skill-registry-injection-20k-tokens-per-call
The full skill-registry (~170 skills) is injected as a system reminder with every tool call return, costing approximately 20k context tokens per round-trip. This severely limits multi-file bulk editing sessions and causes context exhaustion mid-task. Large-scope skill edits (300+ sections across 65+ files) must be planned across multiple sessions or delegated to parallel subagents with bounded scopes.
Related
- graphiti
- docker
- clawteam-openclaw-multi-agent-swarm-evaluation
- oracle
- long-running-mcp-tools-must-return-run-id-immediately-not-bl
- skill-registry-system-reminder-bloats-context-20k-per-turn
- skill-registry-injection-costs-20k-tokens-per-tool-call
- vague-verb-mass-replacement-causes-regressions-in-skill-file
- skill-registry-system-reminder-context-cost
- skill-registry-system-reminder-context-burn
- skill-registry-system-reminder-costs-20k-tokens-per-tool-cal
- skill-registry-system-reminders-burn-20k-tokens-per-tool-cal
- fusion-invariants-lint-substance-check-via-template-generato
- skill-registry-reminder-costs-30k-tokens-per-tool-call