Claude Code v2.1.178 adds tool parameter matching and nested directory wins

Claude Code `v2.1.178`, authored by ashwin-ant, delivers several notable feature additions. The new `Tool(param:value)` permission syntax lets users match a tool's input parameters — including `*` wildcard support — so rules like `Agent(model:opus)` can block specific subagent model choices. Skills in nested `.claude/skills` directories now load automatically when working on files in those directories; on a name clash, the nested skill is surfaced as ` : ` so both versions remain accessible. For nested `.claude/` directories broadly, the agent, workflow, and output-style closest to the working directory now takes precedence when names collide, and project-scope workflow saves target the closest existing `.claude/workflows/` folder.

The `/doctor` command gains a consistent flat tree layout with clearer section status icons and highlighted command names.

Auto mode receives a safety improvement: subagent spawns are now evaluated by the classifier before launch, closing a gap where a subagent could request a blocked action without review. The `/doctor` command gains a consistent flat tree layout with clearer section status icons and highlighted command names. The workflow prompt keyword now uses a purple shimmer highlight and triggers only on explicit phrases like "run a workflow" or "workflow:", rather than any mention of the word. Remote Control error messages are improved, with connection failures showing a persistent red "/rc failed" footer indicator and the "not yet enabled" error now explaining whether the cause is a gate, a check failure, stale entitlement, or org policy. The `/bug` command now requires a description before submission and no longer uses model-refusal text as the GitHub issue title.

On the fix side, the release resolves an out-of-memory crash when the CLI inherits a stale websocket/OAuth file-descriptor environment variable from a parent process, and fixes Claude in Chrome silently failing to connect when the OAuth token belongs to a different account than the Claude Code login. Subagent reliability is improved: viewing a subagent's transcript now shows tool results and live progress, messages sent while it finishes its turn are no longer dropped, and backgrounding a running subagent with ctrl+b no longer restarts it from scratch. Additional fixes address `401 Invalid bearer token` errors for `claude agents` workers using `ANTHROPIC_BASE_URL` and `ANTHROPIC_AUTH_TOKEN` with a custom API gateway, compaction not honoring `--fallback-model`, stale cached request configurations causing persistent auth errors after credential refresh, MCP server-level specs in subagent `disallowedTools` being silently ignored, vim mode undo merging commands into a single step, and a VSCode issue where pressing Esc to dismiss a CJK IME candidate window canceled the running Claude task.

Key facts

1. 01New `Tool(param:value)` syntax for permission rules matches a tool's input parameters, with `*` wildcard support — e.g. `Agent(model:opus)` blocks Opus subagents.
2. 02Nested `.claude/` directories now use a proximity rule: the agent, workflow, or output-style closest to the working directory wins on name collisions.
3. 03Skills in nested `.claude/skills` directories load automatically; name clashes surface both as ` : `.
4. 04Auto mode now evaluates subagent spawns via the classifier before launch, closing a gap where blocked actions could be requested without review.
5. 05Fixed an out-of-memory crash when the CLI inherits a stale websocket/OAuth file-descriptor environment variable from a parent process.
6. 06Fixed `401 Invalid bearer token` errors for `claude agents` workers using a custom API gateway via `ANTHROPIC_BASE_URL` and `ANTHROPIC_AUTH_TOKEN`.
7. 07Fixed compaction not honoring `--fallback-model`, and fixed MCP server-level specs in subagent `disallowedTools` being silently ignored.

Topics