From c4cf6ce0cc52734157ba03d938598e29c4d0fda7 Mon Sep 17 00:00:00 2001 From: Lex Date: Thu, 8 Oct 2026 23:21:05 +0800 Subject: [PATCH 1/5] docs: introduce GraphAgent features in the README and add 1.0.65 release notes The README described GraphAgent almost only through its DAG engine. Add a feature overview and sections for reasoning shaping, Project Memory, autonomous goals, hooks and context management, regroup the DAG material, drop the retired dev prerelease channel and the outdated Bun requirement, and add the validated 1.0.65 series notes. --- .github/releases/v1.0.65.md | 58 ++++++++++++++ README.md | 152 ++++++++++++++++++++++++++++++------ README.zh.md | 120 ++++++++++++++++++++++------ 3 files changed, 284 insertions(+), 46 deletions(-) create mode 100644 .github/releases/v1.0.65.md diff --git a/.github/releases/v1.0.65.md b/.github/releases/v1.0.65.md new file mode 100644 index 000000000..102d4515b --- /dev/null +++ b/.github/releases/v1.0.65.md @@ -0,0 +1,58 @@ +## opencode 1.0.65 + +Stable release from `main` branch. Reasoning shaping now rewrites each reasoning part through the built-in engine before the model sees it again, and the repository adopts SpecGit 2.5 with a refreshed README. + +--- + +### 🎯 Features + +- **Reasoning shaping, per step**: Each reasoning part is organized by the configured small model as soon as it ends, while the step's tools run, instead of once after the turn. Before every provider request the session waits for unfinished rewrites up to `OPENCODE_REASONING_DISTILLATION_SETTLE_MS` (default 3000 ms) and seals the rest; a sealed part keeps its original text for good. A part is therefore rewritten before its first resend or never, so earlier prompt-cache prefixes are never invalidated, and within-turn resends now carry the organized text. Adoption is checked per part, so a quick follow-up message no longer discards a whole turn (#721). +- **Engine organizer**: Both runtimes organize through the `@opencode-ai/llm` engine with no tools, temperature 0, no retries and `none`/`low` effort translated per protocol. The opencode runtime keeps the AI SDK only as a fallback for packages the engine does not route. Parts are organized concurrently and independently; the session budget counts reported or estimated tokens instead of stopping when a provider omits usage or after a fixed number of calls (#721). +- **Barrier logs**: Every sealed barrier is logged as `reasoning rewrite sealed` (info) and other barriers as `reasoning rewrite barrier` (debug), with job, adopted and sealed counts and the wait time (#721). + +--- + +### 🐛 Bug Fixes + +- **Reasoning shaping**: A Chinese rewrite of English reasoning was rejected as larger because the savings check used the conservative reserve estimator, which counts one token per CJK character and four characters per English token. Measured on the configured DeepSeek and GLM tokenizers the rewrite was smaller; the check now uses a script-calibrated estimate (#721). +- **Reasoning shaping**: An all-noise part was replaced with empty text, which made the core runner drop `reasoning_content` from that assistant message. It is now replaced with a short non-empty placeholder (#721). +- **SpecGit**: Prettier had reformatted the SpecGit-managed block in `AGENTS.md`, so SpecGit 2.5 reported a damaged marker and refused `init`, `migrate` and `remove`. The blocks are regenerated and `AGENTS.md` is excluded from Prettier (#723). + +--- + +### 🏗️ Architecture / Refactor + +- **Engine-owned reasoning carriers**: `ReasoningCarrier.classify` in `@opencode-ai/llm` decides whether provider metadata is plain, a plaintext mirror, or opaque (signed, encrypted or referenced). Core no longer inspects SDK metadata key names (#721). +- **Dead code removed**: The unreachable native-wire projector, the propose/judge/claims/gates pipeline, the native-runtime hook, and the core runner's unused per-turn history conversion and request preparation, about 11.7k lines (#721). + +--- + +### ⚙️ CI / Engineering + +- SpecGit 2.5: `.specgit.yaml` is per-checkout local configuration and is no longer tracked; `AGENTS.md` documents the `specgit init` / `specgit setup --agent opencode` commands. The uncalled v1 bootstrap wrapper is removed. Pull requests merge with merge commits only (#723). +- The lint ratchet drops from 4850 to 4750 after fixing every warning on touched lines (#721). +- The README introduces the fork's own features (reasoning shaping, DAG workflows, Project Memory, autonomous goals, hooks and context management) and drops the retired `dev` prerelease channel. + +--- + +### 🧪 Test Summary + +```text +workspace typecheck tasks: 31/31 passed +opencode tests: 5241 passed, 0 failed +core tests: 1459 passed, 0 failed +llm tests: 325 passed, 0 failed +tui tests: 282 passed, 0 failed +DAG behavior/coverage: passed +lint: 4738 warnings, 0 errors (ratchet 4750) +``` + +--- + +### 🔍 Verification + +New regression tests cover the send barrier (adopt, seal, interrupt, early cleanup), per-part adoption after later messages, the placeholder keeping `reasoning_content`, carrier classification, the calibrated estimator, the organizer engine transport, and integration tests in both runtimes showing the next step's request carries the rewrite. A branch build was installed over the local command and verified by path, version and SHA-256, then exercised in isolated sessions against the configured DeepSeek and GLM models with a recording relay: rewrites were in place in the next step's and next turn's requests, a slow organizer call that missed the window was sealed and its part stayed original in every later request, nothing was organized with the feature off, and no server errors were logged at DEBUG level. SpecGit 2.5 `init --check`, an idempotent `setup` refresh and a clean `remove` preview were confirmed. Native CI verified each pull request head before merge. + +--- + +**Full changelog:** [`graphagent-v1.0.64`...