创意生成单元

Creativity Generation Unit (CGU)

An honest creativity scaffold for LLM agents, shipped as an MCP server (SDK 2) plus a portable Agent Plugin.

🌐 繁體中文

v0.8.0 is a breaking rewrite. The old 24-tool server (generate_ideas, spark_collision, deep_think, …) is gone, with no aliases. See Migrating from 0.6.

v0.9.0 adds opt-in inquiry memory and caller-model distillation of creative material. Install sources are pinned to v0.9.0; release notes and packages.

Why

An LLM can already brainstorm. What it cannot do alone is the part around brainstorming:

CGU does not pretend to be creative. By default (CGU_PROVIDER=passthrough) it never calls an LLM: it returns work orders for the model you are already using, then validates, stores, measures and fences what comes back. Numbers are produced only by code, and every float travels inside a Measurement that names its method.

No additional local LLM is required. Default CGU_EMBEDDING=ngram also avoids probing Ollama. The calling agent's model performs reasoning, naming, semantic review and creative judgment; that agent may use a local or cloud model. Optional semantic embeddings improve comparisons, not reasoning.

With inquiry memory enabled, a hook can collect user prompts. cgu_status reports pending reviews; after 20 prompts by default, the agent uses cgu_inquiry(organize) and submits distill to store reusable questions, constraints, assumptions, analogies, observations and ideas with source IDs and hashes. materials retrieves them across sessions. Storage defaults to ~/.cgu, shared across clients. Without an active agent, prompts accumulate until the next review. See the contract.

Install

Requirements: uv (for uvx), git, network access to GitHub on first start. The first start builds the package from git and takes about a minute; later starts are cached.

GitHub Copilot CLI

copilot plugin marketplace add u9401066/creativity-generation-unit
copilot plugin install cgu@creativity-generation-unit

Verify with copilot plugin list and, in a session, ask for cgu_status. Installing from a local clone also works: copilot plugin install ./plugins/cgu (Copilot CLI marks direct local-path installs as deprecated in favor of marketplaces).

VS Code (GitHub Copilot) and Codex

Codex CLI (verified install and MCP registration on Codex CLI 0.160.0; a full model session was not verified because it needs a login):

codex plugin marketplace add u9401066/creativity-generation-unit
codex plugin add cgu@creativity-generation-unit
codex mcp list    # shows the cgu server

VS Code implements the same Agent Plugins 1.0 standard (Copilot reads .github/plugin/marketplace.json). Add this repository as a plugin marketplace and install cgu; menu names vary by client version. VS Code is not yet verified by us. Details: plugins/cgu/README.md.

Any MCP client (server only, no skills)

{
  "mcpServers": {
    "cgu": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "git+https://github.com/u9401066/creativity-generation-unit@v0.9.0", "cgu-server"],
      "env": { "CGU_PROVIDER": "passthrough" }
    }
  }
}

From a clone: uv sync then uv run cgu-server; uv run cgu doctor prints what is available.

What is inside the plugin

Copilot also gets four thin agents (creative-facilitator, frame-auditor, independent-ideator, adversarial-critic). The plugin does not automatically install hooks; the optional Copilot collector is installed separately by the user.

The 11 tools

Every tool returns ToolResult{ok, data, work_order, work_orders, provenance, error}. Domain errors are ok=false, never exceptions. Also exposed: 5 resources (cgu://methods/*, cgu://operators, rubrics, triggers) and 4 prompts for clients without skills. Full contract: docs/architecture.md.

Example (medical product development)

"We are an 8-person startup building a home-care product to cut 30-day readmissions. Not another SpO2 wearable plus app."

With the plugin, a mid-tier model opens a session, writes the 8 typical answers, extracts the shared hidden assumptions (home sensing → alert → someone acts), and rewrites three of them with operators. For example, it negates "the alert is the product" into "the product is who receives the alert, under an SLA". It then fans each rewrite out as an independent idea card with payer vs. user, regulatory class (marked as a guess to confirm with the regulator), required validation and the cheapest MVP. Each card carries its derivation path (typical answer → assumption → operator → frame id).

Inquiry memory and the optional hook

Recording starts off. The portable path is for the agent to forward selected prompts through capture. To collect Copilot CLI prompts continuously, install the pinned CLI, inspect the hook and enable recording:

uv tool install git+https://github.com/u9401066/creativity-generation-unit@v0.9.0
cgu inquiry install-hook --print
cgu inquiry settings --enable --yes
cgu inquiry install-hook

Restart Copilot CLI after installation. The hook records user prompt text, timestamps, channel and the final project-directory name. It does not capture assistant replies or full transcripts. Installing the CLI first keeps the hook command local. Hook payload and installation were tested for Copilot CLI; Codex and VS Code hooks have not been verified.

