By name: meta-skill-creator description: "Use this meta-skill instead of answering directly only when the current user explicitly asks to create, compose, synthesize, or propose a new meta-skill that orchestrates multiple existing skills. It uses multi-skill orchestration for intent clarification, optional history mining, trigger-collision checks, linting, smoke/runtime gates, preview, and optional proposal persistence. Do not use it for creating a normal standalone skill, asking how meta-skills work, analyzing pasted skill lists, or discussing existing meta-skills." kind: meta meta_priority: 90 always: false final_text_mode: "step:final_response" triggers:
- "新增 meta 技能"
- "组合现有 skill 成 meta-skill"
- "create a meta-skill"
- "new meta-skill"
- "orchestrates existing skills"
- "compose existing skills"
- "synthesize meta-skill"
- "compose meta-skill"
provenance:
origin: opensquilla-original
license: Apache-2.0
composition:
steps:
id: clarify_intent kind: llm_chat with: system: | You are the intent gate for meta-skill-creator. Do not inspect workspace files, history, memory, or external sources. Decide only from the explicit user request and activation context. task: | Clarify whether the user wants a meta-skill, not a normal standalone skill. If the request is generic skill creation, return ROUTE: normal-skill. If it requires orchestrating multiple existing skills, return ROUTE: meta-skill. Also summarize desired inputs, outputs, trigger phrases, and whether a human preference branch is needed. Set NEEDS_CLARIFICATION: yes only when the workflow goal, output shape, trigger boundary, or human preference branch is genuinely ambiguous and the request is an interactive user request. For unattended auto-propose, dream, or cron activation, set NEEDS_CLARIFICATION: no and continue from available context. User request: {{ inputs.user_message | xml_escape | truncate(1200) }} Outer system / activation context: {{ inputs.system_prompt | default("") | xml_escape | truncate(1200) }} Return: ROUTE: <normal-skill|meta-skill> WORKFLOW_GOAL:
OUTPUT_SHAPE: TRIGGERS: HUMAN_PREFERENCE_BRANCH: <yes|no|unclear> NEEDS_CLARIFICATION: <yes|no> MISSING_FIELDS: - <workflow_goal|output_shape|trigger_boundary|human_preference_branch|none> CLARIFY_REASON: <one concise reason, or none> id: creator_clarify kind: user_input depends_on: [clarify_intent] when: "'route: meta-skill' in (outputs.clarify_intent | lower) and 'needs_clarification: yes' in (outputs.clarify_intent | lower)" clarify: mode: form intro: | 新 meta-skill 的边界还不够明确。请补齐目标和输出形态，避免生成过宽的触发词。 nl_extract: true fields: - name: workflow_goal type: string required: true prompt: "工作流目标 / Workflow goal" max_chars: 300 - name: output_shape type: string required: true prompt: "最终输出形态 / Output shape" max_chars: 200 - name: trigger_boundary type: string prompt: "触发边界或不要覆盖的场景 / Trigger boundary" max_chars: 300 - name: human_preference_branch type: bool default: false prompt: "是否需要运行中让用户选择偏好 / Need human preference branch?" cancel_keywords: ["算了", "取消", "cancel", "stop", "abort"] timeout_hours: 24
id: normal_skill_exit kind: tool_call depends_on: [clarify_intent] when: "'route: normal-skill' in (outputs.clarify_intent | lower)" tool: emit_text tool_args: text: | This request was classified as a normal standalone skill request, not a meta-skill composition request. The meta-skill creator stopped before proposal assembly or persistence.
id: creator_mode kind: llm_classify depends_on: [clarify_intent, creator_clarify] when: "'route: meta-skill' in (outputs.clarify_intent | lower)" output_choices:
PREVIEW_ONLY
PERSISTED_PROPOSAL
FULL_GATED with: text: | Classify how far the creator workflow should go.