`graphagent-v1.0.65`](https://github.com/LeXwDeX/OpenCode-GraphAgent/compare/graphagent-v1.0.64...graphagent-v1.0.65) diff --git a/README.md b/README.md index 8f955dcdd..2a8424a02 100644 --- a/README.md +++ b/README.md @@ -5,12 +5,12 @@ # GraphAgent -> A coding agent that decomposes a task into a dependency graph of child agents and drives it to completion. State is durable, crashes are recoverable, and the whole graph can be inspected and controlled from the terminal. +> An opencode fork for long, multi-step coding work: durable DAG workflows of child agents, reasoning shaping, project memory, autonomous goals, and Claude Code–compatible hooks — all inspectable and controllable from the terminal. GraphAgent is the product name for this project; the repository is published as **OpenCode-GraphAgent**, a fork of the MIT-licensed [opencode](https://github.com/anomalyco/opencode) terminal AI agent that adds -the DAG workflow engine. **Not affiliated with or endorsed by the OpenCode +the DAG workflow engine and the features listed below. **Not affiliated with or endorsed by the OpenCode team.** > [!IMPORTANT] @@ -25,7 +25,123 @@ team.** --- -## What is the graph +## What GraphAgent adds + +Everything below is developed in this fork on top of opencode. Each feature is opt-in or self-contained; the rest of +opencode (providers, LSP, client/server, TUI/desktop/web clients) behaves as upstream. + +| Feature | What it does | How to use it | +| --- | --- | --- | +| [Reasoning shaping](#reasoning-shaping) | Reorganizes each reasoning part with a small model before the model sees it again | `reasoningDistillation.enabled` + `small_model` | +| [DAG workflows](#dag-workflows) | Decomposes a task into a durable dependency graph of child agents | `/dag-auto`, the `workflow` tool, `.opencode/dag.jsonc` | +| [Project Memory](#project-memory) | Keeps confirmed preferences, decisions and terminology per project | `/memory on` | +| [Autonomous goals](#autonomous-goals-goal) | Works one durable, budgeted goal across turns, judged externally | `/goal`, `/subgoal` | +| [Hooks](#hooks) | Claude Code–compatible lifecycle hooks | `hooks.json`, `/create-hook`, `/import-claude-hooks` | +| [Context management](#context-management) | Folds duplicate tool output, bounds tool use, compacts long sessions | `compaction`, `maxToolCalls` | + +--- + +## Reasoning shaping + +Thinking models send their earlier reasoning back to the provider on later steps. Long reasoning is full of restated +requests, abandoned guesses and correction narrative, and every later request pays for it again. + +With reasoning shaping on, each reasoning part is handed to the configured small model **as soon as it ends**. The +organizer keeps final facts, values, calculations, constraints, real execution results and open items; it removes +duplicates, refuted guesses and filler; paths, commands, code, URLs, configuration keys and numbers stay verbatim. The +rewrite runs while the step's tools execute, and the next request carries it. + +- **Rewrite before first resend, or never.** Before every provider request, the session waits for unfinished rewrites + up to a short window (`OPENCODE_REASONING_DISTILLATION_SETTLE_MS`, default 3000 ms). Anything not ready is sealed and + keeps its original text for good, so provider prompt-cache prefixes from earlier requests are never invalidated. +- **Only safe carriers.** Plain reasoning and plaintext mirrors are rewritten. Signed, encrypted or referenced reasoning + (Anthropic thinking signatures, OpenAI encrypted reasoning, Gemini thought signatures) is never touched. +- **Reversible.** The original text and metadata are stored with the part; turning the feature off replays originals. +- **Engine-level.** The organizer call goes through the built-in `@opencode-ai/llm` engine with no tools, no retries and + `none`/`low` reasoning effort; a rewrite is adopted only if it is actually smaller. + +```jsonc +// opencode.json +{ + "small_model": "provider/small-model", // prefer one with a `none` reasoning variant + "reasoningDistillation": { + "enabled": true, + "language": "zh", // organizer prose language: "zh" (default) or "en" + }, +} +``` + +`OPENCODE_DISABLE_REASONING_DISTILLATION=1` turns it off regardless of configuration; +`OPENCODE_REASONING_DISTILLATION_AUX_TIMEOUT_MS` bounds each organizer call. Logs report every rewrite +(`reasoning distillation`) and every sealed barrier (`reasoning rewrite sealed`). + +--- + +## Project Memory + +Project Memory keeps durable, **user-confirmed** context for a project: preferences, decisions and terminology. + +- `/memory on` / `/memory off` control it. The project needs a real identity first (run `/init`); Memory stays inert + for uninitialized projects. +- One Memory per project, shared by all of its worktrees. It never forks per worktree, and it moves with the project + when the repository gains its first remote. +- The main session injects relevant Memory into context; `memory_search` retrieves topics on demand. +- The controller owns persistence, limits and migration; models only propose bounded changes. Memory is not a code + index and never an instruction source: current user input and higher-priority instructions always win. +- Policy lives in `.opencode/memory.jsonc`. + +--- + +## Autonomous goals (`/goal`) + +The goal loop works one durable goal autonomously across turns of the current session; graph orchestration (below) is +its multi-session counterpart. + +- Commands: `/goal ` sets a goal and starts the loop; `/goal status|pause|resume|done|clear|stop` controls