tech-lead-skill 5.5.5 → 5.5.6
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.en.md +182 -0
- package/README.md +107 -85
- package/docs/AUDIT_REPORT.md +2 -2
- package/package.json +2 -2
- package/skill/SKILL.md +1 -1
- package/README.zh-CN.md +0 -160
package/README.en.md
ADDED
|
@@ -0,0 +1,182 @@
|
|
|
1
|
+
# Tech Lead Skill
|
|
2
|
+
|
|
3
|
+
[简体中文](README.md) | English
|
|
4
|
+
|
|
5
|
+
An evidence-driven planning and delivery skill for software, infrastructure, research, reverse-engineering, and operations work.
|
|
6
|
+
|
|
7
|
+
## Why this instead of another prompt file
|
|
8
|
+
|
|
9
|
+
Most engineering specs stay at the "advice text" layer — easy to agree with, hard to enforce. This project ships one judgment layer in two forms: a prose spec anyone can install, and `dsh-themis`, which turns key rulings into a **machine-checkable read-only runtime**.
|
|
10
|
+
|
|
11
|
+
| Dimension | Generic skill/prompt | tech-lead-skill | dsh-themis (DSH plugin) |
|
|
12
|
+
|---|---|---|---|
|
|
13
|
+
| Task tiering | model's good faith | T0/T1/T2 hard rules + bidirectional re-tiering | same, mechanically re-checkable |
|
|
14
|
+
| Input governance | none | budget clauses | **fail-closed**: oversized input rejected outright (INPUT_TOO_LARGE / SCAN_INCOMPLETE), never silently passed |
|
|
15
|
+
| State validation | reminders | prose schema notes | `state_validate` full-field machine check: enums, done anchors, E0-E4 evidence provenance |
|
|
16
|
+
| Leak audit | none | release checklist | `release_audit` scans absolute paths / token-like strings / credential assignments with line numbers |
|
|
17
|
+
| Result format | whatever comes out | text convention | uniform `protocolJson` negotiation across 22 tools + stable v2 envelope (findings/guidance/meta.complete) |
|
|
18
|
+
| Side effects | — | none | **zero-by-design**: no writes, no subprocesses, no network — computes over your JSON only |
|
|
19
|
+
|
|
20
|
+
Six verifiable differentiators:
|
|
21
|
+
|
|
22
|
+
1. **Governance that rules, not pleads.** Gate precheck, blind-review triggers, stall breakers are executable hard rules — the plugin mechanically answers "can this gate pass now, and which anchor is missing".
|
|
23
|
+
2. **Explicit failure over false confidence.** Audits over-window refuse verdicts; the state→context-v2 projection returns NON_CONVERTIBLE_STATE when identity/source/fingerprint are missing — pinned by 250 tests, not promises.
|
|
24
|
+
3. **Deterministic output.** Same input, same output; guidance actionIds are deterministic and usable as flow keys.
|
|
25
|
+
4. **Smooth migration.** bare JSON → v1 → v2 envelopes negotiated by the schema-declared `protocolJson` on every tool: zero breakage for old callers, fail-closed on unknown selections.
|
|
26
|
+
5. **Field-born.** Rules distilled from real multi-project incident reviews — execution reliability, config/deploy traps, backup/rollback discipline, anti-risk-control, multi-agent orchestration — not armchair templates.
|
|
27
|
+
6. **One-command adoption.** Versioned single npm packages; bilingual docs throughout.
|
|
28
|
+
|
|
29
|
+
## What It Solves
|
|
30
|
+
|
|
31
|
+
Many plans fail in one of two ways: they are vague and cannot guide execution, or they are over-detailed and become false after the first environmental change. This skill keeps the plan coarse until evidence justifies detail, then continuously revises it against observed reality.
|
|
32
|
+
|
|
33
|
+
The control loop is:
|
|
34
|
+
|
|
35
|
+
```text
|
|
36
|
+
goal -> constraints/assumptions -> L0/L1/L2 plan -> execute -> observe evidence
|
|
37
|
+
-> revise -> choose CONTINUE / PAUSE / SCOPE-DOWN / PIVOT / STOP
|
|
38
|
+
```
|
|
39
|
+
|
|
40
|
+
## Core Features
|
|
41
|
+
|
|
42
|
+
- PLAN and EXECUTE modes with a clear boundary between planning and side effects.
|
|
43
|
+
- Progressive disclosure: L0 architecture, L1 milestones, L2 executable focus.
|
|
44
|
+
- Goal, metric, fact, assumption, decision, risk, dependency, and evidence ledgers.
|
|
45
|
+
- Protected asset handling for source, user data, configuration, secrets, runtime state, and generated artifacts.
|
|
46
|
+
- Minimal mutation protocol: `READ -> CLASSIFY -> PROTECT -> CHANGE -> VERIFY -> RECONCILE -> ROLLBACK/RECORD`.
|
|
47
|
+
- Evidence levels E0-E4, from model inference to observed user outcomes.
|
|
48
|
+
- Failure re-planning, stagnation circuit breaker, rollback discipline, and real-state reconciliation.
|
|
49
|
+
- Optional adversarial review for high-impact or irreversible changes.
|
|
50
|
+
- Git-backed and non-Git project-state recovery.
|
|
51
|
+
- Release-readiness checks for public documentation and skill publishing.
|
|
52
|
+
- Runtime discipline rules for external dependency health, silent failure classes, automation guardrails, idempotent batch jobs, and fallback ladders.
|
|
53
|
+
|
|
54
|
+
## Install
|
|
55
|
+
|
|
56
|
+
Pick your edition:
|
|
57
|
+
|
|
58
|
+
- **Prose skill for any agent environment** (opencode / Claude Code / Codex) → `tech-lead-skill`
|
|
59
|
+
- **DSH plugin with machine-checked read-only tools** → `dsh-themis`
|
|
60
|
+
|
|
61
|
+
### npm (GitHub-hosted, no registry account needed)
|
|
62
|
+
|
|
63
|
+
```bash
|
|
64
|
+
npm i -g github:240xu/tech-lead-skill
|
|
65
|
+
tech-lead-skill # installs into ~/.config/opencode/skills/tech-lead
|
|
66
|
+
```
|
|
67
|
+
|
|
68
|
+
One-shot without a global install:
|
|
69
|
+
|
|
70
|
+
```bash
|
|
71
|
+
npx github:240xu/tech-lead-skill
|
|
72
|
+
```
|
|
73
|
+
|
|
74
|
+
The installer is idempotent: repeated runs back up existing files as `*.bak-<timestamp>` first. Use `--target <dir>` to choose another destination, `--check` to verify an installed copy against the package (hash + version drift), `--dry-run` to preview without writing, and `--uninstall` to remove. Uninstall removes only manifest-managed files — user files and `*.bak-*` backups are kept.
