v2.6.2 — Focused tests and reusable verification

Buddy keeps testing focused on distinct uncovered failures.

• Reuses or extends existing tests, with new cases for meaningful boundaries and partial failures.
• Reuses valid passing checks across agents and verification stages; reruns checks when relevant changes invalidate them.

v2.6.1 — Review effort that fits your change

Buddy now matches review and model effort to the size of your change. Small wording edits stay with the host, narrow behavior changes get a balanced review, and substantial work keeps deeper checks.

Review that fits the change

• Trivial edits need no subagent, independent review, or specification amendment. Each new request after a completed run is classified on its own.
• Small changes get one balanced review. Low-only fixes use host verification; another optional review needs your agreement after a short cost note. Security-sensitive work retains stronger review.
• Requests such as “no review,” “just change the text,” and “quick fix” skip review and subagent dispatch for that request. Required checks still apply.
• Validation starts with affected tests for trivial and small changes. Full suites run for substantial work or shared-code changes, subject to repository requirements.

v2.5.2 — Blocked heredocs no longer stall your agent

Agents no longer stall when Buddy's shell guard blocks a heredoc. Instead of stopping the task, the agent now gets a safe way to finish it.

• Blocked heredocs no longer stop your work: when the guard blocks a shell heredoc, the agent writes the file with its file-edit tool and runs the other commands without the heredoc. Commit messages, pull request bodies, and release notes go through a file with git commit -F, --body-file, or --notes-file.
• Same protection: the guard blocks the same heredoc forms as before. Only the guidance to the agent changed.

See the README for update instructions.

v2.5.1 — Research briefs that stay accurate

This patch keeps Buddy's research briefs accurate as your questions change, and the README now walks first-time readers through Buddy step by step.

• Research briefs stay current: when new evidence invalidates an earlier finding, the brief corrects or removes it and says so in the update. Earlier questions that still affect your work stay in the brief when you ask a new one.
• No lost questions at the length limit: a question that is still relevant moves to a linked note instead of being dropped.
• Clearer README and workflow guide: the README is restructured for first-time readers, and a new workflow guide explains how Buddy carries work from research to verified code.

See the README for update instructions.

v2.5.0 — Live Buddy progress in Claude Code

Buddy now shows its progress live in Claude Code. While develop or any Buddy skill runs, you can see the stage in work, spec phase progress, busy subagents, and the run's cost without leaving the prompt.

Buddy mod for Claude Code

• Live progress above the prompt: the stage track from Research to Review, skipped optional stops, and a progress bar for the spec phase in work.
• Subagents and cost: the subagents on the stage, plus tokens and cost per stage and per phase. Subagent costs are estimates at API list prices.
• Details pane: select Details or run /buddy-mod to see each phase's task checklist and each subagent's tokens and cost.
• Desktop and terminal: an animated pixel card in the Desktop app, the same information as text in the terminal. The mod adds nothing to your prompts or context.

Requirements

The mod needs Claude Code v2.1.287 or later in the terminal, or v2.1.286 or later in the Desktop app. It is on by default; to turn it off, set the Buddy option Show the Buddy mod to off. See the README for details and update instructions.

v2.4.4 — Fewer false heredoc blocks in the shell guard

This patch stops the shell guard from blocking common agent commands that only pass text through a heredoc. The guard still checks every command that runs.

• Commit messages, pull request bodies, and release notes from heredocs: git commit -F -, gh pr and gh issue create, edit, and comment, gh pr review, and gh release create and edit can now read their text from a heredoc. gh api and other subcommands that act on that input stay blocked.
• Python stdin scripts with flags and arguments: commands such as python3 -I - README.md <<'PY' now work with the safe flags -I, -E, -s, -B, and -u. Options that supply code, such as -c and -m, stay blocked.
• Pipelines after a heredoc: commands after the last heredoc can now use pipes and stderr redirection, for example git diff --stat | tail -3.
• << in quoted text: a quoted pattern such as grep -c '<<<<<<<' README.md is no longer treated as a heredoc, also in pipelines and inside bash -lc.

See the README for update instructions.

v2.4.3 — Uninterrupted implement runs and fewer heredoc blocks

This patch keeps implement runs moving when Cursor refuses a native Goal, and lets the shell guard accept common heredoc file writes that it blocked before.

• Implement runs continue without a Goal: if Cursor refuses the Goal or the Goal tool is missing, Buddy keeps the outcome as a checklist and continues. Workers are no longer stopped and restarted, and a missing Goal confirmation no longer ends the run.
• Fewer false heredoc blocks: agents can now write a file from a heredoc, including after a cd into a folder, without a false block. The guard still checks every command.

See the README for update instructions.

v2.4.2 — Better heredoc handling and focused research

Buddy now handles supported quoted heredoc input without false shell-syntax denials, keeps research findings short and current, and provides clearer plugin installation and update instructions.

• Fewer false command blocks: direct cat, python -, and python3 - commands can use supported quoted heredocs containing apostrophes. The guard still checks commands after the closing delimiter.
• Stricter heredoc checks: expanding or ambiguous heredocs, execution wrappers, and shell or database consumers are blocked for manual review. Existing commands that use these forms can now require manual execution; see the README for supported syntax and limits.
• Focused research and innovation: research artifacts aim for 300 words and stay within 500 unless you request a longer report. Updates replace stale findings and remove resolved questions; innovation options share the same document limit.
• Clearer plugin updates: the README explains installation and updates for Codex desktop and CLI, Claude Code, and Cursor using each harness's native workflow.

See the README for update instructions and the supported heredoc syntax.

v2.4.1 — More focused skill instructions

Buddy's skills now use shorter, more focused instructions, leaving more room for your project context during development, research, planning, review, and reporting.

Repeated guidance has been consolidated while keeping workflow boundaries, permission rules, model preferences, and verification requirements intact. Supporting references are loaded when the task needs them.

v2.4.0 — Automatic model selection, with your choices first

Use Buddy without maintaining a separate model configuration. Skills guide Codex, Claude Code, and Cursor to select from their available models for fast, balanced, and frontier work.

• General work recommends balanced; specification and code review recommend frontier. Your coding tool can choose another tier when the task warrants it.
• Override any skill or stage by naming an available agent, tier, model, or reasoning effort. Your scoped request takes precedence, including a lower-capability choice.
• If your tool cannot honor an explicit choice, Buddy reports the limitation instead of silently substituting another setting.

Model preferences

Specify preferences directly when invoking a skill. Saved Buddy model profiles do not control selection, and the configure-models and model-policy skills are no longer available.