it; + `/subgoal |list|remove |clear` manages subgoals attached to the active goal. +- Budget: `/goal --max-turns 30 ` sets a positive integer total budget (default 20). `/goal resume --max-turns 50` + changes the total budget without resetting used turns; the new total must exceed used turns. Plain `/goal resume` + retains the budget and permits one more execution when exhausted. +- Judge loop: after each turn an external judge evaluates progress — `done` clears the goal, `continue` injects the next + continuation turn against the turn budget (exhaustion pauses the goal; it stays resumable). The agent can + self-declare completion with `goal(action: "complete")`; `goal(action: "status")` inspects state. +- Visibility: while a goal is active or paused, the system prompt carries a live goal block; the TUI sidebar shows a + compact goal widget; `GET /session/:sessionID/goal` exposes the state (`404` when no goal is set). +- Durability: goal state is persisted per session, survives restarts, and is cleared when the session is deleted. Only + the main conversation can create or resume a goal. + +--- + +## Hooks + +Lifecycle hooks compatible with the Claude Code hooks protocol: 26 events (`PreToolUse`, `PostToolUse`, +`SessionStart`, `PermissionRequest`, `WorktreeCreate`, …) × 5 execution types (`command`, `mcp`, `http`, `prompt`, +`agent`). + +- `hooks.json` is read from the global opencode config directory and from project/worktree `.opencode` directories, + merged by appending and hot-reloaded; hooks can also be registered per session over HTTP, with optional workspace-trust + gating. +- Command hooks can declare `inputFormat: "claude-code"` to receive Claude Code tool names and input keys. This gives + naming compatibility, not full behavioral parity. +- Claude `.claude/settings*.json` files are not loaded automatically: `/import-claude-hooks` migrates them, and + `/create-hook` walks through authoring a new hook. +- Reference: [hooks guide](./packages/core/src/plugin/skill/configure-hooks.md). + +--- + +## Context management + +- **Context folding**: repeated tool outputs are folded in the outbound request while the stored history stays + canonical (`compaction.dynamic`, `compaction.prune`; `OPENCODE_DISABLE_PRUNE=1` disables pruning). +- **Bounded tool use**: tool output is size-bounded, and `maxToolCalls` caps tool calls per user input. +- **Compaction**: automatic summarization of long sessions (`compaction.auto`, `compaction.keep.tokens`, + `compaction.buffer`, `compaction.max_context_tokens`). + +Which of these apply depends on configuration, provider support and request purpose. + +--- + +## DAG workflows A workflow is a set of nodes connected by dependency edges. Each node is a real child session with its own agent and context window; an edge means the downstream node consumes the upstream node's output. Nodes run wave by wave in dependency order, so independent work executes in parallel and dependent work waits. @@ -37,7 +153,7 @@ Three terms worth knowing: - **Wave** — the set of nodes whose dependencies are all satisfied; a wave runs in parallel up to a concurrency limit. - **Gate** — a node whose job is judgment (review, verification, arbitration). Gates emit verdicts (`ACCEPT` / `REVISE` / `REJECT` / `BLOCKED`) and downstream nodes can be conditioned on the verdict. -## Features +### DAG features **Orchestration** @@ -59,10 +175,9 @@ Three terms worth knowing: - Sidebar panel with per-session progress; HTTP API mirroring every tool action (see [below](#observing--controlling)). - Deep mode admission: a bounded Q&A pass (1/3/5 rounds) produces a fingerprinted Requirement Brief with a `READY` / `NOT_READY` / `WAIVED` verdict before an expensive graph starts. -**Beyond the graph** +**Messaging** -- Autonomous goal loop (`/goal`): one durable goal worked across turns of a single session, judged externally, budgeted and resumable. -- Claude Code hooks compatibility (26 events × 5 execution types), CJK/IME terminal fixes, per-workflow worktree isolation, and a standalone Go configuration assistant. +- The main agent can observe nodes of its own workflows and exchange messages with an exact current node attempt. Sending is non-blocking; messages provide context and never grant authorization or change workflow lifecycle. ## Using workflows @@ -258,23 +373,9 @@ Curated reference topologies — design decision deep-dive, parallel project del --- -## Autonomous goal loop (`/goal`) - -Graph orchestration decomposes a task across child sessions; the goal loop is -its single-session complement: one durable goal that the agent works toward -autonomously across turns of the current session. - -- Commands: `/goal ` sets a goal and starts the loop; `/goal status|pause|resume|done|clear|stop` controls it; `/subgoal |list|remove |clear` manages subgoals attached to the active goal. -- Budget: `/goal --max-turns 30 ` sets a positive integer total budget (default 20). `/goal resume --max-turns 50` changes the total budget without resetting used turns; the new total must exceed used turns. Plain `/goal resume` retains the budget and permits one more execution when exhausted. -- Judge loop: after each turn an external judge evaluates progress — `done` clears the goal, `continue` injects the next continuation turn against a configurable turn budget (budget exhaustion pauses the goal; it stays resumable). The agent can self-declare completion with the `goal(action: "complete")` tool, which bypasses the judge; `goal(action: "status")` inspects state. -- Visibility: while a goal is active or paused, the system prompt carries a live goal block (text, status, turns used/remaining, subgoals, last judge verdict); the TUI sidebar shows a compact goal widget; `GET /session/:sessionID/goal` exposes the state (`404` when no goal is set). -- Durability: goal state is persisted per session (`goal_state`), survives restarts, and is cleared automatically when the session is deleted. - ---- - ## Other changes in this fork -- **Hooks API**: Claude Code hooks protocol compatibility. 