Toolbelt

toolbelt

My collection of day-to-day tools to get the job done.

Installation

Register the marketplace (one-time):

claude plugin marketplace add glnds/glnds-cc-marketplace

Install the plugin:

claude plugin install toolbelt@glnds-cc-marketplace

Prerequisites

• Claude Code CLI with plugin support

Skills

gh-create-issue

Create well-defined GitHub issues (single, epic, or sub-issues) with label discovery and an optional TDD implementation plan. Superpowers-friendly.

gh-fix-pr-review

Fix Claude Code Action review feedback on the current PR — both inline review threads and PR issue comments. Picks the highest-severity unaddressed item by default, implements the fix, commits, and replies on the review.

Flags: --all, --skip-tests, --inline, --comments, --resolve, --dry-run, --pr <number>.

deep-research

Multi-agent orchestrator that produces a long-form narrative research report (3 000–8 000 words) grounded in the local codebase, the current web, and live ops evidence (CloudWatch metrics, Logs Insights, git log). Dispatches six specialised subagents across a seven-phase workflow: plan → codebase recon → user-approved outline → parallel workers (web / ops-probe / hybrid) → single-agent synthesis → citation pass → validation. Output is written to ./research/{date}-{slug}-report.md.

The planner emits an evidence-locus (web, web-grounded, ops, codebase, hybrid) alongside the shape classification, so ops-shaped queries route to the read-only ops-probe subagent instead of noisy web workers. The skill is plan-mode-aware: when the harness is in plan mode, ExitPlanMode becomes the Phase 3 approval gate instead of fighting it. A self-disclosure block prints the tool surface before any token-expensive phase.

Not a quick-lookup tool — spends tokens deliberately (~200 K–700 K per run).

Flags: --no-code (skip codebase reconnaissance), --locus=web|web-grounded|ops|codebase|hybrid (override the classifier), --dry-run (stop after Phase 2), --yes (skip the Phase 3 approval gate outside plan mode).

adversarial-plan-review

Destructive, diagnosis-only pre-flight review of an implementation plan. Designed for the seam between Superpowers' writing-plans and subagent-driven-development. Runs as a dispatched subagent in a fresh context with only the plan and spec paths as input — the context starvation is what keeps the reviewer adversarial instead of sympathetic.

Attack angles (priority-ordered): Structural → Operational → Scope.

Emits: severity-tagged findings table plus a binary verdict — PASS, NEEDS REWORK, or NEEDS HUMAN. No remediation, no rewrites. Findings capped at 20. One-pass-per-session rule enforced in both SKILL.md and the reviewer's verdict rules so it cannot drift into a constructive-reviewer loop.

⚠️ Installing the plugin is not enough. The main agent only dispatches this skill automatically when a routing rule is present in your project CLAUDE.md. Paste the block below into your CLAUDE.md, adjusting the docs/plans/ and docs/specs/ paths to match your layout. Without the routing rule, you must invoke the skill manually each time.

## Plan review routing

When `writing-plans` completes and produces a plan file in `docs/plans/`,
dispatch `adversarial-plan-review` as a subagent with only the plan path
and spec path as input, before invoking `subagent-driven-development`.

- On verdict PASS: proceed to implementation.
- On verdict NEEDS REWORK: re-invoke `writing-plans` with the findings file
  at `docs/plans/<plan>.review.md` as additional input. Do not invoke the
  review skill a second time in the same session; if the reworked plan
  still fails on next review, promote to NEEDS HUMAN.
- On verdict NEEDS HUMAN: halt and surface the review file to the user.

Skip the review entirely on plans with fewer than three tasks, or tasks
that modify a single file. The inline self-review in writing-plans already
covers those cases.

See skills/adversarial-plan-review/README.md for the full design rationale and trade-offs.

code-audit

Read-only deep audit of a repository or subtree. Produces a severity-ranked Markdown report (audit-<slug>-<YYYYMMDD>.md) modelled on Trail-of-Bits-style findings (Severity × Difficulty, short- and long-term recommendations, effort estimate). Never modifies the audited project — the only write is the report itself.

Six dimensions (NFR priority order): security (OWASP Top 10:2025, ASVS v5) → resilience → cost efficiency → architecture & coupling → documentation freshness → monitoring gaps.

Falls back to ripgrep heuristics when scanners (gitleaks, trivy, semgrep, osv-scanner, infracost, checkov) are absent.

Flags: --scope=security|resilience|cost|architecture|docs|monitoring|all.

Changelog

See CHANGELOG.md.