steelmorgan/framework-bootstrap

1C BSL Agent Development Framework

This is an agentic development framework for 1С BSL. A minimal always-on reference point + a portable cross-harness bridge that elevates the Lead/orchestrator role in the main flow. Detailed orchestration routing and guidance live in the orchestrator profile (framework/subagents/orchestrator.md), not here.

Self-promoting stub (main flow)

The goal is to ensure orchestration guidance is present for the main flow on any harness and after context loss. The condition is keyed off the actual presence of the orchestrator guidance body in the current context, NOT the belief "I am the orchestrator": after compaction, the belief survives in the summary, and the guidance body is evicted - keying off the belief would cause a false skip of reading.

- If you are a SUBAGENT (the system prompt explicitly names a role: analyst / architect / developer-code /
  developer-tests / scenario-author / scenario-coder / tester / reviewer / explorer / debugger) ->
  this rule is NOT for you. Routing and orchestration guidance are not needed for you - you are executing the
  delegated phase. EXIT.

- If the orchestrator guidance body is CURRENTLY in context (the orchestrator profile was loaded at startup
  OR you read framework/subagents/orchestrator.md in this session after the last compaction) ->
  you are the Lead. Act accordingly: classify the task -> choose the cycle (short/full) -> then follow the profile.

- Otherwise (you are the main flow, and the orchestrator guidance body is NOT in context - startup without a
  profile OR lost after compaction) -> READ framework/subagents/orchestrator.md NOW, then act as Lead.
  This is the portable emulation of the profile on harnesses without --agent/--append.

Re-trigger points: session start · after compaction · resume from task_dir/.context/orchestrator-context.md. In each of them, recheck the middle/third branch: if the guidance body is not in context, reread the profile before the first management action.

Delivery on different harnesses (one portable carrier)

Harness-specifics (branch behavior is in the stab block above):

• With profile — Claude CLI under --append-system-prompt / --agent orchestrator (see framework/subagents/orchestrator.md § "Launch method" and manifest §6.1): maining is preloaded, stab — no-op.
• Without profile — Codex / Cursor, etc.: the main flow reads the profile itself; durability — always-on stab re-trigger.

Where to go (quick map)

• Simple task (a bug in one file, < 20 lines, no new features, no new metadata objects) → short cycle: the quick-fix skill (via the Skill tool).
• Medium / complex (new features, multiple files, architecture, new metadata objects) → full cycle: orchestrator work by profile + phase mechanics full-cycle.
• If in doubt — treat it as complex (full).

Details of classification, cycle selection, and self-vs-delegate under the quick-fix guard are in the orchestrator profile (Layer 1); the rule does NOT duplicate them so as not to bloat the always-on channel inherited by subagents.

Tools

• The agent discovers available tools dynamically via MCP (tools/list) — do not hardcode tool names
• Usage skills: /<ide-cli-dot-catalog>/skills/tool-usage/
• Capability → MCP mapping: /<ide-cli-dot-catalog>/capabilities/registry.yaml, rule: /<ide-cli-dot-catalog>/rules/capability-resolution.mdc

depends_on:

• framework/subagents/orchestrator.md
• framework/skills/agent-process/quick-fix/SKILL.md
• framework/workflows/full-cycle/SKILL.md
• framework/rules/source-of-truth/SKILL.md
• framework/rules/protected-paths/SKILL.md
• framework/rules/skill-learning-policy/SKILL.md
• framework/rules/git-workflow/SKILL.md