project-settings

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Project Settings v2.5

项目设置 v2.5

Some decisions belong to the project rather than to a scene or a plan. POV, tense, and how much the agent may decide alone are configuration, not prose. Keeping them in a plan means they get re-litigated every session.
有些决策应归属于项目,而不是某个场景或计划。POV、时态,以及 Agent 可以独自决定的事项范围,都属于配置,而不是叙事性文本。把它们放在计划中,意味着每次会话都会重新讨论这些决定。

Record

记录

Configuration lives in
PROJECT_SETTINGS.md
. Every field the author has decided is authoritative; every field they have not is
UNRESOLVED
, and an unresolved setting is never guessed.
配置存放在
PROJECT_SETTINGS.md
中。作者已经作出决定的每个字段均为权威值;尚未决定的每个字段均为
UNRESOLVED
,并且绝不会猜测未解决的设置。

Fields that change agent behaviour

改变 Agent 行为的字段

POV model and tense — read by
narrative-voice
and
scene-writing
. An unset value means the narrator has not been chosen yet, which is different from third limited.
Approval strictness — the most consequential field, because it governs how much proceeds unattended:
LevelMEDIUM riskHIGH risk
strict
Ask before actingAsk before acting
standard
Proceed when clearly local and reversibleAsk before acting
permissive
ProceedAsk before acting
HIGH-risk work always requires approval at every level. Permissive relaxes only MEDIUM, and never below the guarantee that canon and structure are not changed unattended. There is no setting that waives the HIGH gate.
Autonomy ceiling — caps what may proceed even when a risk level would allow it. Useful when an author wants narrow scope for a while.
Enabled engines — restricts which engines may run. When an author wants canon frozen mid-draft, disabling
canon-manager
and
version-control
is clearer than relying on instructions not to use them.
POV 模型与时态 — 由
narrative-voice
和
scene-writing
读取。未设置的值表示尚未选定叙述者,这与第三人称限知视角不同。
审批严格度 — 这是后果最严重的字段,因为它决定哪些操作可以在无人询问的情况下继续:
级别
MEDIUM
风险
HIGH
风险
strict
执行前询问执行前询问
standard
如果操作明确属于局部操作且可逆,则继续执行前询问
permissive
继续执行执行前询问
无论处于哪个级别,
HIGH
风险工作始终需要审批。
permissive
仅放宽
MEDIUM
风险限制,并且绝不会低于确保 canon 和结构不会在无人监督的情况下被更改这一底线。不存在任何可以豁免
HIGH
风险门槛的设置。
自主权上限 — 即使某项操作的风险等级原本允许其继续,也会限制可以继续执行的操作范围。当作者希望在一段时间内收窄范围时,此设置很有用。
已启用的引擎 — 限制可以运行的引擎。当作者希望在草稿写到中途时冻结 canon,禁用
canon-manager
和
version-control
比仅依赖“不要使用它们”的指示更清晰。

Rules

规则

  • Settings never override the author. A setting is a standing instruction that the author can change at any time, and a current instruction beats a stored one.
  • A setting never overrides canon.
    PERMANENT_CANON
    outranks any configuration.
  • Record why a setting changed. Approval strictness is usually changed because something nearly went wrong, and the reason is the useful part.
  • Do not set a field to make a validation pass. Unresolved is a valid state.
  • Settings are project-scoped. Cross-project behaviour belongs in
    author-profile
    .
  • 设置永远不会凌驾于作者之上。设置是一项长期指令,作者可以随时更改;当前指令优先于已存储的设置。
  • 设置永远不会覆盖 canon。
    PERMANENT_CANON
    的优先级高于任何配置。
  • 记录设置发生变更的原因。审批严格度通常是因为某件事差点出错而被更改,而原因才是最有用的信息。
  • 不要为了让校验通过而设置某个字段。
    UNRESOLVED
    是有效状态。
  • 设置的作用范围限定在项目内。跨项目行为应归入
    author-profile
    。

Coordination

协调

Read by
novel-orchestrator
for every routing decision, by
canon-manager
for classification, by
narrative-voice
and
scene-writing
for POV and tense, and by
quality-gate
when judging whether an action was in scope. Report a setting that contradicts observed practice as a finding rather than silently reconciling it.
novel-orchestrator
会在每次路由决策时读取这些设置;
canon-manager
会读取它们以进行分类;
narrative-voice
和
scene-writing
会读取 POV 和时态;
quality-gate
会读取它们,以判断某项操作是否在允许范围内。如果某项设置与观察到的实际行为相矛盾,应将其作为一项发现报告,而不是悄悄地自行调和。