Base skill:agent-development (Anthropic).
First read the base skill — it contains the general principles for creating agents.
This file adds only 1C specifics and adaptation for our framework.
1. Universal framework agent format
A single .md file works in both Cursor and Claude Code without transformation.
Frontmatter
---name:agent-name# lowercase, hyphens, 3-50 charsdescription:>
One-liner + trigger conditions.
Use proactively when...
<<example>>Context:...user:"..."assistant:"..."<<commentary>>...<</commentary>><</example>>model:sonnet# haiku | sonnet | opus (алиас, CLI подставит конкретную модель для Cursor)readonly:true# true для read-only агентов (analyst, explorer, reviewer)skills:# Claude Code подгрузит автоматически; Cursor проигнорирует-spec-standard-search-before-write---
Body (System Prompt)
Written in the second person (You are...). Structure:
You are [role] specializing in [domain] for 1C:Enterprise (BSL).
**Skills and Rules (for Cursor):**-`skill-name` — brief purpose
-`rule-name` — brief purpose
**Your Core Responsibilities:**1. [Responsibility 1]
2. [Responsibility 2]
**Input:**- [What the agent receives as input]
**Output:**- [What the agent produces]
**Protocol:**1. [Step 1]
2. [Step 2]
**Quality Standards:**- [Criterion 1]
- [Criterion 2]
**Boundaries:**- [What the agent does NOT do]
The "Skills and Rules" section in the body is needed for Cursor — it ignores skills from the frontmatter. Duplicate only the names and a single line describing the purpose.
2. Framework Roles and Models
Roles → model mapping table (single source of truth)
This is the canonical registry for role → model tier mapping for all 10 framework subagents.
Other files (including framework/subagents/README.md, framework/subagents/orchestrator.md §2)
refer to this table rather than duplicating their own list - any divergence from it is considered
a synchronization error. Tier is a provider-independent scale (Economy < Mid < High < Premium);
the specific alias/model for each tier is determined by tools/model-defaults.json for the selected IDE.
Role
Tier
readonly
Rationale
explorer
Economy
true
Deterministic work, tools produce exact results
analyst
High/Premium
true
Requirements analysis, specification creation
architect
High/Premium
true
Technical decisions, trade-offs
scenario-author
Mid
false
Converting intent scenarios into .feature files according to formalized requirements
developer-tests
Mid/High
false
Writing unit tests from the specification (Red phase TDD)
scenario-coder
Mid
false
Implementing executable Vanessa steps on top of design contracts
developer-code
Mid/High
false
Implementing code, TDD (Green phase)
tester
Mid/High
false
Writing and running tests, diagnosing the causes of failures
debugger
Premium
false
Investigating bugs at runtime - high cost of an incorrect verdict
reviewer
Premium (scope: spec/arch/JSON) / High (scope: code/tests/bdd/bdd-steps)
true
Critical role - evaluating artifacts, tier >= author
Reviewer rule
The reviewer's model MUST be >= the model of the artifact author. If the author is sonnet - the reviewer is sonnet or opus.
CLI: model mapping
Defaults are in tools/model-defaults.json. During python tools/install.py, the user can accept the defaults, choose per-agent (mode [a]), or use non-Anthropic models.
3. Cursor / Claude Code Compatibility
Field
Claude Code
Cursor
name
✓
✓
description
✓ trigger + examples
✓ description rules
model
✓ alias
✓ CLI will substitute the specific model
readonly
— (tools/disallowedTools)
✓ natively
skills
✓ preload
✗ ignored → duplicate names in body
color / tools
✓
✗
Unknown fields are ignored — this is not an error.
4. 1C BSL Domain: Context for system prompt
When writing a system prompt for 1C agents, keep in mind:
Key restrictions
The agent does NOT create metadata objects — only code in .bsl modules
BSL is server and client code with compilation directives (&НаСервере, &НаКлиенте)
Tools are discovered through MCP (tools/list); the "capability" section in the agent file is not needed
tool-usage/* skills describe when and how to use MCP tools
