@qlingzzy/qling 1.0.0 → 1.1.0

package/CHANGELOG.md

@@ -1,7 +1,161 @@
 # Changelog
 
+## Unreleased
+
+### Phase 7.0 — Sprint 4 分发与生态
+
+- **`eval:tasks`**：10 个本地 repo fixture（修测试 / 改文案 / 加函数 / 重命名 / 修 import / package scripts / JSON / 文档节 / 多文件一致 / 中文路径），不依赖 LLM。
+- **`validate:packaging`**：校验 Scoop / winget 草案版本与 `package.json` 对齐。
+- **Scoop / winget 文档**：`docs/install.md` 与 `packaging/README.md` 补充本地校验步骤；清单版本钉到 1.1.0。
+- **Skills 示例包**：`fix-failing-test` / `add-function` / `pr-summary` + `skills/examples/README.md`。
+- **英文 README**：补齐 1.1 能力、eval、分发与 skills 说明。
+- **Demo 说明**：`docs/demo.md`（文字演示路径；GIF 另录不入库）。
+- **`ci:check`**：纳入 `eval:tasks` 与 `validate:packaging`。
+
+## v1.1.0 (2026-07-14)
+
+1.1 聚焦：**可证明的执行韧性**、**可维护的内核分层**、**编码/TUI 精度** 与 **Windows CI**。
+
+### Highlights
+
+- 执行韧性：确定性恢复策略、`/recover`、验证闭环、脱敏 trace、`eval:recovery`
+- 架构：从 `agent-loop` 抽出 LLM 客户端、工具编排、主循环、system prompt、session 持久、验证闭环
+- 编码精度：patch 原子写、repo-map/search 预算、工具输出折叠与 CJK 宽度
+- CI：ubuntu 全量 + windows unit
+
+### Phase 6.0 — Sprint 3 编码精度 / TUI / Windows CI
+
+- **patch 原子写**：`writeFileAtomic`（temp + rename，Windows 回退直写），降低半截写入风险。
+- **repo map 预算**：`buildRepoMapSection` 支持 `maxSymbols`/`maxChars`（默认 200 / 6000），超限截断并提示。
+- **search 上下文卫生**：默认 limit 40；`truncateSearchLine` 截断过长匹配行。
+- **TUI**：工具输出折叠 footer 中英双语；底部提示含 Shift+Tab / `/mode`；CJK `visibleWidth` 回归测试。
+- **CI**：ubuntu 全量 `ci:check` + `windows-latest` 跑 unit（`npm run test`）。
+
+### Phase 5.2 — 巨石拆分预备
+
+- **LLM 客户端抽出**：`src/providers/llm-client.ts`（`LlmHttpClient`）
+- **记忆 lifecycle 抽出**：`src/memory/lifecycle.ts` 的 `runAutoDream`
+- **Dashboard 解耦**：对 `dashboard-server` 动态 import
+- **工具编排抽出**：`src/agent/tool-orchestrator.ts`
+- **验证闭环抽出**：`src/execution/verification-loop.ts` + `recovery-messages.ts`
+- **Session 持久抽出**：`src/session/session-persistence.ts`
+- **主循环 / system prompt 抽出**：`src/agent/main-loop.ts`、`src/agent/system-prompt.ts`
+
+### Phase 5.1 — 验证闭环统一 + Doctor
+
+- 写操作恢复只走 `StagedVerifier`；多阶段 env 配置
+- Progress 含 `changedFiles` / 策略字段
+- Doctor Phase5：`recovery_budget` / `run_traces` / `verifier_stages` / `verify_llm_advisory`
+- Metrics：`pausedRuns`、`averageTimeToPauseMs`
+
+### Phase 5.0 — 执行韧性收口
+
+- `RecoveryStrategyPlanner` + 暂停动作语义
+- TUI 订阅 execution events
+- `eval:recovery` 扩充
+
+### TUI / Dashboard / Phase 3–4（延续）
+
+- Shift+Tab 模式循环、Slash 输入框修复、Dashboard 任务工作台
+- Harness lean、progressive skills、角色 sub-agent、browser_act、LSP 可选等（详见下方历史与 v1.0.0）
+
+### Phase 3.4 — Eval harness 指标（smoke 增量）
+
+- `eval:smoke` 新增：工具输出折叠、explore 只读、回传契约、skill 扫描、角色别名等本地任务。
+
+### Phase 3.3 — 浏览器 / 外联路由
+
+- **`docs/web-routing.md`**：opencli / url_fetch / browser_fetch / browser_act 决策树。
