agent-tools

Commit task changes from a worktree onto the base branch (read from KANBAN_BASE_BRANCH env var). Use when finishing a kanban task to stage, commit, and cherry-pick changes back to the base ref.

Recover orphaned/dangling commits from kanban headless runs where the AI agent committed in the worktree but the kanban-commit agent failed to cherry-pick to the target branch (commit exists as a dangling object unreachable from any ref). Use when a ai-task-start-kanban-headless run completed/finished but the commit is missing from the target branch, or when investigating "nothing to commit — working tree is clean" in kanban-commit agent output.

Capture a new feature, bug, or tweak idea as an issue file in the project's kanban repo at <project>-kanban/issues/<id>-<slug>.backlog.md. Runs a short grilling pass against the existing domain model so the entry is sharper than a one-liner. Uses sequential IDs (via flock(1)) so files sort naturally even with concurrent agents.

Plan and execute a full git history squash job — analyze commits between two refs, group them into ~10 semantically meaningful squashes, drop irrelevant directories or paths, cherry-pick each group with conflict resolution onto a new branch, and verify the final tree matches the original modulo dropped paths. Use when the user says "reduce commits", "squash branch history", "clean up git history", "rewrite branch", "compress commits", or wants to consolidate a messy branch before merge.

DEPRECATED — The backlog/issue distinction no longer exists. All work items are issues (<project>-kanban/issues/<id>-<slug>.<state>.md) from creation. Use backlog-add to capture an issue, to-prd to create a PRD. This skill is kept for reference only.

Read the last pi agent session for the current project, extract what was discussed or done, and create issue files (as <project>-kanban/issues/<id>-<slug>.backlog.md) for any work items, bugs, or follow-ups it uncovered. Use when a session ended without wrapping everything up, when the user wants to capture session outcomes, or when you need to preserve context across sessions.

ELIMINATED — Issues are created directly via backlog-add or manually. PRDs reference issues by ID in their frontmatter. This file is kept for reference.

Turn the current conversation context into a PRD with its own sequential ID and save it as <project>-kanban/prds/<id>-<slug>.md. Preserves history — old PRDs are never overwritten, they coexist. Use when user wants to create a PRD from context.

Write, test, and model-check Quint formal specifications for code. Use when the user asks to create a Quint spec, formally verify logic, model-check state machines, or validate code correctness with Quint.

Check whether a Unity project's C# scripts currently compile by rerunning the cached Bee `ScriptAssemblies` target from the latest Unity Editor DAG. Use when an agent needs a fast build verdict for Unity scripts, wants compiler errors without launching a full editor-driven build, or needs to confirm whether recent code changes introduced C# compilation failures in a Unity project.

Add a simple Unity Edit Mode unit test to an existing project, run it through the repository's `./unity-test` daemon workflow, inspect `status`, `results`, and `events`, and report whether the new test definitely ran and passed. Use when a user asks Codex to create Unity tests and verify them through `./unity-test` instead of only editing files.

Apply the current Git working-tree diff (including untracked files) into the main repository directory resolved from git-common-dir. Use when the user asks to mirror, sync, or apply local changes from the current nested repository/worktree into the parent repository with a generated patch.

Run evidence-first debug investigations that require runtime logs before code fixes. Use when a user reports a bug, exception, stale-reference issue, lifecycle race, or other behavior that cannot be fixed safely from static code inspection alone, especially when the workflow should generate hypotheses, add targeted instrumentation, collect reproduction logs, evaluate hypotheses with cited log evidence, verify the fix with before/after logs, and keep instrumentation until success is proven.

Iterate on UI visuals by editing code, hot reloading, capturing screenshots, and refining in loops.

Ensure implementation plans and code changes include automated tests (unit/integration) for every new scenario or use case; use when asked to guarantee tests, enforce coverage, or create plans that must include tests.

Code review and design critique using TigerBeetle's TigerStyle principles (safety, performance, developer experience, zero technical debt). Use when the user asks for a "tiger style" review of unstaged changes, a commit, or a branch comparison, or explicitly requests TigerStyle guidance. Apply across languages and adapt Zig-specific rules to the project's conventions.