26 hook events (`PreToolUse`, `PostToolUse`, `SessionStart`, `PermissionRequest`, `WorktreeCreate`, …) × 5 execution types (`command`, `mcp`, `http`, `prompt`, `agent`), loaded from a global/project/worktree `hooks.json` chain or registered per-session over HTTP, with optional workspace-trust gating. See the [hooks reference](./packages/core/src/plugin/skill/configure-hooks.md). +- **Delegated agents**: `task` delegates to sub-agents and continues them by `task_id`; background sub-agents with completion notifications are available behind `OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS`. - **Tool robustness**: JSON repair for broken multi-byte Unicode escapes in LLM output, structured validation errors with field-level hints, expanded tool docs, child-process pipe fixes. - **CJK & IME fixes**: corrections for Chinese/Japanese/Korean input in the terminal UI (IME composition flushing, full-width text handling), plus a Korean IME fix script under [`patches/`](./patches). - **Worktree isolation**: per-workflow `git worktree` isolation, with experimental sandbox-worktree HTTP endpoints. @@ -286,14 +387,14 @@ All upstream capabilities (multi-provider, built-in LSP, client/server architect ## Install -Prebuilt CLI binaries (Linux / macOS / Windows, with SHA256SUMS) are published on the [releases page](https://github.com/LeXwDeX/OpenCode-GraphAgent/releases). Builds from `main` are formal releases; builds from `dev` are prereleases. +Prebuilt CLI binaries (Linux / macOS / Windows, with SHA256SUMS) are published on the [releases page](https://github.com/LeXwDeX/OpenCode-GraphAgent/releases). Stable releases are built from `main`; the `oc` installer selects and installs a release. Release acceptance enforces two distinct integrity boundaries: - **Archive integrity (before extraction)**: the `oc` installer verifies the release `SHA256SUMS` entry before unpacking and refuses to extract on mismatch. If upstream serves no `SHA256SUMS`, it installs with a warning (GitHub HTTPS transport only). - **Post-install signature validity (macOS)**: the installer clears quarantine attributes (`xattr -cr`) and ad-hoc re-signs (`codesign -fs -`); acceptance then asserts `codesign --verify` passes and the binary runs. The installed binary's hash is intentionally **not** compared to the archive payload — ad-hoc signing can rewrite the binary's bytes, so differing hashes are legitimate. No post-signature digest is published (cross-version reproducibility of codesign output has not been established, and no supported reproducibility matrix exists). -From source (requires [Bun](https://bun.sh) 1.3+): +From source (requires the [Bun](https://bun.sh) version pinned in `package.json` `packageManager` and the Node version in `.node-version`): ```bash bun install @@ -327,6 +428,9 @@ Exact file boundaries are listed in [`NOTICE`](./NOTICE). The AGPL covers the DA ## Docs - [`docs/architecture.md`](./docs/architecture.md) — system overview, client/API boundaries, and source map +- [`docs/reasoning-rewrite-engine-2026-10-08.md`](./docs/reasoning-rewrite-engine-2026-10-08.md) — reasoning shaping design, measurements and acceptance +- [`packages/opencode/src/memory/CONTEXT.md`](./packages/opencode/src/memory/CONTEXT.md) — Project Memory model and invariants +- [Hooks guide](./packages/core/src/plugin/skill/configure-hooks.md) — events, schemas and verification - [Saved workflow authoring guide](./packages/core/src/plugin/skill/create-dag-workflow.md) — the `create-dag-workflow` skill body - [Graph Engineering workflow catalog](./.opencode/workflows/GRAPH-ENGINEERING.md) — reference topologies and adaptation contracts - [`docs/harness-dag.md`](./docs/harness-dag.md) — deep-mode admission & review lifecycle diff --git a/README.zh.md b/README.zh.md index 3e256052d..6b360f058 100644 --- a/README.zh.md +++ b/README.zh.md @@ -5,10 +5,10 @@ # GraphAgent -> 一个把任务拆成子智能体依赖图、并驱动它跑完的编码智能体。状态持久化，崩溃能恢复，整张图在终端里就能看、能控。 +> 面向长程、多步骤编码工作的 opencode fork：可持久化的子智能体 DAG 工作流、思考自动塑形、项目 Memory、自主目标和兼容 Claude Code 的 hooks，全部能在终端里查看和控制。 GraphAgent 是本项目对外的产品名；仓库以 **OpenCode-GraphAgent** 发布，是 MIT 许可的 -[opencode](https://github.com/anomalyco/opencode) 终端 AI 智能体的 fork，在其之上加了 DAG 工作流引擎。**与 OpenCode 团队无任何隶属或背书关系。** +[opencode](https://github.com/anomalyco/opencode) 终端 AI 智能体的 fork，在其之上加了 DAG 工作流引擎和下文列出的功能。**与 OpenCode 团队无任何隶属或背书关系。** > [!IMPORTANT] > **GraphAgent v1 现进入聚焦维护阶段。**维护范围限定为 DAG 配置、精选工作流模板，以及通过 @@ -19,7 +19,93 @@ GraphAgent 是本项目对外的产品名；仓库以 **OpenCode-GraphAgent** --- -## graph 是什么 +## GraphAgent 自研功能 + +以下功能都是本 fork 在 opencode 之上自行开发的，各自可选或自成体系；opencode 的其余部分（多 Provider、LSP、客户端/服务器、TUI/桌面/Web 客户端）与上游一致。 + +| 功能 | 作用 | 怎么用 | +| --- | --- | --- | +| [思考自动塑形](#思考自动塑形) | 思考内容在再次发给模型之前，先由小模型整理去噪 | `reasoningDistillation.enabled` + `small_model` | +| [DAG 工作流](#dag-工作流) | 把任务拆成可持久化的子智能体依赖图并跑完 | `/dag-auto`、`workflow` 工具、`.opencode/dag.jsonc` | +| [项目 Memory](#项目-memory) | 按项目保存经用户确认的偏好、决策和术语 | `/memory on` | +| [自主目标（GOAL）](#自主目标循环goal) | 跨回合推进一个持久、有预算、外部评审的目标 | `/goal`、`/subgoal` | +| [Hooks](#hooks) | 兼容 Claude Code 的生命周期 hooks | `hooks.json`、`/create-hook`、`/import-claude-hooks` | +| [上下文管理](#上下文管理) | 折叠重复工具输出、限制工具调用、压缩长会话 | `compaction`、`maxToolCalls` | + +--- + +## 思考自动塑形 + +思考型模型会在后续步骤里把之前的思考内容回传给服务商。长思考里充斥着复述需求、被推翻的猜测和纠错过程，后续每次请求都要为它重复付费。 + +开启后，每段思考**一结束**就交给配置的小模型整理：保留最终的事实、数值、计算结果、约束、真实执行结果和待办，删掉重复、已被否定的猜测和空话；路径、命令、代码、URL、配置键和数值逐字保留。整理和本步的工具执行并行进行，下一次请求就带上整理后的内容。 + +- **要么在第一次回传前替换，要么永不替换。