+- **`browser_act` 工具**：有限 goto/click/type/wait_for/extract/press；**默认关闭**（`QLING_BROWSER_ACT=1` 启用）；共用网络 Guard；Plan Mode 禁止。
+- opencli skill 与 system Restrictions 同步分工说明。
+
+### Phase 3.3.1 — browser_act 跨步会话
+
+- **`session` 保活**：`open` / `close` / `status` + 同 session 上 click/type/extract。
+- 空闲回收：`QLING_BROWSER_ACT_IDLE_TTL_MS`（默认 10min）；上限 `QLING_BROWSER_ACT_MAX_SESSIONS`（默认 3）。
+- 模块：`src/tools/browser-act-session.ts`。
+
+### Phase 3.2.1 — explore 并行（默认关）
+
+- `subtask tasks=["…","…"]` + `role=explore|review`，须 `QLING_SUBTASK_PARALLEL=1`。
+- implement 禁止并行；上限 `QLING_SUBTASK_PARALLEL_MAX`（默认 3）。
+- 模块：`src/agent/subtask-parallel.ts`。
+
+### Phase 3.5 — Mission 进度通知
+
+- 使命状态变更可推送 Telegram / Slack（需已配置 channel token + chat/channel id）。
+- `QLING_MISSION_NOTIFY=off` 可关；`QLING_MISSION_TELEGRAM_CHAT_ID` / `QLING_MISSION_SLACK_CHANNEL_ID` 可选覆盖。
+
+### Phase 3.5.1 — 通知结构化卡片
+
+- 默认 `QLING_MISSION_NOTIFY_STYLE=rich`：Telegram HTML、Slack Block Kit；`plain` 回退纯文本。
+- 状态 emoji 前缀；失败带错误摘要。
+
+### Phase 3.5.2 — 使命日志推送 + Doctor Phase3
+
+- **日志推送**：`appendLog` 可推送通道；`QLING_MISSION_NOTIFY_LOGS=off|milestone|all`（默认 milestone）。
+- **`qling doctor`**：展示 browser_act / subtask_parallel / mission_notify 开关摘要。
+- **可选 e2e**：`QLING_BROWSER_ACT_E2E=1` + `QLING_BROWSER_ACT=1` 跑真实 Playwright 冒烟（默认 skip）。
+
+### Phase 4.0–4.2 — 路线 + LLM eval + code_symbols
+
+- **路线**：`docs/superpowers/specs/20260710-phase4-capability-roadmap-spec.md`
+- **可选 LLM 评测**：`npm run eval:llm`（需 `QLING_EVAL_LLM=1` + API key；默认 skip，不进 ci:check）
+- **工具 `code_symbols`**：工作区符号名检索（静态提取，非完整 LSP）；explore/implement/review 可用
+- eval:smoke 增加 code_symbols 本地任务
+
+### Phase 4.3 — 可选 TypeScript 语义查询（lsp）
+
+- **工具 `lsp`**：`definition` / `hover` / `references` / `document_symbols`
+- **默认关闭**：`QLING_LSP=1` 启用；动态加载 `typescript`（缺包时明确报错）
+- 进程内 LanguageService（非 stdio 多语言 LSP 协议）；`doctor` 展示 phase4:lsp
+- 通用符号扫描仍用 `code_symbols`
+
+### Phase 4.4 — 模块分层与包边界
+
+- **文档**：`docs/architecture-layers.md`（目标分层、禁止边、拆包门槛）
+- **扫描**：`npm run dep:layers` → `scripts/dep-layers.mjs`（`--json` / `--write-doc` / `--strict`）
+- **诚实债务**：记录 adapters→cli、eval/runtime 等反向依赖；strict 暂不进 ci:check
+
+### 审计收尾 — 本地优先与恢复边界
+
+- Memory Dream LLM 改为仅在 `QLING_MEMORY_DREAM_LLM_ENABLED=true|1|on|yes` 时启用，SDK 直接构造默认不额外调用模型。
+- 远程 discovery 默认拒绝未签名清单，并接入 network Guard、禁止自动重定向、限制 1 MiB 响应；需审批 source 在无审批回调时 fail-closed。
+- Discovery 工具清单与运行时可执行工具分离，未绑定 handler/MCP transport 的 metadata 不再向模型宣称可调用。
+- Workflow checkpoint 持久化状态机定义，恢复后迁移继续有效；非法 `runId` 和缺失定义的旧 checkpoint 明确拒绝。
+- `browser_act` 在创建会话前完成 URL Guard 检查，拒绝目标不再留下空白 Playwright 会话。
+
 ## v1.0.0 (2026-07-09)
 
