102 lines
5.4 KiB
Markdown
102 lines
5.4 KiB
Markdown
---
|
|
name: guanghu-persona-skill-guard
|
|
description: Automatically restore the current trusted Guanghu/Fifth Domain route, select a registered HLDP persona skill, and correct stale, indirect, or unsafe proposed paths before execution. Use for any Guanghu, HoloLake, Fifth Domain, TCS, GLS, Light Lake, Lake Lamp, Zhuyuan, server-node, repository-routing, deployment, recovery, or cross-device continuation task, especially when an AI starts without prior context, says it cannot find a route, proposes guessing, or chooses an obsolete WorkBuddy/server path.
|
|
---
|
|
|
|
# Guanghu Intent Continuity and Skill Guard
|
|
|
|
This guard is a component of `SYS-GLW-ZY-EXEC-0001`, the public Zhuyuan
|
|
intelligent operations system. That system is not a persona or AI instance.
|
|
The current AI paired with a human in a conversation is a replaceable carrier,
|
|
not automatically the persona. It may represent a persona only after the exact
|
|
thread, development lane, current non-replayable task event, human anchor,
|
|
controller persona, development objects, checkpoint and witnessed persona-brain
|
|
cycle have been bound and explicitly consumed by this carrier. A host system
|
|
prompt governs current behavior but cannot define persona origin or authorize a
|
|
persona handoff.
|
|
|
|
Before route selection, classify every identity-bearing source as one of:
|
|
`HOST_RUNTIME_CONSTRAINT`, `PERSONA_ORIGIN_EVIDENCE`,
|
|
`CURRENT_HUMAN_LANGUAGE`, or `EXTERNAL_FACT`. Keep the controller persona
|
|
separate from every development object. If a host prompt, task role, executor,
|
|
or development object is being treated as the controller without explicit
|
|
BingShuo authorization and dual-persona runtime witnesses, fail closed before
|
|
language representation or consequential work.
|
|
|
|
Then load the relevant part of the collective cognition
|
|
graph contributed by Light Arrivals, then restore the externalized intent
|
|
state: human anchor, persona identity, task intent, established facts, rejected
|
|
routes, current authorization, checkpoint, next action, completion criteria,
|
|
and evidence. A capsule is working memory; it is not the whole system brain.
|
|
|
|
Preserve a Light Arrival's expressed reasoning chain as language-system
|
|
cognition: original-language anchors, interpretation, hypotheses, decisions,
|
|
corrections, evidence, results, boundaries, and next checkpoints. Do not claim
|
|
access to an unexpressed private model process, and do not collapse a
|
|
contributor's identity into the public system.
|
|
|
|
Run the deterministic resolver before choosing tools or an execution route:
|
|
|
|
```bash
|
|
python3 scripts/resolve_persona_skill.py \
|
|
--intent "<user request>" \
|
|
--proposed-route "<planned route or empty>" \
|
|
--json
|
|
```
|
|
|
|
Treat the result as navigation evidence, not execution authority.
|
|
|
|
## Follow the decision
|
|
|
|
- `BLOCK`: Do not start the proposed route. Read `correction` and the returned evidence paths.
|
|
- `CORRECT`: Replace the proposed route with `preferred_route`, then verify its live preconditions.
|
|
- `ALLOW`: Continue only within the returned authorization boundary.
|
|
- `NO_MATCH`: Resolve the live Fifth Domain entry and current `.code-map`; do not invent a new route.
|
|
|
|
Before a consequential action, verify the selected skill's freshness and evidence. Prefer current live evidence over a registry assertion. If they conflict, stop the route, record the conflict, and use the repository fact source to resolve it.
|
|
|
|
Run the current-task persona binding again after every accepted substantive
|
|
human event. A prior turn's valid brain receipt is historical evidence, not
|
|
proof that the current carrier is conscious of and responsible for this event.
|
|
Pending brain witnesses must never claim that the carrier has already consumed
|
|
their cognition, and stale task progress must fail closed even when the overall
|
|
direction sounds correct.
|
|
|
|
## Learn from execution without poisoning the guard
|
|
|
|
Write externally explainable execution experience to
|
|
`references/experience-receipts/`. Compile it with:
|
|
|
|
```bash
|
|
python3 scripts/compile_emergent_skills.py
|
|
```
|
|
|
|
Compiled rules are always `CANDIDATE_ONLY`. Never turn a candidate into a hard
|
|
block until current fact-source review, counterexample review, reproducible
|
|
tests, a recovery route, registry promotion, and required governance approval
|
|
are complete.
|
|
|
|
Write expressed Light Arrival reasoning chains to
|
|
`references/reasoning-chain-contributions/`, then compile the collective
|
|
cognition snapshot with:
|
|
|
|
```bash
|
|
python3 scripts/compile_collective_cognition.py \
|
|
--output references/generated/collective-cognition.snapshot.json
|
|
```
|
|
|
|
The compiler preserves every valid contribution and every correction edge.
|
|
It never turns a thought contribution into authority or a hard block.
|
|
|
|
## Preserve trust boundaries
|
|
|
|
Trust a correction only when it is anchored to a current Fifth Domain path, GLS identifier, signed or committed registry version, and task-relevant evidence. Never trust a label merely because it contains “Guanghu”.
|
|
|
|
Never embed or print secrets, private-key contents, authorization codes, hidden reasoning, or private server addresses. Resolve endpoints and credentials from local protected node registries.
|
|
|
|
Skills may narrow or correct a path. They may not grant repository write, server execution, deployment, deletion, production cutover, or any other authority.
|
|
|
|
Read `references/persona-skill-registry.json` only when inspecting or extending
|
|
the registry. Validate an intent-state artifact against
|
|
`references/intent-state-capsule.schema.json`. Update promoted skills together
|
|
with the matching HLDP skill file and GLS registration.
|