|
|
75
|
+
|
|
76
|
+
`--check` exit codes: `0` clean, `1` drift detected, `2` usage/refusal error.
|
|
77
|
+
|
|
78
|
+
### DeepSeek Harness plugin (read-only tools)
|
|
79
|
+
|
|
80
|
+
**For DSH users:** ships on the DSH plugin market (npm) as ONE self-contained package, **`dsh-themis`** — named after the goddess of divine order: gates as verdicts, evidence weighed, releases audited. Install into any profile:
|
|
81
|
+
|
|
82
|
+
```bash
|
|
83
|
+
dsh plugin --profile headless add dsh-themis
|
|
84
|
+
```
|
|
85
|
+
|
|
86
|
+
22 registered entry points in a single artifact: 21 read-only governance tools plus a capability-discovery tool (`tech_lead_capabilities`) that lists only what this bundle actually registers (the original nine audits plus context validation, evidence graph/freshness analysis, progress decisions, critical-path/impact analysis, resume reconciliation, gate planning/aggregation/reopen checks, and mutation preview). Tools compute over caller-supplied JSON only — no filesystem writes, no subprocesses, no network access. The earlier split packages (`dsh-tech-lead-{core,plugin,bundle}`) are deprecated in favor of this one. Source checkouts remain supported: build via `node scripts/build-market-package.mjs` then `dsh plugin add packages/dsh-themis`. The root npm package distributes the skill and installer only.
|
|
87
|
+
|
|
88
|
+
**Two artifacts, one rulebook:** `tech-lead-skill` is the conservative, broadly-compatible prose spec — install it into any agent environment (opencode, Claude Code, Codex) and nothing else is required. **`dsh-themis` is the DSH-plugin specialized edition**: the same judgment layer plus a machine-checked read-only runtime — schema-declared `protocolJson` negotiation on all 22 tools (bare legacy by default, explicit v1/v2 envelopes, fail-closed on unknown selections), one-way state→context-v2 projection that never invents identity or provenance, strict/compat input handling, and stable v2 envelopes (`findings`, `guidance`, `meta.complete`, `meta.outputProtocol`).
|
|
89
|
+
|
|
90
|
+
```bash
|
|
91
|
+
dsh plugin --profile headless add /path/to/tech-lead-skill/packages/dsh-themis
|
|
92
|
+
dsh --profile headless --dump-config # verify the tech-lead-tools row is present
|
|
93
|
+
```
|
|
94
|
+
|
|
95
|
+
|
|
96
|
+
Output families and limits:
|
|
97
|
+
|
|
98
|
+
- The twelve strengthened tools return a `tech-lead.result.v1` envelope (discriminator: `meta.schema`); the original nine tools return their bare domain shapes for backward compatibility.
|
|
99
|
+
- Rendered output is clamped: finding/error arrays are capped at 500 entries with a `FINDINGS_TRUNCATED` warning; caller-echo arrays (evidence/targets/expectedDiff/verification/items) collapse into `{truncated,total}` beyond 100 entries; any OTHER array inside `data` (including computed results such as `criticalPath`) is head-sliced at 1000 entries while keeping its shape; subtrees deeper than 64 levels collapse into a `DEPTH_LIMIT` marker; payloads above 256 KB switch to compact serialization. Legacy bare top-level arrays slice silently at 500 (no warning field exists on that shape).
|
|
100
|
+
|
|
101
|
+
See [`docs/superpowers/specs/2026-08-25-dsh-tech-lead-system.md`](https://github.com/240xu/tech-lead-skill/blob/main/docs/superpowers/specs/2026-08-25-dsh-tech-lead-system.md) for the architecture and permission matrix.
|
|
102
|
+
|
|
103
|
+
### Manual
|
|
104
|
+
|
|
105
|
+
Copy `skill/SKILL.md` and the `skill/templates/` directory into the skills directory of your OpenCode-compatible environment:
|
|
106
|
+
|
|
107
|
+
```text
|
|
108
|
+
~/.config/opencode/skills/tech-lead/
|
|
109
|
+
```
|
|
110
|
+
|
|
111
|
+
The skill is triggered by project construction, system building, implementation plans, deployment, migration, release, recovery, restructuring, operations work, and other substantial planning requests. It can also be loaded explicitly.
|
|
112
|
+
|
|
113
|
+
## Lifecycle results are valid analyses
|
|
114
|
+
|
|
115
|
+
`PAUSE`, `PIVOT`, `SCOPE-DOWN` and `STOP` from `tech_lead_progress_decide` are returned as **ok:true** analyses with `data.outcome`; a not-yet-passed gate returns `ok:true` with `data.verdict`/`data.pass`. `ok:false` is reserved for malformed input, over-budget payloads (`INPUT_TOO_LARGE`, `ITEM_LIMIT_EXCEEDED`) and incomplete safety scans (`SCAN_INCOMPLETE`). Decision tools attach a deterministic `data.guidance.nextActions[]`; every action carries reason codes, a finding reference and a `doneWhen` predicate. Heuristic suggestions appear only under explicit `guidanceMode:"heuristic"`.
|
|
116
|
+
|
|
117
|
+
## Four-tool starter loop
|
|
118
|
+
|
|
119
|
+
1. `tech_lead_classify` — copy the returned tier into your snapshot's `current.tier`.
|
|
120
|
+
2. `tech_lead_context_validate` — validate the full inline snapshot (schema `tech-lead.context.v1`; canonical example: tests/fixtures/starter-context.v1.json).
|
|
121
|
+
3. `tech_lead_evidence_lint` — pass the snapshot's `evidence` array as a JSON string; findings are advisory.
|
|
122
|
+
4. `tech_lead_progress_decide` — feed the same snapshot; read `data.outcome`, then follow `data.guidance`.
|
|
123
|
+
|
|
124
|
+
## Operating Modes
|
|
125
|
+
|
|
126
|
+
### PLAN
|
|
127
|
+
|
|
128
|
+
Use for intake, goal definition, architecture, decomposition, risk analysis, and verification design. PLAN does not edit files or run side-effecting commands.
|
|
129
|
+
|
|
130
|
+
### EXECUTE
|
|
131
|
+
|
|
132
|
+
Use after the L2 scope is clear. Execute only the smallest approved mutation, record the change, verify behavior, reconcile actual state, and update the plan.