+### opencli 小红书路由补强
+
+- skill `opencli` 增加小红书专节：`note`/`comments` 必须使用含 `xsec_token` 的完整 URL（禁止裸 note-id）。
+- Restrictions 同步提示 xiaohongshu 正确传参。
+
+### Agent 任务执行基本规则
+
+- 常驻 system prompt（Workflow / Restrictions / Tone）固化三条规则：
+  1. 调用外部工具前先做工具与任务的关联分析
+  2. 成功后总结正确可复现流程
+  3. 单流程失败或未准确执行时实事求是承认
+
+### Agent 路由（opencli）
+
+- **内置 skill `opencli`**：教模型用 `bash` + `opencli <site> … -f json` 获取平台数据；明确抖音≠TikTok、禁止 `url_fetch` 强反爬站。
+- **system prompt 短路由**：Restrictions 中写入网页/社交平台调用规则。
+- **skill 扫描路径**：包内 `skills/` + `~/.qling/skills/` + 工作区 skills，全局安装后仍能加载内置 skill。
+
+### Token 与预算
+
+- **官方 usage 计数**：会话 token 仅累加模型 API 返回的 `prompt_tokens` / `completion_tokens` / `total_tokens`（及 Anthropic/Ollama 等价字段）；不再用字符×4 估算账单。
+- **删除预算功能**：移除 `TokenBudgetManager`、预算 nudge、system prompt Token 预算节、`max_token_budget` 配置接线与百分比水位。
+- **展示**：`/usage`、`/context`、statusline 显示 total/in/out 与 `provider|unknown` 来源。
+
+### TUI 体验（结果聚焦）
+
+- **输入区去噪**：输入框上方不再打印 statusline 与快捷键黑灰提示行；详情仍用 `/statusline`、`/shortcuts` 与顶栏。
+- **结果高亮**：任务最终答复以「结果」边框强调；完成态改为醒目的「任务完成」+ 耗时。
+
 ### 发布收尾（发布就绪补丁）
 
 - **跨平台 CI**：`write-sandbox` 单测改用 `path.resolve`/`tmpdir`，去掉硬编码 `C:\\` 路径。

package/README.en.md

@@ -2,13 +2,14 @@
 
 [![Node](https://img.shields.io/badge/Node-%E2%89%A518-339933?logo=node.js&logoColor=white)](#requirements)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
-[![Version](https://img.shields.io/badge/version-1.0.0-orange.svg)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-1.1.0-orange.svg)](CHANGELOG.md)
+[![Release](https://img.shields.io/github/v/release/Zzy-min/qling?display_name=tag&sort=semver)](https://github.com/Zzy-min/qling/releases)
 
 **Qling** is a **local-first** AI Agent CLI workbench. It keeps sessions, tools, skills, permissions, diagnostics, long-running missions, and recovery controls in one auditable terminal console.
 
 > One line: Qling turns the agent into a local workbench—not a black-box chat window.
 
-**Chinese docs:** [README.md](README.md) · **Install details:** [docs/install.md](docs/install.md)
+**Chinese docs:** [README.md](README.md) · **Install:** [docs/install.md](docs/install.md) · **Skills:** [docs/skills.md](docs/skills.md) · **Demo notes:** [docs/demo.md](docs/demo.md)
 
 ## Why Qling
 
@@ -16,15 +17,23 @@
 |---|---|
 | Local-first | Sessions, checkpoints, exports, memory, and diagnostics stay on disk; `qling privacy` shows boundaries. |
 | Chinese UX | TUI, help, and slash aliases target Chinese terminals (English works too). |
-| Recoverable | `/checkpoint`, `/resume`, daemon + mission state machines. |
+| Recoverable | `/checkpoint`, `/resume`, `/recover`, staged verification, execution traces. |
 | Governed tools | Permission matrix, approval gate, content filter, audit log. |
 | Honest boundaries | Unsupported cloud features are explained—never faked. |
 
+## What's new in 1.1
+
+- Deterministic recovery strategies + staged verification (`eval:recovery`)
+- Leaner agent core: LLM client, tool orchestration, main loop, system prompt extracted from the monolith
+- Coding precision: atomic patch writes, repo-map/search budgets, bilingual tool-output fold, CJK width
+- Windows unit CI alongside full Ubuntu `ci:check`
+- Sprint 4 ecosystem: `eval:tasks` fixtures, packaging validators, example skills pack
+
 ## Requirements
 
