Software Orchestrator
5769c1cd22
- Auto-Registration-Bug behoben (register_project/get_project_id/resolve_project Trennung) - 25 Tests gruen (Pytest) - block_compactor-Tool refactored (Option B: Soft-Check statt Hard-Block) - 4 Restbaustellen gefixt - DB-Schema: plugin_settings-Tabelle hinzugefuegt - 3 Schattenprojekte aus DB geloescht - Plan v3 + Refactor-Plan + Worklog dokumentiert
1.2 KiB
1.2 KiB
name, description, tags
| name | description | tags | |||
|---|---|---|---|---|---|
| context-budgeting | Manager context small, raw data to files, token estimation workflow for the orchestrator agent. |
|
Context Budgeting
Purpose
Keep the main agent's context small by storing raw data in files and using includes/references instead of inline data.
When to Use
- When a tool returns more than 80 lines of text
- When the user provides large files for processing
- When the conversation history grows large
- When a subordinate returns a long report
When Not to Use
- For short tool outputs (< 30 lines)
- For data that must be reasoned about inline
Procedure
- Receive large output from tool or subordinate
- Save to a file (e.g., /a0/usr/workdir/.md)
- Reference the file in the response:
See §include(/path/to/file.md) - Optionally summarize the file in a few lines for context
- Use raw file only when deep analysis is needed
Token Estimation Rule
- 1 line of code/markdown ≈ 10-20 tokens
- 80 lines of tool output ≈ 1000-1500 tokens
- 1000 lines of conversation ≈ 15000-20000 tokens
- Compress when output exceeds 1000 tokens
See Also
- Tool:
context_compactor(plugin tool for structured compression)