/test/daemon.test.ts, src/lattice/daemon/__init__.py, src/lattice/daemon/server.py, src/lattice/daemon/dispatch.py, src/lattice/daemon/__main__.py, src/lattice/daemon/protocol_gen.py, src/lattice/graph/writer.py, src/lattice/graph/facts.py, tests/test_daemon_protocol.py, tests/test_daemon_crash.py; wave: 3
+  - `Task-4.1`: 冻结前穷举 daemon op 面：图读（`cmds/read.py:build_table` 全条目）、magics 只读视图、会话 create/exec/close（含握手帧携带 role/caps/write_paths）、写 op、lease/commit、snapshot/restore、health/version；`schema/protocol.json`（JSON Schema + `schema_version` + durable `request_id` 字段 + replayable 标记）；双侧现成生成器（json-schema-to-typescript / datamodel-code-generator）+ 提交产物 + 再生成 diff 零一致性用例
+  - `Task-4.2`: Python `daemon/`：`__main__.py --stdio --schema-version N` 为唯一 spawn 契约；stdio JSONL 服务端（stdout 仅帧）；分派按会话解析 per-agent `GraphHandle`（`assembly.py` ensure_agents），一切 op 走 handle 公共方法、绝不直投 `writer.submit`；test/gate op 强制走两阶段提交缝（`commit_prepare` 离环执行 + `commit_finalize`），不阻塞读循环与心跳；durable request_id 贯穿 `PendingItem` 与 facts（`_persist_intent`/`_persist_done` 均落 id），启动时对账开放 id、重复 id 返回已记录结果；租约重建语义写入协议文档；`_apply_one()` 保持无 `await`
+  - `Task-4.3`: `daemon.ts` 客户端：spawn、请求 id 生成与幂等、超时与未确认队列、SIGKILL 后重启 + 按 request_id 对账重放、背压上限；冷启动计时埋点（含 snapshot 恢复）；RTT 预算限定读类帧（机器/采样法入 evidence）
+- `Task-5`: S40 agent loop 与 plan 链上移 — deps: Task-3, Task-4; files: packages/host/src/loop.ts, packages/host/src/chain.ts, packages/host/src/roles.ts, packages/host/src/transport/record_replay.ts, packages/host/src/ui/cards.ts, packages/host/test/loop.test.ts, packages/host/test/chain.test.ts, packages/host/test/fixtures/cassettes/; wave: 4
+  - `Task-5.1`: `loop.ts`：基于 pi-agent-core 的工具调用循环 + daemon 工具桥（图读经协议；权限由 Python 侧 handle 强制）
+  - `Task-5.2`: `chain.ts`：plan 链 revision 生命周期（envelope.json 由 TS 直写 `.lattice/plans/<id>/`；reviewer 冷启动不含自己上一轮原文、跨 revision `must_survive` 仍被引用）；`roles.ts` 承接 ROLE_NAMES
+  - `Task-5.3`: 真实 provider 录制装置：录制缝 = pi-ai transport shim；以现有 `tests/cassettes/chain.json` 形状为 cassette schema 种子；fixtures 默认合成、真实转录提交前脱敏扫描（无凭据/无绝对家路径/无 org id）；revision 闭环用例跑在录制 cassette 上
+  - `Task-5.4`: cards 接线真实 chain 数据（wave 状态接线归 Task-6）
+- `Task-6`: S41 扇出调度与预算 — deps: Task-5; files: packages/host/src/schedule.ts, packages/host/src/budget.ts, packages/host/src/ui/panels.ts, packages/host/test/schedule.test.ts; wave: 5
+  - `Task-6.1`: 波次划分：同 wave 文件集两两不相交；N executor 生命周期（spawn/回收/失败隔离）；panels 接线 wave/budget 真状态
+  - `Task-6.2`: 三类硬预算（token / 时长 / 写次数）越限即停；`budget.ts` 承接预算超限错误类型
+  - `Task-6.3`: 越界写拒绝：租约与乐观 hash 校验留 Python 侧经 daemon 强制（改未租约文件被拒并回错误帧）；TS 只发请求不做本地放行
+- `Task-7`: S42 picker 与键位 — deps: Task-3; files: packages/host/src/ui/picker.ts, packages/host/src/keys.ts, packages/host/test/picker.test.ts; wave: 4
+  - `Task-7.1`: `/login` 两级屏（认证方式 → (provider, method) 行 + 凭据徽章）对接自管 auth；完成登录仅在未设或已是该 provider 时落 `default` 角色；仅列出免 OAuth 的认证方式
+  - `Task-7.2`: 模型选择器两级（provider 页 → model 页，type-to-filter；`←→` 调 effort、`Tab` 循环 role、`Ctrl+P` 唤起；effort 草稿 Enter 提交 Esc 丢弃）；数据源 `catalog/`（context_window + effort 矩阵）
