context

name	context
description	Build a bounded, cited context pack from the living DeepWiki/code-map before broad source reads. Use in brownfield lanes and whenever source orientation could otherwise require large file reads.
argument-hint	"question" [--budget N]
context	fork

Context Pack

Use /context to retrieve exact files and line ranges from the current DeepWiki/code-map before reading source broadly.

Command

node .claude/scripts/context-pack.js "$ARGUMENTS"

Optional flags:

node .claude/scripts/context-pack.js --budget 1600 "$ARGUMENTS"
node .claude/scripts/context-pack.js --diff --budget 1600 "$ARGUMENTS"
node .claude/scripts/context-pack.js --depth 2 --diff --budget 1600 "$ARGUMENTS"
node .claude/scripts/context-pack.js --no-receipt --budget 1600 "$ARGUMENTS"

Unified facade (pack + graph + impact + semantic + co-change + refresh):

node .claude/scripts/nav-query.js pack --budget 1600 "$ARGUMENTS"
node .claude/scripts/nav-query.js symbol validateSession
node .claude/scripts/nav-query.js callers py:src/auth/session.py
node .claude/scripts/nav-query.js impact --files src/auth/session.py
node .claude/scripts/nav-query.js cochange src/auth/session.py
node .claude/scripts/nav-query.js semantic "session validation"
node .claude/scripts/nav-query.js refresh   # TF-IDF index + co-change + concept pages

Optional MCP: merge .claude/templates/mcp-nav.snippet.json into .mcp.json before long runs (tools: nav_pack, nav_symbol, nav_callers, nav_impact, nav_cochange, nav_semantic).

Flag	Meaning
--budget N	Max estimated pack tokens (default 1200)
--diff	Boost files from git status + graph dirty list
--depth N	Graph neighbor expansion depth (default 2)
--root DIR	Project root (default cwd)
--json-out PATH	Also write the pack JSON to PATH
--no-receipt	Do not write .claude/state/context-pack-last.json

Reading The Result (schema v2)

The JSON result contains:

• schema_version: 2
• status: ok, low_confidence, no_match, missing, or placeholder
• confidence / confidence_reasons: how safe it is to act without clarifying
• results[]: cited file paths, line ranges, symbols, reasons, scores, sources
• read_next[]: exact source reads to perform next
• task_map: edit_candidates, must_not_break, tests_to_run, clusters, clarify_options
• estimated_tokens: approximate context-pack size
• fallback: when narrow rg is allowed
• warnings[]: stale/missing/no-match guidance

A successful run also writes .claude/state/context-pack-last.json (session receipt). token_governor.context_search_required uses that receipt so agents are steered to pack before production source reads.

Rules

• Read the read_next ranges before reading entire files.
• Prefer task_map over front-loading brownfield essays.
• If status is placeholder, use the planned component map until source exists.
• If status is missing, run /code-map or /brownfield.
• If status is no_match or low_confidence, use task_map.clarify_options or one narrow rg, then re-pack if you find source that was not indexed.
• Change-family skills (/change, /feature, /refactor, /vibe) require this pack when a real code-graph exists — see their Context-first Iron Law.