wantsong
/
writing-workbench
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
writing-workbench/docs/ccpe-rebuild-context.md

4.6 KiB
Raw Blame History

CCPE Rebuild Context From Writing Workbench

Use this file as input context when starting a CCPE optimization or rebuild session.

Why This Exists

Writing Workbench was rebuilt on 2026-06-15 because the earlier attempt treated deep writing as an end-to-end article production runtime. That failed in practice.

The new Writing Workbench starts only after the user has already produced a Web initial outline. CCPE should therefore stop optimizing for a fully autonomous article production pipeline and instead supply the concrete agents, protocols, and safety rules needed by this narrower workflow.

Current Writing Workbench Model

Web discussion and source digestion
  -> Web initial outline v1.0
  -> Writing Workbench project
  -> outline review preparation
  -> user-named reviewer invocation
  -> user reflection and decision
  -> user-directed lead writer repair
  -> revised outline v2.0+
  -> user-directed section/chapter drafting

The user decides every stage transition.

What CCPE Should Supply First

CCPE is currently most useful as a supplier for Writing Workbench.

Priority capabilities:

  1. Lead writer agents

    • 韩愈 for academic essay / structured prose.
    • 蒙田 for essayistic reflection.
    • Other lead writers only when the user requests them.

  2. Reviewer agents

    • 张辽 red-team review.
    • 巨人认知 macro-model review.
    • 认知显影 cognitive-structure inspection.
    • Any additional reviewer only after a concrete Writing Workbench requirement exists.

  3. Invocation protocols

    • reviewer invocation packet;
    • lead writer repair packet;
    • returned-output record;
    • session/report contract;
    • no-simulation rule.

  4. Context preparation rules

    • how to prepare source digests under context-window limits;
    • how to avoid compressing away material not represented in the initial outline;
    • how to distinguish source-grounded content from model-supplied additions.

  5. Evaluation and decision support

    • review report contract;
    • conflict synthesis for user decision;
    • repair instruction builder;
    • outline versioning rules.

Hard Boundaries For CCPE

CCPE should not store article project process records.

Do not put these in CCPE:

  • full article project folders;
  • source digests made for one article;
  • reviewer session.md files for one article;
  • draft chapters;
  • user decision logs;
  • publication metadata for one essay.

These belong in writing-workbench or knowledge-vault.

No-Simulation Requirement

The earlier workflow failed because the controller session simulated reviewer outputs after reading prompts.

For CCPE:

  • dispatch packet is not execution;
  • prompt-to-send is not execution;
  • a controller summary is not a reviewer report;
  • a formal reviewer or writer output requires real invocation;
  • missing capability must block the project or produce a supplier request.

Formal simulation, when used for planning, must be labeled:

simulation_only: true
formal_output: false
excluded_from_synthesis: true

Supplier Relationship

Writing Workbench sends requests to CCPE through:

writing-workbench/requirements/ccpe/

CCPE should treat those files as client requirements.

CCPE may produce:

  • agent specs;
  • runtime protocols;
  • invocation contracts;
  • model cards;
  • evaluation rubrics;
  • usage notes for Writing Workbench.

CCPE should not implement deterministic scripts that belong in skills-vault.

Recommended CCPE Rebuild Focus

For the next CCPE optimization pass, prioritize:

  1. Replace the old full Article Production Runtime assumption with a narrower outline-review-and-repair runtime.
  2. Encode Writing Workbench as the first concrete consumer of article-writing assets.
  3. Add a clear intake rule: CCPE builds only assets requested by a project, proven by recurring use, or required for invocation safety.
  4. Separate expert-agent assets from deterministic automation skills.
  5. Make the no-simulation boundary visible in every article-writing agent and runtime protocol.
  6. Add usage notes showing how Writing Workbench should call each agent without copying canonical specs into project folders.

Summary For A New CCPE Session

Writing Workbench is not asking CCPE for a complete autonomous writing machine.

It needs a reliable supplier layer for:

  • real lead writer invocation;
  • real reviewer invocation;
  • source/context preparation under context limits;
  • review synthesis for user decision;
  • user-directed outline repair;
  • careful section-level drafting support.

The user remains the director. CCPE supplies the actors, contracts, and guardrails.