> ## Documentation Index > Fetch the complete documentation index at: https://platform.stepfun.ai/docs/llms.txt > Use this file to discover all available pages before exploring further. # Release Notes ## v0.1.6 October 23, 2026 ### New features * Background subagents: the subagent tool gains `run_in_background`. A background task returns a channel ID immediately and posts a notification with a summary, usage, and the sub-session path when it finishes. `agent_send` operates by channel ID and supports reply, stop, and status. The previous aliases, grouping, and fan-out addressing have been removed. The concurrency limit is 8, idle tasks stop after 10 minutes, and `[subagent] background = false` returns to foreground execution. * Lark remote connection (experimental): `/connect lark` connects a terminal session to Lark. After authorization, send a first message to the bot to bind the session, and mention the bot in group chats. Tool approvals appear as Lark cards so you can approve or reject remotely. Remote approval requires completing one callback configuration after the connection succeeds. `/connect lark off` disconnects. * One-off side questions: `/btw ` asks and answers in an overlay. It does not write to the session file or enter the main conversation's model context, and the answer is generated from a copy of the current conversation branch. Side-question history is kept in memory only (20 entries) and is used for follow-up questions. * Working directory selection: the new `-C` / `--cwd ` option pins the project directory for scripts and IDE integrations. Session storage and settings discovery follow that path, and an invalid path fails at startup. * Terminal tab busy indicator: the terminal tab shows a busy indicator while the agent works or compacts context. It is enabled by default on Otty and Ghostty, and other terminals can force it in `/settings`. Failed tool rows no longer show a recovery hint row. * Recent models: the `/model` picker gains a Recent section at the top listing recently used models, up to 10, with repeats moved to the top. The list is saved globally and shared across sessions. * Session search index: a new SQLite full-text session index powers Web and desktop search. JSONL remains the only data source and can be rebuilt at any time, and subagent sessions are excluded from search. The CLI adds `step session-index status / reindex / clear`. * Feedback reports with evidence: `/feedback` automatically attaches diagnostic evidence: a session summary with model, tokens, cost, and cache hit rate; request duration and interruption phase; archived heads and tails of very long conversations; and tool output. Everything is redacted uniformly, version and environment are recorded, and a confirmation page appears before submission. ### Improvements * Plan mode no longer invalidates the prompt cache: the cache hits on the common prefix of a request, with the tool list at the front of that prefix. Entering plan mode previously reordered the tool list and invalidated the cache for the entire conversation, measured at 233k tokens read for 1.6k returned, about 0.7%. The default tool list now stays in place, and read-only limits are enforced by an execution-time guard that blocks out-of-scope writes from editing tools. * Task status snapshots are appended only when they change: context status hints are also pinned to their original position, which keeps request prefixes more stable and improves cache reuse. * `/model` saves on Enter: Enter now selects the model and saves it as the default, and `/model ` saves as well. Ctrl+S has been removed. Previously Enter applied only to the current session, and the two paths could drift apart silently. * `run_command` titles truncate by width: when collapsed, the title shows only the first line of the command, truncated to the terminal width, so long scripts and heredocs no longer fill the session record. Expanding still shows the full command. ### Fixes * `step update` always failed on Windows: the update failed even after the download and checksum succeeded, because the updater looked for `step.exe` in the outer directory of the Unix package while the Windows zip is flat. The root directory is now determined from the actual location of the binary. * Update failures are diagnosable: installation failures include the last 20 lines of installer output, noting how many lines were omitted, instead of only an exit code. * Non-interactive commands return early: in `-p` / JSON mode, commands such as `/plan ` destroyed the session mid-turn, reporting ctx is stale and not writing the session file. They now wait for the turn to finish before returning. * `search_files` line-number misjudgment: matched text containing `:number:` was treated as a line number, reading the wrong source line or reporting no match. Structured match results are now read directly instead of re-parsing rendered text. * `find_files` dropped paths: paths starting with `[` or `WARNING:` were mistaken for notifications and discarded. The structured match list is now read first. * workflow `phase()` ran zero times silently: `phase(title)` is only a milestone label, so passing agent calls to it inside a callback ran nothing while reporting completion. A static check now rejects this and points to the line number. * Compaction retries: compaction retries on incomplete response streams when SSE disconnects early. If it still fails, redacted diagnostics are written to the session file. ## v0.1.5 October 16, 2026 ### New features * Usage panel: the new `/usage` command provides Overview and Account tabs. Overview shows a recent activity heatmap and daily token usage, while Account queries your Step Plan remaining allowance and reset time. `/usage overview` and `/usage account` open the corresponding tab directly. * Status panel: the new `/status` command shows the current model, thinking level, permission mode, working directory, and account information on one screen. * Task self-review: in `-p`/JSON non-interactive mode, `--completion-review` has the agent review its work against the original task before wrapping up: whether everything requested was done, whether interfaces and boundaries were covered, and whether tests ran. Issues are fixed before the summary. ### Improvements * Post-edit verification: `then_run` for `edit_file` and `write_file` now uses the same execution path as `run_command`, with the same timeout, output limits, and approval rules. * Plan approval without an interface: in `-p`/JSON non-interactive mode, submitting a plan when the user cannot be asked directly now returns the plan content and stops. ### Fixes * Fix startup spam when the same skill is installed in both `~/.agents/skills` and `~/.claude/skills`. * Fix sign-in for accounts without a subscription: these accounts could not use Step Code after signing in and are now guided to use an API key instead. * Fix workflow retry and rate-limit misclassification by judging error type rather than error text. Also fix workflow and parallel subagents potentially exceeding the concurrency limit when running together. * Fix a single call failure during resume causing all subsequent calls to rerun. Failed nodes and callback errors now appear in the final result. ## v0.1.4 