BenchBox

Shared protocol for review-shaped actions, authorization scope, defect routing, L2 audit handling, and local-only capture.

Use when the user asks to "test TPC-H", "check compliance", "review architecture", "run quality checks", "check binaries", "test dialect translation", "compare implementations", "run live platform tests", or "plan and execute" a benchmark feature.

Use for "commit code", "review code", "fix lint/type error", "improve performance", "compare code", "shrink code", "generate spec from code", "investigate code", "debug an error", "triage a bug", "iterate to green", or "create handoff prompt".

Use when the user asks to "create documentation", "build docs", "review docs", "compare documents", "compress docs", "adversarial review docs", or "commit docs".

Use when the user asks to "sync skills", "set up skill-sync", "check skill status", "validate skills", "preview skill changes", "diagnose skills", "pin a skill", "unpin a skill", "prune skills", or "promote skill changes".

Use when the user asks to "run tests", "create tests", "fix failing test", "add test coverage", "fix slow tests", or "commit test changes".

Consolidate accumulated permission grants across Claude Code, Codex, and Gemini: move trusted commands into project settings, clean garbage entries, verify cross-agent consistency, commit project-level configs.

Use when the user asks to "create a TODO", "show TODO items", "manage TODOs", "prioritize TODOs", "implement a TODO", "implement a batch of TODOs", "batch implement TODOs", "complete a TODO", "cleanup TODOs", "review TODO quality", "claim a TODO", "what's ready" / "ready queue", "defer this work", "promote a deferral", "dismiss a deferral", "block"/"unblock a TODO", "create TODOs from a spec", or "todo stats". The production database-backed tracker; all tracker state lives in the shared DB and flows through the `todo` CLI.

Use when the user asks to "ideate on an idea", "refine an idea", "brainstorm", "write a spec", or "create a specification". Idea -> spec authoring that precedes tracked work; all TODO tracker actions (create/claim/implement/complete/defer/batch/...) belong to the `todo-db` skill.

Unified identify-verify-stage-commit workflow for skill-driven commits.

Compare two artifacts for semantic and behavioral equivalence.

Defines context trust levels, confusion protocol, and anti-patterns for agents during multi-step work.

Systematic root-cause debugging workflow from reproduction through regression-test verification.

Three-layer planning review framework for challenging obvious answers, surfacing blind spots, and reframing problems.

Pre-edit investigation workflow for understanding code and behavior before changes.

Validation-driven compression workflow that requires semantic comparison before approval.

Build multi-file changes in thin vertical slices: implement, test, verify, commit, repeat.

Post-edit verification workflow for lint, typecheck, tests, and spot checks.