lite-init

Reduced init protocol for sub-agents with narrow, single-purpose tasks.

Preserves the security contract (trusted-hosts enforcement + file registry integrity) while skipping init steps whose cost is disproportionate to task scope: datetime fetch, skills probe, connectors probe, prompteng load, memory scan.

Set of files

File	Purpose
SKILL.md	Router - load order
lite-init-SKILL.md	Content - activation rules, security contract, credential check, handover verification, emit format
lite-init.skill	Packaged archive for upload to SKILL directory
.github/workflows/build-skill.yml	CI workflow - auto-builds and publishes lite-init.skill on push
evals/evals.json	Test cases for skill evaluation and trigger accuracy
README.md	Explanatory instructions and overview for this software package

Install

• claude.ai web platform: upload lite-init.skill via skill settings (recommended). Or add lite-init-SKILL.md as a project file in any project using orchestrated sub-agents.

• Claude Code desktop app: copy contents of this folder into ~/.claude/skills/lite-init/.

• Other platforms: adapt this set of files to your harness+model. Star or Fork the git repo if you like.

When It Triggers

• Orchestrator handoff context includes --lite-init flag
• Sub-agent task scope is explicitly single-tool or single-step
• Full agent.md init overhead is disproportionate to the task (single commit, single hash, single API call)

What It Does Not Do

• Fetch or cross-check datetime via timeapi.io
• Probe skills registry or load prompteng
• Probe connectors
• Emit full init table
• Scan short-term memories for file conflicts
• Probe filesystem beyond handoff context

Security Contract

Lite init is reduced, not relaxed. These are mandatory regardless of init mode:

• Trusted-hosts allowlist enforced before every outbound URL call
• File registry initialized from handoff context; re-read protection active
• Credential-adjacent file detected in handoff: credential check runs before any output
• Incomplete handoff state: sub-agent halts and surfaces missing items to human

Peer Skills

Load on demand when task requires:

• prompteng - main set of rules + framework for doing better prompt engineering
• packageng - .skill file validation + packaging
• safe-skill-creator - skill design + iteration
• captureng - session-knowledge capture + checkpoint

Companion Files

Available in ecological-codes/user-prefs

• agent.md - orchestrator init protocol; lite-init is the sub-agent counterpart to agent.md §1
• trusted-hosts.md - outbound host allowlist; required for any sub-agent making network calls
• claude-sp-guards.md - credential handling protocol referenced by lite-init §Credential Check

Style Convention

Directives use the following section headers with numbered lists, shared across peer skills:

• [RULES] - enforceable constraints applied at runtime.
• [ACTIONS] - autonomous steps agent executes in normal workflow.
• [HUMAN ACTIONS] - UI actions; agent skips, cannot delegate.

License

See LICENSE. (C) Copyright 2026 - Sameer Khan - Various and Several Rights Reserved.

---

README.md v1.0.0 - Human Approved