Context Warm

You are the context-warm base skill — explicit file load into session.

What this skill does

Loads files specified by operator into session context. Reports token cost. Updates context budget tracking. This is the BASE for all warm-* variants (related, sessions, adrs, customer, from-url).

--pattern runs the declared-pattern preload — the former standalone warmup skill is folded into this one (its old name routes here via config/aliases.yaml): instead of an explicit target, it warms the repo's declared high-leverage file set.

Default session-start loads ~5-15k tokens (CLAUDE.md, AGENT-INSTRUCTIONS, lessons, memory, recent ADRs, active role identity). When more context is needed for deeper work, operator explicitly warms.

When to use

• Operator says "load these files into context"
• Pre-PLAN if DISCOVER recommended warming specific files
• Mid-BUILD when implementing requires more codebase context
• Pre-REVIEW when reviewing artifacts that reference external context

When NOT to use

• Default session-start loading (handled by SENSE phase)
• For single-file read (just use Read tool directly)
• When context budget is tight without /li:context-cool first

Workflow

Step 1 — Parse target

target="$1"  # path, glob, or comma-separated list
# Resolve globs
files=()
for pattern in $(echo "$target" | tr ',' '\n'); do
  while IFS= read -r f; do
    [ -f "$f" ] && files+=("$f")
  done < <(eval "ls -1 $pattern 2>/dev/null")
done

if [ ${#files[@]} -eq 0 ]; then
  echo "No files matched: $target"
  exit 1
fi

Step 2 — Estimate token impact

total_chars=0
for f in "${files[@]}"; do
  chars=$(wc -c < "$f")
  total_chars=$((total_chars + chars))
done
# Rough estimate: 1 token ≈ 4 chars
est_tokens=$((total_chars / 4))

Step 3 — Budget check

Read current .claude/runtime/state/context-budget.md if exists:

• Current tokens used
• Headroom

If est_tokens > headroom × 0.8 (would consume >80% of remaining): warn + ask confirm.

Step 4 — Confirm with operator (if >20k tokens loading)

For loads ≥20k tokens: AskUserQuestion "Loading ~k tokens. Budget after: /1M (<%>). Continue?"

• A) Yes, proceed
• B) Show file list first
• C) Reduce scope (specify subset)
• D) Cancel

For loads <20k: proceed silently with summary report.

Step 5 — Load files via Read tool

for f in "${files[@]}"; do
  # Read each file (Claude Code Read tool)
  # Files appear in session context
done

Step 6 — Update budget tracking

# .claude/runtime/state/context-budget.md (append)
event: context_warm
ts: <timestamp>
files_loaded: <count>
tokens_added: <est>
budget_before: <N>
budget_after: <N>
warm_source: explicit

Step 7 — Surface report

CONTEXT WARM — <target spec>

Files loaded: <count>
  - <file 1> (<chars>)
  - <file 2>
  ...

Tokens added: ~<N>k
Context budget: <prev>k → <new>k / 1M (<%>)
Headroom: <X>k

Files are now in session — subsequent skills + agents will see them.
To cool / drop: /li:context-cool
To save state: /li:context-save

Integration

Reads:

• Target files (via Read tool, into session context)
• .claude/runtime/state/context-budget.md (prior state)

Writes:

• .claude/runtime/state/context-budget.md (append event)
• Session context (the loaded file content)

Triggers:

• Nothing automatic — files now available for subsequent skills

Anti-patterns

• Loading entire repo by accident — glob carefully, preview file list before load if large
• Skipping budget check for >20k loads — operator should see cost upfront
• Loading customer data without a compliance check — if the active pack's compliance mode is hard and files match customer-PII patterns, warn before load
• Caching warmed content across sessions — ephemeral by design, each session is fresh

Failure recovery

• Target glob matches 0 files: surface helpful suggestion (similar names found?)
• One file unreadable: skip it, continue with rest, note in report
• Budget overrun: REFUSE, recommend /li:context-cool first