dev-flow-deepseek 0.8.9 → 0.9.1

package/README.md

@@ -67,6 +67,11 @@ dsh --profile $ProfileName --dump-config
 
 ## 启动一个 Task
 
+仓库调查遵循当前用户指令和适用的 `AGENTS.md`。需要项目索引时，DeepSeek 在当前 Workspace Root
+和权限内只读检查索引、候选项目说明及相关代码与配置，形成完整候选范围，再逐仓确认并准备工作树；
+Task 创建后范围固定。选择代码检索工具时，这些指令优先于 `host_preferences.deepseek.codebase_memory`
+默认偏好。
+
 先把开发请求正常发给 DSH。Adapter 只读检查候选实现、调用关系、测试、配置和 Git 状态，返回
 `small|standard|large|uncertain`、已经找到的影响面、未知项和建议；这一轮不调用 Core、不修改 Git、
 不运行测试，也不创建 Task。即使第一条消息已经包含 `/dev-flow`，新请求也不能跳过评估和确认。
@@ -115,12 +120,14 @@ Adapter 在 TASKS 完成需求、设计、工作拆分、影响面和现有测
 留在 TEST。为了更全面、提高信心、保险起见或“还有预算”都不能支持增加。
 
 每次完整套件前都重新判断广泛影响、定向/包级检查是否足够、待补的具体风险和仓库当前检查点要求，
-本次理由写入 `full_suite_reason`，小修复后不自动沿用旧理由。测试代码只为稳定行为、公共合同、重要
+本次理由写入 `full_suite_reason`，小修复后不自动沿用旧理由。测试代码只为稳定行为、公开接口规范、重要
 失败路径或真实回归保留；一次性 README 词语要求只做一次文本搜索。
 
 普通实现后的复核只看当前 diff、直接或间接影响和验收所需路径。修复复核发现后只检查原问题、相关
 回归和对应定向检查，不重启全仓库审计。显式 code review 阶段只读，交付完整发现后等待单独修复授权。
 
+普通修改后的复核和交付只报告与当前改动有因果关系的问题，无关历史问题不进入报告。
+
 ## 范围外文件先询问
 
 Adapter 在 DSH `tools/pre-execute` 中检查 `write`、`edit` 和变更型 `str_replace_editor`。当前 direct
@@ -170,7 +177,7 @@ dsh --profile "$PROFILE" --dump-config
 ## DeepSeek 权限与边界
 
 - DSH 启动时的 canonical Workspace Root 是权限边界；仓库和 symlink 解析结果必须位于其中；
-- Dev Flow 不扩大 Workspace Root，也不会通过索引发现并加入相邻仓库；需要 sibling worktree 时重启到
+- Dev Flow 保持当前 Workspace Root 权限边界，索引结果不改变已创建 Task 的 Scope；需要 sibling worktree 时重启到
   Coordinator 返回的新 Root；
 - Core 只读观察 Git，不执行 fetch、branch、worktree、commit、push、merge、rebase、tag 或 publish；
 - DeepSeek 负责用户确认后的 fetch、branch、worktree、文件修改和命令执行；Host gate 检查列出的

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dev-flow-deepseek",
-  "version": "0.8.9",
+  "version": "0.9.1",
   "private": false,
   "description": "Explicit DeepSeek Harness adapter for the Dev Flow process graph.",
   "license": "Apache-2.0",

package/skills/dev-flow/SKILL.md

@@ -17,6 +17,14 @@ inspect Git without changing it, and inspect candidate implementation, callers,
 and package manifests. Do not edit files, run tests or builds, install dependencies, call any Dev Flow
 Core or Host tool, fetch, create a branch or worktree, or write a receipt.
 
