context-health-visual

name	context-health-visual
disable-model-invocation	true
description	Diagnose Claude Code environment health — context budget, description obesity, trigger collisions, hooks, MCP, plugins, CLAUDE.md, memory, and skill-security scan. Use when asked to audit the environment, check context budget, review plugins, or scan installed skills for risky patterns.
argument-hint	[--format html|md] [--lang <code>] [--paste-context] [--use-instructions-loaded-hook]
allowed-tools	Read, Glob, Grep, Agent, AskUserQuestion, Bash(node *), Bash(open *), Bash(rm -rf /tmp/env-health-*)

Environment Health

Diagnose the user's Claude Code environment health. Outputs either an inline markdown report or a self-contained interactive HTML dashboard. Covers 10 diagnostic areas — 6 graded scores against official thresholds (§4 splits into §4a at-rest and §4b post-compact; §9 Skill Security Scan uses security baselines rather than docs-cited thresholds) plus 5 observational areas (raw numbers, no tier).

Instructions

Input Parsing

Parse these arguments:

Flag	Values	Default	Meaning
--format	html | md	html	Output mode. html generates a full interactive dashboard (default — recommended). md produces an inline markdown report for non-browser contexts or chat pasting
--lang	ISO code (en, ko, fr, etc.)	detected	Report language. Falls back to detecting the user message language, then en
--paste-context	(flag)	off	Ask the user to paste their /context output and use it to correct the estimated startup load
--use-instructions-loaded-hook	(flag)	off	Guide the user through temporarily enabling the InstructionsLoaded hook for file-level ground-truth data, then offer to revert it

Phase 1 — Data Collection

Determine the context window size. The scan subprocess cannot detect the active session's window from process.env. Derive it from the active model ID:

• Model ID contains [1m] (e.g. claude-opus-4-6[1m]) → 1000000
• Any other model ID → 200000

Pass this as --window-size=<N> so the scan records it and downstream formulas use the correct denominator.

Run the scan:

node ${CLAUDE_PLUGIN_ROOT}/skills/context-health-visual/scripts/env-health-scan.js --window-size=<N>

The script writes a JSON blob to stdout with these sections:

• scan_date, context_window_size, env_and_settings (normalized enable_tool_search per mcp.md's 5-value table, agent_teams_enabled, anthropic_base_url)
• installed_plugins ({plugins, orphan_count, orphans}), disabled_plugins
• installed_skills, installed_commands, local_skills (each entry carries desc_chars, when_to_use_chars, combined_chars, disabled, user_invocable)
• skill_bodies (at-rest 500-line flag + post-compact 5K-token flag)
• hook_inventory (type counts, event collisions, llm_hooks — includes inline hooks from plugin.json's hooks field: schema_issues and schema_issue_counts for hook schema validation)
• skill_security (scanned_count, findings with confidence levels, counts_by_severity, counts_by_category — covers 6 pattern categories: prompt_injection, data_exfil, destructive, hardcoded_credential, obfuscation, safety_override)
• context_metrics (MCP server count + source scopes — includes plugin .mcp.json files AND inline mcpServers from plugin.json)
• plugin_components (bin / monitors / lsp_servers / output_styles / channels per plugin + totals)
• subagent_preloads (agents with skills: frontmatter that preload bodies at start)
• plugin_options (per-plugin option keys only, values omitted — sensitive)
• claude_md (walks cwd → filesystem root + ~/.claude/CLAUDE.md, plus enumerates lazy-loaded nested files below cwd; respects claudeMdExcludes; each entry carries scope, load_mode, compact_resilient)
• rules,

Save to /tmp/env-health-<pid>/scan.json.

Optional ground truth refinement:

• If --paste-context is set: use AskUserQuestion to ask for the /context output, parse the reported always-loaded token counts, and override the estimates in the report. Clear the estimate-caveat when doing so.
• If --use-instructions-loaded-hook is set: walk the user through adding a temporary command-type InstructionsLoaded hook to ~/.claude/settings.json that logs to /tmp/env-health-<pid>/instructions-loaded.log. Instruct them to start a new Claude Code session so the hook fires, then parse the log for exact per-file loading. Offer to revert the hook after reading the log.

