agent-command-center

Use for enforcing hard limits on cached artifacts, summaries, indexes, and other compact metadata files so they stay bounded and high-signal.

Use for maintaining a rolling summary of only the last 10 conversations and refreshing matched history entries instead of appending unbounded cache.

Manage an explicitly approved rename, merge, supersession, discouragement, deprecation, or removal with compatibility routing and evidence-based exit criteria. Do not activate for ordinary edits, speculative consolidation, or a direct correction with no lifecycle impact.

Design, add, or amend tests for changed behavior with surface-appropriate coverage. Use when test design or test-file changes are part of the task; do not activate merely to run an existing test command, and use Playwright only when a browser-visible flow changed.

Perform an explicitly requested repository file-hygiene review or maintenance pass focused on factual accuracy, freshness, consistency, and duplicate or stale file prevention; read-only audits do not authorize corrections.

Use for designing and maintaining repository file structure for clarity, modularity, discoverability, and low-duplication across app and documentation surfaces.

Generate, validate, and enforce content-bound governance plans and exact-head release attestations for governed work. Use when governance tooling or CI enforcement is being run or debugged; do not activate for policy-only decisions, ordinary project tests, or routine validation.

Use for building and maintaining compact chat/session indexes (`docs/chat-history-index.md`) so long conversations can be queried precisely without re-reading broad transcript history.

Audit whether a task descriptor, selected skills, authority, gates, artifacts, execution, and completion claims semantically match user intent and actual effects. Use for policy-system audits or high-confidence governed review; do not use for ordinary implementation or mechanical schema checks.

Govern release-affecting, policy-changing, externally mutating, destructive, or otherwise high-risk work through typed effects, mandatory evidence, independent safety gates, and an explicit go/no-go decision. Do not use for ordinary answers, read-only inspection, tiny reversible edits, or a task made "multi-skill" only by optional process helpers.

Use for periodically reviewing which skills are overused, underused, missing, or creating friction by examining recent governance artifacts, handoffs, instruction trackers, and task outcomes.

Use only when the user explicitly asks to record or audit instructions, when a durable cross-task directive changes in a repository that already opts into tracking, or when maintaining an existing instruction ledger is the requested work. Do not activate for ordinary task requests, transient acceptance criteria, status questions without a ledger, or merely because a tracker file is missing.

Update existing canonical documentation when an authorized change alters behavior, interfaces, commands, setup, operations, policy, or user-facing usage. Use when documentation would otherwise become inaccurate; do not create unrelated README, tracker, architecture, or runbook artifacts merely because the skill activated.

Use only when the user explicitly requests bounded adjacent exploration or when a task already has multiple active branches that need resumable branch control. Do not activate for routine, single-threaded, answer-only, or one-command work.

Coordinate an authorized change whose coupled files, contracts, flows, or data paths must move together to remain valid. Use when a local edit would leave a known dependent surface inconsistent; do not use merely because several files are touched independently.

Optional compact-notation controls for explicit `/internal-lang on|off` and `/internal-lang --response on|off` commands, or when the user directly requests shorthand scratch or response notation. Do not activate for routine tasks, generic planning, or general concision.

Use only for explicit multi-step work with real dependencies, mixed or unsafe requested order, hidden prerequisites, or parallel branches that require sequencing. Do not activate for single-step tasks, routine command execution, or obvious local ordering.

Deprecated compatibility wrapper for explicit legacy invocation of `process-budget-controller`. Do not auto-activate; use the active router's minimum-skill and process-budget policy instead.

Establish or verify backup readiness for an explicitly authorized operation that could irreversibly alter external, production, or otherwise non-reconstructable state. Use for backup implementation/audits or real data-loss exposure; do not use for ordinary local source edits, source-only migration design, or topics such as authentication without an external-state operation.

Run an authorized browser or native-GUI workflow that requires runtime observation and adaptive interaction rather than a deterministic API, shell command, or existing test runner. Use for dynamic authenticated flows, browser extraction, or native GUI control; exclude ordinary Playwright test execution and deterministic file/API work.

Implement or review non-trivial code changes while preserving contracts, repository conventions, user work, and affected behavior. Use for logic, refactors, dependencies, APIs, authentication, data, runtime, or tooling changes and for qualitative code review; exclude copy-only edits, explicit scoring requests, and test design that is fully owned by effective-testing-methods.

Convert materially ambiguous or conflicting requests into a typed task contract before mutation. Use when uncertainty could change user-visible behavior, compatibility, security, data handling, authorization, or acceptance criteria; do not use for clear bounded work or to force clarification of harmless implementation details.

Prove that an identified backup can restore an explicitly scoped external or disaster-recovery target within required recovery objectives. Use for requested restore exercises or recovery-sensitive production operations; do not require a drill for ordinary local code changes or source-only migration design.

Use for deterministic shell automation and setup tasks (for example environment bootstrap, service orchestration, batch file operations, and repeatable local command pipelines with explicit validation).

Use only when the user explicitly asks to rate, score, grade, benchmark, or assign a numeric or weighted quality rating. Do not activate for a plain review, audit, assessment, evaluation, comparison, critique, or feedback request unless the user also requests scoring or a formal rubric.

Model end-user goals, expected baseline behavior, and product tradeoffs when a feature request requires product judgment. Use for feature planning or authorized implementation with meaningful user-experience choices; do not use for fixed mechanical work, and never treat a plausible enhancement as authorized scope.

Use when the user explicitly requests shorter or lower-token output, or when long context is causing concrete retrieval or response overhead. Limit this skill to context selection and output efficiency; do not activate it for routine tasks, project intake, testing ownership, deployment policy, or repository bookkeeping.

Use only when the user explicitly requests the Spatial Canvas design system or when repository evidence establishes Spatial Canvas as the existing interface architecture being modified. Do not activate for generic frontend, UI, UX, layout, navigation, interaction, or visual-system work.

Use when the user invokes --quizme to persistently require exhaustive task clarification before substantive execution, with intuitive options for multiple-choice questions, one-at-a-time questioning, explicit confirmation, and durable contract recording.

Use for debugging or remediation work when reported symptoms may not be the true root cause; verify the cause independently before applying fixes, even when the user suggests one.

Use for building or reviewing UI design quality using principles synthesized from trusted UX sources (Apple HIG, Google Material, MIT HCI, Nielsen heuristics, WCAG).