Global Coding-Agent Assets

Use this reference when a readiness run, screenshot, or user request points to Cursor, Qoder, Codex, Claude, Qwen, Copilot, Kimi Code, or DeepSeek Harness (DSH) settings, installed assets, global skills, user hooks, commands, agents, plugins, MCPs, or memories. Treat this as a configured asset inventory, not a session behavior report.

Scope

• Project assets: .cursor, .qoder, .codex, .claude, .agents, .github, .kimi-code, .kimi, project rules, skills, agents, commands, hooks, workflows, settings, and MCP config.
• User/global assets: ~/.{cursor,qoder,codex,claude,qwen,copilot,kimi-code} skills, hooks, commands, agents, rules, settings, and MCP config.
• Plugin/marketplace assets: provider plugin caches and install evidence under ~/.cursor, ~/.qoder, ~/.codex, ~/.claude, ~/.qwen, ~/.copilot, and ~/.kimi-code, including plugin-declared Skills, MCPs, Commands, Hooks, Rules, and Subagents.
• Memories: ~/.qoder/memories/** plus Qoder SharedClientCache app-config.json memory keys and cache/db/*.db* file presence; and Codex generated-memory metadata under ~/.codex/memories/ plus supported config.toml memory settings.
• Session observed behavior: cite session-analysis.mjs sources/facets separately; configured assets do not prove runtime use.

Inventory Command

Run the read-only inventory when user-home or installed assets are in scope:

<node> <better-harness-root>/scripts/agent-customize/cli.mjs inventory --provider <cursor|qoder|codex|claude|qwen|copilot|kimi|dsh> --workspace <absolute-target-path>
<node> <better-harness-root>/scripts/coding-agent-practices/inventory.mjs <cursor|qoder|codex|claude|qwen|copilot|kimi> --workspace <absolute-target-path> --include-user-home --include-memories --format markdown
<cli> coding-agent-practices asset-integrity <cursor|qoder|codex|claude|qwen|copilot|kimi> --workspace <absolute-target-path> --language <en|zh-CN> --json [--include-memories] [--include-user-home]

Use --cursor-home <path>, --qoder-home <path>, --codex-home <path>, --claude-home <path>, --qwen-home <path>, --copilot-home <path>, --kimi-home <path>, --dsh-home <path>, --claude-state <file>, --codex-app-path <path>, or --shared-cache <path> for fixtures, alternate installs, or non-standard homes. Use the agent-customize command as the provider-specific configured asset source of truth; use the coding-agent-practices wrapper when the report also needs the matrix shape or Qoder/Codex memory metadata summary. For Qoder reports, project summary.practiceCoverageRows; it resolves ~/.qoder assets separately from the active SharedClientCache MCP home, excludes runtime-only project MCP metadata, and omits zero-count rows.

For Claude-specific settings/state/Plugin precedence and privacy boundaries, continue with Claude Code Configured Assets. For Kimi Code configured-asset locations and evidence boundaries, continue with Kimi Code Configured Assets.

For DSH's filesystem-only winners, cwd-sensitive Instruction sources, privacy default, and unresolved runtime boundary, continue with DeepSeek Harness Configured Assets.

The provider-labelled asset-integrity command reuses that inventory for a lightweight second pass. It checks Memory filename-title collisions/similarity, enabled Plugin canonical/display-name and non-empty capability-fingerprint overlap, and Hook count/exact-duplicate/fan-out pressure when those surfaces are available. Treat exact name and fingerprint matches as owner-review candidates and similar E2E/P7/P8-style Plugin names as advisory families. Never edit a Plugin cache or disable a variant from naming similarity alone.

Evidence Levels

• Configured: path or settings entry exists.
• Inspected: metadata, frontmatter, settings sections, or plugin index was opened.
• Observed: session, audit, hook, permission, or tool event shows runtime use.
• Backstopped: CI, host policy, sandbox, or hook tests enforce the practice.

Report configured and inspected global assets as evidence, but do not claim a Skill, Hook, Command, MCP, Plugin, Agent, Workflow, or Memory was used unless session or audit evidence shows it. For the cross-asset Exists -> Routed -> Applied -> Effective check, continue with Knowledge-Asset Review.

Presence, Use, And Cleanup Eligibility

Customization Checkup projects the evidence ladder into these reader states:

• observed: mapped runtime use exists in the bounded scope;
• configured-only: configuration was inspected but runtime use was not in scope;
• unobserved: a bounded window did not contain mapped use;
• candidate: sufficiently covered, outside a new-install grace period, unshadowed, owner-mapped, and safe for a disable-first plan;
• healthy: configured/effective state and observation are consistent;
• unavailable: active source, identity, or coverage cannot be established.

No observed use in one workspace cannot make a user/global asset a cleanup candidate. User/global cleanup needs an explicitly enabled user-global session pass and a declared time/session window. Plugin use can be demonstrated by an observed plugin-owned Skill, MCP tool, Command, Hook, or Agent only when stable metadata maps the child to the plugin. Ambiguous ownership stays unobserved.

Keep generated runtime state, caches, merged MCP snapshots, logs, and encrypted diagnostics read-only. Plan changes against author-owned files or supported host CLI operations.

Memory Boundary

Do not read raw memory text, private transcripts, auth files, token stores, model catalogs, or database rows by default. For memories, summarize only:

• memory root existence and category/path counts
• memory-related config keys such as Qoder memory.fetch.enable, memory.retrieve.enable, memory.file_memory.prompt.enabled, and memory.topictree.enable, or Codex features.memories, memories.generate_memories, and memories.use_memories
• database file presence such as local.db, memory.db, WAL, and SHM files
• drift or risk notes such as stale, duplicated, or sensitive-looking category names without printing content

If raw memory content is explicitly requested, treat it as sensitive user data and ask for a bounded scope before opening files. For recall, stale-memory, adoption, or learning-loop claims, continue with Memory Review And Learning-Loop Evidence; metadata and configuration never prove that a task retrieved or applied a memory.

Report Projection

In AI Agent Practices, surface a compact matrix with these rows when the inventory is in scope: Project assets, User/global assets, Plugin/marketplace assets, Session observed behavior, and Memories. Each row should show configured/inspected/observed status, strongest evidence, main gap, and Low, Medium, or High confidence.

Current Harness coverage rows remain inventory-only. Project material asset-integrity observations as ordinary grouped findings[] for Memories, Plugins, and Hooks; do not add health status or evaluation fields to the coverage rows.