+  - `Task-7.3`: 禁用态带原因（`disabled_levels`/`disabled_reason`）；键位冲突表为零
+- `Task-9`: S44 会话持久化与 resume — deps: Task-8; files: packages/host/src/session.ts, packages/host/src/ui/resume.ts, packages/host/src/cli.ts, packages/host/test/session.test.ts; wave: 6
+  - `Task-9.1`: 会话 JSONL 落 `$LATTICE_HOME/sessions/`，定位为**纯重放 transcript（非权威）**：envelope.json + 图 snapshot 是唯一权威；恢复时对话从 transcript 重放、plan/图状态从权威源重建
+  - `Task-9.2`: `lattice resume` 列表选择恢复（cli.ts 注册）；断言单改 transcript 不改变恢复结果；重启后图等价
+- `Task-10`: S45 打包与分发 — deps: Task-6, Task-8, Task-9; files: scripts/build-binaries.sh, scripts/install.sh, scripts/bootstrap-python.sh, Makefile, package.json, pyproject.toml, README.md, tests/isolation/; wave: 7
+  - `Task-10.1`: `bun build --compile` 五平台（darwin-arm64/x64、linux-arm64/x64、windows-x64；x64 走 baseline target；pi-tui 原生加速器为可选件缺失即 JS 回退，不内嵌 `.node`）
+  - `Task-10.2`: sidecar wheel：`pyproject.toml` bump 0.4.0，`uv build` 产 wheel 随二进制入 dist；bootstrap：首启检测 `python3.11+` → 引导 uv 至 `$LATTICE_HOME` → `uv venv $LATTICE_HOME/daemon-venv` + 离线安装 sidecar wheel；`kernel_version` 握手协商，不匹配给类型化明确报错
+  - `Task-10.3`: 安装脚本：二进制+sidecar 落 `$LATTICE_HOME/bin`，打印 PATH 建议不自动改；干净机器（无 Node/Bun/pi）验证；isolation 套件对编译二进制复跑；README 安装文档改写与形态一致
+
+## Verification Checks
+
+- [ ] `VC-095` covers `Task-4`; pass condition: daemon 被 SIGKILL 后 host 重启并按 request_id 对账重放——无重复应用（受影响文件 sha256 不变 + fact 计数不变），重复 id 返回已记录结果，租约重建语义有文档与用例；evidence: `tests/test_daemon_crash.py`（含 fsync 后 SIGKILL 场景）；metric: ≥2 用例通过。
+- [ ] `VC-102` covers `Task-3`; pass condition: rail 与输入框前缀颜色 token 不同、单色下字形不同；evidence: 快照 + 断言；metric: 彩色/单色各 ≥1。
+- [ ] `VC-104` covers `Task-3`; pass condition: status bar 弹性三段在 8–130 列 model 段贴右误差 ≤1、定宽防抖、Ctx 段数据源 `catalog/`；evidence: 宽度扫描用例；metric: 0 溢出 + 误差 ≤1 列。
+- [ ] `VC-106` covers `Task-5`; pass condition: 真实 provider 录制 cassette 下完整 revision 闭环；reviewer 冷启动不含自己上一轮原文；跨 revision `must_survive` 仍被引用；evidence: `bun test` 链路用例 + cassette fixture；metric: 三项各 ≥1。
+- [ ] `VC-107` covers `Task-6`; pass condition: 同 wave 文件集运行时两两不相交（executor 预检拒绝交集）；N=3 真并发（`wall < 0.6 × Σ`）；三类预算超限即停；改未租约文件被拒（经 daemon 校验回错误帧）；evidence: `bun test` + daemon 拒绝帧断言；metric: 四项各 ≥1。
+- [ ] `VC-111` covers `Task-9`; pass condition: transcript 为纯重放（单改 transcript 不改变恢复结果）；resume 后对话与图状态等价；evidence: resume 往返用例 + 图等价断言；metric: 两项各 ≥1 用例。
+- [ ] `VC-118` covers `Task-4`; pass condition: `writer._dispatch` 每分支与 `cmds/read.py:build_table` 每条目均映射到 schema 帧（op 面无缺口）；evidence: 映射完备性检查用例（工作树可判）；metric: 0 缺口。
+- [ ] `VC-119` covers `Task-5`; pass condition: 提交的 cassette 无凭据、无绝对家路径、无 org id；evidence: 脱敏扫描用例；metric: 0 命中。
+- [ ] `VC-121` covers `Task-10`; pass condition: bootstrap 离线安装 sidecar wheel 后 `kernel_version` 握手匹配且 daemon 可用；evidence: 干净 venv 安装 + 握手用例；metric: ≥1 用例。
+- [ ] `VC-122` covers `Task-2` and `Task-7`; pass condition: picker 列出的每个模型都有 `context_window` 与 effort 矩阵，且 ≥1 可见禁用级（含原因）；evidence: `bun test` 数据完备性用例（对 `catalog/` 全量断言）；metric: 0 缺失。
+- [ ] `VC-123` covers `Task-10`; pass condition: README 安装文档与新分发形态（二进制 + sidecar + uv 自举）一致；evidence: 文档审读 + 安装步骤复跑；metric: 0 处失配。