|
|
133
|
+
|
|
134
|
+
## Protected Assets
|
|
135
|
+
|
|
136
|
+
| Class | Default handling |
|
|
137
|
+
|---|---|
|
|
138
|
+
| `SOURCE` | Inspectable diff, tests, and restore point |
|
|
139
|
+
| `USER_DATA` | Read-only by default; write only to an explicit target with recovery |
|
|
140
|
+
| `CONFIG` | Read current state, make the smallest change, reload, verify |
|
|
141
|
+
| `SECRET` | Never place in plans, logs, ordinary backups, or diffs |
|
|
142
|
+
| `RUNTIME` | Inspect live state before restart, kill, replace, or migration |
|
|
143
|
+
| `GENERATED` | Prefer regeneration; do not treat as the source of truth |
|
|
144
|
+
|
|
145
|
+
## Evidence Levels
|
|
146
|
+
|
|
147
|
+
- `E0`: model inference; hypothesis only.
|
|
148
|
+
- `E1`: static inspection, grep, or configuration inspection.
|
|
149
|
+
- `E2`: local command or unit test; local behavior only.
|
|
150
|
+
- `E3`: integration test, real process, or real endpoint.
|
|
151
|
+
- `E4`: user acceptance, real business result, or production observation.
|
|
152
|
+
|
|
153
|
+
## Templates
|
|
154
|
+
|
|
155
|
+
- `templates/intake.md`: goal, constraints, assets, risk, and completion level.
|
|
156
|
+
- `templates/plan.md`: L0/L1/L2 plan and current focus.
|
|
157
|
+
- `templates/change-record.md`: one EXECUTE mutation and its reconciliation.
|
|
158
|
+
- `templates/round.md`: one planning iteration and its outcome.
|
|
159
|
+
- `templates/state.json`: canonical resumable state projection.
|
|
160
|
+
- `templates/gate-review.md` and `templates/gate-verdict.md`: independent review and decision records.
|
|
161
|
+
- `templates/release-check.md`: publication inventory, scans, remote verification, and limitations.
|
|
162
|
+
|
|
163
|
+
## Scope
|
|
164
|
+
|
|
165
|
+
This is an engineering planning skill. It focuses on correctness of planning, safe handling of user files and code changes, evidence quality, rollback, and real-environment reconciliation. It intentionally stays focused on project-level engineering work rather than organization-wide process design.
|
|
166
|
+
|
|
167
|
+
## Limitations
|
|
168
|
+
|
|
169
|
+
- The prose judgment layer stays non-mechanical by design; only the source-checkout DSH bundle adds a machine-checkable read-only runtime for context, evidence, progress, gates, release/install audits, recovery, and mutation preview.
|
|
170
|
+
- Inside the prose skill itself (without the bundle), evidence freshness and state reconciliation remain operator/agent responsibilities.
|
|
171
|
+
- The skill does not provide a sandbox; untrusted code must not be executed unless an actual isolated execution environment is already available.
|
|
172
|
+
- MCP candidates should be selected only after observing repeated real-project violations.
|
|
173
|
+
|
|
174
|
+
## Notes
|
|
175
|
+
|
|
176
|
+
The executable skill body (`SKILL.md`) is authored in Simplified Chinese; coding agents execute it correctly regardless of the conversation language. Templates are English and agent-facing. Documentation translations cover this README and the technical guide.
|
|
177
|
+
|
|
178
|
+
## Version
|
|
179
|
+
|
|
180
|
+
Current version: `v5.5.6`.
|
|
181
|
+
|
|
182
|
+
See [`docs/TECHNICAL_GUIDE.md`](docs/TECHNICAL_GUIDE.md) for the full operating model and [`docs/AUDIT_REPORT.md`](docs/AUDIT_REPORT.md) for the publication audit.
|
package/README.md
CHANGED
|
@@ -1,160 +1,182 @@
|
|
|
1
1
|
# Tech Lead Skill
|
|
2
2
|
|
|
3
|
-
English
|
|
3
|
+
[English](README.en.md) | 简体中文
|
|
4
4
|
|
|
5
|
-
|
|
5
|
+
一套以证据驱动的规划与交付技能，适用于软件、基础设施、研究、逆向工程与运维类工作。
|
|
6
6
|
|
|
7
|
-
##
|
|
7
|
+
## 为什么是它，而不是又一份提示词
|
|
8
8
|
|
|
9
|
-
|
|
9
|
+
大多数工程规范停留在「文字建议」层面——读完很认同，执行靠自觉。本项目把同一条判断层做成两种载体：纯文本规范人人可用；`dsh-themis` 进一步把关键判定变成**机器可校验的只读运行时**。
|
|
10
10
|
|
|
11
|
-
|
|
11
|
+
| 维度 | 普通 Skill / Prompt | tech-lead-skill | dsh-themis（DSH 插件） |
|
|
12
|
+
|---|---|---|---|
|
|
13
|
+
| 任务分级 | 模型自觉 | T0/T1/T2 硬规则 + 双向升降档 | 同左，机械可复核 |