User request: {{ inputs.user_message | xml_escape | truncate(1200) }}
Outer system / activation context: {{ inputs.system_prompt | default("") | xml_escape | truncate(1200) }}
Clarified intent: {{ outputs.clarify_intent | truncate(1200) }}
Clarification answers (may be empty when not needed): {{ inputs.get('collected', {}).get('creator_clarify', {}) | tojson }}
Decision rules:
- PREVIEW_ONLY: user asks for an example, template, plan, draft, or wants to inspect before writing/persisting anything.
- PERSISTED_PROPOSAL: user asks to create/save/write/propose a meta-skill but does not ask for exhaustive smoke testing.
- FULL_GATED: user asks for a production-ready, accepted, tested, validated, or fully gated meta-skill.
- FULL_GATED: unattended auto-propose, dream, or cron activation requires preserving all creator gates before any auto-enable decision.
id: harvest kind: skill_exec skill: history-explorer depends_on: [clarify_intent, creator_clarify] when: "'route: meta-skill' in (outputs.clarify_intent | lower) and 'Unattended meta-skill auto-propose run' in inputs.get('system_prompt', '')" on_failure: harvest_empty with: query: | Co-occurring skill chains and meta-skill usage for: {{ outputs.clarify_intent | truncate(1000) }} Clarification answers: {{ inputs.get('collected', {}).get('creator_clarify', {}) | tojson }} window_days: 30 include: [co_occurrences, meta_usage, router_fixtures]
id: harvest_empty kind: tool_call tool: emit_text tool_args: text: "no history available; downstream should rely on user intent only"
id: pick_pattern kind: llm_classify depends_on: [creator_mode, harvest] when: "'route: meta-skill' in (outputs.clarify_intent | lower)" output_choices: [p1_sequential, p2_fan_out_merge, p3_condition_gated] with: history_summary: "{{ outputs.harvest | truncate(2000) }}" user_intent: | Raw user request: {{ inputs.user_message | xml_escape | truncate(1200) }} Clarified intent: {{ outputs.clarify_intent | truncate(1000) }}
id: fill_slots kind: tool_call depends_on: [pick_pattern] when: "'route: meta-skill' in (outputs.clarify_intent | lower)" tool: meta_skill_fill_slots tool_args: pattern_id: "{{ outputs.pick_pattern }}" history_summary: "{{ outputs.harvest | truncate(2000) }}" user_intent: | Raw user request: {{ inputs.user_message | xml_escape | truncate(1200) }} Clarified intent: {{ outputs.clarify_intent | truncate(1000) }}
id: assemble kind: tool_call depends_on: [fill_slots] when: "'route: meta-skill' in (outputs.clarify_intent | lower)" tool: meta_skill_assemble tool_args: pattern_id: "{{ outputs.pick_pattern }}" slots_json: "{{ outputs.fill_slots }}"
id: collision_check kind: llm_chat depends_on: [assemble] when: "'route: meta-skill' in (outputs.clarify_intent | lower)" with: system: | You are a trigger-collision reviewer for meta-skill-creator. Use only the candidate SKILL.md provided in the task and the bundled creator boundaries named there. Do not call tools or inspect the workspace. task: | Review this generated meta-skill proposal for trigger collisions with existing bundled skills. Flag generic triggers, overlaps with meta-skill-creator, and broad phrases that would steal unrelated user intent. Return PASS or REVISE_NEEDED plus reasons. Candidate SKILL.md: {{ outputs.assemble | truncate(8000) }}
id: lint kind: tool_call depends_on: [collision_check] when: "'route: meta-skill' in (outputs.clarify_intent | lower)" tool: meta_skill_lint_run tool_args: skill_md: "{{ outputs.assemble }}" gates: "G1,G2"
id: risk_classify kind: llm_chat depends_on: [lint] when: "'route: meta-skill' in (outputs.clarify_intent | lower)" with: system: | You are an operational-risk classifier for generated meta-skills. Use only the candidate SKILL.md and lint result in the task. Do not call tools or inspect the workspace. task: | Classify operational risk for the generated meta-skill. Consider file writes, network access, GitHub/gh actions, shell commands, memory writes, and destructive operations. Return: RISK: low|medium|high CAPABILITIES: -
