From a2c896d0aea80e6ecfcc9c385d69eaaf796247ae Mon Sep 17 00:00:00 2001 From: murashit Date: Mon, 20 Jul 2026 19:08:58 +0900 Subject: [PATCH] docs(review): require proportional evidence for findings --- AGENTS.md | 2 +- docs/design.md | 2 ++ 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/AGENTS.md b/AGENTS.md index e5b85276..06430e9f 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -2,7 +2,7 @@ This repository contains the Codex Panel Obsidian plugin. ## Working Principles -- Start from evidence. Reproduce defects before fixing them; for product changes, reconstruct the current workflow, user need, alternatives, and losses before deciding what should change. +- Start from evidence. Reproduce defects before fixing them; for product changes, reconstruct the current workflow, user need, alternatives, and losses before deciding what should change. Treat theoretical failure modes and reviewer agreement as hypotheses, not evidence of meaningful user or operational harm. - Use design documents as context, not as a substitute for current behavior and user expectations. Prefer coherent user-facing behavior and clear ownership, and remove needless abstraction or obsolete compatibility instead of preserving it by default. - Treat implementation, tests, documentation, lint policy, and final history as one deliverable: inspect each for drift, but update only the durable contracts that changed. Documentation should record user-facing behavior, long-term design, or reusable workflow; do not add it merely to narrate an implementation change or duplicate tests and tooling. - Parallelize only substantial concerns that can be implemented and tested independently with little shared-file contention. Keep small, tightly coupled, sequential, or coordination-heavy work together. diff --git a/docs/design.md b/docs/design.md index 87561cfa..be7a5499 100644 --- a/docs/design.md +++ b/docs/design.md @@ -10,6 +10,8 @@ Keep the panel thin. Codex Panel owns the Obsidian experience around Codex: pane Codex owns runtime behavior and thread state: models, credentials, sandboxing, approval policy, MCP servers, hooks, providers, agent-initiated network access, thread history, archived state, goals, and runtime settings. The panel should read and update that state through `codex app-server`, not redefine it. +The panel may provide a narrow management surface when direct configuration of Codex-owned state is impractical and app-server still owns the semantics. Hook controls serve this role: help users recognize and manage hooks they added or changed without presenting the panel as a code-audit surface. + The panel may acquire prompt context through an explicit user action. `/web` fetches the requested page through Obsidian and attaches the extracted content as untrusted turn context; it is not a search surface or agent network policy. Panel-acquired context remains reference material rather than current user authorization. Send its bounded payload as untrusted app-server context, while persisting only a tightly bounded, context-body-free metadata descriptor needed to reconstruct panel display and archive metadata. Treat that fixed-schema envelope as inert reference metadata rather than user instructions. Apply physical byte and part limits at the app-server boundary; keep source-specific selection and truncation policy with the context producer.