package/tests/fixtures/lattice-code-blocked/state.json

@@ -0,0 +1,158 @@
+{
+	"_source": {
+		"run": "lattice-code 20261002T080333Z-v0-4-ts-host-python-daemon",
+		"checkpoint_slices": [
+			"execution.tasks (only the covered tasks)",
+			"execution.audit.rounds",
+			"execution.audit.lastResult",
+			"execution.audit.undeterminable"
+		],
+		"real_state": "round 1 rolled back the failed checks' coverage; the repair round re-closed everything except the parent Task-2, which is still pending here",
+		"note": "no `execution.blocked` key: the checkpoint predates the feature, which is exactly what the backfill must handle"
+	},
+	"tasks": {
+		"Task-2": {
+			"status": "pending",
+			"evidence": "提交 ce36d13（22 文件，+1101/−34）。make gate 全绿（frozen install + pytest 902 + bun test 11）；ruff 干净；门禁三查（更新版白名单）clean。VC-089/090/091/112 的种子用例就位（wave-2 形态：bun run 宿主）。"
+		},
+		"Task-2.1": {
+			"status": "complete",
+			"evidence": "修复轮补强（78ba6c1）：provider.ts overlay 层不变；隔离层经 F-001/F-014 相关修复复验（tests/isolation 对 bun-run 与新编译二进制各 8 用例全绿——含重跑的三条 manifest/env/共存证明）。VC-089~093 复验通过。"
+		},
+		"Task-2.2": {
+			"status": "complete",
+			"evidence": "隔离层原样复验：六变量 + sanitizedChildEnv + auth 0600 + XDG_CACHE_HOME 重定向；daemonPython 增 dev-venv 探测；bootstrap.ts 子进程沿用 sanitizedChildEnv。env.test 全绿。"
+		},
+		"Task-2.3": {
+			"status": "complete",
+			"evidence": "catalog/ 中立基座不变；statusContextFromCatalog（app.ts）新增：Ctx 段数据源接 catalog context_window（F-015 修复），VC-104 catalog-fed 用例（dual_marks.test）+ picker 全量 catalog 完备性用例（VC-122，见 Task-7 重开证据）。"
+		},
+		"Task-2.4": {
+			"status": "complete",
+			"evidence": "隔离套件对重建后的编译二进制复跑全绿（doctor 链路 + 预置 pi 共存 + 六变量）；manifest diff 证明不变。"
+		},
+		"Task-3": {
+			"status": "complete",
+			"evidence": "提交 5e18b99 + 修复轮 78ba6c1（挂载 + 数据源 + 差异化）。bun test 115/115、gate 全绿。"
+		},
+		"Task-3.1": {
+			"status": "complete",
+			"evidence": "四区装配现有生产挂载（F-002 修复）：mountTui/runTui 进 cli.ts 默认路径（TTY 时真跑 pi-tui 主屏 + Editor 焦点 + Ctrl+C 退出）；statusRegion 由 statusContextFromCatalog 供数（F-015）。app.test 4 用例 + dual_marks 全绿。"
+		},
+		"Task-3.2": {
+			"status": "complete",
+			"evidence": "banner 梯子不变；双 ⌗ 差异化修复（F-012）：input 前缀改 accent_dim 非加粗（与内核侧 prompt_glyph 奇偶校验一致），六相位逐一断言 rail token ≠ accent_dim、三档样式互异、ascii 档 [X] 码 ≠ [#]（dual_marks.test 6 用例）。"
+		},
+		"Task-3.3": {
+			"status": "complete",
+			"evidence": "rail 六态不变；statusbar 补 VC-104 断言：catalog-fed Ctx 真值渲染（Ctx 12.5%/200k）、模型组贴行尾 40–130 全扫、占位槽零抖动。"
+		},
+		"Task-3.4": {
+			"status": "complete",
+			"evidence": "卡片/输入线不变；宽度扫描 8–130 与样式逐行断言维持全绿。"
+		},
+		"Task-4": {
+			"status": "complete",
+			"evidence": "提交 2dbb046 + 修复轮 78ba6c1。pytest 218（含新 lease-less CONFLICT 帧、bootstrap 握手 2 例）、bun 115、gate 全绿。"
+		},
+		"Task-4.1": {
+			"status": "complete",
+			"evidence": "schema 增 read.peek/commit.abort/graph.reparse（F-007），双侧再生成（sha 25e658d1…）+ --check 双绿。"
+		},
+		"Task-4.2": {
+			"status": "complete",
