guardrails-coding-agent

Plugin guardrails-coding-agent from marketplace software-agents-team in gagandeepp/software-agent-teams. Commands read from the repo's manifest — 1 month ago.

Skills (15)

• audit-trail

  Use after every guardrail gate invocation — writes a structured JSONL record to .guardrails-audit.jsonl with timestamp, agentFamily, skill, check, result, detail, userStoryId, sessionId, and resolvedBy. Always-on; cannot be disabled.

• budget-monitor

  Use at invocation start and poll mid-session — enforces token budget per session, rate limit per user per hour, circuit breaker after consecutive failures, and user story cost ceiling. This skill is the sole enforcer of budget ceilings; cost-tracker only aggregates.

• commit-push-pr

  Use when committing, pushing, and raising a PR — stages changed files, writes a Conventional Commits message, pushes to the remote branch, creates a GitHub PR via gh CLI, then posts the PR URL to configured Slack or Teams channels for review. Channel targets are configurable via .commit-push-pr.json

• context-compact

  Use when a session is approaching token budget limits, processing large codebases, multi-ticket sprint runs, or multi-scanner security triages where context growth could exhaust the token budget mid-task.

• cost-tracker

  Use at session end — aggregates token cost for the current session, appends to .guardrails-costs.jsonl, maintains a per-userStoryId rollup, and exposes the accumulated total to budget-monitor for ceiling enforcement. Does not enforce ceilings itself.

• drift-detection

  Use at session start and when PreToolUse hook detects writes to dependency/config files — detects config drift from git baseline, schema drift between DB schema and ORM models, dependency version drift against pinnedVersions, and guideline file drift. Re-ingests guidelines automatically on drift; fa

• error-escalation

  Use when any guardrail gate returns BLOCK or WARN — classifies severity (P0-P3), routes notifications to Teams/Slack/Jira per config, applies auto-fix for eligible P1 types, and enforces mandatory human-in-loop on ambiguous results. Routing is overridable via p0Action/p1Action config arrays.

• harm-prevention

  Use during and after code generation — scans for forbidden patterns, hardcoded credentials, dangerous infra operations, prototype pollution, and path traversal. Pattern IDs defined in defaults/forbidden-patterns.md. Adapters may override individual patterns using patternOverride parameter.

• idempotency-guard

  Use when operationType is migration (pre-action phase) or migration/infra/seed/ddl (post-action phase) — detects duplicate migrations before writing, and checks re-run safety, infra idempotency, and seed idempotency after writing. Mandatory human-in-loop if operationType is ambiguous.

• input-validation

  Use at invocation start before any file read or code action — validates prompt safety (injection detection), classifies intent, checks action-level permissions against config, and verifies file scope. Mandatory human-in-loop on any ambiguous result.

• jsonl-db-sync

  PostToolUse hook that incrementally syncs JSONL audit records to SQL Server. Fires after Write/Edit tool completions to ensure records that failed DB dual-write (network issues, DB down) are eventually synced. Idempotent and non-blocking.

• memory-reader

  Use to fetch recent memory entries for an agent family — reads the last N lines of the family's JSONL memory file and returns a formatted markdown block for <system-reminder> injection (used by run-session-start hooks and domain-invoker). Filter by userStoryId when present.

• memory-writer

  Use at the end of every domain-agent session (post-work phase) to append a rich memory record with patternsChosen, teamConventions, and lessonsLearned to the agent family's JSONL memory file under .claude/agent-memory/{family}/. Optional SQL Server dual-write when memory.persist is "sqlserver" or "b

• output-validation

  Use after code is written, before declaring task complete — checks LSP diagnostics (zero errors), test coverage threshold, diff size limit, secrets scan, and lint. Adapters pass lintTool and coverageTool; pass null to skip the respective check. Project-level overrides via outputQuality config keys.

• policy-enforcement

  Use after config-resolver, before any coding action — enforces org policy (approved languages/frameworks, agent-identity permissions, mandatory guidelines, LSP, org policy endpoint). Supersedes per-agent config-resolver governance checks when .guardrails-agent.json is present.

Slash commands (3)

• /guardrails-coding-agent:guardrails-db-init

  Initialise the SQL Server audit database -- creates the guardrails schema and all four audit tables (AuditLog, CostLog, RateLog, CircuitBreakerState) if they do not already exist. Safe to re-run (idempotent). Uses sqlcmd on Windows; Python/pyodbc on Linux/macOS.

• /guardrails-coding-agent:guardrails-memory-reindex

  Rebuild the ChromaDB vector index from JSONL memory files. Iterates all agent-family JSONL files and .story-memory.jsonl, skips hook-only records, and upserts each surviving record into the correct Chroma collection. Idempotent — safe to re-run at any time.

• /guardrails-coding-agent:guardrails-reset

  Reset the guardrails circuit breaker counter for the current workspace. Use after investigating and resolving the cause of 3 consecutive P0/ceiling failures.

Hooks (3)

• SessionStart

  "${CLAUDE_PLUGIN_ROOT}/hooks/run-guardrails-hook" session-start

• PostToolUse

  "${CLAUDE_PLUGIN_ROOT}/hooks/run-posttooluse-sync"

• Stop

  "${CLAUDE_PLUGIN_ROOT}/hooks/run-guardrails-hook" stop

Similar plugins