AgentSkillsCN

artifacts

为用户的“任务请求”创建并管理仓库本地工作成果，生成 `.artifacts/artifacts/{request_id}/`，包含 `plan.md`、`todo.md` 和 `walkthrough.md`，然后将完成的工作索引到存储在 `.artifacts/chroma_db/` 中的 ChromaDB，以便 7 天内召回。当用户想要 (1) 跟踪/确认执行前的工作流程，(2) 最小化最终答案并链接到 walkthrough，或 (3) 问“我们之前用 artifacts 技能做了什么？”时使用。

SKILL.md
--- frontmatter
name: artifacts
description: Create and manage repo-local work artifacts for a user “task request” by generating `.artifacts/artifacts/{request_id}/` with `plan.md`, `todo.md`, and `walkthrough.md`, then indexing completed work into ChromaDB stored at `.artifacts/chroma_db/` for 7-day recall. Use when the user wants (1) a tracked/confirm-before-execute workflow, (2) a minimal final answer that links to a walkthrough, or (3) to ask “what did we do before with the artifacts skill?” and you should answer by querying the ChromaDB artifacts index.

Artifacts

Workflow

Goal: leave a durable, linkable trace of work in .artifacts/artifacts/{request_id}/, and answer future questions by querying the .artifacts/ ChromaDB index (retained for 7 days).

0) Bootstrap the required environment under .artifacts/ (always)

  1. •Create .artifacts/.venv/ and install chromadb there:
    • •python3 "$CODEX_HOME/skills/artifacts/scripts/artifacts_bootstrap_env.py"
  2. •After bootstrap, run all other commands using .artifacts/.venv/bin/python (required).

1) Initialize artifacts (always)

  1. •Run the initializer script (from repo root or anywhere inside the repo):
    • •.artifacts/.venv/bin/python "$CODEX_HOME/skills/artifacts/scripts/artifacts_init.py" --request "<user request>"
  2. •The script prints paths for:
    • •.artifacts/artifacts/{request_id}/plan.md
    • •.artifacts/artifacts/{request_id}/todo.md
    • •.artifacts/artifacts/{request_id}/walkthrough.md
  3. •Reply to the user with:
    • •The plan.md + todo.md paths
    • •A single confirmation question: “이대로 진행할까?”
  4. •Do not start implementation until the user explicitly confirms.

2) Execute work (after user confirmation)

  1. •Update .artifacts/artifacts/{request_id}/todo.md by checking items as they complete.
  2. •Do the requested work in the repo (code changes / analysis / commands).
  3. •Write .artifacts/artifacts/{request_id}/walkthrough.md as the detailed record (plan snapshot, steps, commands, files, validation). Ensure it is sufficient for a reviewer to reconstruct the work without the chat.

3) Index + retention (end of task)

  1. •Run indexing (this also prunes anything older than 7 days, both index + folders):
    • •.artifacts/.venv/bin/python "$CODEX_HOME/skills/artifacts/scripts/artifacts_index.py"
  2. •The index is stored at .artifacts/chroma_db/ (repo-local, derived).
  3. •If .artifacts/chroma_db/ is missing or chromadb is not installed, initialize by running the indexer once with a Python environment located under .artifacts/ (required).

4) Answering “what did we do before?” (query mode)

  1. •Query the artifacts index:
    • •.artifacts/.venv/bin/python "$CODEX_HOME/skills/artifacts/scripts/artifacts_query.py" "<question>"
  2. •Use the returned results as the basis for your answer.
  3. •In your reply, include the todo.md + walkthrough.md paths for the most relevant result(s).

Output rules (important)

When responding after work is done, keep the answer minimal:

  1. •Include clickable paths to:
    • •.artifacts/artifacts/{request_id}/todo.md
    • •.artifacts/artifacts/{request_id}/walkthrough.md
  2. •Keep prose to 1–3 short sentences: what instruction you received + what you did.
  3. •Do not paste the walkthrough into the chat; the walkthrough is the source of detail.

Reference

  • •File format + retention details: references/artifacts_format.md