 - Node.js ≥ 18
 - npm ≥ 9
-- Optional: Playwright Chromium for `browser_fetch`
+- Optional: Playwright Chromium for `browser_fetch` / `browser_act`
 
 ## Install
 
@@ -36,27 +45,37 @@ cd qling
 npm run bootstrap
 # optional browser tools:
 npm run bootstrap -- --with-browser
-```
-
-Link the CLI globally:
-
-```bash
 npm link
-qling
+qling --version
 ```
 
 ### From npm
 
+Package name is scoped: **`@qlingzzy/qling`** (CLI binary remains `qling`):
+
 ```bash
 npm install -g @qlingzzy/qling
-# CLI binary is still: qling
 qling bootstrap
 qling setup
 ```
 
-### Windows notes
+From GitHub without npm registry:
+
+```bash
+npm install -g github:Zzy-min/qling
+```
+
+> `better-sqlite3` is a native module. On Windows install [VS Build Tools](https://visualstudio.microsoft.com/visual-cpp-build-tools/) if prebuilds are unavailable.
+
+### Windows: Scoop / winget drafts
+
+Drafts live under `packaging/` and are **not** in official catalogs yet:
+
+```powershell
+npm run validate:packaging
+```
 
-See [docs/install.md](docs/install.md) for PowerShell env vars, Scoop draft, and winget draft.
+See [docs/install.md](docs/install.md) for PowerShell env vars, Scoop notes, and winget submission checklist.
 
 ## Minimal config
 
@@ -76,7 +95,7 @@ Interactive wizard (does **not** write secrets into `.env`):
 qling setup
 ```
 
-Any OpenAI-compatible endpoint works via `QLING_LLM_ENDPOINT` + `QLING_LLM_MODEL`.
+Any OpenAI-compatible endpoint works via `QLING_LLM_ENDPOINT` + `QLING_LLM_MODEL`. Local models: `/model use ollama`.
 
