Changelog

This page is the source of truth for what’s in each Skep release. If a concept or reference page says “the current release” without a version number, that’s because the answer lives here.

v0.2.0 — current

HTTP API sidecar. Each daemon now listens on a random loopback port (127.0.0.1:<port>) alongside its Unix socket. The URL and a randomly-generated bearer token are written to .skep/http.json (mode 0600). Clients that cannot speak the Unix-socket protocol — VS Code extensions, webhook receivers, and future remote surfaces — hit the HTTP endpoints with the same JSON verbs the socket accepts. Everything stays on 127.0.0.1; there is no network exposure.

Step-level task execution. When a task is approved, its plan is materialized into a new task_steps table keyed on (task_id, seq). The executor then shells out per step instead of once per task, with a per-step retry, per-step commit-SHA capture, and a first-failure-stops-task policy. skep task show <id> renders each step with status glyphs (✓/✗/→/·), verb, target file, commit SHA, duration, and retry count. See Task lifecycle for the full flow.

Per-verb model routing (infrastructure). New step_model_by_verb config key — a map from plan verb to model name — rewrites the per-step command template to inject --model <value> at dispatch. Empty default routes everything to model; populate the map once you have data on which verbs need which model. See Config reference.

MCP parity for local task verbs. Five new tools in skep mcp: show_task, approve_task, reject_task, clarify_task, and delete_task. Classifying LLMs can now drive the whole local task lifecycle over MCP instead of dropping to the CLI mid-session. See MCP tools reference.

CLI polish. --json is now supported on every write verb (task create / approve / reject / clarify / delete), not just reads. skep task create - reads the task description from stdin. skep completion <bash|zsh|fish> emits a shell completion script you can source or drop into the appropriate completions directory.

Integration test harness. Internal: a multi-repo fixture runner lives in internal/testkit and drives the CLI end-to-end against temporary workspaces. Not user-facing; unblocks future releases.

Known limitations (still).

• Claude Code is the only wired LLM backend. Gemini / Codex presets remain gated on end-to-end testing — the preset plumbing is in internal/llm/presets.go and ready to flip on.
• Windows is not supported natively. Use WSL2.

v0.1.0

LLM backend. Ships with Claude Code as the configured LLM CLI. Classifier defaults to claude-haiku-4-5-20251001, plan-gen and executor default to the model you set with skep config model (typically sonnet or opus).

Task lifecycle. Eleven states: created, classified, pending, pending_clarification, approved, queued, executing, done, failed, interrupted, rejected. See Task lifecycle for the full state machine.

MCP tools. Eleven tools exposed by skep mcp serve: get_overview, search_symbols, get_call_graph, get_file_context, list_tasks, create_task, dedup_task, create_remote_task, get_remote_task, approve_remote_task, wait_remote_task. See MCP tools reference.

Indexer. Native tree-sitter for Go, TypeScript, JavaScript, Python, Kotlin, Java. Line-based parsers for YAML, Terraform, Dockerfile, JSON. universal-ctags fallback for Rust, Ruby, Elixir, Lua, C, C++, C#, PHP, Swift, R, Shell, Bash, SQL, Protobuf, Haskell, Scala, Perl, OCaml, Nim, Crystal, Make, CMake (22 languages via ctags when the binary is available; without ctags, those files are indexed by filename only).

Cockpit. skep cockpit setup writes a tmux config snippet to ~/.tmux.conf.d/skep.conf. See The workspace cockpit.

Earlier

v0.1.0 is the first public release. Internal releases before v0.1.0 are not documented here.