avav25/context-load

/context-load — Per-Role Context Slice

Extract role-relevant excerpts from project documentation. Trim to a token budget. Wrap in <untrusted_content> envelope per G1. Output: paste-ready Markdown for subagent prompt assembly.

When to use

• Inside an orchestrator about to spawn a subagent — load just the slice that role needs
• During /feature-design Wave 1 — feed each agent its targeted context (vs full project dump)
• Before any G7 spawn payload assembly — populates state_slice.related_artefacts

Reduces per-agent input tokens vs sending full CLAUDE.md to every subagent.

Invocation

/context-load --for db-engineer
/context-load --for ui-ux-designer --max-tokens 2000
/context-load --for security-engineer --include-memory

Arguments

| Flag | Default | Effect | |---|---|---| | --for <role> | required | One of the 26 plugin agent names | | --max-tokens N | 2000 | Hard cap on output token count | | --include-memory | off | Also include relevant L4 learnings.md excerpts | | --include-architecture | on | Include ARCHITECTURE.md if exists |

Behavior

1. Validate --for <role> matches one of plugin/agents/<name>.md (else error with list of valid roles).
2. Read source files (G1 wrap each per untrusted-content-wrapping.md):
   • <repo>/CLAUDE.md — extract sections matching role keywords
   • <repo>/AGENTS.md (if exists) — extract role-specific subsection
   • <repo>/ARCHITECTURE.md (if exists, --include-architecture) — extract role-relevant parts
   • L4 learnings.md (if --include-memory) — entries tagged with role or with role-domain keywords
3. Role-keyword extraction map (examples):
   • db-engineer → sections matching database, schema, migration, query, index, transaction
   • frontend-engineer → frontend, UI, component, routing, bundler, accessibility
   • security-engineer → auth, authentication, authorization, secrets, encryption, vulnerability, OWASP
   • devops-engineer → deploy, docker, kubernetes, helm, terraform, CI, pipeline
4. Trim to --max-tokens:
   • First trim: drop excerpts below relevance threshold
   • Second trim: truncate longest excerpts proportionally
   • Always preserve at least one snippet per source file (unless source had nothing role-relevant)
5. Format output:

# Project context for `<role>` (loaded by /context-load, ~N tokens)

## From CLAUDE.md
<wrapped excerpt>

## From AGENTS.md
<wrapped excerpt>

## From ARCHITECTURE.md
<wrapped excerpt>

## From L4 memory (if --include-memory)
<wrapped excerpt>

---
Sources: CLAUDE.md (lines X-Y), AGENTS.md (section "Z"), ARCHITECTURE.md (sections A,B), .ai-skills-memory/learnings.md (3 entries)

Failure modes

• Invalid role: error with full list of 26 valid agent names
• No source files exist: return "No project context available — skipping context load" (not an error)
• Token budget too small for any meaningful excerpt: return shortest possible plus warning that budget was insufficient

Memory writes

This skill is read-only — no memory writes.

Integration

• Reads: <repo>/CLAUDE.md (L2), <repo>/AGENTS.md, <repo>/ARCHITECTURE.md, L4 learnings.md (if --include-memory)
• Validates against: plugin/agents/*.md for role name lookup
• Rules: untrusted-content-wrapping (G1 wrap on every source excerpt)
• Used by: orchestrator skills (/feature-design, /develop, /team-bugfix) — feeds state_slice of G7 spawn payloads
• Companion: /subagent-spawn (constructs full G7 payload using context-load output)
• Layer model: plugin/docs/concepts/memory.md documents the L0–L5 layers (L0 Cowork host, L1 plugin templates, L2 CLAUDE.md, L3 session run-logs, L4 project memory, L5 user-global memory). This skill reads from L2 + L4.