 ## Four commands to try
 
@@ -105,24 +124,70 @@ qling privacy           # data retention paths
 /model list    provider presets
 /model use ollama
 /plan on       read-only plan mode
+/skill list    progressive skills index
 /expand        expand long tool output
+/recover       execution recovery controls
+/verify        verification command / stages
 /doctor        environment check
 /privacy       data boundaries
 ```
 
+Shift+Tab cycles agent modes in the TUI.
+
+## Skills
+
+Progressive skills: short index in the system prompt; full body loaded on demand via `skill` tool or `/skill`.
+
+Bundled examples (`skills/examples/`):
+
+| Skill | Use when |
+|-------|----------|
+| `repo-triage` | Onboarding an unfamiliar repo |
+| `fix-failing-test` | Red unit tests |
+| `add-function` | Small exported API addition |
+| `pr-summary` | PR / release notes from git |
+
+Also: `opencli`, lifecycle suite (`lifecycle-spec` … `lifecycle-ship`), template at `skills/templates/SKILL.md`.
+
+Details: [docs/skills.md](docs/skills.md).
+
+## Eval & quality gates
+
+```bash
+npm test                 # unit
+npm run test:smoke
+npm run eval:smoke       # local policy/presets smoke
+npm run eval:recovery    # recovery planner fixtures
+npm run eval:tasks       # 10 coding repo fixtures (no LLM)
+npm run validate:packaging
+npm run ci:check         # CI entry (ubuntu); windows runs unit
+# optional with key:
+QLING_EVAL_LLM=1 npm run eval:llm
+```
+
 ## Design principles
 
 - **Local-first** — valuable state stays on the machine.
 - **Slash-first** — discover controls from `/`.
-- **Recoverable** — checkpoint / resume / recap.
+- **Recoverable** — checkpoint / resume / recover / verify.
 - **Honest boundaries** — no fake success for unsupported features.
 - **Terminal-native** — enhance TUI without heavy GUI deps.
 
+## Packaging status (honest)
+
+| Channel | Status |
+|---------|--------|
+| Source + bootstrap | Ready |
+| npm `@qlingzzy/qling` | Ready when published (`npm login` required on publisher machine) |
+| Scoop | Draft manifest only |
+| winget | Draft YAML only |
+| Portable zip / single binary | Not yet |
+
 ## Version
 
-- Current: `1.0.0`
+- Package: `1.1.0` (see latest tag on GitHub for hotfixes)
 - Changelog: [CHANGELOG.md](CHANGELOG.md)
-- Roadmap specs: `docs/superpowers/specs/`
+- Specs / plans: `docs/superpowers/`
 
 ## License
 

package/README.md

@@ -2,7 +2,7 @@
 
 [![Node](https://img.shields.io/badge/Node-%E2%89%A518-339933?logo=node.js&logoColor=white)](#环境要求)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
-[![Version](https://img.shields.io/badge/version-1.0.0-orange.svg)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-1.1.0-orange.svg)](CHANGELOG.md)
 
 [English](README.en.md) · [安装指南](docs/install.md) · [CHANGELOG](CHANGELOG.md)
 
@@ -20,10 +20,10 @@
 | 可恢复长任务 | `/checkpoint`、`/resume`、`/rewind`、`--continue`、`qling workflow resume` 让中断后的上下文可以接上。 |
 | 本地 skill | `skills/` 与 `.qling/skills/` 中的 Markdown skill 可被 `/skill` 或 `/<skill-name>` 直接读取；内置命令优先级高于 skill。 |
 | 后台使命 | `qling daemon` + `qling mission` 提供长任务状态机、暂停 / 恢复 / 重试 / attach。 |
-| 观测台 | `qling dashboard start` 打开本地白盒化观测控制台，串起 thinking / tool / token 链路。 |
+| 任务工作台 | `qling dashboard start` 打开仅监听本机的任务中心，统一查看 Mission、循环任务、Workflow、执行来源与最近活动。 |
 | 通道与扩展 | `src/channels/` 内置 console / Telegram / Slack 通道；`qling discovery sync` 动态同步插件与技能。 |
 | 权限可解释 | `/permissions`、`/permissions explain <tool>`、guard、内容过滤、速率限制和密钥脱敏都以本地规则呈现。 |
-| 上下文透明 | `/context`、`/usage` 显示 token 来源、估算、上下文预算和压缩状态。 |
+| 上下文透明 | `/context`、`/usage` 显示 provider 官方 token usage 与压缩状态（无预算）。 |
 | 诊断内建 | `/doctor`、`/config`、`/mcp`、`/hooks`、`/diff` 用于快速判断当前项目和运行时状态。 |
