Codeep

Two big quality-of-life additions: Anthropic prompt caching is on by default (60–90% cheaper on cache-eligible input), and /plan lets you preview an agent's full plan before any file gets touched. Run /go to execute, or /plan <revised task> to refine.

Added — Anthropic prompt caching, automatic

• Two cache breakpoints per request: the system prompt (and embedded
  skills catalog / project intelligence) and the tools array. Cache hits
  bill at 0.1× the input rate; cache writes at 1.25×. Net win after the
  second same-shape request, which is every iteration in an agent loop.
  Below 1024 input tokens Anthropic silently skips caching — no error
  path. Applies to the agent chat path, the agent fallback path, and
  the chat() path used by /agent and inline replies. Also propagates
  through OpenRouter → Anthropic routes (caching headers honoured
  upstream).
• TokenUsage.cacheCreationTokens + cacheReadTokens fields
  surfaced on every record. getCacheStats() aggregates per-session
  cache hits, misses, and estimated USD savings vs running without
  caching. /cost (and /stats) renders a new "Prompt caching"
  section when at least one cached call landed.

Added — Plan mode (/plan + /go)

• /plan <task> — generates a numbered plan for the task (no tool
  calls, no file changes), surfaces it as a Markdown message so you can
  review what the agent would do, which files it would touch, what
  commands it would run, and the risk level it self-assesses. Holds
  the (task, plan) pair as the pending plan, scoped to the current
  process. Re-running /plan <revised task> replaces the pending plan
  with a new one (you pay one extra LLM call but get readable revision
  history in the chat).
• /go — executes the pending plan: hands the task + approved plan
  as a single prompt to the regular agent loop, so all MCP tools,
  lifecycle hooks, verification, permissions, and skill bundles apply
  unchanged. Includes an explicit anti-improvisation clause in the
  injected prompt — if any step turns out to be wrong mid-execution
  the agent must stop and report rather than silently rewriting the
  plan.
• Available in both the TUI and ACP clients (Zed, VS Code). ACP
  /plan streams the plan back via session/update; ACP /go runs
  the agent inline and streams iterations through onChunk.
• Surfaced in /help ("Agent Mode" section) and / autocomplete.

Fixed

• Anthropic streaming usage extraction missed cache fields. Both
  the agent stream handler (utils/agentStream.ts) and the chat
  stream handler (api/index.ts) now pick up
  cache_creation_input_tokens and cache_read_input_tokens from the
  message_start event, so cached requests no longer undercount
  prompt tokens or display $0 savings.

Notes

• OpenAI-format providers (OpenAI direct, Z.AI, DeepSeek, MiniMax,
  Ollama) don't expose explicit cache markers — those providers
  generally apply automatic prefix caching server-side. No code change
  on our end needed; cost reports stay accurate via standard
  prompt_tokens accounting.