+			"evidence": "daemon 修复：graph.test 改经 asyncio.to_thread 离环跑门（F-003——读循环/心跳不再被 120s 测试阻塞）；commit.finalize 校验 staged ticket、commit.abort 走 handle.abort_commit（F-006）；handle 增 abort_commit/reparse；EOF 排空竞态修复（in-flight 响应先写后关）；read.peek/graph.reparse 接线。VC-116/117 维持，VC-118 升级为 _dispatch 分支 + build_table 条目动态枚举映射（正则抽源 + 临时仓建 store）。"
+		},
+		"Task-4.3": {
+			"status": "complete",
+			"evidence": "F-001 修复：request id 每逻辑请求恰铸一次，重放复用同一 rid（send 显式收 rid；requestIds 日志）；专测：在飞 session.exec 被 kill → 重启重放 → requestIds 切片 ≥2 次发送且去重后恰 1 个 id。VC-095 主机半边成立（daemon 半边 tests/py 两用例原有）。"
+		},
+		"Task-5": {
+			"status": "complete",
+			"evidence": "提交 db1f0b8 + 修复轮 78ba6c1。VC-106（cassette 闭环 + 冷启动 + must_survive）与 VC-119（脱敏零命中）用例就位。"
+		},
+		"Task-5.1": {
+			"status": "complete",
+			"evidence": "F-011 修复：RoleLoop 重建于 pi-agent-core `Agent`（transcript 状态、工具执行、turn 生命周期皆库驱动；我们的 stream 适配器经 createAssistantMessageEventStream 供帧，LoopTools 以真 AgentTool 声明注入，错误经 stopReason=error 呈现为抛出）。headless 冒烟双态复绿。"
+		},
+		"Task-5.2": {
+			"status": "complete",
+			"evidence": "chain/envelope/冷启动规则不变；错误从适配器透传（errorMessage → throw）。"
+		},
+		"Task-5.3": {
+			"status": "complete",
+			"evidence": "F-009/F-010 修复：提交 cassette 产物 fixtures/cassettes/revision-loop.json（4 条目：拒绝→修订→批准）；扫描器增 org-/org_/acct_ 三类 org-id 模式并计入 redact；cassette 驱动全修订闭环用例（仅凭录制驱动 ChainDriver 到 approved、rev=2、双产物在盘）+ 提交产物零泄漏断言 + 脏样本捕获断言。"
+		},
+		"Task-5.4": {
+			"status": "complete",
+			"evidence": "cards 接线不变，维持全绿。"
+		},
+		"Task-6": {
+			"status": "complete",
+			"evidence": "提交 0a7dcaf + 修复轮 78ba6c1。schedule.test 14 用例全绿。"
+		},
+		"Task-6.1": {
+			"status": "complete",
+			"evidence": "F-008 修复：runWave 增运行时预检 assertDisjointWave（WaveOverlapError 携带碰撞对，./ 归一，executor 起跑前拒绝）+ 调度产物逐波过检用例；无租约写经真 daemon 拒绝用例（tests/py：graph.write 无 lease → ok:false CONFLICT \"holds no lease\"）。"
+		},
+		"Task-6.2": {
+			"status": "complete",
+			"evidence": "三轴预算与隔离语义不变，全绿维持。"
+		},
+		"Task-6.3": {
+			"status": "complete",
+			"evidence": "租约强制留 kernel 侧不变；CONFLICT 帧现由真 daemon 用例钉死（VC-107 第四项闭环）。"
+		},
+		"Task-7": {
+			"status": "complete",
+			"evidence": "提交 353edad + 修复轮 78ba6c1。picker.test 全绿。"
+		},
+		"Task-7.1": {
+			"status": "complete",
+			"evidence": "/login 两级与徽章语义不变（env 态维持刻意缺席）。"
+		},
+		"Task-7.2": {
+			"status": "complete",
+			"evidence": "选择器 FSM 不变；F-004 修复：catalog effort 的 null 值视为显式禁用（\"maps no native parameter\" 原因），缺失级维持 \"exposes no native parameter\"；VC-122 全量断言：catalog_v1.json 每模型 context_window≥1 且 effort≥1 级，禁用级模型 ≥1（picker.test 新增全量用例）。"
+		},
+		"Task-7.3": {
+			"status": "complete",
+			"evidence": "禁用级带原因 + 键位零冲突维持（含 null-effort 模型渲染 disabled 列表）。"
+		},
+		"Task-9": {
+			"status": "complete",
+			"evidence": "提交 c5b9def + 修复轮 78ba6c1。VC-111 两项闭环。"
+		},
+		"Task-9.1": {
+			"status": "complete",
+			"evidence": "F-005 修复：生产路径写会话（runHeadless 建 session、user/assistant/end 事件落 $LATTICE_HOME/sessions）；lattice resume 起真 DaemonClient（state.health 图恢复报告）后再回放 transcript；VC-111 强化：真 daemon 图等价往返用例（恢复前后 read.ls id 集恒等 + 篡改只显于回放）。"
+		},
+		"Task-9.2": {
+			"status": "complete",
+			"evidence": "resume 半边同上（daemon 恢复 + transcript 纯显示）；session.test 5 用例全绿。"
