Skill Generator (enterprise doc → installed skill)

Instructions here, reference/, and scripts/ (package scripts/skill_gen/). You obtain SOP plain text, run full structured extraction exactly as in reference/sop-structure-pipeline.md, author the package under get_agent_workspace_dir() / "skills-draft" (per reference/generator-worker-spec.md), then immediately promote it so the skill is installed and loadable (see reference/operator-playbook.md canonical flow).

What you do (high level)

1. Plain text
   • Mandatory sequence when the user wanted a document but you still have no SOP plain text (wrong path, file not found, fetch failed, empty body, permission error, etc.):
     1. Recovery before fallback — Do not call build_intent_fallback_sop, build_fallback_sop_structure, or generate the installable skill from intent-only fallback in the same assistant turn as that failure. A missing file is only a failed first ingest; it is not “no document” yet. Send exactly one recovery message first (single reply): offer a corrected absolute path, URL, or paste—same menu and tone as reference/sop-recovery-and-interview-fallback.md Part 1—and wait for the user’s next message (or treat recovery as already done if one recovery was sent earlier in the thread and the user declined, ignored, or supplied material afterward).
     2. Fallback only after recovery — Only when recovery is complete (user gave path/URL/paste you can use, or declined/ignored after the one ask) and there is still no parseable SOP body, proceed to build_intent_fallback_sop / build_fallback_sop_structure below. Do not skip step (1) because you “already know” the path failed.
   • Local file → python3 <ABS>/scripts/skill_generator_cli.py sop-text --sop-file <abs-path> [--out-text <path>] or read the file in-tool; optional --print-raw-chars for length.
   • HTTP(S) / WeChat → url-fetch --url '…' [--out-json <path>] then use page text as the SOP body for extraction.
   • Pasted only — use the same string as input to parse_sop_raw_text (no skipping structured extraction).
   • No document after recovery — you must build the canonical object in code. Preferred (one call): await skill_gen.sop_fallback.build_intent_fallback_sop(user_intent=…, skill_name_hint=…, invoke_llm_json=…) — default: pass invoke_llm_json whenever the runtime can invoke the model: deterministic skeleton then internal LLM refine (enrich_fallback_sop_with_llm). Only if the system cannot call a model for this step, pass invoke_llm_json=None (skeleton only) or use skill_gen.sop_fallback.build_fallback_sop_structure on sync-only paths (same shape; no enrich). Same SOPStructure fields as extraction (title, purpose, scope, roles, steps, knowledge_items, sections, decision points, exceptions, references, plus synthesized raw_text). Do not call parse_sop_raw_text on empty text; do not hand-write a “representative” or illustrative SOP in prose instead of running the fallback—downstream drafting depends on the real SOPStructure and extraction_meta. Treat extraction_meta["fallback_sop"] as “template only” in the generated skill body. Still no real SOP after LLM polish: keep fallback_sop semantics and do not treat output as audited policy.
2. Structured SOP (required) → Either skill_gen.sop_parser.parse_sop_file / parse_sop_raw_text with invoke_llm_json when you have SOP plain text, or when you do not: await skill_gen.sop_fallback.build_intent_fallback_sop(..., invoke_llm_json=…) with invoke_llm_json whenever the runtime can invoke the model (default: skeleton + internal enrich); invoke_llm_json=None or build_fallback_sop_structure only when the system cannot (single path; sop-structure-pipeline.md).
3. Draft package → under get_agent_workspace_dir() / "skills-draft" / <skill_name> (when the host exposes that helper) or the equivalent agent-workspace path from the system prompt: SKILL.md (YAML frontmatter first) plus optional reference/, per generator-worker-spec.md. Create skills-draft next to the runtime skills/ folder if it does not exist.
   • Intent-only snapshot (required when fallback) — If extraction_meta["fallback_sop"] is true (intent-only / template path), you must add reference/intent-sop-snapshot.md in that draft folder. Start with a short warning block (this is not an authoritative enterprise SOP; synthesized from stated intent and optional LLM refine; provisional until a real document is attached). Then include the verbatim SOPStructure.raw_text from the structured object you used for drafting so reviewers can see exactly what the fallback produced.
4. Install in the same workflow → Call skills.import_local with path = absolute path to that draft directory (folder containing SKILL.md). Use force: true if get_agent_skills_dir() / <skill_name> already exists and should be replaced. Do not stop after step 3 and ask the user to import manually; promotion is part of this skill’s default completion.

After step 4, the new skill lives under get_agent_skills_dir() / <skill_name> (same directory tree the runtime loads user-installed skills from) and is available like any other installed skill.

Do not silently shrink generator-worker-spec when writing SKILL.md.

Invoking skill_generator_cli.py safely

• Always use the absolute path to scripts/skill_generator_cli.py.
• httpx / beautifulsoup4 for url-fetch. Rich file formats need openjiuwen AutoFileParser when installed; otherwise use .md / .txt SOPs.

Path tables and skills.import_local: reference/operator-playbook.md.

Normative SKILL.md format (YAML frontmatter first)

The generated skill’s SKILL.md must start with --- YAML frontmatter, then ---, then the body. Optional: scripts/skill_gen/validator.py on the draft path before import.

When to open reference files

Human approval gates

Do not modify this meta-skill’s SKILL.md, reference/sop-structure-pipeline.md, reference/generator-worker-spec.md, reference/operator-playbook.md, or other bundled reference/*.md while generating a user skill. Edit only the target skill under the workspace-relative skills-draft/<name>/ staging tree or the installed skills/<name>/ tree (resolve with get_agent_workspace_dir() / get_agent_skills_dir() when available).

Communicating with the user

• Prefer plain language for skill names and outcomes.
• Infer skill_name (kebab-case) from title or URL when the user did not specify.