v2.14.0

Released July 2, 2026 · View on GitHub →

Claude Sonnet 5 replaces Sonnet 4.6 in the Anthropic picker (1M context, the full low→max reasoning-effort range), plus correctness fixes from a full-project audit: the ACP edit-approval diff, per-session naming, Grok effort gating, OpenAI-protocol cache accounting, and a /copy edge case.

Added

Anthropic model list (and the OpenRouter seed). Same $3/$15 pricing and 1M context window; graded /thinking effort (lowmax). Sonnet 5 rejects non-default sampling params, so a custom /temperature is now omitted for it (it would otherwise 400) — matching the Fable 5 / Opus 4.7+ handling.

(and the OpenRouter seed) now that it's available again — Anthropic's most capable model, for the hardest reasoning and long-horizon agentic work. 1M context, full lowmax /thinking effort, $10/$50 pricing in /cost. The sampling-param and effort handling were already wired up, so this just re-lists it in the picker.

Fixed

VS Code branched on old_string/new_string, but the edit_file tool emits old_text/new_text — so in Manual mode you approved an edit seeing only the file path, not the change. Now reads the real parameter names.

derived display name wasn't cleared on /new, so syncSession/reportStats attributed the fresh, unrelated session to the old name on the dashboard.

default) and grok-code-fast are coders, not reasoning models — the param 400s, and a 400 there silently dropped the turn into the weaker text-tool fallback. They're now excluded from the effort gate.

ignored prompt_tokens_details.cached_tokens, so DeepSeek/OpenAI cache hits were estimated at the full input rate instead of the ~0.1× cache-read rate.

cleared the clipboard instead of showing "Invalid block number"; now guarded.

stream_options.include_usage was only requested for the literal openai provider, so DeepSeek/Kimi/Grok/Qwen/GLM/… streamed with no usage block and the whole turn logged 0 tokens / $0.00 in /cost and the dashboard. Now requested for every OpenAI-compatible provider (as the plain-chat path already does).

Token tracking was destructively reset at the start of each run to compute the cloud-telemetry delta, so the status bar and /cost lost all prior session usage (and in the ACP server, /cost after a 2nd prompt showed only the last one — or nothing). Now uses a non-destructive marker: cloud stats still get only the run's delta while the session-cumulative total survives.

one process-wide record buffer, so two sessions on one process (e.g. VS Code "New chat" while another turn streams) clobbered and cross-reported totals. Each ACP session now accumulates into its own buffer (via AsyncLocalStorage), isolating usage while keeping the single-session TUI path unchanged.

Downloads

Install with npm install -g codeep@2.14.0 or brew install vladoivankovic/codeep/codeep.