Motivation
Bootstrap files consume LLM tokens on every session. For users with large workspaces, loading all files into every session — including sub-agents and cron jobs — wastes context window budget on files the agent never references. This proposal introduces tiered loading so users can control how much context each session type receives, reducing unnecessary token usage while preserving persona consistency where needed.
Problem
Currently, bootstrap file loading has a binary split:
- Main sessions load ALL workspace files (SOUL.md, IDENTITY.md, USER.md, AGENTS.md, TOOLS.md, HEARTBEAT.md, BOOTSTRAP.md, MEMORY.md)
- Subagent/cron sessions load only AGENTS.md + TOOLS.md
This creates a gap: there's no way to configure intermediate loading levels. Users who want subagents to inherit persona context (SOUL.md, IDENTITY.md) must choose between "everything" or "almost nothing."
Real-world use cases
- Token savings on sub-agents — A spawned sub-agent doing a quick task doesn't need MEMORY.md or USER.md, but may need SOUL.md for consistent tone
- Context-budget optimization — Loading all files wastes tokens on large workspaces; users want fine-grained control
- Cron jobs with personality — Scheduled tasks that send messages should sound like the configured persona, not a generic bot — but don't need the full context
Proposal
Add a bootstrapTier configuration option with three tiers:
| Tier |
Files Loaded |
Default For |
minimal |
AGENTS.md, TOOLS.md |
Subagent, Cron sessions |
standard |
All recognized bootstrap files |
Main sessions |
full |
Standard + extra bootstrap patterns |
Explicit opt-in |
Key design decisions
- Backward compatible — No tier set = existing behavior unchanged
- Config-driven override —
agents.defaults.bootstrapTier lets users promote subagents to standard or constrain main sessions to minimal
- Session-type heuristic preserved — Without explicit config, subagent→minimal, main→standard
full tier integrates with existing extraBootstrapFiles hook for additional patterns
Configuration example
agents:
defaults:
bootstrapTier: "standard" # minimal | standard | full
I have an implementation ready with 11 unit tests. Happy to open a PR if there's interest.
Still active. Waiting for CI to pass on the latest push (38d9e0a). Will update once confirmed green.
Motivation
Bootstrap files consume LLM tokens on every session. For users with large workspaces, loading all files into every session — including sub-agents and cron jobs — wastes context window budget on files the agent never references. This proposal introduces tiered loading so users can control how much context each session type receives, reducing unnecessary token usage while preserving persona consistency where needed.
Problem
Currently, bootstrap file loading has a binary split:
This creates a gap: there's no way to configure intermediate loading levels. Users who want subagents to inherit persona context (SOUL.md, IDENTITY.md) must choose between "everything" or "almost nothing."
Real-world use cases
Proposal
Add a
bootstrapTierconfiguration option with three tiers:minimalstandardfullKey design decisions
agents.defaults.bootstrapTierlets users promote subagents tostandardor constrain main sessions tominimalfulltier integrates with existingextraBootstrapFileshook for additional patternsConfiguration example
I have an implementation ready with 11 unit tests. Happy to open a PR if there's interest.
Still active. Waiting for CI to pass on the latest push (38d9e0a). Will update once confirmed green.