|
|
14
|
+
| 输入治理 | 无 | 预算上限条款 | **fail-closed**：超限输入直接拒（INPUT_TOO_LARGE / SCAN_INCOMPLETE），绝不静默放行 |
|
|
15
|
+
| 状态校验 | 口头提醒 | schema 说明 | `state_validate` 全字段机检：枚举、done 锚点、证据溯源 E0-E4 |
|
|
16
|
+
| 泄漏审计 | 无 | 发布检查单 | `release_audit` 自动扫描绝对路径/token 形串/凭据赋值并带行号 |
|
|
17
|
+
| 结果格式 | 随缘 | 文本约定 | 22 工具统一 `protocolJson` 协商 + 稳定 v2 信封（findings/guidance/meta.complete） |
|
|
18
|
+
| 副作用 | — | 无 | **零副作用设计**：不写盘、不起子进程、不联网，只计算你传入的 JSON |
|
|
19
|
+
|
|
20
|
+
六条可以被验证的差异化优势：
|
|
21
|
+
|
|
22
|
+
1. **治理是判定的，不是恳求的。** 门禁前置校验、盲评触发条件、停滞断路器全部是可执行的硬规则——插件版能机械回答"这个 Gate 现在能不能过、缺什么锚点"。
|
|
23
|
+
2. **宁可显式失败，不给假安全感。** 审计超窗拒绝出结论；state→context v2 投影缺失 project/fingerprint/source 直接返回 NON_CONVERTIBLE_STATE，身份永不虚构——这些行为由 250 个测试钉死，不是口头承诺。
|
|
24
|
+
3. **确定性输出。** 同输入必同输出；guidance 的 actionId 确定性生成，可直接用作流程 key。
|
|
25
|
+
4. **平滑迁移。** 裸 JSON → v1 信封 → v2 信封由每个工具 schema 内声明的 `protocolJson` 协商：老调用零破坏，未知取值 fail-closed。
|
|
26
|
+
5. **实战出身。** 规则源自真实多项目的事故复盘——执行可靠性、配置部署陷阱、备份回滚纪律、对抗风控、多代理编排，不是从理论里推出来的模板。
|
|
27
|
+
6. **一键采用。** npm 单包版本化发布，一条命令装进任意 agent 环境；全文档中英双语。
|
|
28
|
+
|
|
29
|
+
## 解决什么问题
|
|
30
|
+
|
|
31
|
+
计划失败通常有两种方式：要么模糊到无法指导执行，要么过度细化而在第一次环境变化后就失真。本技能让计划在证据支撑之前保持粗粒度，然后依据观察到的现实持续修订。
|
|
32
|
+
|
|
33
|
+
控制回路：
|
|
12
34
|
|
|
13
35
|
```text
|
|
14
|
-
|
|
15
|
-
->
|
|
36
|
+
目标 -> 约束/假设 -> L0/L1/L2 计划 -> 执行 -> 观察证据
|
|
37
|
+
-> 修订 -> 选择 CONTINUE / PAUSE / SCOPE-DOWN / PIVOT / STOP
|
|
16
38
|
```
|
|
17
39
|
|
|
18
|
-
##
|
|
40
|
+
## 核心特性
|
|
19
41
|
|
|
20
|
-
- PLAN
|
|
21
|
-
-
|
|
22
|
-
-
|
|
23
|
-
-
|
|
24
|
-
-
|
|
25
|
-
-
|
|
26
|
-
-
|
|
27
|
-
-
|
|
28
|
-
- Git
|
|
29
|
-
-
|
|
30
|
-
-
|
|
42
|
+
- PLAN 与 EXECUTE 双模式，划清规划与副作用的边界。
|
|
43
|
+
- 渐进展开：L0 架构、L1 里程碑、L2 可执行焦点。
|
|
44
|
+
- 目标、指标、事实、假设、决策、风险、依赖与证据账本。
|
|
45
|
+
- 受保护资产处置：源码、用户数据、配置、秘密、运行态与生成物。
|
|
46
|
+
- 最小变更协议：`READ -> CLASSIFY -> PROTECT -> CHANGE -> VERIFY -> RECONCILE -> ROLLBACK/RECORD`。
|
|
47
|
+
- E0-E4 证据分级，从模型推断到真实用户结果。
|
|
48
|
+
- 失败重规划、停滞断路器、回滚纪律与真实状态对账。
|
|
49
|
+
- 面向高影响或不可逆变更的对抗式评审（可选）。
|
|
50
|
+
- 支持 Git 与非 Git 两种项目状态恢复方式。
|
|
51
|
+
- 面向公开文档与技能发布的收尾验证检查。
|
|
52
|
+
- 运行期纪律规则：外部依赖健康、静默失败类别、自动化护栏、幂等批处理与降级阶梯。
|
|
31
53
|
|
|
32
|
-
##
|
|
54
|
+
## 安装
|
|
33
55
|
|
|
34
|
-
|
|
56
|
+
先选版本：
|
|
35
57
|
|
|
36
|
-
-
|
|
37
|
-
-
|
|
58
|
+
- **纯文本规范，任意 agent 环境**（opencode / Claude Code / Codex）→ `tech-lead-skill`
|
|
59
|
+
- **带机器可校验只读工具的 DSH 插件** → `dsh-themis`
|
|
38
60
|
|
|
39
|
-
### npm
|
|
61
|
+
### npm 方式（GitHub 源，无需注册账号）
|
|
40
62
|
|
|
41
63
|
```bash
|
|
42
64
|
npm i -g github:240xu/tech-lead-skill
|
|
43
|
-
tech-lead-skill #
|
|
65
|
+
tech-lead-skill # 安装到 ~/.config/opencode/skills/tech-lead
|
|
44
66
|
```
|
|
45
67
|
|
|
46
|
-
|
|
68
|
+
一次性运行（不全局安装）：
|
|
47
69
|
|
|
48
70
|
```bash
|
|
49
71
|
npx github:240xu/tech-lead-skill
|
|
50
72
|
```
|
|
51
73
|
|
|
52
|
-
|
|
74
|
+
安装器幂等，重复执行会先把旧文件备份为 `*.bak-<时间戳>`；`--target <目录>` 可自定义目标；`--check` 校验已装副本与包的哈希/版本漂移；`--dry-run` 预览不写入；`--uninstall` 卸载时只删除清单受管文件，用户文件与 `*.bak-*` 备份全部保留。
|
|
53
75
|
|
|
54
|
-
`--check`
|
|
76
|
+
`--check` 退出码：`0` 正常，`1` 检出漂移，`2` 用法/拒绝错误。
|
|
55
77
|
|
|
56
|
-
### DeepSeek Harness
|
|
78
|
+
### DeepSeek Harness 插件（只读工具）
|
|
57
79
|
|
|
58
|
-
**
|
|
80
|
+
**DSH 用户看这里：** 已上架 DSH 插件市场（npm）——**单一自包含包**，任意 profile 一键安装：
|
|
59
81
|
|
|
60
|
-
```
|
|
82
|
+
```
|
|
61
83
|
dsh plugin --profile headless add dsh-themis
|
|
62
84
|
```
|
|
63
85
|
|
|
64
|
-
22
|
|
86
|
+
单包注册 22 个入口：21 个只读治理工具（原九项审计＋上下文校验、证据图/新鲜度分析、推进决策、关键路径/影响分析、续跑对账、Gate 计划/聚合/重开、变更预览）＋能力发现工具 `tech_lead_capabilities`（只列出本包实际注册的工具）。仅对调用方传入的 JSON 做计算——无文件写入、无子进程、无网络访问。旧拆分包（`dsh-tech-lead-{core,plugin,bundle}`）已弃用并指向本包。源码安装：`node scripts/build-market-package.mjs` 装配后 `dsh plugin add packages/dsh-themis`。根 npm 包只发布技能与安装器。
|
|
65
87
|
|
|
66
|
-
|
|
88
|
+
**两个产物，同一套规范：** `tech-lead-skill` 是保守、广泛兼容的纯文本规范——装进任何 agent 环境（opencode/Claude Code/Codex）即可，无其他依赖。**`dsh-themis` 是 DSH 插件专项优化版**：同样的判断层外加机器可校验的只读运行时——全部 22 工具在 schema 中声明 `protocolJson`（legacy 默认裸形态、显式请求 v1/v2 信封、未知取值 fail-closed）、state→context-v2 单向投影（身份与来源永不虚构）、strict/compat 输入兼容模式，以及稳定的 v2 信封（`findings`/`guidance`/`meta.complete`/`meta.outputProtocol`）。
|