+		},
+		"Task-10": {
+			"status": "complete",
+			"evidence": "提交 4325370 + 修复轮 78ba6c1。VC-093/121/123 证据链补齐。"
+		},
+		"Task-10.1": {
+			"status": "complete",
+			"evidence": "五平台 + sidecar 不变（修复轮后 make dist 重建，二进制嵌入修复后 TS）。"
+		},
+		"Task-10.2": {
+			"status": "complete",
+			"evidence": "F-013 修复：tests/py/test_bootstrap_handshake.py —— 临时 venv 离线优先装 dist/kernel wheel（uv 缺席走 ensurepip+pip 回退），断言 wheel==installed==pyproject 0.4.0，且安装的 daemon 应答 ping（2 用例实跑通过；无 wheel 时 skip）。"
+		},
+		"Task-10.3": {
+			"status": "complete",
+			"evidence": "F-014 修复：README 全面改写（Quickstart 旧流程/ui.json/tmux/--glob 全清；新 Install/How-it-works/Development 与分发一致）；tests/py/test_readme_shape.py 三用例钉死（必含 $LATTICE_HOME/install.sh/-p，七类陈旧片段零命中，dev 命令在）。"
+		}
+	},
+	"audit": {
+		"rounds": 1,
+		"lastResult": "VC-095,VC-102,VC-104,VC-106,VC-107,VC-111,VC-118,VC-119,VC-121,VC-122,VC-123",
+		"undeterminable": [
+			"VC-109"
+		]
+	}
+}

package/tests/refine-ui.test.ts

@@ -331,12 +331,12 @@ describe("refine overlay wiring", () => {
 	it("threads the chrome language through both overlay construction sites (issue #3, VC-007)", () => {
 		const refineSource = fs.readFileSync(path.join(process.cwd(), "tools", "refine.ts"), "utf8");
 		assert.match(refineSource, /const overlayLang = uiLanguageFromTag\(config\.language\.tag\)/);
-		assert.match(refineSource, /new RefineOverlayController\("reviewer", lanes, relayAbort, lang\)/);
+		assert.match(refineSource, /new RefineOverlayController\(\s*"reviewer",\s*lanes,\s*relayAbort,\s*lang,/);
 		assert.match(refineSource, /modelLabel,\s*\n\s*overlayLang,/);
 		const refsSource = fs.readFileSync(path.join(process.cwd(), "tools", "analyze-refs.ts"), "utf8");
 		assert.match(
 			refsSource,
-			/new RefineOverlayController\("refs", batch\.map\(\(job\) => \(\{ id: job\.laneId, label: job\.laneId \}\)\), relayAbort, resolveUiLanguage\(workdir\)\)/,
+			/new RefineOverlayController\(\s*"refs",\s*batch\.map\(\(job\) => \(\{ id: job\.laneId, label: job\.laneId \}\)\),\s*relayAbort,\s*resolveUiLanguage\(workdir\),/,
 		);
 	});
 });
@@ -533,4 +533,27 @@ describe("refine overlay kitty and fallback key handling (issue #2)", () => {
 			__clearPiTuiForTests();
 		}
 	});
+
+	it("forwards unhandled keys instead of swallowing them (Ctrl+Shift+T while focused)", () => {
+		// pi-tui routes input ONLY to the focused component (no bubbling), so
+		// an open overlay used to swallow every global shortcut. Unhandled keys
+		// must reach the onUnhandledKey hook so callers can re-dispatch them.
+		const calls: string[] = [];
+		const component = new RefineOverlayComponent(
+			fakeTheme,
+			"auditor",
+			[readyLane("lane-1", "reviewer-1", "output")],
+			() => {},
+			undefined,
+			undefined,
+			"en",
+			(data) => calls.push(data),
+		);
+		// kitty CSI-u form of Ctrl+Shift+T — not an overlay key.
+		component.handleInput("\x1b[84;6u");
+		assert.deepEqual(calls, ["\x1b[84;6u"], "unhandled keys are forwarded");
+		// Handled keys (arrow down) are NOT forwarded.
