❄️ Cold Auditor Phase — Add Support for Context‑Isolated Sub‑Agent Audits

Aperta
#259 3 commenti 0 reazioni 0 assegnatari Vedi su GitHub

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
5/5
Tempo stimato
Più di una settimana
Idoneità per principianti
35/100
Tipo di issue
Funzionalità
Chiarezza
Abbastanza chiara
Stato di attività
Tranquilla
Stack tecnologico
javascript
Ambito
ai, tooling

Direzione di ricerca

La proposta non indica file di implementazione, test o punti di ingresso. Inizia individuando il workflow e l’implementazione delle fasi esistenti di PromptKit, quindi segui come vengono gestiti i sub-agents e gli input espliciti; il lavoro è completato quando sono presenti una primitiva di audit isolata dal contesto, un workflow di esempio e una documentazione che copra i criteri di accettazione elencati.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

❄️ Cold Auditor Phase — Add Support for Context‑Isolated Sub‑Agent Audits

Summary
Introduce a Cold Auditor Phase to PromptKit: a workflow primitive that spawns a context‑isolated sub‑agent to perform adversarial audits of earlier artifacts.
This sub‑agent receives only the artifacts and instructions explicitly passed to it, ensuring a cold read with zero inherited assumptions, memory, or narrative bias.

This pattern emerged organically during real‑world use and proved extremely effective at tightening specs, exposing drift, and stabilizing multi‑phase workflows.


Motivation
In multi‑phase workflows, later phases often inherit the model’s earlier assumptions.
This creates several problems:

  • Early‑phase fuzziness gets “grandfathered in”
  • The model avoids contradicting itself
  • Drift accumulates across phases
  • Audits become polite instead of adversarial
  • The system loses its ability to self‑correct

A context‑reset audit eliminates these issues.
When a sub‑agent evaluates artifacts with no prior context, it behaves like a fresh engineer joining the project:

  • no sunk‑cost bias
  • no narrative loyalty
  • no continuity assumptions
  • no memory of earlier reasoning

This produces dramatically stronger audits and more stable specs.


Proposal
Add a workflow primitive that allows a phase to declare:

“Run this audit using a context‑isolated sub‑agent.
Provide only the specified artifacts and instructions.
Do not inherit any prior conversation or reasoning.”

Key properties

  • Isolation: sub‑agent receives no parent context
  • Explicit inputs only: artifacts + instructions
  • Adversarial stance: evaluate as if encountering the spec for the first time
  • Deterministic: no hidden state, no bleed‑through
  • Model‑agnostic: works with any model combination

Benefits

  1. True adversarial auditing
    The sub‑agent is free to tear apart earlier phases without self‑consistency bias.

  2. Drift resistance
    Specs converge toward correctness instead of drifting with model personality.

  3. Robustness to context loss
    If a model loses its window, the workflow still behaves correctly.

  4. Multi‑model synergy
    Cheap/creative models can propose; precise models can audit.

  5. Stronger invariants
    Cold reads force vague requirements to become crisp, explicit, and enforceable.


Example Workflow Addition (conceptual)

- phase: protocol_audit
  mode: cold_subagent
  inputs:
    - requirements
    - spec
  instructions: |
    Evaluate these artifacts as if you have never seen them before.
    Identify inconsistencies, missing invariants, and drift.
    Produce a corrected, tightened version of the spec.

Acceptance Criteria

  • Ability to spawn a sub‑agent with no inherited context
  • Ability to pass explicit artifacts and instructions
  • Deterministic isolation: no cross‑phase memory leakage
  • Clear documentation of the “Cold Auditor Phase” pattern
  • Example workflow demonstrating usage

Why This Belongs in PromptKit
PromptKit already enforces structure, determinism, and artifact‑bounded reasoning.
A Cold Auditor Phase extends this by adding a powerful correctness mechanism:

self‑disagreement as a feature, not a bug.

This turns PromptKit workflows into systems that:

  • generate
  • critique
  • refine
  • converge

…without relying on any single model’s continuity or assumptions.


Notes
This pattern emerged organically during real use when a context window reset caused an audit phase to evaluate the spec with zero preconceptions.
The result was dramatically better than a continuity‑preserving audit.
Formalizing this behavior would make it a first‑class tool for all PromptKit workflows.

Lingua principale
JavaScript
Stelle
105
Fork
23
Metriche di merge delle PR
Nessuna PR unita negli ultimi 30g

Guida per i contributori

Apri la guida per i contributori

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di microsoft/PromptKit

Tutte le issue di microsoft/PromptKit

Issue simili

Altre issue su JavaScript

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.