Phase 2 — Analysis

Read ${CLAUDE_PLUGIN_ROOT}/skills/context-health-visual/references/health-criteria.md for the full threshold specification. Apply it as follows:

Compute the effective description budget:

effective_budget = env_and_settings.desc_budget_override
                ?? max(8000, floor(context_window_size * 0.01))

Graded areas (§3, §4a, §4b, §7, §8, §9): classify into 🟢 healthy / 🟡 attention / 🔴 critical using the rules in health-criteria.md. Every threshold you apply must cite its docs source. (§9 Skill Security Scan is an exception — graded on security baselines without docs-cited thresholds; see health-criteria.md §9.)

Observational areas (§1 Plugin Inventory, §2 Startup Context Budget aggregate, §5 Trigger Collisions, §6 Hook Complexity, §10 Plugin Components): do NOT assign a tier. Emit raw numbers and info-level notes only. Delegate individual component grading in §2 to the owner area per the status-delegation table.

Trigger collisions (§5): delegate to the trigger-collision-inspector subagent. Build the input inventory by concatenating every installed skill and command description, one per line:

[plugin-name] skill-name: description text

Invoke the subagent via the Agent tool with subagent_type set to trigger-collision-inspector. The subagent returns {total_descriptions_analyzed, collisions: [...]}. Surface the pairs verbatim — DUPLICATE / OVERLAP classification is shown as reported, without aggregating into a tier (the prior 1-2 vs 3+ OVERLAP thresholds had no official basis and were removed in favor of raw observation).

Do NOT implement Jaccard or pairwise comparison in the orchestrator or the scan script. The subagent owns the entire comparison (direct adoption of Waza inspector-context.md:113).

Top lever computation: rank possible actions by raw numeric impact (chars freed, tokens saved) — not by severity — so observational areas can still surface a lever. Examples:

• Adding disable-model-invocation: true to the N skills with the largest combined_chars → frees sum(combined_chars) (% of budget). Prefer skills users invoke manually (/commit, /deploy, /release)
• Trimming entries over 1,536 chars → frees combined_chars - 1536 per entry
• Moving body content of a 5K+ token SKILL.md to references/ → removes it from both at-rest and compact budgets

Pick the single lever with the largest projected savings and promote it to the header + recommendations top card.

Phase 3 — Report Generation

Principle: HTML is the default because a dashboard renders the 10 diagnostic areas as KPI cards, bars, and collision heatmaps that markdown can't match. Only choose md when running in a non-browser context (cowork, headless CI), when the user asks for markdown explicitly, or when pasting into a chat thread is more useful than opening a file.

Mermaid generation force-loads Layer 0's semantic-tokens.md (tokens) / diagram-type-selection.md (type mapping) / diagram-density-rules.md (budget) / taste-gate.md (checklist), and is auto-validated by scripts/taste-gate.js.

Markdown mode (--format md):

Emit an inline markdown report with this structure:

# Environment Health — <scan_date>

**Graded:** N 🟢 / N 🟡 / N 🔴 (6 areas) · **Observational:** Plugin Inventory, Context Budget, Trigger Collisions, Hook Complexity, Plugin Components

**Top lever:** <one sentence>

**Estimated startup load:** ~N tokens (X% of <window> window) — *estimate, run `/context` for ground truth*

## §1 Plugin & Skill Inventory ℹ️
<raw tables, orphan/7-day-grace notes, plugin-option keys>

## §2 Startup Context Budget ℹ️
<component breakdown, delegation references>

## §3 Skill Description Obesity <status>
<numbers, truncated entries, disable-model-invocation candidates, subagent preload totals>

## §4 Skill Body Size <status>
<4a at-rest, 4b post-compact — report separately>

## §5 Trigger Collisions ℹ️
<subagent-returned DUPLICATE/OVERLAP pairs, verbatim, no tier>

## §6 Hook Complexity ℹ️
<type breakdown, event collisions, inline-vs-file sources>

## §7 MCP Overview <status>
<server count, ENABLE_TOOL_SEARCH 5-value table, effective mode, proxy fallback>

## §8 CLAUDE.md & Memory Health <status>
<file list with scope/load_mode/compact_resilient, MEMORY.md capacity, nested lazy-loaded totals>

## §9 Skill Security Scan <status>
<findings list with confidence; safe/likely_safe collapsed; hook schema_issues as info notes>

## §10 Plugin Components ℹ️
<bin / monitors / lsp_servers / output_styles / channels per plugin>

## Recommendations
<grouped by severity, top lever promoted>

Cite sources inline where a threshold fires. Keep it under 200 lines.