+		component.handleInput("\x1b[B");
+		assert.deepEqual(calls, ["\x1b[84;6u"], "handled keys stay internal");
+	});
 });

package/tests/resume.test.ts

@@ -8,7 +8,7 @@ import * as path from "node:path";
 import { after, before, describe, it } from "node:test";
 import { listResumeCandidates, pickDefaultCandidate } from "../src/resume.ts";
 import { migrateRunIntoCurrentWorktree, resumePlansCommand } from "../src/resume-command.ts";
-import { createCheckpoint, loadCheckpoint, mutateCheckpoint, applyQuestionAsked, applyQuestionAnswered } from "../src/workflow-state.ts";
+import { createCheckpoint, loadCheckpoint, mutateCheckpoint, applyExecutionApproved, applyExecutionProgress, applyPlanWritten, applyQuestionAsked, applyQuestionAnswered, planIdentityOf, resolveHeadAt } from "../src/workflow-state.ts";
 import { acquireOwnership, processStartOf } from "../src/run-ownership.ts";
 import { resetRunBindingForTests } from "../src/run-context.ts";
 import { initState, setArtifactRoot, setRunStatus, startRun, StateError } from "../src/state.ts";
@@ -399,3 +399,47 @@ describe("impl-review crash-window recovery (removed in v0.6.1)", () => {
 	});
 });
 
+describe("execution resume brief with a blocked review (v0.9.2)", () => {
+	const BLOCKED_PLAN = `# PLAN_v1 - blocked brief
+
+## Tasks
+
+- Task-1: engine — files: lib/engine.js; wave: 1
+- Task-2: host — files: lib/host.js; wave: 1
+
+## Verification Checks
+
+- [ ] \`VC-001\` covers \`Task-1\`; pass condition: engine works; evidence: tests; metric: green.
+- [ ] \`VC-002\` covers \`Task-2\`; pass condition: host works; evidence: tests; metric: green.
+`;
+
+	it("names the open rolled-back task and how to unblock the review", async () => {
+		resetRunBindingForTests();
+		// A matching HEAD is required: an unverifiable head re-opens every task
+		// (reverifyAll) and legitimately drops the blocker with the progress.
+		const workdir = setupRepo("blocked-brief", { commit: true });
+		const run = startRun(workdir, { topic: "blockedbrief", skill: "plan-normal", requestText: "b" }).run;
+		const planPath = path.join(run.artifact_dir, "PLAN_v1.md");
+		fs.mkdirSync(run.artifact_dir, { recursive: true });
+		fs.writeFileSync(planPath, BLOCKED_PLAN, "utf8");
+		createCheckpoint(workdir, { runId: run.run_id, originWorkdir: workdir, workdir });
+		mutateCheckpoint(workdir, run.run_id, (cp) => {
+			const plan = planIdentityOf(planPath, 1);
+			let next = applyPlanWritten({ ...cp, nextAction: "accept-execute" }, plan);
+			next = applyExecutionApproved(next, { plan, worktree: workdir, headAtApproval: resolveHeadAt(workdir), approvedAt: next.updatedAt });
+			return applyExecutionProgress(next, {
+				tasks: { "Task-1": { status: "complete", evidence: "e" }, "Task-2": { status: "pending" } },
+				blocked: { rolledBack: ["Task-2"], tasks: ["Task-2"], round: 1, escalatedRounds: 2, since: next.updatedAt },
+			});
+		});
+		setRunStatus(workdir, run.run_id, "executing");
+		const mock = makeCtx(workdir);
+		await resumePlansCommand(ctxAdapter(mock) as never, BASE_DIR);
+		assert.equal(mock.userMessages.length, 1, "one resume brief");
+		const brief = mock.userMessages[0]!;
+		assert.match(brief, /PI-PLANS RESUME/);
+		assert.match(brief, /Review blocked: Task-2 was reopened by execution review round 1 and is still open/);
+		assert.match(brief, /close it with plans_update_task/);
+		assert.match(brief, /no review round is running while these are open\./i);
+	});
+});

package/tests/review-budget.test.ts

@@ -0,0 +1,201 @@
+/**
+ * Execution-review budget tests (v0.9.3): the budget arithmetic, the shared
+ * review-pause predicate, the unlimited no-progress valve, and the picker's
+ * fallback matrix. The loop-level behaviour (resolution timing, tolerated
+ * completion, grants) lives in tests/exec-review-loop.test.ts.
+ */
+
+import * as assert from "node:assert/strict";
+import { describe, it } from "node:test";
+import {
+	askReviewBudget,
+	budgetExhausted,
+	bumpNoProgress,
+	formatReviewBudget,
+	isReviewPauseReason,
+	noProgressSignature,
+	noProgressTripped,
+	resolveStoredBudget,
+	reviewBudgetPanelAvailable,
+	unlimitedHardCapCeiling,
+	DEFAULT_REVIEW_BUDGET,
+	LEGACY_REVIEW_MAX_ROUNDS,
+	NO_PROGRESS_MAX_STREAK,
+	REVIEW_BUDGET_CHOICES,
+	UNLIMITED_HARD_CAP,
+} from "../src/review-budget.ts";
+import { reviewBudgetChrome } from "../src/ui-language.ts";
+
+describe("review budget arithmetic (v0.9.3)", () => {
+	it("resolves a stored budget, the legacy bound, and the undecided state", () => {
+		assert.equal(DEFAULT_REVIEW_BUDGET, 3);
+		assert.equal(LEGACY_REVIEW_MAX_ROUNDS, 5);
+		assert.equal(UNLIMITED_HARD_CAP, 50);
+		assert.deepEqual([...REVIEW_BUDGET_CHOICES], [1, 2, 3, 5, "unlimited"]);
+		// Undecided: a fresh run has not picked yet.
