session-governance

Reduce common coding mistakes with explicit assumptions, surgical edits, bounded ablation, and verifiable success criteria. Use for non-trivial implementation or refactoring. Do not use for read-only code review or concrete software failures.

Route implementation, high-risk review checkpoints, read-only code review, and completion-time verification to this plugin's first-party method Skills. Use for non-trivial coding, review requests, or before claiming work is done. Concrete software failures belong to a dedicated debugging workflow, not this plugin.

Read-only review of a code or configuration change for correctness, regressions, security, compatibility, and missing tests. Use when the user asks to review, audit, or assess a diff before merge. Do not edit files, implement fixes, or use this for diagnosing an already observed concrete failure.

Coordinate one bounded read-only reviewer after a coherent high-risk implementation slice, or when the user explicitly requests an engineering review checkpoint or asks to summon an engineering critic. Use for cross-module, public-contract, security, persistence, migration, concurrency, data-integrity, deployment, runtime-state, recovery, or observability risk. Do not use for simple local edits or an initial read-only audit.

Select the smallest sufficient verification scope and require fresh command evidence before claiming complete, fixed, or passing. Use before completion claims, commits, or pull requests.

Change execution-discipline thresholds for edit loops, command retries, polling budgets, exemptions, or bypass markers in .execution-discipline.mjs.

Rebuild an unclear concept, system, proposal, or decision from essential constraints instead of inherited labels and conventions. Use when the user asks for first principles, underlying logic, the real problem, a mental model, why something exists, or a from-scratch explanation. Prefer reasoning-methods when the main job is proving an answer, diagnosing a concrete failure, or choosing under quantified uncertainty.

Run a user-requested Socratic deep interview that stress-tests a plan, decision, or idea one question at a time. Use only when the user explicitly invokes $grill-me; never auto-trigger it for ordinary ambiguity.

Front-load intent, repository context, plausible interpretations, and adversarial checks before acting on a first request or a materially new task. Use when hook context marks a first-turn discovery pass, when the same conversation starts a new outcome or deliverable, or when the user explicitly asks to explore context and assumptions up front. Do not rerun for continuations, follow-ups, or corrections. Keep simple requests direct; use bounded generic subagents only when independent evidence can change the approach.

Change language-output defaults: response and artifact profiles, drift thresholds, tool feedback, and Stop language gate in .language-output.mjs.

Select a compact, task-specific reasoning and verification structure for exact problems, causal diagnosis, consequential decisions, or factual synthesis where a plausible answer is not enough. Use when correctness depends on quantifiers, boundaries, competing hypotheses, load-bearing assumptions, external evidence, counterexamples, or calibrated uncertainty. Do not use for simple lookup, translation, routine implementation with a direct test oracle, creative writing, or requests whose answer is already mechanically determined.

Teach a user-requested topic through adaptive Socratic questions, small hints, counterexamples, and transfer. Use only when the user explicitly invokes $socratic-tutor; do not replace ordinary explanations with questioning.