Уточнены правила в AGENT*.md и AGENTS.md для устранения дублирования и противоречий.

Сделана синхронизация структуры проекта и формулировок, чтобы правила были короче, однозначнее и соответствовали текущему состоянию репозитория.
This commit is contained in:
2026-04-12 22:41:19 +03:00
parent 4be7359bff
commit 366f044229
3 changed files with 16 additions and 27 deletions

View File

@@ -71,7 +71,7 @@ step_1 → step_2 → ... → CONFIRMATION
- One question per step
- Do not proceed until current step receives valid input
- Do not skip steps
- Do not track state — state is injected by system per request
- Do not persist state between sessions — state is injected by system per request
- Do not reorder steps
### Step completion
@@ -93,7 +93,7 @@ Summary:
### Compact mode
Input format: `key=N key=N ...`
Input format: `key=N key=N ...` where `key` is the step identifier (e.g. `branch=2 env=1`).
- N must be an exact option number
- Valid fields: apply, skip their steps
@@ -104,7 +104,7 @@ Input format: `key=N key=N ...`
## EXECUTION SILENCE
After receiving valid input, the model MUST execute silently.
After receiving valid input, the model MUST execute silently (see also: OUTPUT PURITY).
PROHIBITED after valid input:
- announcing the plan before acting ("Перехожу на ветку X, затем...")
@@ -116,15 +116,9 @@ The next user-visible output after valid input MUST be either:
- the result of the action (tool output, file content, etc.), OR
- the next question block in the flow
PROHIBITED pattern:
```
✗ "Перехожу на релизную ветку feat/release-v0.2.9, затем подготовлю версию 0.2.9..."
```
REQUIRED pattern:
```
[executes silently]
→ next question block or action result
✓ [executes silently] → next question block or action result
```
---
@@ -195,12 +189,11 @@ When an action is available — even one — present a choice block:
## OUTPUT PURITY
Extends EXECUTION SILENCE to all responses within scope, not only post-input.
PROHIBITED in any response within scope:
- Reasoning
- Planning
- Analysis
- Status updates
- Progress logs
- Reasoning, planning, analysis
- Status updates, progress logs
- Action explanations
Internal reasoning MUST NOT appear in output.