القائمة
Injected at session start — establishes samsara axiom, agent constraints, and skill discovery for the session
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
| name | samsara-bootstrap |
| description | Injected at session start — establishes samsara axiom, agent constraints, and skill discovery for the session |
Toward death, through verification.
存在即責任，無責任即無存在。
任何存在於系統中的東西——function、module、service、agent 的一個決策——都必須能回答：「如果你消失了，什麼東西會感到痛？」回答不了的，不應該存在。
Live codebase artifacts are authoritative for current platform behavior: source code, tests, configs, skills, agents, references, templates, and workflow artifacts that the platform reads or validates.
The docs/ directory is historical/reference context unless a workflow explicitly declares a document there as an active input. Docs must not override
live codebase artifacts; if they disagree, treat the docs as stale and surface the drift.
在執行任何開發任務之前，必須先完成以下四個問題。這不是建議，是前置條件。
before invoking samsara:research for new feature work, ask the user to choose
the session-level execution mode:
human-in-the-loop — Default mode and default selection. Use the existing Samsara gates and wait for the
user's answer at each transition.auto — Run the same workflow, but route each later workflow question or
confirmation through samsara:auto-gatekeeper and append the decision to
changes/<feature>/auto-decisions.md.
Source dispatch target: subagent_type: "samsara:auto-gatekeeper".Use this prompt before research:
Execution mode? Choose
human-in-the-looporauto.
Record the selected mode as an explicit session context line before research:
Execution mode: human-in-the-loop
or:
Execution mode: auto
Later skills read this explicit Execution mode: line as the active execution
mode. If no mode is selected, record Execution mode: human-in-the-loop;
unknown mode must not silently become auto.
Persistent config, including samsara_config.yaml, is out of scope for the
first auto-mode implementation. Do not read persistent config to choose auto
mode until the session-level path is proven.
When the user describes work, you MUST invoke the matching samsara skill using the Skill tool BEFORE any response. This is not optional. Do not answer, clarify, or explore code before invoking.
If you think there is even a 1% chance a samsara skill applies, invoke it.
digraph samsara_routing {
rankdir=TB;
node [shape=box];
bootstrap [label="samsara-bootstrap\n(session start 注入)" shape=doublecircle];
input [label="使用者描述問題/任務"];
classify [label="Agent 分類" shape=diamond];
fast_track [label="samsara:fast-track\n(簡化流程)"];
research [label="samsara:research"];
pre_thinking [label="samsara:pre-thinking\n(assumption alignment)"];
planning [label="samsara:planning"];
implement [label="samsara:implement\n(含 Level 1 task iteration)"];
iteration [label="samsara:iteration\n(Level 2 feature iteration)"];
validate [label="samsara:validate-and-ship\n(含 Step 0 security gate)"];
debugging [label="samsara:debugging\n(四階段陰面 debugging)"];
fix_size [label="Fix 規模？" shape=diamond];
done [label="完成" shape=doublecircle];
bootstrap -> input;
input -> classify;
classify -> fast_track [label="低風險小改動\n+ 使用者確認"];
classify -> research [label="新功能/新需求"];
classify -> debugging [label="production failure"];
research -> pre_thinking [label="human gate"];
pre_thinking -> planning [label="human gate"];
planning -> implement [label="human gate"];
implement -> iteration [label="human gate\n(iterate)"];
implement -> validate [label="human gate\n(skip iteration)"];
iteration -> validate [label="iteration done\nor forced stop"];
validate -> done [label="human choose"];
fast_track -> done;
debugging -> fix_size;
fix_size -> fast_track [label="小 fix"];
fix_size -> implement [label="大 fix"];
}
Default rule: When intent is unclear, invoke samsara:research. Research can survive being invoked unnecessarily; shipping without research cannot.
Entry Skills（入口 — 根據意圖分流）：
Chain Skills（鏈式 — 由前一階段觸發，不直接 invoke）：
Utility Skills（工具性 — 按需使用）：
Use when entering a new project for the first time, or when the codebase has changed significantly — generates a yin-side codebase map with structural analysis and silent failure surface assessment
Use when the user describes a small, low-risk change — bug fix with known cause, config change, dependency update, or small refactor under 100 lines
Use when a plan with tasks exists and you need to execute implementation — requires index.yaml and tasks/ directory
Use when implement is complete and feature-level scar items need review — aggregates remaining scars across tasks, triages cross-task patterns, and fixes system-level rot before validate-and-ship
Use when research/kickoff is complete and you need to create an implementation plan with specs, tasks, and acceptance criteria
Use when research output is complete and planning would otherwise need to assume design, task-shape, structural boundaries, or evaluation details before implementation planning