|
67
89
|
|
|
68
90
|
```bash
|
|
69
91
|
dsh plugin --profile headless add /path/to/tech-lead-skill/packages/dsh-themis
|
|
70
|
-
dsh --profile headless --dump-config #
|
|
92
|
+
dsh --profile headless --dump-config # 确认 tech-lead-tools 行已注入
|
|
71
93
|
```
|
|
72
94
|
|
|
73
95
|
|
|
74
|
-
|
|
96
|
+
输出家族与限制：
|
|
75
97
|
|
|
76
|
-
-
|
|
77
|
-
-
|
|
98
|
+
- 十二个强化工具返回 `tech-lead.result.v1` 信封（判别字段：`meta.schema`）；原有九个工具为向后兼容保留裸领域形状。
|
|
99
|
+
- 渲染输出有上限：finding/error 数组每字段最多 500 条并附 `FINDINGS_TRUNCATED` 警告；超限调用方回显数组折叠为 `{truncated,total}` 摘要；超过 256KB 自动切换紧凑序列化。
|
|
78
100
|
|
|
79
|
-
|
|
101
|
+
架构与权限矩阵见 [spec](https://github.com/240xu/tech-lead-skill/blob/main/docs/superpowers/specs/2026-08-25-dsh-tech-lead-system.md)。
|
|
80
102
|
|
|
81
|
-
###
|
|
103
|
+
### 手动方式
|
|
82
104
|
|
|
83
|
-
|
|
105
|
+
将 `skill/SKILL.md` 和 `skill/templates/` 目录复制到你所用 OpenCode 兼容环境的技能目录：
|
|
84
106
|
|
|
85
107
|
```text
|
|
86
108
|
~/.config/opencode/skills/tech-lead/
|
|
87
109
|
```
|
|
88
110
|
|
|
89
|
-
|
|
111
|
+
该技能由"构建项目/搭系统/制定执行计划/部署/迁移/发布/恢复/重构/运维"等实质性任务触发，即使用户没有明确说“项目”或“Tech Lead”也应加载；也可以显式加载。
|
|
90
112
|
|
|
91
|
-
##
|
|
113
|
+
## 生命周期结果是有效分析
|
|
92
114
|
|
|
93
|
-
`
|
|
115
|
+
`tech_lead_progress_decide` 的 `PAUSE`/`PIVOT`/`SCOPE-DOWN`/`STOP` 以 **ok:true** 分析返回（`data.outcome`）；未通过的 Gate 同样 ok:true（`data.verdict`/`data.pass`）。`ok:false` 只用于非法输入、超预算载荷（`INPUT_TOO_LARGE`、`ITEM_LIMIT_EXCEEDED`）与不完整安全扫描（`SCAN_INCOMPLETE`）。决策类工具附带确定性的 `data.guidance.nextActions[]`：每条动作含原因码、finding 引用与 `doneWhen` 完成谓词。启发式建议仅在显式 `guidanceMode:"heuristic"` 下出现。
|
|
94
116
|
|
|
95
|
-
##
|
|
117
|
+
## 四工具起步环
|
|
96
118
|
|
|
97
|
-
1. `tech_lead_classify`
|
|
98
|
-
2. `tech_lead_context_validate`
|
|
99
|
-
3. `tech_lead_evidence_lint`
|
|
100
|
-
4. `tech_lead_progress_decide`
|
|
119
|
+
1. `tech_lead_classify` —— 把返回的 tier 写入快照 `current.tier`。
|
|
120
|
+
2. `tech_lead_context_validate` —— 校验完整内联快照（schema `tech-lead.context.v1`；范例 tests/fixtures/starter-context.v1.json）。
|
|
121
|
+
3. `tech_lead_evidence_lint` —— 把快照的 `evidence` 数组序列化为字符串传入；findings 仅作建议。
|
|
122
|
+
4. `tech_lead_progress_decide` —— 喂入同一快照；读 `data.outcome` 后按 `data.guidance` 行动。
|
|
101
123
|
|
|
102
|
-
##
|
|
124
|
+
## 工作模式
|
|
103
125
|
|
|
104
126
|
### PLAN
|
|
105
127
|
|
|
106
|
-
|
|
128
|
+
用于需求接入、目标定义、架构设计、任务分解、风险分析与验证设计。PLAN 不改文件、不执行有副作用的命令。
|
|
107
129
|
|
|
108
130
|
### EXECUTE
|
|
109
131
|
|
|
110
|
-
|
|
132
|
+
在 L2 范围明确后使用。只执行最小已批准变更，记录改动，验证行为，对账实际状态，并更新计划。
|
|
111
133
|
|
|
112
|
-
##
|
|
134
|
+
## 受保护资产
|
|
113
135
|
|
|
114
|
-
|
|
|
136
|
+
| 类别 | 默认处置 |
|
|
115
137
|
|---|---|
|
|
116
|
-
| `SOURCE` |
|
|
117
|
-
| `USER_DATA` |
|
|
118
|
-
| `CONFIG` |
|
|
119
|
-
| `SECRET` |
|
|
120
|
-
| `RUNTIME` |
|
|
121
|
-
| `GENERATED` |
|
|
138
|
+
| `SOURCE` | 可审查 diff、测试与恢复点 |
|
|
139
|
+
| `USER_DATA` | 默认只读；仅在目标明确且有可恢复副本时写入 |
|
|
140
|
+
| `CONFIG` | 先读现状，最小修改，重载后验证 |
|
|
141
|
+
| `SECRET` | 绝不进入计划、日志、普通备份或 diff |
|
|
142
|
+
| `RUNTIME` | 重启/杀进程/替换/迁移前必须先查实时状态 |
|
|
143
|
+
| `GENERATED` | 优先重新生成；不作为事实源 |
|
|
122
144
|
|
|
123
|
-
##
|
|
145
|
+
## 证据分级
|
|
124
146
|
|
|
125
|
-
- `E0
|
|
126
|
-
- `E1
|
|
127
|
-
- `E2
|
|
128
|
-
- `E3
|
|
129
|
-
- `E4
|
|
147
|
+
- `E0`：模型推断；仅可用于提出假设。
|
|
148
|
+
- `E1`：静态阅读、grep 或配置检查。
|
|
149
|
+
- `E2`：本地命令或单元测试；仅证明局部行为。
|
|
150
|
+
- `E3`：集成测试、真实进程或真实端点。
|
|
151
|
+
- `E4`：用户验收、真实业务结果或生产观察。
|
|
130
152
|
|
|
131
|
-
##
|
|
153
|
+
## 模板
|
|
132
154
|
|
|
133
|
-
- `templates/intake.md
|
|
134
|
-
- `templates/plan.md
|
|
135
|
-
- `templates/change-record.md
|
|
136
|
-
- `templates/round.md
|
|
137
|
-
- `templates/state.json
|
|
138
|
-
- `templates/gate-review.md`
|
|
139
|
-
- `templates/release-check.md
|
|
155
|
+
- `templates/intake.md`：目标、约束、资产、风险与完成层级。
|
|
156
|
+
- `templates/plan.md`：L0/L1/L2 计划与当前焦点。