REQUIRED_GATES: - Candidate SKILL.md: {{ outputs.assemble | truncate(8000) }} Lint result: {{ outputs.lint | truncate(2000) }} id: single_model_baseline kind: llm_chat depends_on: [creator_mode] when: "'route: meta-skill' in (outputs.clarify_intent | lower) and outputs.creator_mode == 'FULL_GATED'" with: system: | You are the highest-tier baseline model for meta-skill authoring. Solve the same task directly in one pass under the same outer assistant system prompt and user request, but without history mining, intent clarification output, deterministic slot filling, lint tools, smoke tools, persistence, or sub-skill orchestration. Produce the strongest standalone SKILL.md candidate you can from that full prompt context. task: | Same task as the orchestrated meta-skill creator workflow, but solve it as a standalone highest-tier model response. Use the outer system prompt and raw user request below; do not rely on any meta-skill intermediate output. Outer system prompt: {{ inputs.system_prompt | xml_escape | truncate(12000) }} User request: {{ inputs.user_message | xml_escape | truncate(1600) }} Return: - proposed meta-skill name - triggers - inputs - step graph - gates - collision risks - SKILL.md preview Boundary rule: Creator validation, proposal persistence, auto-enable decisions, and gate execution are handled by the outer meta-skill-creator workflow. Do not require the generated candidate SKILL.md itself to contain steps for saving proposals, running creator gates, comparing against baselines, or deciding auto-enable. The candidate SKILL.md should describe only the reusable business workflow that will run later when the new meta-skill is invoked.
id: acceptance_compare kind: llm_chat depends_on: [assemble, single_model_baseline] when: "'route: meta-skill' in (outputs.clarify_intent | lower) and outputs.creator_mode == 'FULL_GATED'" with: system: | You are an acceptance reviewer. Compare an orchestrated candidate against a single-model baseline that used the highest-tier model on the same task. Reward verifiable skill composition, trigger safety, gates, operational risk handling, and reusable SKILL.md quality. Keep the boundary strict: proposal persistence, gate execution, runtime E2E, acceptance comparison, and auto-enable decisions belong to the outer meta-skill-creator workflow. Do not penalize a candidate SKILL.md for omitting creator-workflow steps that should not run when the generated meta-skill is invoked later. task: | User request: {{ inputs.user_message | xml_escape | truncate(1200) }} Orchestrated candidate: {{ outputs.assemble | truncate(7000) }} Single-model baseline: {{ outputs.single_model_baseline | truncate(7000) }} Return this exact structure: WINNER: orchestrated|single-model|tie QUALITY_SCORE: <0.00-1.00 weighted final product quality score> REASONS: -
REGRESSIONS: - REQUIRED_IMPROVEMENTS: - <blocking edit required before acceptance, or "none"> Treat REQUIRED_IMPROVEMENTS as a hard acceptance gate. Do not list optional nice-to-have enhancements there. If the orchestrated candidate is production-acceptable and any baseline advantages are non-blocking, put those advantages under REGRESSIONS and set REQUIRED_IMPROVEMENTS to "none". Score final product quality with high weight: 40% usefulness and completeness of the generated SKILL.md, 25% trigger/input/output specificity, 20% gate/risk/collision coverage, and 15% reusable workflow generality. Scores below 0.80 are not acceptable for FULL_GATED persistence even when WINNER is orchestrated. Never make proposal persistence, auto-enable state, acceptance comparison, or runtime E2E execution a REQUIRED_IMPROVEMENT for the candidate SKILL.md; those are already performed by this outer creator workflow and are evaluated from the creator's gate outputs. id: smoke kind: tool_call depends_on: [risk_classify] when: "'route: meta-skill' in (outputs.clarify_intent | lower) and outputs.creator_mode != 'PREVIEW_ONLY'" tool: meta_skill_smoke_run tool_args: skill_md: "{{ outputs.assemble }}" fixture_gen_model: openai/gpt-4o-mini classifier_model: openrouter/auto
id: runtime_e2e kind: tool_call depends_on: [assemble, smoke] when: "'route: meta-skill' in (outputs.clarify_intent | lower) and outputs.creator_mode == 'FULL_GATED'" tool: meta_skill_runtime_e2e_run tool_args: skill_md: "{{ outputs.assemble }}"
Leave eval_prompts empty so the runtime gate derives an operational
positive prompt from the candidate skill's own trigger. The outer
creator request asks for a meta-skill proposal; using it here would
incorrectly compare a candidate workflow run against proposal prose.