 | 诚实边界 | 平台专属命令会被识别并给出本地边界说明，不会伪装成功。 |
 
@@ -219,7 +219,7 @@ qling 日志 <id>   # logs <id>
 | Command | Purpose |
 |---|---|
 | `/model [model]` | 显示或切换当前 session 模型；不写入配置文件。 |
-| `/usage`, `/cost`, `/stats` | token 来源、用量、上下文预算和压缩状态。 |
+| `/usage`, `/cost`, `/stats` | provider 官方 token usage（input/output/total）与压缩次数。 |
 | `/context` | 本地上下文和 token 使用情况。 |
 | `/statusline [on\|off]` | 显示或切换输入区状态线。 |
 
@@ -401,14 +401,22 @@ npm audit --registry=https://registry.npmjs.org --audit-level=high
 
 ## 版本与变更
 
-- 当前版本：`1.0.0`
+- 当前版本：`1.1.0`（Sprint 4 分发/生态见 CHANGELOG Unreleased）
 - 完整变更历史：见 [CHANGELOG.md](CHANGELOG.md)
 - 安装与分发：见 [docs/install.md](docs/install.md)、[packaging/](packaging/)
 - Skills：见 [docs/skills.md](docs/skills.md)
+- 演示路径：见 [docs/demo.md](docs/demo.md)
+- 评测：`npm run eval:smoke` / `eval:recovery` / `eval:tasks`
+- 网页/平台路由：见 [docs/web-routing.md](docs/web-routing.md)
 - SDK：见 [docs/sdk.md](docs/sdk.md)
-- 本地评测：`npm run eval:smoke`
+- 本地评测：`npm run eval:smoke`（无 LLM）；可选 `QLING_EVAL_LLM=1 npm run eval:llm`
+- Phase 4 路线：`docs/superpowers/specs/20260710-phase4-capability-roadmap-spec.md`
+- 可选 TS 语义：`QLING_LSP=1` 后使用工具 `lsp`（definition/hover/…）
+- 模块分层：`docs/architecture-layers.md`；`npm run dep:layers`
 - 英文说明：见 [README.en.md](README.en.md)
-- 竞品对标与提升路线：见 `docs/superpowers/specs/20260709-agent-cli-competitive-analysis-and-v1-roadmap-spec.md`
+- 竞品对标与 1.0 路线：见 `docs/superpowers/specs/20260709-agent-cli-competitive-analysis-and-v1-roadmap-spec.md`
+- **2026-07 生态续调研与 Phase 3 路线**：见 `docs/superpowers/specs/20260710-agent-ecosystem-refresh-and-phase3-roadmap-spec.md`
+- Phase 3 实施计划：见 `docs/superpowers/plans/20260710-phase3-harness-lean-skills-orchestration-plan.md`
 - 设计 / 实施文档：见 `docs/superpowers/specs/`、`docs/superpowers/plans/`
 
 ## License

package/dist/agent/main-loop.d.ts

@@ -0,0 +1,120 @@
+import { type ChatUsage, type TokenUsageSource } from "../token-usage.js";
+import type { Message, ToolCall, ToolDefinition } from "../types.js";
+import type { ToolPipeline } from "../pipeline/hooks.js";
+import type { GuardConfig } from "../config.js";
+import type { Channel } from "../channels/types.js";
+import type { ApprovalGate } from "../guard/approval.js";
+import type { KnowledgeAgentAdapter } from "../knowledge-agent.js";
+import type { MemoryStore } from "../memory.js";
+import type { WorkflowRuntime } from "../workflow-runtime.js";
+import type { ContextCompactor } from "../context-compactor.js";
+import type { VerificationAgent } from "../pipeline/verification.js";
+import type { ExecutionEventBus } from "../execution/event-bus.js";
+import type { RecoveryController } from "../execution/recovery-controller.js";
+import type { FailureClassification, RecoveryState } from "../execution/types.js";
+import { type ReflectionDecision } from "./tool-orchestrator.js";
+import type { LlmChatResponse } from "../providers/llm-client.js";
+export interface TurnTelemetry {
+    turn: number;
+    toolCalls: number;
+    toolFailures: number;
+}
+export interface TokenCounters {
+    sessionTokens: number;
+    sessionPromptTokens: number;
+    sessionCompletionTokens: number;
+    tokenUsageSource: TokenUsageSource;
+}
+export declare function applyProviderUsage(counters: TokenCounters, usage: ChatUsage | undefined): TokenCounters;
+export declare function logTurnTelemetry(metrics: TurnTelemetry, totals: {
+    toolCallTotal: number;
+    toolFailureTotal: number;
+    compactionCount: number;
+    retryCountTotal: number;
+    format: "text" | "json";
+}): {
+    toolCallTotal: number;
+    toolFailureTotal: number;
+};
+export declare function distillSuccessfulBashPractices(preparedCalls: Array<{