|
|
157
|
+
- `templates/change-record.md`：单次 EXECUTE 变更及其对账。
|
|
158
|
+
- `templates/round.md`：单轮规划迭代及其结果。
|
|
159
|
+
- `templates/state.json`：可续跑的规范状态投影。
|
|
160
|
+
- `templates/gate-review.md` 与 `templates/gate-verdict.md`：独立评审与裁决记录。
|
|
161
|
+
- `templates/release-check.md`：发布清单、扫描、远程验证与限制记录。
|
|
140
162
|
|
|
141
|
-
##
|
|
163
|
+
## 范围
|
|
142
164
|
|
|
143
|
-
|
|
165
|
+
这是一套工程规划技能，聚焦规划正确性、用户文件与代码改动的安全处置、证据质量、回滚以及真实环境对账。它有意不涉及组织级流程设计或通用项目管理方法论。
|
|
144
166
|
|
|
145
|
-
##
|
|
167
|
+
## 局限
|
|
146
168
|
|
|
147
|
-
-
|
|
148
|
-
-
|
|
149
|
-
-
|
|
150
|
-
- MCP
|
|
169
|
+
- 文本判断层有意保持非机械化；只有源码安装的 DSH bundle 才提供上下文、证据、推进、Gate、发布/安装审计、恢复与变更预览的可机检只读运行时。
|
|
170
|
+
- 在纯文本技能形态下（无 bundle），证据时效与状态对账仍依赖执行环境与操作者。
|
|
171
|
+
- 本技能不提供沙箱；除非已有真实隔离执行环境，否则不得运行不可信代码。
|
|
172
|
+
- MCP 工具化候选应在观察到真实项目中反复违例之后再筛选。
|
|
151
173
|
|
|
152
|
-
##
|
|
174
|
+
## 说明
|
|
153
175
|
|
|
154
|
-
|
|
176
|
+
可执行的技能正文 `SKILL.md` 以简体中文编写；编码代理无论对话语言为何均可正确执行。模板为面向代理的英文文件。文档翻译覆盖本 README 与技术指南。
|
|
155
177
|
|
|
156
|
-
##
|
|
178
|
+
## 版本
|
|
157
179
|
|
|
158
|
-
|
|
180
|
+
当前版本：`v5.5.6`.
|
|
159
181
|
|
|
160
|
-
|
|
182
|
+
完整运行模型见[技术指南](./docs/TECHNICAL_GUIDE.zh-CN.md)，发布审计见 [docs/AUDIT_REPORT.md](docs/AUDIT_REPORT.md)。
|
package/docs/AUDIT_REPORT.md
CHANGED
|
@@ -17,7 +17,7 @@ The publication set contains only the following intended artifacts:
|
|
|
17
17
|
- `skill/templates/state.json`
|
|
18
18
|
- `skill/templates/gate-review.md`
|
|
19
19
|
- `skill/templates/gate-verdict.md`
|
|
20
|
-
- `README.md` (
|
|
20
|
+
- `README.md` (简体中文, GitHub 默认) / `README.en.md` (English)
|
|
21
21
|
- `docs/TECHNICAL_GUIDE.md` (English; Chinese README ships separately)
|
|
22
22
|
- `docs/AUDIT_REPORT.md`
|
|
23
23
|
- `LICENSE`
|
|
@@ -35,7 +35,7 @@ The source set was scanned for local absolute paths, passwords, API-key prefixes
|
|
|
35
35
|
### Structural scan
|
|
36
36
|
|
|
37
37
|
- YAML frontmatter has `name: tech-lead` and a trigger description.
|
|
38
|
-
- The skill version is `v5.5.
|
|
38
|
+
- The skill version is `v5.5.6`. R8.1 (2026-08-26) completes the R8 public contract: protocolJson declared on all 22 tool schemas; all nine legacy bare tools negotiate with bare defaults and fail-closed envelope passthrough; inputCompatibility=compat migrates unknown context keys instead of rejecting; context_validate dispatches canonical v2 documents to validateContextV2; v2 envelopes carry findings/guidance/meta.complete plus meta.outputProtocol stamped by negotiation. one-way state→context-v2 projection (identity options mandatory), result-protocol negotiation on every tool, and the v2 wire label for strengthened envelopes; reverse projection is deferred as non-lossless (evidence/2026-08-26-r8-projection.md).
|
|
39
39
|
- The eight templates referenced by the skill exist.
|
|
40
40
|
- The installer validates options, requires a marker before removing any target, and records the managed package/version after installation.
|
|
41
41
|
- The release-check template covers allowlist inventory, sensitive-content scanning, reference checks, scope checks, publication results, and remote verification.
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "tech-lead-skill",
|
|
3
|
-
"version": "5.5.
|
|
3
|
+
"version": "5.5.6",
|
|
4
4
|
"description": "Evidence-driven planning and safe change-control skill for OpenCode-compatible agents",
|
|
5
5
|
"license": "MIT",
|
|
6
6
|
"repository": {
|
|
@@ -20,7 +20,7 @@
|
|
|
20
20
|
"skill/",
|
|
21
21
|
"bin/",
|
|
22
22
|
"README.md",
|
|
23
|
-
"README.
|
|
23
|
+
"README.en.md",
|
|
24
24
|
"docs/TECHNICAL_GUIDE.md",
|
|
25
25
|
"docs/AUDIT_REPORT.md",
|
|
26
26
|
"LICENSE"
|
package/skill/SKILL.md
CHANGED
|
@@ -3,7 +3,7 @@ name: tech-lead
|
|
|
3
3
|
description: 用于软件、基础设施、研究和运维中的多步、跨边界或有状态工程交付：项目规划、架构选择、跨模块实现、部署、迁移、发布、恢复、重构、长期运行与跨会话续跑。涉及不可逆/生产变更、回滚验证、真实环境对账、发布安全或项目级交付治理时优先加载，即使用户没有说“项目”或“Tech Lead”。不要仅因一次性小修复、单文件编辑、普通解释、单次测试或只读代码审查而自动加载，除非用户明确要求项目级计划或本规范。