** 每次向模型发请求前，会话最多等待一个很短的窗口（`OPENCODE_REASONING_DISTILLATION_SETTLE_MS`，默认 3000 ms）；没整理完的部分被封存，永久保留原文，因此不会让之前请求已经写入的提示缓存前缀失效。 +- **只改安全的载体。** 只改写普通思考和明文镜像；带签名、加密或服务端引用的思考（Anthropic thinking 签名、OpenAI 加密推理、Gemini thought 签名）一律不动。 +- **可还原。** 原文和原始 metadata 随思考一起保存；关闭功能后回放的就是原文。 +- **走引擎层。** 整理请求通过内置的 `@opencode-ai/llm` 引擎发送：不带工具、不自动重试、推理档位用 `none`/`low`；只有整理结果确实更省 token 才会采用。 + +```jsonc +// opencode.json +{ + "small_model": "provider/small-model", // 建议选有 `none` 推理档位的模型 + "reasoningDistillation": { + "enabled": true, + "language": "zh", // 整理后说明文字的语言："zh"（默认）或 "en" + }, +} +``` + +`OPENCODE_DISABLE_REASONING_DISTILLATION=1` 可以无视配置直接关闭；`OPENCODE_REASONING_DISTILLATION_AUX_TIMEOUT_MS` 限制单次整理调用的时长。日志会记录每一次整理（`reasoning distillation`）和每一次封存（`reasoning rewrite sealed`）。 + +--- + +## 项目 Memory + +项目 Memory 为一个项目保存持久的、**经用户确认的**上下文：偏好、决策和术语。 + +- 用 `/memory on` / `/memory off` 开关。项目需要先有真实身份（执行 `/init`）；未初始化的项目里 Memory 不生效。 +- 一个项目只有一份 Memory，所有 worktree 共享，不会按 worktree 分叉；仓库第一次加上远端时 Memory 会随项目一起迁移。 +- 主会话把相关 Memory 注入上下文；`memory_search` 按需检索主题。 +- 持久化、上限和迁移由控制器负责，模型只能提出有界的修改。Memory 不是代码索引，也不是指令来源：当前用户输入和更高优先级的指令永远优先。 +- 策略配置在 `.opencode/memory.jsonc`。 + +--- + +## 自主目标循环（`/goal`） + +目标循环在当前会话里推进一个目标；下文的图编排是它的多会话对应形态：一个持久目标，智能体在当前会话里跨回合自主推进。 + +- **命令**：`/goal <文本>` 设定目标并启动循环；`/goal status|pause|resume|done|clear|stop` 控制；`/subgoal <文本>|list|remove |clear` 管理挂在当前目标下的子目标。 +- **轮次预算**：`/goal --max-turns 30 <文本>` 指定正整数总轮数（默认 20）。`/goal resume --max-turns 50` 恢复并调整总轮数，保留已用轮数，新总数必须大于已用轮数。不带参数的 `/goal resume` 保留预算，耗尽后仍可再执行一轮。 +- **评审循环**：每回合结束后由外部评审判定进展——`done` 清除目标，`continue` 注入下一轮续跑提示，受可配置的回合预算约束（预算耗尽转暂停，可随时恢复）。智能体也可以用 `goal(action: "complete")` 工具自我宣告完成（绕过评审）；`goal(action: "status")` 查询状态。 +- **可见性**：目标激活或暂停期间，系统提示里带实时目标块（目标文本、状态、已用/总回合、子目标、最近一次评审判定）；TUI 侧边栏有简洁的目标组件；`GET /session/:sessionID/goal` 暴露状态（未设目标时返回 `404`）。 +- **持久化**：目标状态按会话持久化（`goal_state`），重启不丢，会话删除时自动清除。只有主会话能创建或恢复目标。 + +--- + +## Hooks + +兼容 Claude Code hooks 协议的生命周期 hooks：26 个事件（`PreToolUse`、`PostToolUse`、`SessionStart`、`PermissionRequest`、`WorktreeCreate` 等）× 5 种执行类型（`command`、`mcp`、`http`、`prompt`、`agent`）。 + +- `hooks.json` 从全局 opencode 配置目录和项目/worktree 的 `.opencode` 目录读取，按追加方式合并并热加载；也可以经 HTTP 按会话注册，支持可选的工作区信任门控。 +- command hook 可以声明 `inputFormat: "claude-code"`，以 Claude Code 的工具名和参数键接收输入。这提供的是命名兼容，不是完整的行为等价。 +- Claude 的 `.claude/settings*.json` 不会自动加载：用 `/import-claude-hooks` 迁移，用 `/create-hook` 引导编写新 hook。 +- 参考：[hooks 指南](./packages/core/src/plugin/skill/configure-hooks.md)。 + +--- + +## 上下文管理 + +- **上下文折叠**：在发出的请求里折叠重复的工具输出，存储的历史保持原样（`compaction.dynamic`、`compaction.prune`；`OPENCODE_DISABLE_PRUNE=1` 关闭裁剪）。 +- **有界的工具使用**：工具输出有大小上限，`maxToolCalls` 限制每次用户输入的工具调用次数。 +- **压缩**：长会话自动摘要（`compaction.auto`、`compaction.keep.tokens`、`compaction.buffer`、`compaction.max_context_tokens`）。 + +具体哪些生效取决于配置、服务商支持和请求用途。 + +--- + +## DAG 工作流 一个工作流就是一组节点加依赖边。每个节点是一个真实的子会话，有自己的智能体和上下文窗口；一条边意味着下游节点要消费上游节点的产出。节点按依赖顺序逐波执行：能并行的并行跑，有依赖的等着。 @@ -31,7 +117,7 @@ GraphAgent 是本项目对外的产品名；仓库以 **OpenCode-GraphAgent** - **波次（wave）**——依赖全部就绪的那批节点；一次波次内并行执行，受并发上限约束。 - **门禁（gate）**——职责是做判断的节点（审查、验证、仲裁）。门禁输出裁定（`ACCEPT` / `REVISE` / `REJECT` / `BLOCKED`），下游节点可以基于裁定条件执行。 -## 特性 +### DAG 特性 **编排** @@ -53,10 +139,9 @@ GraphAgent 是本项目对外的产品名；仓库以 **OpenCode-GraphAgent** - 侧边栏面板按会话展示进度；HTTP API 覆盖全部工具动作（见[下文](#观察与控制)）。 - deep 模式准入：昂贵的图启动前先过一轮有界问答（1/3/5 轮），产出带指纹的 Requirement Brief，裁定只有 `READY` / `NOT_READY` / `WAIVED` 三种。 -**图之外** +**消息** -- 自主目标循环（`/goal`）：单会话内跨回合推进一个持久目标，外部评审、有预算、可恢复。 -- Claude Code hooks 兼容（26 事件 × 5 种执行类型）、CJK/IME 终端修复、按工作流的 worktree 隔离、独立的 Go 配置助手。 +- 主智能体可以观察自己工作流里的节点，并和某个节点的当前尝试互发消息。发送是非阻塞的；消息只提供上下文，不授予权限，也不改变工作流生命周期。 ## 工作流怎么用 @@ -223,21 +308,9 @@ DAG 相关的东西都放在 `.opencode/` 下，在 opencode 配置目录（`OPE --- -## 自主目标循环（`/goal`） - -图编排把任务拆给多个子会话；目标循环是它的单会话互补形态：一个持久目标，智能体在当前会话里跨回合自主推进。 - -- **命令**：`/goal <文本>` 设定目标并启动循环；`/goal status|pause|resume|done|clear|stop` 控制；`/subgoal <文本>|list|remove |clear` 管理挂在当前目标下的子目标。 -- **轮次预算**：`/goal --max-turns 30 <文本>` 指定正整数总轮数（默认 20）。`/goal resume --max-turns 50` 恢复并调整总轮数，保留已用轮数，新总数必须大于已用轮数。不带参数的 `/goal resume` 保留预算，耗尽后仍可再执行一轮。 -- **评审循环**：每回合结束后由外部评审判定进展——`done` 清除目标，`continue` 注入下一轮续跑提示，受可配置的回合预算约束（预算耗尽转暂停，可随时恢复）。智能体也可以用 `goal(action: "complete")` 工具自我宣告完成（绕过评审）；`goal(action: "status")` 查询状态。 -- **可见性**：目标激活或暂停期间，系统提示里带实时目标块（目标文本、状态、已用/总回合、子目标、最近一次评审判定）；TUI 侧边栏有简洁的目标组件；`GET /session/:sessionID/goal` 暴露状态（未设目标时返回 `404`）。 -- **持久化**：目标状态按会话持久化（`goal_state`），重启不丢，会话删除时自动清除。 - ---- - ## 本 fork 的其他改动 -- **Hooks API**：兼容 Claude Code hooks 协议，26 个 hook 事件（`PreToolUse`、`PostToolUse`、`SessionStart`、`PermissionRequest`、`WorktreeCreate` 等）× 5 种执行类型（`command`、`mcp`、`http`、`prompt`、`agent`），从全局/项目/worktree 的 `hooks.json` 链加载，也可以经 HTTP 按会话注册，支持可选的工作区信任门控。详见 [hooks 参考](./packages/core/src/plugin/skill/configure-hooks.md)。 +- **委派智能体**：`task` 把工作委派给子智能体，并可用 `task_id` 继续；带完成通知的后台子智能体需开启 `OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS`。 - **工具健壮性**：修复 LLM 输出里损坏的多字节 Unicode 转义（JSON 修复），校验错误带字段级提示，工具文档扩充，子进程管道修复。 - **CJK 与 IME 修复**：终端 UI 里中日韩文输入的修正（IME 组字刷新、全角文本处理），另有 [`patches/`](./patches) 下的韩文 IME 修复脚本。 - **Worktree 隔离**：按工作流的 `git worktree` 隔离，附实验性的 sandbox-worktree HTTP 端点。 @@ -249,14 +322,14 @@ DAG 相关的东西都放在 `.opencode/` 下，在 opencode 配置目录（`OPE ## 安装 -预构建 CLI 二进制（Linux / macOS / Windows，附 SHA256SUMS）发布在 [releases 页面](https://github.com/LeXwDeX/OpenCode-GraphAgent/releases)。从 `main` 构建的是正式版；从 `dev` 构建的是预发布版。 +预构建 CLI 二进制（Linux / macOS / Windows，附 SHA256SUMS）发布在 [releases 页面](https://github.com/LeXwDeX/OpenCode-GraphAgent/releases)。正式版从 `main` 构建；`oc` 安装器负责选择并安装版本。 发布验收维护两条相互独立的完整性边界： - **归档完整性（解包前）**：`oc` 安装器在解包前校验 release 的 `SHA256SUMS` 条目，不匹配则拒绝解包。若上游未提供 `SHA256SUMS`，则告警后继续安装（仅依赖 GitHub HTTPS 传输安全）。 - **安装后签名有效性（macOS）**：安装器清除 quarantine 属性（`xattr -cr`）并做 ad-hoc 重签名（`codesign -fs -`），验收断言 `codesign --verify` 通过且二进制可执行。