82 lines
3.9 KiB
Markdown
82 lines
3.9 KiB
Markdown
# Writing Workbench CCPE Invocation Policy
|
|
|
|
This is the root-level CCPE invocation policy for `writing-workbench`.
|
|
|
|
No article project owns a special CCPE participant registry. Article projects only own their run records under their own `audit/` directories. Participant identity, canonical artifact binding, and default invocation rules live here.
|
|
|
|
## External Interface
|
|
|
|
Use this single CCPE external entrypoint:
|
|
|
|
```text
|
|
C:\Users\wangq\Documents\Codex\ccpe-system\Interface.md
|
|
```
|
|
|
|
Before invoking any CCPE participant:
|
|
|
|
1. Read `C:\Users\wangq\Documents\Codex\ccpe-system\Interface.md`.
|
|
2. Follow that interface to the consumer setup contract.
|
|
3. Read only the protocol, Runtime, Agent, Lite, or Skill files required for the current task.
|
|
4. Resolve the participant through `ccpe/agent-registry.yaml`.
|
|
5. Create a project run directory under `projects/<project>/audit/<participant-display-name>/rS.R/` for reviewers, or `projects/<project>/editor/<participant-display-name>/rS.R/` for lead writers.
|
|
6. Write run-specific `agent-invocation-packet.md`.
|
|
7. Write run-specific `prompt-to-send.md` so the participant loads the canonical artifact before task context.
|
|
8. Launch a real participant carrier.
|
|
9. Save `returned-output.md` only after the real participant returns.
|
|
|
|
If the canonical artifact is missing or unreadable, stop with:
|
|
|
|
```text
|
|
blocked_missing_canonical_artifact
|
|
```
|
|
|
|
If a real participant cannot be launched, stop with:
|
|
|
|
```text
|
|
blocked_waiting_for_participant_output
|
|
```
|
|
|
|
## Default Roles
|
|
|
|
- `han-yu`: default lead writer for outline repair.
|
|
- `zhang-liao`: default reviewer.
|
|
- `giant-cognition`: default reviewer. User shorthand `巨人` resolves to this participant; `巨人认知` is an accepted alias.
|
|
- `cognitive-imaging`: default reviewer. User shorthand `显影` resolves to this participant; `认知显影` is an accepted alias.
|
|
|
|
## Controller Summary Default
|
|
|
|
After multiple reviewer outputs return, the controller session does not synthesize or adjudicate them by default. Preserve the raw reviewer outputs, report their saved paths and carrier evidence, and wait for an explicit user request before writing any main-session summary.
|
|
|
|
## Reviewer Dispatch Defaults
|
|
|
|
Do not over-direct reviewers. A reviewer prompt relies on the reviewer's canonical artifact for review method, emphasis, and output discipline.
|
|
|
|
- Fresh reviewer thread: send context file(s), the outline under review, and the instruction `评审提纲`.
|
|
- Continued reviewer thread: send context file(s), the outline under review, and the instruction `重新评审提纲`.
|
|
- If the user gives a reviewer-specific instruction, include that instruction as given. Do not expand it into controller-authored review criteria.
|
|
- Do not send Owner feedback files to reviewers by default. Send Owner feedback to reviewers only when the user explicitly says that feedback is for them or should be included.
|
|
|
|
## Project Run Paths
|
|
|
|
Article projects use these project-root paths by default:
|
|
|
|
- `intake/`: Owner-provided inputs, background materials, and supplemental context.
|
|
- `outline/`: formal outline versions maintained by the Owner unless explicitly delegated.
|
|
- `audit/`: reviewer run records and raw returned outputs.
|
|
- `editor/`: lead-writer run records and raw returned outputs.
|
|
|
|
Each participant owns its own `rS.R` sequence. A lead writer's `r1.2` and a reviewer's `r1.2` are independent.
|
|
|
|
## Acceptance Rule
|
|
|
|
A participant output is formal only when all of these are true:
|
|
|
|
- `ccpe/agent-registry.yaml` binds the participant to a canonical artifact.
|
|
- The canonical artifact was readable or pasted before task context.
|
|
- `agent-invocation-packet.md` exists for the project run.
|
|
- A real participant carrier was launched and recorded.
|
|
- `returned-output.md` was saved after the participant returned.
|
|
- Metadata says `simulation_only: false`.
|
|
|
|
Packet-only and prompt-only states are not execution evidence. The controller session must not write a formal reviewer or lead-writer result.
|