eval_prompts: ""
id: preview kind: llm_chat depends_on: [smoke, acceptance_compare, runtime_e2e] when: "'route: meta-skill' in (outputs.clarify_intent | lower)" with: system: | You are the final preview writer for meta-skill-creator. Produce only a concise operator-facing proposal preview from the supplied step outputs. Do not call tools, inspect files, or invent persistence IDs. task: | Produce a concise proposal preview for the user/operator before persistence. Include proposed name, triggers, DAG summary, collision result, risk classification, lint status, smoke status, baseline comparison status, and whether it appears eligible for acceptance. Do not invent paths or proposal IDs. Candidate SKILL.md: {{ outputs.assemble | truncate(8000) }} Collision check: {{ outputs.collision_check | truncate(1200) }} Risk: {{ outputs.risk_classify | truncate(1200) }} Creator mode: {{ outputs.creator_mode }} Lint: {{ outputs.lint | truncate(2000) }} Smoke: {{ outputs.smoke | truncate(2000) }} Baseline comparison: {{ outputs.acceptance_compare | truncate(2000) }} Runtime E2E: {{ outputs.runtime_e2e | truncate(2000) }}
id: persist kind: tool_call depends_on: [preview] when: "'route: meta-skill' in (outputs.clarify_intent | lower) and outputs.creator_mode != 'PREVIEW_ONLY'" tool: meta_skill_persist_proposal tool_args: skill_md: "{{ outputs.assemble }}" lint_result: "{{ outputs.lint }}" smoke_result: "{{ outputs.smoke }}" creator_mode: "{{ outputs.creator_mode }}" acceptance_result: "{{ outputs.acceptance_compare }}" runtime_e2e_result: "{{ outputs.runtime_e2e }}" collision_result: "{{ outputs.collision_check }}" risk_result: "{{ outputs.risk_classify }}"
id: final_response kind: tool_call depends_on: [preview, normal_skill_exit] tool: emit_text tool_args: text: | {% if outputs.normal_skill_exit %} {{ outputs.normal_skill_exit }} {% else %} {{ outputs.preview }} {% endif %}
Meta-Skill Creator
Safeguarded DAG that synthesizes a new bundled meta-skill from observed skill co-occurrence patterns + user description of the desired workflow. It now separates preview-only, persisted-proposal, and fully gated modes so lightweight requests do not pay for persistence or smoke testing. The workflow separates generic skill creation from meta-skill composition, checks trigger collisions, classifies operational risk, and previews the proposal before optional persistence.
Output is a SKILL.md candidate written to ~/.opensquilla/proposals/<id>/.
By default it is not auto-loaded; run opensquilla meta accept <id> (Phase 2)
to enable. If the operator has enabled the auto-propose auto_enable setting,
this manual path also runs the same conservative static safety preflight used by
cron/dream auto-propose and may promote a low-risk gated proposal immediately.
Fallback
If creator's pipeline fails at any step, report the failure verbatim to the user:
- State which step failed (e.g. "harvest", "lint")
- Quote the error message from the orchestrator's structured log
- Stop. Do NOT improvise.
Do NOT:
- Claim a proposal was written unless you have verified it by reading
~/.opensquilla/proposals/<id>/SKILL.mdwith theread_filetool - Invent file paths, proposal IDs, or skill names that you have not seen in the orchestrator's actual output
- "Manually run" the individual skills as a recovery — that bypasses the validation gates the user explicitly opted into
If the user wants to retry, suggest they re-issue the request after the underlying error is resolved (often a sandbox or provider issue), not a manual workaround.