# Writing Workbench

Writing Workbench is the local tool-calling and process-record workspace for article writing after a Web-generated initial outline already exists.

It is not the idea base, not the long-context material discussion space, and not an autonomous article production runtime.

## Current Position

```text
knowledge-vault      = upstream discussions, source materials, published essays, public-facing archive
writing-workbench    = local project workspace for outline review, agent invocation, records, drafts
ccpe-system          = expert agents, runtime protocols, model cards, governed cognitive assets
skills-vault         = reusable automation skills and deterministic tools
```

## Entry Rule

A project starts here only when the user provides:

- a Web-generated initial outline, normally version `1.0`;
- source paths for the article materials;
- the lead writer identity, such as `韩愈` or `蒙田`;
- an explicit instruction to create or continue the project.

If there is no initial outline, do not create a Writing Workbench project. Continue discussion and material development in `knowledge-vault`.

## Mission

This workspace helps the user save effort after the article has entered writing:

- register source paths in `source-map.md`;
- preserve the initial outline;
- prepare outline-review context under current context-window limits;
- compress or organize source materials without replacing the original sources;
- invoke real reviewer and writer sessions when the user explicitly names them;
- preserve root-level CCPE registry and invocation policy files under `ccpe/`;
- preserve reviewer audit runs under `audit/<reviewer-id>/rS.R/` with agent invocation packets, thread IDs, prompts, and returned outputs, where `S` is the reviewer-session number and `R` is that session's round number;
- preserve complete invocation records and final returned outputs;
- preserve user decisions, reflections, outline versions, and draft sessions;
- use lightweight drafting records by default after the user approves drafting: `drafting/drafting.md` plus `drafting/sessions/<unit>.prompt-N.md` and `<unit>.returned-N.md`;
- create capability request tickets when required agents or skills are missing.

## Hard Rules

1. No initial outline, no project.
2. No user-named reviewer, no reviewer invocation.
3. No real sub-session, no formal reviewer or writer output.
4. No root-level registry binding to a readable CCPE canonical artifact, no CCPE participant invocation.
5. No `agent-invocation-packet.md`, no accepted reviewer audit record.
6. No recorded thread ID and returned output, no accepted reviewer audit record.
7. No user decision, no stage advancement.
8. No full-article generation by default.
9. Missing capability creates a supplier ticket and pauses the project.
10. Drafting is not an audit system by default; do not create per-section packets, metadata, proofread logs, or aggregate draft files unless the user asks for them.

## Directory Overview

```text
docs/                    Operating rules for this workspace
projects/                Article project workspaces, one folder per article
requirements/ccpe/       Supplier tickets for CCPE agents/runtimes/model assets
requirements/skills-vault/ Supplier tickets for automation skills
templates/               Project and requirement templates
```

Old pre-rebuild scaffold assumptions were discarded on 2026-06-15. The current model is: the user brings a Web outline and source paths; Codex helps compress, invoke, record, and organize, but does not decide how the article proceeds.