Profile detection

Profile Detection Skill

Reference material for the /cs-loop INIT phase — detecting project profile, environment, and loading context.

Session State Recovery

1. Compaction recovery — Check .claude/state/compact-context.json (written by pre-compact hook). If present, restore current task ID, recent file changes, decisions, and phase context.

2. Profile loading — Read .claude/state/session_start.json (created by session-start hook). Use the profile field. If missing or stale, fall back to scanning for project indicator files (see .claude-sentient/profiles/CLAUDE.md). For gate commands, read .claude-sentient/profiles/{name}.yaml.

Python Environment Detection

Indicator	Environment	Command Prefix
environment.yml	Conda	conda run -n <env> --no-capture-output
.venv/, venv/	Virtualenv	Activate first or use .venv/bin/python
poetry.lock	Poetry	poetry run
pdm.lock	PDM	pdm run

Report: [INIT] Environment: conda (myenv) or [INIT] Environment: system python

Web Project Detection

Indicators	Profile	Auto-load
next.config, vite.config, react, vue, svelte	TypeScript Web	ui-ux-design
templates/, django, flask, jinja2	Python Web	ui-ux-design

Rule Auto-Loading

Rules with paths: frontmatter in .claude/rules/ are loaded automatically by Claude Code based on file context. For supplementary keyword-based loading, see .claude/rules/_index.md for the full keyword-to-rule mapping. The INIT phase does keyword matching first, then a semantic pass over .claude/rules/_index.md to catch logically relevant rules that keywords missed.

MCP Context Gathering

Server	What to Load	When
context7	Library docs for detected imports	Always (scan task-related files)
github	Issue/PR requirements, recent commits	When task references #issues or PRs
memory	Prior decisions/patterns matching task keywords	Always (search nodes)
memory	Cross-project global/org learnings	Always (search scope:global, scope:org)

Dependency Changelogs

Trigger keywords: "update", "upgrade", "migrate", "bump". Use WebFetch to load CHANGELOG.md or release notes for the dependency being changed. Prevents breaking changes from surprising you.

Governance Files

Check these exist, create from .claude-sentient/templates/ if missing: STATUS.md, CHANGELOG.md, DECISIONS.md, .claude/rules/learnings.md.

Check for CLAUDE.md — if missing, report: [INIT] No CLAUDE.md found. Run /cs-init to create context architecture.

Check for CLAUDE.local.md in project root. If missing and first session, note: [INIT] Tip: Create CLAUDE.local.md for personal preferences (gitignored). Template: .claude-sentient/templates/CLAUDE.local.md

• Cross-model workflow: For high-stakes changes, consider Claude + Codex QA review. See references/cross-model.md

Gotchas

• Conda env not activated: If environment.yml exists but commands run in system Python, the conda run -n <env> prefix was not applied. Always check for this indicator before running any Python commands.
• nvm FUNCNEST explosion: On zsh systems with nvm, bare node triggers recursive shell function loading (500+ _nvm_load command not found lines). Hook commands must use absolute node paths via process.execPath. The fixHookPaths() function in session-start.cjs self-heals this, but if hooks fail with FUNCNEST errors, check that settings.json hook commands use absolute paths.
• session_start.json stale or missing: The profile field in session state may be from a previous session or absent on first run. Always fall back to scanning indicator files (pyproject.toml, tsconfig.json, go.mod, etc.) when the state file seems stale.
• Two-step rule loading: Keyword matching alone misses logically relevant rules (e.g., error-handling for a debug task that uses no keyword triggers). The semantic pass over rules/_index.md catches these — don't skip it.
• Windows MCP servers: MCP servers require cmd /c npx wrapper on Windows, not direct npx. See learnings.md for the add-json pattern.