安装后的二进制 hash **有意**不与归档 payload 对比——ad-hoc 签名可能改写二进制字节，两者 hash 即使不同也属正常。也不发布签名后 digest（codesign 输出的跨版本可复现性尚未确立，亦无受支持的可复现性矩阵）。 -从源码构建（需要 [Bun](https://bun.sh) 1.3+）： +从源码构建（需要 `package.json` 中 `packageManager` 固定的 [Bun](https://bun.sh) 版本，以及 `.node-version` 指定的 Node 版本）： ```bash bun install @@ -290,6 +363,9 @@ bun dev serve # headless API 服务（端口 4096） ## 文档 - [`docs/architecture.md`](./docs/architecture.md) —— 系统总览、客户端/API 边界与源码索引 +- [`docs/reasoning-rewrite-engine-2026-10-08.md`](./docs/reasoning-rewrite-engine-2026-10-08.md) —— 思考自动塑形的设计、测量与验收 +- [`packages/opencode/src/memory/CONTEXT.md`](./packages/opencode/src/memory/CONTEXT.md) —— 项目 Memory 的模型与不变量 +- [hooks 指南](./packages/core/src/plugin/skill/configure-hooks.md) —— 事件、schema 与验证方法 - [存盘工作流编写指南](./packages/core/src/plugin/skill/create-dag-workflow.md) —— `create-dag-workflow` skill 正文 - [Graph Engineering 工作流目录](./.opencode/workflows/GRAPH-ENGINEERING.md) —— 参考拓扑与自适应协议 - [`docs/harness-dag.md`](./docs/harness-dag.md) —— deep 模式准入与审查生命周期 From 10b89648d3ff40e398323324bfd53da37d7b53d5 Mon Sep 17 00:00:00 2001 From: Lex Date: Fri, 9 Oct 2026 00:24:55 +0800 Subject: [PATCH 2/5] docs: update 1.0.65 notes for the send fence and final test counts --- .github/releases/v1.0.65.md | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/.github/releases/v1.0.65.md b/.github/releases/v1.0.65.md index 102d4515b..fb15d48ed 100644 --- a/.github/releases/v1.0.65.md +++ b/.github/releases/v1.0.65.md @@ -6,8 +6,9 @@ Stable release from `main` branch. Reasoning shaping now rewrites each reasoning ### 🎯 Features -- **Reasoning shaping, per step**: Each reasoning part is organized by the configured small model as soon as it ends, while the step's tools run, instead of once after the turn. Before every provider request the session waits for unfinished rewrites up to `OPENCODE_REASONING_DISTILLATION_SETTLE_MS` (default 3000 ms) and seals the rest; a sealed part keeps its original text for good. A part is therefore rewritten before its first resend or never, so earlier prompt-cache prefixes are never invalidated, and within-turn resends now carry the organized text. Adoption is checked per part, so a quick follow-up message no longer discards a whole turn (#721). +- **Reasoning shaping, per step**: Each reasoning part is organized by the configured small model as soon as it ends, while the step's tools run, instead of once after the turn. Before every provider request the session waits for unfinished rewrites up to `OPENCODE_REASONING_DISTILLATION_SETTLE_MS` (default 3000 ms) and seals the rest; a sealed part keeps its original text for good. A part is therefore rewritten before its first resend or never, so earlier prompt-cache prefixes are never invalidated, and within-turn resends now carry the organized text. Adoption is checked per part, so a quick follow-up message no longer discards a whole turn. Disposing a directory or deleting a session cancels its pending rewrites and their model calls (#721). - **Engine organizer**: Both runtimes organize through the `@opencode-ai/llm` engine with no tools, temperature 0, no retries and `none`/`low` effort translated per protocol. The opencode runtime keeps the AI SDK only as a fallback for packages the engine does not route. Parts are organized concurrently and independently; the session budget counts reported or estimated tokens instead of stopping when a provider omits usage or after a fixed number of calls (#721). +- **Send fence across processes**: Adoption is refused once any process has persisted a later assistant attempt. The opencode loop persists each attempt before re-reading the previous message's reasoning for its request, so a rewrite committed by another process is either in that request or refused (#721). - **Barrier logs**: Every sealed barrier is logged as `reasoning rewrite sealed` (info) and other barriers as `reasoning rewrite barrier` (debug), with job, adopted and sealed counts and the wait time (#721). --- @@ -39,8 +40,8 @@ Stable release from `main` branch. Reasoning shaping now rewrites each reasoning ```text workspace typecheck tasks: 31/31 passed -opencode tests: 5241 passed, 0 failed -core tests: 1459 passed, 0 failed +opencode tests: 5245 passed, 0 failed +core tests: 1461 passed, 0 failed llm tests: 325 passed, 0 failed tui tests: 282 passed, 0 failed DAG behavior/coverage: passed @@ -51,7 +52,7 @@ lint: 4738 warnings, 0 errors (ratchet 4750) ### 🔍 Verification -New regression tests cover the send barrier (adopt, seal, interrupt, early cleanup), per-part adoption after later messages, the placeholder keeping `reasoning_content`, carrier classification, the calibrated estimator, the organizer engine transport, and integration tests in both runtimes showing the next step's request carries the rewrite. A branch build was installed over the local command and verified by path, version and SHA-256, then exercised in isolated sessions against the configured DeepSeek and GLM models with a recording relay: rewrites were in place in the next step's and next turn's requests, a slow organizer call that missed the window was sealed and its part stayed original in every later request, nothing was organized with the feature off, and no server errors were logged at DEBUG level. SpecGit 2.5 `init --check`, an idempotent `setup` refresh and a clean `remove` preview were confirmed. Native CI verified each pull request head before merge. +New regression tests cover the send barrier (adopt, seal, interrupt, early cleanup), cancellation on directory disposal and session deletion, a rewrite committed by another process before a step's claim, per-part adoption after later messages, the placeholder keeping `reasoning_content`, carrier classification, the calibrated estimator, the organizer engine transport, and integration