Briefing — Iteration 2¶
Parent: HL-TFW-42 Goal: Map agent capabilities for research subtasks and formalize how TFW guides coordinators in agent selection.
Predecessor Context (from iter 1)¶
Decisions to build on:
- D1: agent field = free-text string, optional → we now need to answer WHAT the coordinator writes in this field and WHY
- D6: Framework guidance = prompt-in-template → we now need to design the actual prompt/guidance content
- D7: Agent purpose = traceability, not dispatch → guidance should inform, not automate
Open threads from iter 1: - FC3 identified 3 agent archetypes (Web Researcher, Code Auditor, Infra Operator) but didn't validate externally - No external research on actual tool capabilities (Claude Code, Codex CLI, Antigravity-class tools) - "Prompt-in-template" (D6) was decided but the template content wasn't designed
Research Plan¶
Gather: - External research: documented capabilities of Claude Code, Codex CLI (Gemini), Cursor, Windsurf, Aider — what makes each unique for research - Map research subtask types to tool strengths (web search, code analysis, file traversal, MCP integration, context window, synthesis) - Analyze: where should guidance live in TFW? (conventions vs template vs workflow)
Extract: - Build capability matrix: tool × research subtask type - Cross-reference with formalization options: where does this matrix live in TFW? - Design concrete template/guidance text
Challenge: - Stress-test: will capability guidance become stale as tools evolve? - Counter-evidence: does prescriptive guidance conflict with TFW's tool-agnostic philosophy? - Edge case: what if a user has only 1 tool available?
Hypotheses (from HL §10)¶
| # | Hypothesis | HL Status |
|---|---|---|
| H1 | Multi-agent orchestration needs agent field in iterations.yaml, not a separate mechanism |
confirmed (iter 1) — iter 2 extends: what guidance accompanies the field? |
Scope Intent¶
- In scope: Tool capability mapping for research. Formalization location (where guidance lives). Concrete template/comment text. Human-choice vs recommendation model.
- Out of scope: Tool installation/setup. Runtime integration. Non-research tool usage (execution, review).
Guiding Questions¶
- Should TFW name specific tools (Claude Code, Codex) or use generic archetypes (Web Researcher, Code Auditor)?
- Where does capability guidance live without becoming a maintenance burden?
- How prescriptive should the guidance be — "consider using X for Y" vs "X is best for Y"?
User Direction¶
User feedback (triggering iter 2): «мы не ответили на вопрос для чего и каких случаев каких агентов лучше использовать claude gemini codex, и как это лучше формализовать, рекомендовать что-то или человек сам выбирает или что?»
Key signal: user wants ACTIONABLE guidance, not just a schema field. The coordinator should know WHEN to consider switching agents and WHY.
Stage complete: YES