+		assert.equal(resolveStoredBudget(undefined, 0), undefined);
+		// Legacy: a pre-feature checkpoint that already spent rounds keeps 5.
+		assert.equal(resolveStoredBudget(undefined, 2), LEGACY_REVIEW_MAX_ROUNDS);
+		// A stored value wins.
+		assert.equal(resolveStoredBudget("unlimited", 2), "unlimited");
+		assert.equal(resolveStoredBudget(1, 0), 1);
+	});
+
+	it("formats the budget the way status lines do", () => {
+		assert.equal(formatReviewBudget(1), "1");
+		assert.equal(formatReviewBudget(3), "3");
+		assert.equal(formatReviewBudget(5), "5");
+		assert.equal(formatReviewBudget("unlimited"), "∞");
+	});
+
+	it("measures numeric budgets per grant and unlimited across the whole run", () => {
+		const counters = { reviewRoundsTotal: 0, reviewCapExtension: 0 };
+		assert.equal(budgetExhausted(3, 2, counters), false);
+		assert.equal(budgetExhausted(3, 3, counters), true);
+		assert.equal(budgetExhausted(1, 1, counters), true);
+		// Unlimited ignores the per-grant counter and uses the cumulative one.
+		assert.equal(budgetExhausted("unlimited", 9999, counters), false);
+		assert.equal(budgetExhausted("unlimited", 0, { reviewRoundsTotal: UNLIMITED_HARD_CAP - 1, reviewCapExtension: 0 }), false);
+		assert.equal(budgetExhausted("unlimited", 0, { reviewRoundsTotal: UNLIMITED_HARD_CAP, reviewCapExtension: 0 }), true);
+		// An explicit unlimited grant lifts the cap by exactly one window.
+		assert.equal(budgetExhausted("unlimited", 0, { reviewRoundsTotal: UNLIMITED_HARD_CAP + 10, reviewCapExtension: 50 }), false);
+		assert.equal(unlimitedHardCapCeiling({ reviewRoundsTotal: 12, reviewCapExtension: 50 }), 100);
+	});
+
+	it("recognizes every review pause reason and nothing else (shared predicate)", () => {
+		assert.equal(isReviewPauseReason("execution review exhausted 3 rounds (failed: VC-002). Only /plans-execute"), true);
+		assert.equal(isReviewPauseReason("execution review exhausted 50 rounds (unlimited budget safety cap)"), true);
+		assert.equal(isReviewPauseReason("completion audit exhausted 5 rounds (legacy prefix)"), true);
+		assert.equal(isReviewPauseReason("execution review stalled — 3 consecutive rounds reported the same outcome"), true);
+		// A blocked wake names a different state and must never be grantable.
+		assert.equal(isReviewPauseReason("blocked: review round 2 cannot start — Task-2 still open"), false);
+		assert.equal(isReviewPauseReason(undefined), false);
+		assert.equal(isReviewPauseReason(""), false);
+	});
+});
+
+describe("no-progress valve (v0.9.3)", () => {
+	it("signs the post-classification triple deterministically", () => {
+		const base = noProgressSignature(["VC-002"], ["VC-001"], ["F-001"]);
+		assert.equal(base, noProgressSignature(["VC-002"], ["VC-001"], ["F-001"]));
+		// Order and duplicates never change the signature.
+		assert.equal(base, noProgressSignature(["VC-002", "VC-002"], ["VC-001"], ["F-001"]));
+		// Any real movement changes it — the valve only fires on a dead loop.
+		assert.notEqual(base, noProgressSignature([], ["VC-001"], ["F-001"]));
+		assert.notEqual(base, noProgressSignature(["VC-002"], [], ["F-001"]));
+		assert.notEqual(base, noProgressSignature(["VC-002"], ["VC-001"], []));
+		assert.notEqual(base, noProgressSignature(["VC-001"], ["VC-002"], ["F-001"]));
+	});
+
+	it("counts consecutive identical outcomes and trips on the third", () => {
+		const sig = noProgressSignature([], ["VC-001", "VC-002"], []);
+		const first = bumpNoProgress(undefined, sig);
+		assert.deepEqual(first, { key: sig, streak: 1 });
+		assert.equal(noProgressTripped(first), false);
+		const second = bumpNoProgress(first, sig);
+		assert.equal(second.streak, 2);
+		assert.equal(noProgressTripped(second), false);
+		const third = bumpNoProgress(second, sig);
+		assert.equal(third.streak, NO_PROGRESS_MAX_STREAK);
+		assert.equal(noProgressTripped(third), true);
+		// Progress resets the streak to one.