October 12, 2026 ### New features * Add CLM (Context Language Models) context management. When a long conversation approaches the context limit, CLM is enabled by default and the model selectively condenses earlier context without overwriting the original session log. If condensing fails, it falls back to the existing summary compaction. `/clm` provides the `status`, `on`, `off`, `diff`, and `reset` subcommands to manage the state, and `/clm-compact` asks the model to reorganize the context. Model usage counts toward session statistics. * Change permission modes to Ask, Approve for Me, and Full Access. Ask prompts before modifying files or running commands. Approve for Me runs routine operations automatically and still asks for confirmation on detected dangerous commands or commands it cannot fully analyze. Full Access runs risky commands directly without asking each time. Enabling Full Access requires acknowledging the risk, and you can apply it only to the current session or remember the choice. ### Improvements * Use `/effort` to set the thinking level. The `/thinking` command has been removed. * Reopening the same session restores its Ultracode setting and status bar indicator. New sessions default to off. ### Fixes * Fix a missing `step` command after reinstalling or upgrading and opening a new terminal. * Fix the same plugin being detected twice when starting from the home directory, which incorrectly reported a plugin override conflict. * Fix MCP servers disabled in the global configuration being rediscovered by a same-named plugin declaration. * Fix tools that had finished while the interface still showed Running, Reading, or Editing. * Fix sessions from other projects appearing in the session list, and `--continue` resuming the wrong project session, when different project paths mapped to the same session directory. * Fix subagents returning incomplete results for multi-step tasks, not continuing after a background task finished, and truncating long reports. Also fix incorrect task status, list flicker, and notification spam. * Fix commands running in a WSL environment when Git is installed in a custom directory, which caused tools installed on Windows to be reported as missing. * Fix background commands in Git Bash on Windows exiting with a failure and producing an empty log. ## v0.1.3 October 9, 2026 ### New features * Add remote Step Plan sign-in. On SSH, remote servers, or containers, `step login --no-browser` returns a full authorization link that you can open in a browser on this or another device. Sign-in no longer requires a local callback port. * Add the optional `then_run` parameter to `edit_file` and `write_file`. After an edit succeeds, the verification command runs immediately and returns its result, so a separate verification call is no longer needed. * Add the optional `--completion-check git-committed` option. In `-p` or JSON mode, it checks whether the changes in the current run are committed, whether uncommitted files remain, and whether a final answer was provided. When a condition is not met, the agent is reminded in the same session to finish the work, with up to two follow-up rounds by default. Combine it with `--completion-review` for a self-review; both share the same follow-up rounds. * Add the `/ultracode` command to enable, disable, and check the status of multi-agent collaboration. The existing `/ultraloop` command still works and behaves the same. ### Improvements * Store plan files outside the workspace with `STEP_CODING_AGENT_PLAN_DIR`. Plan files stay out of code commits and do not affect workspace status checks. ### Fixes * Fix auto-discovery of existing skills. `~/.claude/skills`, `.claude/skills` in trusted projects, and `~/.stepcode/skills` now load without manual path configuration. * `run_command` now uses a 120-second default timeout, adjustable with `timeout_ms`, which reduces stalls caused by commands that stop responding. * Fix lingering child processes on Linux that left tasks waiting, and report commands terminated by a system signal as failures. * Fix truncated command output losing the full log path or exit status. The exit code, timeout, or abort reason is preserved so the agent can read the full log and continue troubleshooting. * Fix earlier history summaries being dropped after repeated context compaction, so long tasks keep using previously summarized work. Also fix summaries that exceeded the model output limit. * Fix inconsistent welcome text and command entry points for Ultracode, missing help and command completion, and the enablement state not being visible at common terminal widths. ## v0.1.2 September 30, 2026 ### New features * Add `--non-interactive-denial continue`. In non-interactive runs, such as `-p` mode, a dangerous operation that requires approval is blocked and the reason is returned to the agent so it can attempt a safer approach and continue the task. The default remains `terminate`, which ends the task. Both policies enforce the existing permission restrictions. ### Improvements * Add search to marketplace plugin lists and remove duplicate entries with the same name. Adding a marketplace in the interactive interface now shows progress, supports cancellation, and opens the plugin list on success. * Show a preview of long tool results with a file path to the full text, reducing model context usage. * Display only the most recent successful task plan update in a conversation, including when resuming a session. * Improve Up Arrow history navigation for single-line input. Add `Ctrl+End` in fullscreen mode to return to the bottom and follow new output. ### Fixes * Fix the thinking level being forced to the highest setting when starting or switching to a Step model. * Fix skill scanning hangs and skill commands failing to recognize arguments entered on a new line. * Fix loaded skill instructions being lost after built-in context compaction. * Retry compaction after trimming older history when the summary request exceeds a limit. * Fix skills and prompt commands from installed plugins not loading. Reload resources to make them available. * Fix plugin MCP configuration recognition, server name recognition, and reuse of login authorization. Local MCP servers now start from the plugin directory by default. * Fix incomplete loading of paginated MCP tool lists and search failing to find available MCP and extension tools loaded later. Tool lists update in response to server notifications. ## v0.1.1 September 24, 2026 ### New features ### Improvements * Reduce redraws in the subagent task list to minimize flicker and improve readability. * Add a post-installation prompt on macOS and Linux to guide users to run `step`. ### Fixes * Fix Workflow and Ultraloop being unavailable in release installations. The workflow sandbox now uses QuickJS WebAssembly instead of `isolated-vm`. * Fix unexpected exits and false timeouts in Workflow involving unawaited asynchronous tasks, long-running parallel tasks, and resource cleanup. * Correct the billing route for Step Plan web search to use the subscription endpoint. * Fix plugin diagnostics incorrectly reporting a missing `STEPFUN_API_KEY` for users who are already signed in. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.