Ask your agent to organize accumulated material. cgu_status.inquiry.maintenance reports pending and due (20 pending prompts by default; change this through cgu_inquiry(action=settings, organize_after=...)). The agent gets a batch with organize, interprets it using its own model, and submits distill with evidence IDs and stored-text SHA-256 hashes. materials retrieves questions, constraints, assumptions, analogies, observations and ideas; fragments_payload can be added to an ideation session. Without an active agent, prompts accumulate until the next review.

Manage records with cgu inquiry settings --disable, cgu inquiry settings --exclude PROJECT, cgu inquiry export, cgu inquiry delete --all --yes, or cgu inquiry install-hook --remove. Removing the hook leaves existing records in place. Deleting a source also deletes material that depends on it. De-identification is best effort; the agent's model and data policies are controlled by your client. Full contract and limits.

Upgrading from 0.8

Storage defaults to ~/.cgu and ignores PLUGIN_DATA. For an existing plugin-specific database, stop the server and back it up before moving cgu.sqlite3 to ~/.cgu, or set CGU_DATA_DIR to the existing directory. There is no automatic move or merge. Ensure all database connections are closed before moving a WAL database. Schema migrations preserve existing sessions.

Default embeddings change from auto to ngram: similarity is lexical (semantic=false) and no model is contacted. Explicit CGU_EMBEDDING=auto or ollama opts into the configured embedding backend. Embeddings improve comparison; the agent's model still performs reasoning.

Configuration

Architecture

src/cgu/
├── domain/          pure rules: Measurement, frames, operators, doubt, fence, judge (no I/O)
├── application/     ports (Protocols) + services per tool
├── infrastructure/  SQLite archive, embeddings, retrieval, optional LLM, Settings
└── interfaces/      mcp/ (SDK 2 server, tools, resources, prompts) and cli

Dependency direction is enforced by a test (domain imports nothing from other layers, nor httpx/sqlite3).

Honest limits

• Novelty is relative to an explicit reference set, and with the n-gram fallback it cannot detect a paraphrase.
• Heuristic numbers (priority, diversity, win rate) are uncalibrated. LLM judges inherit judge-model bias; order disagreement is reported.
• "Philosophical frame audit improves creativity" and "maieutic mode preserves human originality" are hypotheses under test, not conclusions.
• Skill effectiveness depends on the host model following the procedure; mid-tier models can skip steps.
• Creativity is not given an ethics gate here by design; downstream regulation (medical devices, privacy, hospital rules) is surfaced as a risk field on each idea card, not as a filter.

Evidence

We ran an isolated Copilot CLI experiment (plugin installed for real; mid-tier models claude-sonnet-5.5 and gpt-6-luna; research, medical-product and admin-process tasks; blinded pairwise judging by gpt-6-sol and claude-sonnet-5.5, both orders). Full write-up: evals/reports/SUMMARY.md.

What we can and cannot claim

• The plugin does not demonstrably beat plain prompting overall. Cell-level win rate vs. baseline was 27% (exp1, v1), 25% (exp2, v2) and 50% (exp3, v3: 1 win / 1 loss / 4 ties). Every 95% CI spans 50%, so the honest reading is "indistinguishable", with the point estimate unfavorable for v1 and v2.
• It reliably changes what answers look like: strongly preferred on novelty (75–96%) and problem reframing (75–96%) in all three rounds.
• v1 and v2 were disliked on practicality and decidability (17–38%). The judges' own reasons: tool jargon leaking into the answer, and go/no-go thresholds left as "待估". v3 fixed those two things (plain-language output, concrete adjustable thresholds) and moved decidability from ~36% to 79% and practicality from ~19% to 38%, at the cost of a smaller novelty edge. v3 rests on only 6 cells with unstable judge order agreement.
• One of our hypotheses was falsified: adding a constraint ledger and feasibility gate (v2) did not recover practicality.
• Cost: about 2–5× the wall time and 4–40× the input tokens (mostly cached).
• Auto-triggering is unreliable: with an unmodified prompt the skill led to substantial CGU tool use in 9 of 12 cells (Sonnet 3/6, luna 6/6).
• LLM judges only, no human calibration, tiny n, and the v2/v3 changes were tuned from judge reasons, so they may partly be tuned to LLM judges. Not a claim about SOTA models.

Development

uv sync --extra dev
uv run pytest
uv run ruff check src tests evals && uv run ruff format --check src tests evals
uv run mypy src

On Windows, a running cgu-server.exe (for example one started by an editor) locks .venv. Use a separate environment: $env:UV_PROJECT_ENVIRONMENT = "$env:TEMP\cgu-venv".

Documents

Migrating from 0.6

License

Apache-2.0. Author: u9401066 u9401066@gap.kmu.edu.tw.