+		const moved = bumpNoProgress(third, noProgressSignature(["VC-001"], ["VC-002"], []));
+		assert.equal(moved.streak, 1);
+		assert.equal(noProgressTripped(moved), false);
+	});
+});
+
+describe("budget picker (v0.9.3)", () => {
+	const theme = { fg: (_color: string, text: string) => text, bold: (text: string) => text };
+
+	it("falls back to null without a native surface (the default-budget path)", async () => {
+		assert.equal(reviewBudgetPanelAvailable({ mode: "print", hasUI: false }), false);
+		assert.equal(reviewBudgetPanelAvailable({ mode: "tui", hasUI: true, ui: {} }), false);
+		assert.equal(await askReviewBudget({ mode: "print", hasUI: false }, "en"), null);
+		assert.equal(await askReviewBudget({ mode: "tui", hasUI: true, ui: {} }, "en"), null);
+		// RPC: ui.custom exists but resolves undefined (no overlay support).
+		assert.equal(await askReviewBudget({ mode: "rpc", hasUI: true, ui: { custom: async () => undefined } }, "en"), null);
+	});
+
+	it("treats a headless session as panel-less even though ui.select exists (round-1 F-001)", async () => {
+		// The SDK's ExtensionUIContext.select is REQUIRED, so json/print
+		// sessions still carry a function; hasUI is the availability signal.
+		const headless = {
+			mode: "json" as const,
+			hasUI: false,
+			ui: {
+				select: async () => {
+					throw new Error("a headless select must never be asked");
+				},
+				custom: async () => {
+					throw new Error("a headless custom must never be asked");
+				},
+			},
+		};
+		assert.equal(reviewBudgetPanelAvailable(headless), false, "hasUI false means no panel");
+		assert.equal(await askReviewBudget(headless, "en"), null, "and the picker answers null without calling the UI");
+		// A TUI/RPC host with hasUI true keeps the native surfaces.
+		assert.equal(reviewBudgetPanelAvailable({ mode: "rpc", hasUI: true, ui: { select: async () => undefined } }), true);
+		assert.equal(reviewBudgetPanelAvailable({ mode: "tui", hasUI: true, ui: { custom: async () => null } }), true);
+	});
+
+	it("returns the menu selection and maps the label back to a budget", async () => {
+		const titles: string[] = [];
+		const pick = {
+			mode: "rpc",
+			hasUI: true,
+			ui: {
+				select: async (title: string, options: string[]): Promise<string | undefined> => {
+					titles.push(title);
+					return options[3];
+				},
+			},
+		};
+		// options[3] = "5 rounds" (en); the title carries the current value.
+		assert.equal(await askReviewBudget(pick, "en", 3), 5);
+		assert.match(titles[0] ?? "", /current: 3/);
+		// Cancelled select → null (the caller decides: default at first ask,
+		// keep-paused at a grant).
+		assert.equal(await askReviewBudget({ mode: "rpc", hasUI: true, ui: { select: async () => undefined } }, "en"), null);
+	});
+
+	it("builds the TUI panel, returns its pick, and treats Esc as no answer", async () => {
+		const panelHost = {
+			mode: "tui",
+			hasUI: true,
+			ui: {
+				// Construction must succeed with the host's minimal tui/theme —
+				// exactly what the real overlay passes in tests.
+				custom: (render: (tui: unknown, theme: unknown, kb: unknown, done: (value: unknown) => void) => unknown) => {
+					render({ requestRender() {} }, theme, undefined, () => {});
+					return Promise.resolve(2);
+				},
+			},
+		};
+		assert.equal(reviewBudgetPanelAvailable(panelHost), true);
+		assert.equal(await askReviewBudget(panelHost, "zh", 1), 2);
+		const escHost = {
+			mode: "tui",
+			hasUI: true,
+			ui: {
+				custom: (render: (tui: unknown, theme: unknown, kb: unknown, done: (value: unknown) => void) => unknown) => {
+					render({ requestRender() {} }, theme, undefined, () => {});
+					return Promise.resolve(null);
+				},
+			},
+		};
+		assert.equal(await askReviewBudget(escHost as never, "zh", 3), null);
+	});
+
+	it("ships zh and en chrome for the panel and its rows", () => {
+		const zh = reviewBudgetChrome("zh");
+		const en = reviewBudgetChrome("en");
+		assert.match(zh.unlimitedOption, /无上限/);
+		assert.match(zh.roundsOption(3), /3 轮/);
+		assert.match(zh.currentSuffix("3"), /当前 3/);
+		assert.match(en.unlimitedOption, /unlimited/);
+		assert.match(en.roundsOption(1), /^1 round$/);
+		assert.match(en.roundsOption(3), /^3 rounds$/);
+		assert.notEqual(zh.panelTitle, en.panelTitle);
+	});
+});