+Resolve candidate repositories from the current user request and applicable `AGENTS.md` instructions.
+When those instructions require project-index discovery, read the index and each candidate's project
+documentation, then inspect relevant code and configuration to establish the complete proposed
+Repository Scope. Keep discovery read-only and within the current Workspace Root and Host permissions.
+Otherwise use the repositories declared by the user. Every proposed repository must pass the explicit
+worktree confirmation below before Task creation; index results alone do not authorize provisioning
+or change an existing Task's immutable Scope.
+
 Return exactly these developer-readable fields:
 
 ```text
@@ -133,17 +141,20 @@ failed handshake.
 ## Optional code discovery
 
 After the successful handshake, consume only `host_preferences.deepseek.codebase_memory` and the
-capabilities actually visible in this DeepSeek session:
+capabilities actually visible in this DeepSeek session. Current user instructions and applicable
+`AGENTS.md` instructions take precedence over this default preference when choosing code-discovery
+tools. Use the preference rules below only when those instructions do not select a discovery method:
 
-- When the preference is `false`, do not call any codebase-memory tool even when one is visible. Use
-  built-in Git inspection, file reads, file search, and text search, and do not prompt for installation.
+- When the preference is `false`, use built-in Git inspection, file reads, file search, and text
+  search, and do not prompt for installation.
 - When the preference is `true` and codebase-memory is already visible and usable, it may be
-  preferred for cross-repository symbol discovery, relationships, and impact analysis. Repository
-  Scope still comes only from the user's declarations, Workspace Root remains the permission
-  boundary, and file modification uses ordinary Host file tools.
-- When the preference is `true` but the capability is absent, incomplete, or becomes unavailable,
-  notify the user at most once in the current Dev Flow session and immediately fall back to built-in
-  search without blocking Task creation or progress.
+  preferred for symbol discovery, relationships, and impact analysis within the confirmed Task Scope.
+
+Whether selected by user instructions, `AGENTS.md`, or the preference, an absent, incomplete, or
+unavailable index triggers at most one notice in the current Dev Flow session and an immediate
+fallback to built-in search without blocking Task creation or progress. Repository Scope remains
+the complete set confirmed before provisioning and retained by Core; Workspace Root remains the
+permission boundary, and file modification uses ordinary Host file tools.
 
 Never install, configure, upgrade, start, repair, or remove codebase-memory; never call plugin
 management to install it; never change MCP configuration; and never start a daemon. Index results
@@ -531,8 +542,8 @@ restart a repository-wide audit after each edit. Any added review area needs a s
 from the current change.
 
 Fix only defects introduced by the current change or caused in another location by that change.
-Unrelated historical problems are not repaired, tested, or added to this Task; mention them
-separately at delivery and suggest another Task when useful.
+Report findings only when they have that causal relationship. Unrelated historical problems stay
+outside the review, Task work, and delivery summary.
 
 After fixing a review finding, re-check only the original finding, related regressions, affected
 acceptance criteria, and matching targeted checks. Never restart a broad audit because one finding

package/skills/dev-flow/references/method-profiles.md

@@ -115,28 +115,31 @@ For each step returned by Core:
 6. Record completion only after the semantic work actually completes.
 
 For `plain`, render external capability availability as `not_applicable`. When the ordinary work
-actually completes, evidence uses `status="plain_fallback"` and an empty `capability`; plain work is
-never labeled as an external capability completion. The same `plain_fallback` rule applies when a
-selected external capability is unavailable but the equivalent work actually completes.
+actually completes, submit an empty `capability`; Core records `status="plain_fallback"`. The same
+rule applies when a selected external capability is unavailable but the equivalent work actually
+completes. A non-empty `capability` names the actual external capability that completed the step.
 
-When equivalent work remains incomplete, record `unavailable` or `not_run` honestly. Neither status
-can satisfy a required semantic step, so the Adapter must not call apply for that Action.
+When equivalent work remains incomplete, report `unavailable` or `not_run` honestly to the user.
+An incomplete required semantic step prevents submission for that Action.
 
-## MethodEvidence
+## Method results
 
-A normal mutation supplies exactly one `MethodEvidence` item for every current Action method step,
-in Action order. Each item contains only `step_id`, `status`, `capability`, and `summary`.
+A normal submission supplies `method_results` as a closed object keyed by every current Action
+`method_steps[].step_id`. Each value contains only `capability` and `summary`. Core creates the
+internal `MethodEvidence` items in Action order and fills their `step_id` and `status`.
 
-- `completed` requires the non-empty ID of the actual capability that completed the work.
-- `plain_fallback` requires an empty `capability` and completed plain-equivalent work.
-- `unavailable` and `not_run` do not satisfy a required step.
-- Unknown, duplicate, previous-node, reordered, or omitted steps are invalid.
+- Completed external work supplies the actual non-empty capability ID; Core records `completed`.
+- Completed plain-equivalent work supplies an empty `capability`; Core records `plain_fallback`.
+- Incomplete required steps prevent submission; availability and execution status are not input fields.
+- Unknown, duplicate, previous-node, or omitted step keys are invalid. Core determines item order.
 - Capability output cannot substitute for the typed `node_result` or any current Core evidence gate.
 
 ## Artifact references
 
-Submit artifact evidence only as `role`, contract `path`, `digest`, and `summary`. A single-repository
-Task uses an ordinary repository-relative path; a multi-repository Task uses
+Put current-node artifacts in `artifacts.current` when the live schema exposes that slot, and related
+method artifacts in `artifacts.other_process`. Each entry contains only contract `path`, `digest`,
+and `summary`; Core assigns `role` from the slot and current node. A single-repository Task uses an
+ordinary repository-relative path; a multi-repository Task uses
 `<repository-key>::<repository-relative-path>`. Refer
 only to a file actually observed by the Host. Never submit full contents, command output, prompts,
 token data, runtime configuration, or private locations. An artifact digest does not replace the