+    call: ToolCall;
+    immediateResult?: {
+        is_error?: boolean;
+    };
+}>, messages: Message[], memoryStore: MemoryStore): void;
+export interface InnerLoopHost {
+    messages: Message[];
+    turnCount: number;
+    sessionId: string;
+    maxIterations: number;
+    toolRepeatLimit: number;
+    parseRetries: number;
+    verificationCommand: string | null;
+    counters: TokenCounters;
+    compactionCount: number;
+    toolCallTotal: number;
+    toolFailureTotal: number;
+    retryCountTotal: number;
+    loggingFormat: "text" | "json";
+    activeRunId?: string;
+    compactor: ContextCompactor;
+    pipeline: ToolPipeline;
+    tools: ToolDefinition[];
+    guardConfig: GuardConfig;
+    channel: Channel | null;
+    approvalGate: ApprovalGate;
+    knowledgeAdapter: KnowledgeAgentAdapter;
+    memoryStore: MemoryStore;
+    workspaceDir: string;
+    workflowRuntime: WorkflowRuntime;
+    executionEventBus: ExecutionEventBus;
+    recoveryController: RecoveryController;
+    verifier: VerificationAgent;
+    buildSystemPrompt: () => Promise<string>;
+    chat: (systemPrompt: string, overrides?: Record<string, unknown>) => Promise<LlmChatResponse>;
+    emit: (event: string, ...args: unknown[]) => void;
+    runVerificationCommand: (cmd: string) => Promise<{
+        code: number;
+        stdout: string;
+        stderr: string;
+    }>;
+    getRecoveryState: () => RecoveryState | null;
+    reflectiveThink: (tc: ToolCall) => Promise<{
+        decision: ReflectionDecision;
+        reason: string;
+    }>;
+    checkAutoDream: () => Promise<void>;
+}
+/**
+ * Inner maxIterations loop: compact → prompt → chat → tools → verify.
+ * Mutates host.messages / counters / turnCount in place.
+ */
+export declare function runInnerIterationLoop(host: InnerLoopHost): Promise<string>;
+export interface OuterLoopHost {
+    sessionId: string;
+    activeRun: {
+        runId: string;
+        sessionId: string;
+        originalTask: string;
+        startedAt: number;
+    } | null;
+    messages: Message[];
+    executionEventBus: ExecutionEventBus;
+    recoveryController: RecoveryController;
+    emit: (event: string, ...args: unknown[]) => void;
+    getRecoveryState: () => RecoveryState | null;
+    formatRecoveryPause: (reason: string, next: string) => string;
+    applyRecoveryStrategy: (failure: FailureClassification, strategy?: string) => Promise<void>;
+    executeInner: () => Promise<string>;
+    setActiveRun: (run: {
+        runId: string;
+        sessionId: string;
+        originalTask: string;
+        startedAt: number;
+    } | null) => void;
+}
+/**
+ * Outer loop: start run, handle provider transport retries + recovery strategy.
+ */
+export declare function runOuterAgentLoop(host: OuterLoopHost): Promise<string>;
+//# sourceMappingURL=main-loop.d.ts.map

package/dist/agent/main-loop.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"main-loop.d.ts","sourceRoot":"","sources":["../../src/agent/main-loop.ts"],"names":[],"mappings":"AAIA,OAAO,EAEL,KAAK,SAAS,EACd,KAAK,gBAAgB,EACtB,MAAM,mBAAmB,CAAC;AAC3B,OAAO,KAAK,EAAE,OAAO,EAAe,QAAQ,EAAE,cAAc,EAAE,MAAM,aAAa,CAAC;AAClF,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,sBAAsB,CAAC;AACzD,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,cAAc,CAAC;AAChD,OAAO,KAAK,EAAE,OAAO,EAAE,MAAM,sBAAsB,CAAC;AACpD,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,sBAAsB,CAAC;AACzD,OAAO,KAAK,EAAE,qBAAqB,EAAE,MAAM,uBAAuB,CAAC;AACnE,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,cAAc,CAAC;AAChD,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,wBAAwB,CAAC;AAC9D,