Claude Code v2.1.292 Adds Sub-Agent Effort Control, Mod Autocomplete, and One-Step Marketplace Installs

Anthropic shipped Claude Code v2.1.292 on October 6, 2026, delivering three developer-facing capabilities alongside a round of critical security hardening. The release adds sub-agent effort control so you can set a reasoning budget per spawn, a prompt.autocomplete hook that lets mods and plugins enrich the prompt box, and a --marketplace flag on claude plugin install that collapses marketplace discovery and installation into a single step.

Sub-agent effort control. The Agent tool now accepts an effort parameter, decoupling a sub-agent’s reasoning budget from the main session. If your lead agent runs at medium effort but needs a sub-agent to think carefully through a complex refactor, you can spawn it with effort: "high" for that one task without changing the session default. Conversely, a straightforward file-search sub-agent can run at low effort while the main session stays at medium. This puts reasoning-cost control where it belongs: at the granularity of individual agent spawns, not the whole conversation. For teams running fleets of parallel sub-agents, the parameter lets you ration expensive deep-reasoning passes to the tasks that genuinely need them.

Mod autocomplete and prompt caching for mods. The new prompt.autocomplete hook event lets any mod or plugin add rows to the prompt box’s autocomplete list. A plugin can surface recent file paths, saved prompt templates, or project-specific completions right in the typing UI. For example, a documentation plugin could offer boilerplate section headings, or an issue-tracker plugin could complete ticket references as you type.

Paired with this, $.model.complete for mods now supports prompt caching. The prompt and system parameters accept blocks of text with cache: true on a block, caching the request up to it. This cuts token costs on repeated model calls from mods — useful for plugins that make frequent classification or summarization calls with a stable system prompt and changing user input.

One-step marketplace install from the CLI. claude plugin install --marketplace source-name adds a marketplace under the same policy checks as claude plugin marketplace add, then installs the plugin from it. Before this release, adding a marketplace and installing from it were two separate commands. The combined command streamlines team onboarding scripts and README install instructions that previously had to explain both steps.

Workflow agent visibility in hooks. The agent.spawn mod hook now receives workflow agents with their run identifier and index, so a mod can inspect which workflow agent is being spawned — and refuse it. For fleet operators, this means organizational policy can gate agent spawns at the hook level. A plugin can enforce rules like “never spawn a workflow agent with network access in this repository” before the agent even starts.

Security hardening. v2.1.292 patches several serious vectors. Sandboxed commands can no longer read the staged file copies of /ultrareview uploads stored under ~/.claude/seed-admin. A managed sandbox read-deny path that materializes or re-points mid-session now correctly drops project grants inside it and ends credential injection from files it covers. Notebook and PDF reads on macOS and Windows are hardened against mid-read symlink swaps that could return files the user never approved. A tampered on-disk cache of server-managed settings can no longer switch off or unseat the built-in policy plugin while the settings fetch is failing. The PreToolUse hook and auto mode can no longer bypass permission prompts for file reads from network (UNC) paths. And rm -rf on the 8.3 short name or any alternate Windows spelling of the home folder or a drive is now correctly treated as removing it — closing a sandbox escape via filename aliasing.

The release also ships a new CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS environment variable for tuning backoff on 529 overloaded responses, and fixes over two dozen additional issues: NO_PROXY is now respected when HTTPS_PROXY is set, plan mode is properly restored on session resume, one-shot claude -p and SDK runs wait for background commands instead of killing them after five seconds, and MCP tools with names longer than 128 characters no longer break every request. The full list is in the release notes and the official changelog.

To upgrade, native installs auto-update in the background. Homebrew users run brew upgrade claude-code, and WinGet users run winget upgrade Anthropic.ClaudeCode. The release is available across every Claude Code surface — terminal, IDE extensions, desktop app, and web — and the version history is tracked on the releases page. For a deeper dive into Claude Code’s capabilities, see our power user’s field guide.