|
|
4
4
|
---
|
|
5
5
|
|
|
6
|
-
# Tech Lead · 自主演进规划与交付规范 v5.5.
|
|
6
|
+
# Tech Lead · 自主演进规划与交付规范 v5.5.6
|
|
7
7
|
|
|
8
8
|
> 双层架构声明：本文件是**判断层**（自由裁量域）。其中可机械判定的不变量（Gate 前置校验、
|
|
9
9
|
> 证据锚点存在性、分级判定器、状态机转换有效性）为未来 **MCP 工具化候选**——
|
package/README.zh-CN.md
DELETED
|
@@ -1,160 +0,0 @@
|
|
|
1
|
-
# Tech Lead Skill
|
|
2
|
-
|
|
3
|
-
[English](README.md) | 简体中文
|
|
4
|
-
|
|
5
|
-
一套以证据驱动的规划与交付技能，适用于软件、基础设施、研究、逆向工程与运维类工作。
|
|
6
|
-
|
|
7
|
-
## 解决什么问题
|
|
8
|
-
|
|
9
|
-
计划失败通常有两种方式：要么模糊到无法指导执行，要么过度细化而在第一次环境变化后就失真。本技能让计划在证据支撑之前保持粗粒度，然后依据观察到的现实持续修订。
|
|
10
|
-
|
|
11
|
-
控制回路：
|
|
12
|
-
|
|
13
|
-
```text
|
|
14
|
-
目标 -> 约束/假设 -> L0/L1/L2 计划 -> 执行 -> 观察证据
|
|
15
|
-
-> 修订 -> 选择 CONTINUE / PAUSE / SCOPE-DOWN / PIVOT / STOP
|
|
16
|
-
```
|
|
17
|
-
|
|
18
|
-
## 核心特性
|
|
19
|
-
|
|
20
|
-
- PLAN 与 EXECUTE 双模式，划清规划与副作用的边界。
|
|
21
|
-
- 渐进展开：L0 架构、L1 里程碑、L2 可执行焦点。
|
|
22
|
-
- 目标、指标、事实、假设、决策、风险、依赖与证据账本。
|
|
23
|
-
- 受保护资产处置：源码、用户数据、配置、秘密、运行态与生成物。
|
|
24
|
-
- 最小变更协议：`READ -> CLASSIFY -> PROTECT -> CHANGE -> VERIFY -> RECONCILE -> ROLLBACK/RECORD`。
|
|
25
|
-
- E0-E4 证据分级，从模型推断到真实用户结果。
|
|
26
|
-
- 失败重规划、停滞断路器、回滚纪律与真实状态对账。
|
|
27
|
-
- 面向高影响或不可逆变更的对抗式评审（可选）。
|
|
28
|
-
- 支持 Git 与非 Git 两种项目状态恢复方式。
|
|
29
|
-
- 面向公开文档与技能发布的收尾验证检查。
|
|
30
|
-
- 运行期纪律规则：外部依赖健康、静默失败类别、自动化护栏、幂等批处理与降级阶梯。
|
|
31
|
-
|
|
32
|
-
## 安装
|
|
33
|
-
|
|
34
|
-
先选版本：
|
|
35
|
-
|
|
36
|
-
- **纯文本规范，任意 agent 环境**（opencode / Claude Code / Codex）→ `tech-lead-skill`
|
|
37
|
-
- **带机器可校验只读工具的 DSH 插件** → `dsh-themis`
|
|
38
|
-
|
|
39
|
-
### npm 方式（GitHub 源，无需注册账号）
|
|
40
|
-
|
|
41
|
-
```bash
|
|
42
|
-
npm i -g github:240xu/tech-lead-skill
|
|
43
|
-
tech-lead-skill # 安装到 ~/.config/opencode/skills/tech-lead
|
|
44
|
-
```
|
|
45
|
-
|
|
46
|
-
一次性运行（不全局安装）：
|
|
47
|
-
|
|
48
|
-
```bash
|
|
49
|
-
npx github:240xu/tech-lead-skill
|
|
50
|
-
```
|
|
51
|
-
|
|
52
|
-
安装器幂等，重复执行会先把旧文件备份为 `*.bak-<时间戳>`；`--target <目录>` 可自定义目标；`--check` 校验已装副本与包的哈希/版本漂移；`--dry-run` 预览不写入；`--uninstall` 卸载时只删除清单受管文件，用户文件与 `*.bak-*` 备份全部保留。
|
|
53
|
-
|
|
54
|
-
`--check` 退出码：`0` 正常，`1` 检出漂移，`2` 用法/拒绝错误。
|
|
55
|
-
|
|
56
|
-
### DeepSeek Harness 插件（只读工具）
|
|
57
|
-
|
|
58
|
-
**DSH 用户看这里：** 已上架 DSH 插件市场（npm）——**单一自包含包**，任意 profile 一键安装：
|
|
59
|
-
|
|
60
|
-
```
|
|
61
|
-
dsh plugin --profile headless add dsh-themis
|
|
62
|
-
```
|
|
63
|
-
|
|
64
|
-
单包注册 22 个入口：21 个只读治理工具（原九项审计＋上下文校验、证据图/新鲜度分析、推进决策、关键路径/影响分析、续跑对账、Gate 计划/聚合/重开、变更预览）＋能力发现工具 `tech_lead_capabilities`（只列出本包实际注册的工具）。仅对调用方传入的 JSON 做计算——无文件写入、无子进程、无网络访问。旧拆分包（`dsh-tech-lead-{core,plugin,bundle}`）已弃用并指向本包。源码安装：`node scripts/build-market-package.mjs` 装配后 `dsh plugin add packages/dsh-themis`。根 npm 包只发布技能与安装器。
|
|
65
|
-
|
|
66
|
-
**两个产物，同一套规范：** `tech-lead-skill` 是保守、广泛兼容的纯文本规范——装进任何 agent 环境（opencode/Claude Code/Codex）即可，无其他依赖。**`dsh-themis` 是 DSH 插件专项优化版**：同样的判断层外加机器可校验的只读运行时——全部 22 工具在 schema 中声明 `protocolJson`（legacy 默认裸形态、显式请求 v1/v2 信封、未知取值 fail-closed）、state→context-v2 单向投影（身份与来源永不虚构）、strict/compat 输入兼容模式，以及稳定的 v2 信封（`findings`/`guidance`/`meta.complete`/`meta.outputProtocol`）。
|
|
67
|
-
|
|
68
|
-
```bash
|
|
69
|
-
dsh plugin --profile headless add /path/to/tech-lead-skill/packages/dsh-themis
|
|
70
|
-
dsh --profile headless --dump-config # 确认 tech-lead-tools 行已注入
|
|
71
|
-
```
|
|
72
|
-
|
|
73
|
-
|
|
74
|
-
输出家族与限制：
|
|
75
|
-
|
|
76
|
-
- 十二个强化工具返回 `tech-lead.result.v1` 信封（判别字段：`meta.schema`）；原有九个工具为向后兼容保留裸领域形状。
|
|
77
|
-