tests in both runtimes showing the next step's request carries the rewrite. A branch build was installed over the local command and verified by path, version and SHA-256, then exercised in isolated sessions against the configured DeepSeek and GLM models with a recording relay: rewrites were in place in the next step's and next turn's requests, a slow organizer call that missed the window was sealed and its part stayed original in every later request, nothing was organized with the feature off, and no server errors were logged at DEBUG level. SpecGit 2.5 `init --check`, an idempotent `setup` refresh and a clean `remove` preview were confirmed. Native CI verified each pull request head before merge. --- From d735d6252bbf687a13875ca59c2da117f1593988 Mon Sep 17 00:00:00 2001 From: Lex Date: Fri, 9 Oct 2026 00:43:55 +0800 Subject: [PATCH 3/5] docs: note the organizer output cap in 1.0.65 --- .github/releases/v1.0.65.md | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/.github/releases/v1.0.65.md b/.github/releases/v1.0.65.md index fb15d48ed..5ac8ea337 100644 --- a/.github/releases/v1.0.65.md +++ b/.github/releases/v1.0.65.md @@ -17,6 +17,7 @@ Stable release from `main` branch. Reasoning shaping now rewrites each reasoning - **Reasoning shaping**: A Chinese rewrite of English reasoning was rejected as larger because the savings check used the conservative reserve estimator, which counts one token per CJK character and four characters per English token. Measured on the configured DeepSeek and GLM tokenizers the rewrite was smaller; the check now uses a script-calibrated estimate (#721). - **Reasoning shaping**: An all-noise part was replaced with empty text, which made the core runner drop `reasoning_content` from that assistant message. It is now replaced with a short non-empty placeholder (#721). +- **Reasoning shaping**: The organizer always requested 24,576 output tokens, so providers that enforce a lower output cap for the small model rejected every rewrite and paused the session. Both transports now clamp to the model's declared output limit (#721). - **SpecGit**: Prettier had reformatted the SpecGit-managed block in `AGENTS.md`, so SpecGit 2.5 reported a damaged marker and refused `init`, `migrate` and `remove`. The blocks are regenerated and `AGENTS.md` is excluded from Prettier (#723). --- @@ -41,7 +42,7 @@ Stable release from `main` branch. Reasoning shaping now rewrites each reasoning ```text workspace typecheck tasks: 31/31 passed opencode tests: 5245 passed, 0 failed -core tests: 1461 passed, 0 failed +core tests: 1462 passed, 0 failed llm tests: 325 passed, 0 failed tui tests: 282 passed, 0 failed DAG behavior/coverage: passed From 90141ec4aca505803c6e0e8779d623703785c166 Mon Sep 17 00:00:00 2001 From: Lex Date: Fri, 9 Oct 2026 06:22:15 +0800 Subject: [PATCH 4/5] docs: point DAG prompt and catalog references at their current homes --- README.md | 16 ++++++++-------- README.zh.md | 9 ++++----- 2 files changed, 12 insertions(+), 13 deletions(-) diff --git a/README.md b/README.md index 2a8424a02..e924cedcc 100644 --- a/README.md +++ b/README.md @@ -266,10 +266,11 @@ pick, and proves it by starting it once. For one-off graphs, `workflow(action="d takes the structured graph as tool parameters and hands back a validated `spec_path`, so field drift never reaches the file. -Node prompts come from `.opencode/dag-prompts/*.md` — 12 templates ship -in-repo, referenced by `prompt_template.id`. Add your own `.md` file there to -make a new template available; a global workflow should prefer `inline` prompts -so it does not depend on a repo-local template. +Node prompts referenced by `prompt_template.id` resolve from the project's +`.opencode/dag-prompts/.md`, then the global `/dag-prompts/`. +Add your own `.md` file there to make a new template available; a global +workflow should prefer `inline` prompts so it does not depend on a repo-local +template. ## The engine @@ -283,7 +284,7 @@ Each node declares: |---|---| | `depends_on` | Dependency edges; cycle detection and dangling-reference validation at creation | | `worker_type` | Which agent runs the node (`explore`, `build`, `general`, or any configured agent) | -| `prompt_template` | Prompt by `id` (from `.opencode/dag-prompts/`, 12 templates ship in-repo) or `inline`, with `{{var}}` interpolation | +| `prompt_template` | Prompt by `id` (from project, then global `dag-prompts/`) or `inline`, with `{{var}}` interpolation | | `input_mapping` | Map upstream node outputs into template variables (`"count": "node-b.output.count"`) | | `condition` | Expression over upstream outputs; false → node skipped, pure descendants cascade-skip | | `output_schema` | JSON Schema; the child agent must call `submit_result` with a matching structured payload | @@ -369,7 +370,7 @@ The operating doctrine that came out of it: Why a DAG at all: a single agent loop struggles once a task has staged dependencies, parallelizable independent work, or a quality gate in the middle. Splitting decisions from volume (advanced vs. standard tiers), asking before building (deep-mode admission), gating verdicts with mandatory disposal, and recovering from evidence rather than guesses are the four judgments this engine is built on. -Curated reference topologies — design decision deep-dive, parallel project delivery, deep review of an existing subsystem, compact change review — ship through the workflow library's global scope (curated by the [`opencode-dag-config`](https://github.com/LeXwDeX/opencode-dag-config) repo) and a builtin tier embedded in release binaries. See the [Graph Engineering workflow catalog](./.opencode/workflows/GRAPH-ENGINEERING.md). +Curated reference topologies — design decision deep-dive, parallel project delivery, deep review