qe-framework

Use when collecting verified project-local coding guidance into .claude/skills based on the project tech stack, with TTL and user-edit protection.

Shows QE Framework usage overview. With no arg, prints the full catalog. With a skill name arg (e.g., /Qhelp Qcommit), reads that skill's SKILL.md and summarizes it in the user's language. Also supports find/search skills mode via reference/find-skills.md.

QE framework (Query Executor) initial setup. Creates the QE state directory, client-specific instruction/config artifacts, and .gitignore entries in a new project, then auto-analyzes the project. Use when the user wants to initialize a project or set up the framework.

Diagnoses and repairs QE installation health across qe-framework, MCP client config, and the project .qe directory. Use for 'doctor', 'check QE health', dependency verification, corrupted .qe state, or repair guidance. Distinct from Qupdate, which updates installed assets.

Use when executing a TASK_REQUEST/VERIFY_CHECKLIST or when already-changed code needs the test-review-fix quality loop. With no flag it auto-selects sequential or parallel-wave execution; `-verify` runs verification. Use Qautoresearch for iterative code-optimization loops and Qscenario-test for scenario/E2E.

Use when a plan or task needs spec documents: TASK_REQUEST and VERIFY_CHECKLIST. Use Qplan for roadmap/phases; use Qexecute after the spec exists.

Use when the user invokes Qgs or a Qplan handoff needs TASK_REQUEST + VERIFY_CHECKLIST generation. Alias for Qgenerate-spec. Use Qplan for roadmap; use Qexecute after specs exist.

Unified MCP operations for QE: check MCP client/server status, guide MCP server setup, and synchronize generic MCP config across supported clients.

Use when the user wants to plan work, create/update a roadmap, choose phases, or move to the next phase. Use Qgs/Qgenerate-spec for TASK_REQUEST specs; use Qexecute/Qexecute for execution.

Manage the QE shadow-git snapshot store — create, list, diff, restore, and prune isolated shadow snapshots. Use when the user wants to checkpoint, review, or recover working-tree state without touching any real git repo.

View and modify SIVS engine routing configuration. Change which engine (claude/codex) handles each stage (spec/implement/verify/supervise). Use when the user wants to check or change SIVS routing, engine config, model assignment, or effort level.

Updates QE marketplace metadata/cache, the QE Framework body, installed Codex assets, and the codex-plugin-cc bridge using the correct path for the current install. Use for "update plugin", "upgrade", "update qe", "update codex", or "codex plugin".

Delegate a SIVS claude-stage from a Codex session to the local Claude CLI (reverse of /codex:rescue). Use when a Codex base session must hand spec/verify/supervise work to Claude.

Critical thinking verification for SIVS stages. Spawns adversarial sub-agents to stress-test specs, implementations, and merge readiness. Also hosts structured debate and evidence-backed risk proof modes. Use for 'review critically', 'stress test this', 'devil advocate', debates, risk proof, or auto-invoked by Qgenerate-spec, Qexecute -verify, Esupervision.

Test fixture for Qcommit policy compliance — all required sentences present, no forbidden phrases in policy.

Test fixture for Qcommit policy violation — missing required sentence and forbidden phrase present.

Test fixture for the Korean-inversion exploit — negation marker appears AFTER the forbidden phrase, so the line is a mandate, not a prohibition. Must be detected.

Test fixture for the pre-heading exploit — forbidden phrase placed before any policy heading (former section-scoping blind spot). Must be detected.

Unified garbage collection for QE: archive completed tasks, sweep stale .qe artifacts, and analyze code garbage such as doc-code drift, rule violations, and dead code.

Quoted name and description must be treated as unquoted YAML scalars, so this is a valid control (no violation).

First of two skills declaring the same name to trigger the duplicate-name violation.

Second of two skills declaring the same name to trigger the duplicate-name violation.

---

A valid control fixture whose name matches its directory and has a non-empty description.

The name field does not match the directory name to trigger the mismatch violation.

Autonomous experiment loop inspired by Karpathy's autoresearch. Repeatedly modifies target files, runs an experiment, evaluates a single metric, and keeps or discards the change — looping indefinitely until manually stopped. Use when optimizing code through iterative experimentation: ML training, algorithm benchmarks, build optimization, performance tuning, multi-file refactoring. Distinct from Edeep-researcher (web research) and Qexecute -utopia (task confirmation skip) — this skill runs a code-modify-evaluate loop.

Unified QA entry point for planning test documentation, running single-pass scenarios, and orchestrating the multi-agent QA council.

Restores saved context after compaction or session break. Use when resuming work, restoring context, continuing from where you left off, or loading a previous session.

Verify that an implementation and its tests honor a business-logic contract stored under .qe/contracts/active/. Delegates to the Econtract-judge LLM agent on cache miss; returns cached verdict on cache hit. Use when the user says 'verify contract', 'check contract', '/Qverify-contract', or when called from /Qexecute -verify.

Creates git commits that look human-written — no Co-Authored-By, no AI-style language, no emoji. Use when the user says 'commit', 'push', 'save changes', or wants to save/push code changes.

Cross-session learning memory — records and recalls learnings (mistakes, conventions, gotchas, decisions) in .qe/learnings.md, ranked by time-decayed relevance and injected top-N at session start. Superset of Qmistake.

Record mistakes so they are never repeated. When the user points out a mistake, confirms it and saves to .qe/MISTAKE.md. Loaded at every session start.

Manually refreshes project analysis data. Use when refreshing, updating, or syncing .qe/analysis/ files. Also supports --sync mode for synchronizing project source files with a reference/standard project via reference/project-sync.md.

Context preservation and session handoff. Use when the user wants to save state, handoff, save context, end session, or continue later.

File GitHub issues against the qe-framework repo from the CLI. Use for requests like "file a bug", "report an issue", "feature request", "이슈 올려줘", or "버그 리포트". Onboards first-time users with a one-time PAT prompt, then delegates to the `gh` CLI. Default target: `inho-team/qe-framework`.

Folder-aware context memory manager. Creates, views, updates, and refreshes per-folder context files in .qe/context/. Use when the user wants to manage folder-level context, optimize Claude memory, or set up context partitioning.

Manage project memory — add, list, prune cross-session knowledge. Use when asked to remember conventions, gotchas, or decisions.

Manage API keys and tokens through OS-backed secret storage so QE can use them without storing plaintext in the project. Supports metadata-only project/global registries plus secure execution with env injection.

Shows the current QE Framework version. Use when asked 'what version', 'qe version', 'show version', or 'check version'.