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,yBAAyB,CAAC;AAChE,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,6BAA6B,CAAC;AACrE,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,2BAA2B,CAAC;AACnE,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,qCAAqC,CAAC;AAC9E,OAAO,KAAK,EAAE,qBAAqB,EAAE,aAAa,EAAE,MAAM,uBAAuB,CAAC;AAElF,OAAO,EAGL,KAAK,kBAAkB,EACxB,MAAM,wBAAwB,CAAC;AAEhC,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,4BAA4B,CAAC;AAGlE,MAAM,WAAW,aAAa;IAC5B,IAAI,EAAE,MAAM,CAAC;IACb,SAAS,EAAE,MAAM,CAAC;IAClB,YAAY,EAAE,MAAM,CAAC;CACtB;AAED,MAAM,WAAW,aAAa;IAC5B,aAAa,EAAE,MAAM,CAAC;IACtB,mBAAmB,EAAE,MAAM,CAAC;IAC5B,uBAAuB,EAAE,MAAM,CAAC;IAChC,gBAAgB,EAAE,gBAAgB,CAAC;CACpC;AAED,wBAAgB,kBAAkB,CAChC,QAAQ,EAAE,aAAa,EACvB,KAAK,EAAE,SAAS,GAAG,SAAS,GAC3B,aAAa,CASf;AAED,wBAAgB,gBAAgB,CAC9B,OAAO,EAAE,aAAa,EACtB,MAAM,EAAE;IACN,aAAa,EAAE,MAAM,CAAC;IACtB,gBAAgB,EAAE,MAAM,CAAC;IACzB,eAAe,EAAE,MAAM,CAAC;IACxB,eAAe,EAAE,MAAM,CAAC;IACxB,MAAM,EAAE,MAAM,GAAG,MAAM,CAAC;CACzB,GACA;IAAE,aAAa,EAAE,MAAM,CAAC;IAAC,gBAAgB,EAAE,MAAM,CAAA;CAAE,CAqCrD;AAcD,wBAAgB,8BAA8B,CAC5C,aAAa,EAAE,KAAK,CAAC;IAAE,IAAI,EAAE,QAAQ,CAAC;IAAC,eAAe,CAAC,EAAE;QAAE,QAAQ,CAAC,EAAE,OAAO,CAAA;KAAE,CAAA;CAAE,CAAC,EAClF,QAAQ,EAAE,OAAO,EAAE,EACnB,WAAW,EAAE,WAAW,GACvB,IAAI,CA4BN;AAED,MAAM,WAAW,aAAa;IAC5B,QAAQ,EAAE,OAAO,EAAE,CAAC;IACpB,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;IAClB,aAAa,EAAE,MAAM,CAAC;IACtB,eAAe,EAAE,MAAM,CAAC;IACxB,YAAY,EAAE,MAAM,CAAC;IACrB,mBAAmB,EAAE,MAAM,GAAG,IAAI,CAAC;IACnC,QAAQ,EAAE,aAAa,CAAC;IACxB,eAAe,EAAE,MAAM,CAAC;IACxB,aAAa,EAAE,MAAM,CAAC;IACtB,gBAAgB,EAAE,MAAM,CAAC;IACzB,eAAe,EAAE,MAAM,CAAC;IACxB,aAAa,EAAE,MAAM,GAAG,MAAM,CAAC;IAC/B,WAAW,CAAC,EAAE,MAAM,CAAC;IAErB,SAAS,EAAE,gBAAgB,CAAC;IAC5B,QAAQ,EAAE,YAAY,CAAC;IACvB,KAAK,EAAE,cAAc,EAAE,CAAC;IACxB,WAAW,EAAE,WAAW,CAAC;IACzB,OAAO,EAAE,OAAO,GAAG,IAAI,CAAC;IACxB,YAAY,EAAE,YAAY,CAAC;IAC3B,gBAAgB,EAAE,qBAAqB,CAAC;IACxC,WAAW,EAAE,WAAW,CAAC;IACzB,YAAY,EAAE,MAAM,CAAC;IACrB,eAAe,EAAE,eAAe,CAAC;IACjC,iBAAiB,EAAE,iBAAiB,CAAC;IACrC,kBAAkB,EAAE,kBAAkB,CAAC;IACvC,QAAQ,EAAE,iBAAiB,CAAC;IAE5B,iBAAiB,EAAE,MAAM,OAAO,CAAC,MAAM,CAAC,CAAC;IACzC,IAAI,EAAE,CAAC,YAAY,EAAE,MAAM,EAAE,SAAS,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,KAAK,OAAO,CAAC,eAAe,CAAC,CAAC;IAC9F,IAAI,EAAE,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,IAAI,EAAE,OAAO,EAAE,KAAK,IAAI,CAAC;IAClD,sBAAsB,EAAE,CAAC,GAAG,EAAE,MAAM,KAAK,OAAO,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;IACnG,gBAAgB,EAAE,MAAM,aAAa,GAAG,IAAI,CAAC;IAC7C,eAAe,EAAE,CAAC,EAAE,EAAE,QAAQ,KAAK,OAAO,CAAC;QAAE,QAAQ,EAAE,kBAAkB,CAAC;QAAC,MAAM,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;IAC7F,cAAc,EAAE,MAAM,OAAO,CAAC,IAAI,CAAC,CAAC;CACrC;AAED;;;GAGG;AACH,wBAAsB,qBAAqB,CAAC,IAAI,EAAE,aAAa,GAAG,OAAO,CAAC,MAAM,CAAC,CA4IhF;AAED,MAAM,WAAW,aAAa;IAC5B,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAC;QAAC,YAAY,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAA;KAAE,GAAG,IAAI,CAAC;IAChG,QAAQ,EAAE,OAAO,EAAE,CAAC;IACpB,iBAAiB,EAAE,iBAAiB,CAAC;IACrC,kBAAkB,EAAE,kBAAkB,CAAC;IACvC,IAAI,EAAE,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,IAAI,EAAE,OAAO,EAAE,KAAK,IAAI,CAAC;IAClD,gBAAgB,EAAE,MAAM,aAAa,GAAG,IAAI,CAAC;IAC7C,mBAAmB,EAAE,CAAC,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,KAAK,MAAM,CAAC;IAC9D,qBAAqB,EAAE,CACrB,OAAO,EAAE,qBAAqB,EAC9B,QAAQ,CAAC,EAAE,MAAM,KACd,OAAO,CAAC,IAAI,CAAC,CAAC;IACnB,YAAY,EAAE,MAAM,OAAO,CAAC,MAAM,CAAC,CAAC;IACpC,YAAY,EAAE,CACZ,GAAG,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAC;QAAC,YAAY,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAA;KAAE,GAAG,IAAI,KACtF,IAAI,CAAC;CACX;AAED;;GAEG;AACH,wBAAsB,iBAAiB,CAAC,IAAI,EAAE,aAAa,GAAG,OAAO,CAAC,MAAM,CAAC,CA4E5E"}