HTML mode (default, --format html):

Read ${CLAUDE_PLUGIN_ROOT}/skills/context-health-visual/references/section-structure.md for the JSON schema. Then follow ${CLAUDE_PLUGIN_ROOT}/references/report-generation-workflow.md with these parameters:

Parameter	Value
{output-path}	${CLAUDE_PLUGIN_DATA}/reports/<scan_date>-context-health-visual.html
{template-name}	context-health-visual.html
{skill-prefix}	env-health
{expected-sections}	10
{report-title}	"Environment Health — <scan_date>"
{aesthetic-hint}	"Dashboard"
{agent-prompt-data}	The analyzed scan data, subagent collision results, computed tiers per area, top lever, and info notes. Pass the raw scan.json separately so the writer can reference exact numbers.

The sections-data.json must set metadata.report_type = "context-health-visual" so render-sections.js dispatches to the context-health-visual renderers (not the plugin-visual ones).

Privacy guard (both modes): strip any raw file content before rendering. The report emits counts, sizes, line numbers, and file paths — never CLAUDE.md body text, MEMORY.md body text, API keys, or arbitrary file contents. Enforce this as a pre-render pass: walk the sections-data.json tree and remove any field named body, content, raw, text (excluding purposefully-set text fields in info_notes and recommendation action strings, which contain only computed messages).

Cleanup

Remove the temp directory:

rm -rf /tmp/env-health-<pid>

If the user enabled the InstructionsLoaded hook, remind them to revert it now.

Gotchas

• Skill description budget is 1% of context window, not 2%. Earlier drafts used 2% / 16K chars. The official number (skills.md — "Skill descriptions are cut short") is 1% scaling, 8,000-char fallback. Use the effective budget formula: SLASH_COMMAND_TOOL_CHAR_BUDGET ?? max(8000, floor(context_window * 0.01)).
• Skill description loading is a 3-layer model, not a single number. Per skills.md: (A) each entry's combined description + when_to_use is truncated at 1,536 characters — hard per-entry cap. (B) Total across all listing-included entries is capped at effective_budget = SLASH_COMMAND_TOOL_CHAR_BUDGET ?? max(8000, floor(window × 0.01)). When the total exceeds budget, every entry is dynamically shortened (keywords stripped). (C) Even below saturation, a few oversize entries can dominate. Earlier drafts cited "250 chars" — that number does not appear anywhere in the docs. §3 of the report evaluates all three axes separately. disable-model-invocation: true removes a skill from the listing entirely per the frontmatter table, so it contributes zero to the budget.
• 5K/25K skill body limit is POST-COMPACT, not at-rest. The 5,000-tokens-per-skill / 25,000-total limit applies only to re-injection after /compact, and only for skills that were invoked in the session. The at-rest recommendation is separate: keep SKILL.md under 500 lines (skills.md tip). Report them as two distinct sections (§4a and §4b) — do not conflate.
• Skill descriptions are NOT re-injected after compact. Only invoked skills survive compaction. The always-loaded description budget vanishes after compact, which changes what "always-loaded cost" means mid-session. Mention this caveat in the report.
• Auto-compaction can enter a thrashing state and halt. Per how-claude-code-works.md: "If a single file or tool output is so large that context refills immediately after each summary, Claude Code stops auto-compacting after a few attempts and shows an error." The scan cannot detect thrashing directly, but it can surface its ingredients — large SKILL.md bodies, many rules without paths:, and large CLAUDE.md chains all amplify the risk. When all three co-occur, flag it in the report and link users to https://code.claude.com/docs/en/troubleshooting#auto-compaction-stops-with-a-thrashing-error.
• Trigger collision uses a subagent, not inline orchestrator logic. The trigger-collision-inspector subagent owns the entire comparison (Waza-style). Do NOT add Jaccard code in the scanner or do the comparison in the main orchestrator — both would bloat main-session context and duplicate the subagent's job. If accuracy becomes a problem, revise the subagent prompt before adding deterministic pre-filter stages (YAGNI).

Reference Files

• references/health-criteria.md — Grading thresholds (cites every docs source) and recommendation templates
• references/section-structure.md — JSON schema for the 10-section HTML report
• agents/trigger-collision-inspector.md — Subagent spec for trigger collision detection (Waza-adapted)
• scripts/env-health-scan.js — Data collection script
• ../../references/report-generation-workflow.md — Shared HTML generation pipeline (render → assemble → validate → log → open)
• ../../references/design-system/semantic-tokens.md — Color and font semantic tokens
• ../../references/design-system/diagram-type-selection.md — 13-type selection guide
• ../../references/design-system/diagram-density-rules.md — Complexity budget per type
• ../../references/design-system/taste-gate.md — Pre-output quality checklist