Keep always-on agent context small. Load philosophy, SOPs, and skills **on demand**. Prefer **kit-knowledge** and **memory** MCP over dumping files into the prompt.
Targets
| Surface | Target | |---------|--------| | Always-on bootstrap (AGENTS.md + project handshake + thin IDE rules) | **< ~2k tokens** (~8KB chars combined when possible) | | Skill discovery (descriptions only) | Acceptable; do not pre-load skill bodies | | Full SOP / philosophy read on typo or debug routes | **Zero** unless the route needs that procedure | | MCP servers enabled | **One profile** matching the work; match skill mcp: frontmatter |
Split
| Mechanism | Holds | |-----------|-------| | Always-on | Invariants + phase→skill index (AGENTS.md) | | Skills | Role/phase behavior when triggered | | File read / **kit-knowledge** | SOP slices, philosophy sections, handovers | | **memory** MCP | Glossary, SLOs, prefs across sessions (never secrets) | | Other MCP | Live systems (GitHub, Context7, browsers, …) |
Measure
pnpm kit measure-contextReports character/token estimates for always-on surfaces and flags when over budget. Re-run after editing AGENTS.md or project handshake templates. Public write-up: docs/kit.md.
Session checklist
- [ ] Did not bulk-read
CODING_PHILOSOPHY.mdor multiple SOPs before the first edit on a small route - [ ] Loaded at most one phase skill body for the active route
- [ ] Used kit-knowledge or a single SOP path when a procedure was needed
- [ ] Spec/XFN stored durable facts in memory MCP (or N/A in handover)
- [ ] Only one MCP profile active; unused heavy servers not installed globally