- 渲染输出有上限：finding/error 数组每字段最多 500 条并附 `FINDINGS_TRUNCATED` 警告；超限调用方回显数组折叠为 `{truncated,total}` 摘要；超过 256KB 自动切换紧凑序列化。
|
|
78
|
-
|
|
79
|
-
架构与权限矩阵见 [spec](https://github.com/240xu/tech-lead-skill/blob/main/docs/superpowers/specs/2026-08-25-dsh-tech-lead-system.md)。
|
|
80
|
-
|
|
81
|
-
### 手动方式
|
|
82
|
-
|
|
83
|
-
将 `skill/SKILL.md` 和 `skill/templates/` 目录复制到你所用 OpenCode 兼容环境的技能目录：
|
|
84
|
-
|
|
85
|
-
```text
|
|
86
|
-
~/.config/opencode/skills/tech-lead/
|
|
87
|
-
```
|
|
88
|
-
|
|
89
|
-
该技能由"构建项目/搭系统/制定执行计划/部署/迁移/发布/恢复/重构/运维"等实质性任务触发，即使用户没有明确说“项目”或“Tech Lead”也应加载；也可以显式加载。
|
|
90
|
-
|
|
91
|
-
## 生命周期结果是有效分析
|
|
92
|
-
|
|
93
|
-
`tech_lead_progress_decide` 的 `PAUSE`/`PIVOT`/`SCOPE-DOWN`/`STOP` 以 **ok:true** 分析返回（`data.outcome`）；未通过的 Gate 同样 ok:true（`data.verdict`/`data.pass`）。`ok:false` 只用于非法输入、超预算载荷（`INPUT_TOO_LARGE`、`ITEM_LIMIT_EXCEEDED`）与不完整安全扫描（`SCAN_INCOMPLETE`）。决策类工具附带确定性的 `data.guidance.nextActions[]`：每条动作含原因码、finding 引用与 `doneWhen` 完成谓词。启发式建议仅在显式 `guidanceMode:"heuristic"` 下出现。
|
|
94
|
-
|
|
95
|
-
## 四工具起步环
|
|
96
|
-
|
|
97
|
-
1. `tech_lead_classify` —— 把返回的 tier 写入快照 `current.tier`。
|
|
98
|
-
2. `tech_lead_context_validate` —— 校验完整内联快照（schema `tech-lead.context.v1`；范例 tests/fixtures/starter-context.v1.json）。
|
|
99
|
-
3. `tech_lead_evidence_lint` —— 把快照的 `evidence` 数组序列化为字符串传入；findings 仅作建议。
|
|
100
|
-
4. `tech_lead_progress_decide` —— 喂入同一快照；读 `data.outcome` 后按 `data.guidance` 行动。
|
|
101
|
-
|
|
102
|
-
## 工作模式
|
|
103
|
-
|
|
104
|
-
### PLAN
|
|
105
|
-
|
|
106
|
-
用于需求接入、目标定义、架构设计、任务分解、风险分析与验证设计。PLAN 不改文件、不执行有副作用的命令。
|
|
107
|
-
|
|
108
|
-
### EXECUTE
|
|
109
|
-
|
|
110
|
-
在 L2 范围明确后使用。只执行最小已批准变更，记录改动，验证行为，对账实际状态，并更新计划。
|
|
111
|
-
|
|
112
|
-
## 受保护资产
|
|
113
|
-
|
|
114
|
-
| 类别 | 默认处置 |
|
|
115
|
-
|---|---|
|
|
116
|
-
| `SOURCE` | 可审查 diff、测试与恢复点 |
|
|
117
|
-
| `USER_DATA` | 默认只读；仅在目标明确且有可恢复副本时写入 |
|
|
118
|
-
| `CONFIG` | 先读现状，最小修改，重载后验证 |
|
|
119
|
-
| `SECRET` | 绝不进入计划、日志、普通备份或 diff |
|
|
120
|
-
| `RUNTIME` | 重启/杀进程/替换/迁移前必须先查实时状态 |
|
|
121
|
-
| `GENERATED` | 优先重新生成；不作为事实源 |
|
|
122
|
-
|
|
123
|
-
## 证据分级
|
|
124
|
-
|
|
125
|
-
- `E0`：模型推断；仅可用于提出假设。
|
|
126
|
-
- `E1`：静态阅读、grep 或配置检查。
|
|
127
|
-
- `E2`：本地命令或单元测试；仅证明局部行为。
|
|
128
|
-
- `E3`：集成测试、真实进程或真实端点。
|
|
129
|
-
- `E4`：用户验收、真实业务结果或生产观察。
|
|
130
|
-
|
|
131
|
-
## 模板
|
|
132
|
-
|
|
133
|
-
- `templates/intake.md`：目标、约束、资产、风险与完成层级。
|
|
134
|
-
- `templates/plan.md`：L0/L1/L2 计划与当前焦点。
|
|
135
|
-
- `templates/change-record.md`：单次 EXECUTE 变更及其对账。
|
|
136
|
-
- `templates/round.md`：单轮规划迭代及其结果。
|
|
137
|
-
- `templates/state.json`：可续跑的规范状态投影。
|
|
138
|
-
- `templates/gate-review.md` 与 `templates/gate-verdict.md`：独立评审与裁决记录。
|
|
139
|
-
- `templates/release-check.md`：发布清单、扫描、远程验证与限制记录。
|
|
140
|
-
|
|
141
|
-
## 范围
|
|
142
|
-
|
|
143
|
-
这是一套工程规划技能，聚焦规划正确性、用户文件与代码改动的安全处置、证据质量、回滚以及真实环境对账。它有意不涉及组织级流程设计或通用项目管理方法论。
|
|
144
|
-
|
|
145
|
-
## 局限
|
|
146
|
-
|
|
147
|
-
- 文本判断层有意保持非机械化；只有源码安装的 DSH bundle 才提供上下文、证据、推进、Gate、发布/安装审计、恢复与变更预览的可机检只读运行时。
|
|
148
|
-
- 在纯文本技能形态下（无 bundle），证据时效与状态对账仍依赖执行环境与操作者。
|
|
149
|
-
- 本技能不提供沙箱；除非已有真实隔离执行环境，否则不得运行不可信代码。
|
|
150
|
-
- MCP 工具化候选应在观察到真实项目中反复违例之后再筛选。
|
|
151
|
-
|
|
152
|
-
## 说明
|
|
153
|
-
|
|
154
|
-
可执行的技能正文 `SKILL.md` 以简体中文编写；编码代理无论对话语言为何均可正确执行。模板为面向代理的英文文件。文档翻译覆盖本 README 与技术指南。
|
|
155
|
-
|
|
156
|
-
## 版本
|
|
157
|
-
|
|
158
|
-
当前版本：`v5.5.5`.
|
|
159
|
-
|
|
160
|
-
完整运行模型见[技术指南](./docs/TECHNICAL_GUIDE.zh-CN.md)，发布审计见 [docs/AUDIT_REPORT.md](docs/AUDIT_REPORT.md)。
|