of an existing subsystem, compact change review — ship through the workflow library's global scope (curated by the [`opencode-dag-config`](https://github.com/LeXwDeX/opencode-dag-config) repo) and a builtin tier embedded in release binaries. See the [workflow library](https://github.com/LeXwDeX/opencode-dag-config) for the catalog. --- @@ -432,9 +433,8 @@ Exact file boundaries are listed in [`NOTICE`](./NOTICE). The AGPL covers the DA - [`packages/opencode/src/memory/CONTEXT.md`](./packages/opencode/src/memory/CONTEXT.md) — Project Memory model and invariants - [Hooks guide](./packages/core/src/plugin/skill/configure-hooks.md) — events, schemas and verification - [Saved workflow authoring guide](./packages/core/src/plugin/skill/create-dag-workflow.md) — the `create-dag-workflow` skill body -- [Graph Engineering workflow catalog](./.opencode/workflows/GRAPH-ENGINEERING.md) — reference topologies and adaptation contracts +- [`opencode-dag-config`](https://github.com/LeXwDeX/opencode-dag-config) — curated workflows, composable blocks and worker prompts - [`docs/harness-dag.md`](./docs/harness-dag.md) — deep-mode admission & review lifecycle -- [`.opencode/dag-prompts`](./.opencode/dag-prompts) — built-in node prompt templates - [`AGENTS.md`](./AGENTS.md) — contribution & development guide ## Links diff --git a/README.zh.md b/README.zh.md index 6b360f058..22dfcfba5 100644 --- a/README.zh.md +++ b/README.zh.md @@ -207,7 +207,7 @@ config: 内置的 **`create-dag-workflow`** skill 覆盖 spec 编写：工作流库作用域、文件结构、存盘 spec 必须守的规矩（不能钉死模型、`worker_type` 必须存在、模板必需变量必须给全），以及怎么验证。说一句「把这个存成可复用的工作流」，智能体会先跟你确认阶段和门禁，把文件写进你选的作用域，再真跑一次证明它能用。临时图则用 `workflow(action="draft")`：结构化图走工具参数，返回校验过的 `spec_path`，字段漂移到不了文件。 -节点 prompt 来自 `.opencode/dag-prompts/*.md` —— 随仓库附带 12 个，通过 `prompt_template.id` 引用。往那儿加一个 `.md` 就多一个模板；全局工作流建议用 `inline` prompt，否则会依赖某个仓库本地的模板。 +通过 `prompt_template.id` 引用的节点 prompt 先从项目的 `.opencode/dag-prompts/.md` 解析，再找全局 `/dag-prompts/`。往那儿加一个 `.md` 就多一个模板；全局工作流建议用 `inline` prompt，否则会依赖某个仓库本地的模板。 ## 引擎 @@ -221,7 +221,7 @@ config: |---|---| | `depends_on` | 依赖边；创建时做环检测和悬空引用校验 | | `worker_type` | 执行节点的智能体（`explore`、`build`、`general` 或任意已配置 agent） | -| `prompt_template` | 通过 `id` 引用模板（`.opencode/dag-prompts/`，随仓库附带 12 个）或 `inline` 内联，支持 `{{var}}` 插值 | +| `prompt_template` | 通过 `id` 引用模板（先项目、后全局的 `dag-prompts/`）或 `inline` 内联，支持 `{{var}}` 插值 | | `input_mapping` | 把上游节点输出映射为模板变量（`"count": "node-b.output.count"`） | | `condition` | 基于上游输出的表达式；为假则跳过节点，纯依赖它的下游级联跳过 | | `output_schema` | JSON Schema；子智能体必须调用 `submit_result` 提交匹配的结构化结果 | @@ -304,7 +304,7 @@ DAG 相关的东西都放在 `.opencode/` 下，在 opencode 配置目录（`OPE 至于为什么是 DAG：任务一旦涉及分阶段依赖、可并行的独立工作，或者中间需要一道质量门禁，单智能体循环就不太够用了。决策和跑量分开（advanced/standard 分层）、建图前先问清楚（deep 准入）、门禁结论必须有下文（处置契约）、恢复靠证据不靠猜——这四个判断就是这个引擎的地基。 -强约束参考拓扑——设计决策深挖、并行项目落地、已完成子系统深度 Review、轻量变更审查——随工作流库分发：全局作用域由 [`opencode-dag-config`](https://github.com/LeXwDeX/opencode-dag-config) 仓库维护，正式版的二进制里还有内嵌的 builtin 层。入口见 [Graph Engineering 工作流目录](./.opencode/workflows/GRAPH-ENGINEERING.md)。 +强约束参考拓扑——设计决策深挖、并行项目落地、已完成子系统深度 Review、轻量变更审查——随工作流库分发：全局作用域由 [`opencode-dag-config`](https://github.com/LeXwDeX/opencode-dag-config) 仓库维护，正式版的二进制里还有内嵌的 builtin 层。目录见[工作流库](https://github.com/LeXwDeX/opencode-dag-config)。 --- @@ -367,9 +367,8 @@ bun dev serve # headless API 服务（端口 4096） - [`packages/opencode/src/memory/CONTEXT.md`](./packages/opencode/src/memory/CONTEXT.md) —— 项目 Memory 的模型与不变量 - [hooks 指南](./packages/core/src/plugin/skill/configure-hooks.md) —— 事件、schema 与验证方法 - [存盘工作流编写指南](./packages/core/src/plugin/skill/create-dag-workflow.md) —— `create-dag-workflow` skill 正文 -- [Graph Engineering 工作流目录](./.opencode/workflows/GRAPH-ENGINEERING.md) —— 参考拓扑与自适应协议 +- [`opencode-dag-config`](https://github.com/LeXwDeX/opencode-dag-config) —— 精选工作流、可组合块与 worker prompt - [`docs/harness-dag.md`](./docs/harness-dag.md) —— deep 模式准入与审查生命周期 -- [`.opencode/dag-prompts`](./.opencode/dag-prompts) —— 内置节点 prompt 模板 - [`AGENTS.md`](./AGENTS.md) —— 贡献与开发指南 ## 链接 From cae95521cc8e4ba492c7fa0cdd4a4d37d6d48578 Mon Sep 17 00:00:00 2001 From: Lex Date: Fri, 9 Oct 2026 06:28:27 +0800 Subject: [PATCH 5/5] docs: list the global dag-prompts directory in the configuration tables --- README.md | 2 +- README.zh.md | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index e924cedcc..aa3086e8a 100644 --- a/README.md +++ b/README.md @@ -350,7 +350,7 @@ Everything else inherits the main opencode configuration. |---|---|---| | `.opencode/dag.jsonc` | Model tiers (`advanced` / `standard`) and `thinking_depth` for child sessions | `/dag.jsonc`, seeded with comments on first use | | `.opencode/workflows/*.yaml` | Saved workflow specs, startable by name | `/workflows/*.yaml` | -| `.opencode/dag-prompts/*.md` | Node prompt templates referenced by `prompt_template.id` | — (project-scoped) | +| `.opencode/dag-prompts/*.md` | Node prompt templates referenced by `prompt_template.id` (project first) | `/dag-prompts/*.md` | | `.opencode/workflow-reports/` | Node report files (auto-gitignored) | — | Both `dag.jsonc` and the workflow library are read lazily, so an edit applies to diff --git a/README.zh.md b/README.zh.md index 22dfcfba5..ba2a1d12c 100644 --- a/README.zh.md +++ b/README.zh.md @@ -285,7 +285,7 @@ DAG 相关的东西都放在 `.opencode/` 下，在 opencode 配置目录（`OPE |---|---|---| | `.opencode/dag.jsonc` | 模型分层（`advanced` / `standard`）与子会话 `thinking_depth` | `<配置目录>/dag.jsonc`，首次使用时生成带注释的默认文件 | | `.opencode/workflows/*.yaml` | 存盘的工作流 spec，可按名字启动 | `<配置目录>/workflows/*.yaml` | -| `.opencode/dag-prompts/*.md` | 由 `prompt_template.id` 引用的节点 prompt 模板 | ——（仅项目级） | +| `.opencode/dag-prompts/*.md` | 由 `prompt_template.id` 引用的节点 prompt 模板（项目优先） | `<配置目录>/dag-prompts/*.md` | | `.opencode/workflow-reports/` | 节点报告文件（自动 gitignore） | —— | `dag.jsonc` 和工作流库都是惰性读取，改完下一次启动工作流就生效，不用重启。