GTCX Agentic Environment Standard v1

The seven highest-leverage moves

Ranked by leverage: what unlocks the most agent capability across the most repos for the least ongoing maintenance. The first two are structural — nothing else compounds without them.

Stand up a GTCX plugin marketplace

The distribution backbone. Today every skill, hook, and agent lives only in one machine's ~/.claude — nothing propagates to the 17 repos. A single internal marketplace (a git repo of bundled skills + agents + hooks) turns "institutionalize" from copy-paste into enabledPlugins.

Phase 0

Build the custom subagent fleet

~/.claude/agents/ is empty — 0 defined. You run 33 slash commands through the generic agent. Named subagents (scaffolder, reviewer, test-author, security-auditor, migrator) carry their own system prompt, tool allowlist, and model tier — the unit of reusable, routable autonomy.

Phase 0

Codify the model-routing policy

Your MODEL_CAPABILITY_MAP.md still references Opus 4.6 / Sonnet 4.6. Refresh to the current tier (Opus 4.8 · Sonnet 5 · Haiku 4.5 · Fable 5) and bind roles to tiers in agent frontmatter + workflow phases, with fallbackModel for overload.

Phase 0

Institutionalize the hooks catalog

The deletion-staging hook is live and proven — it is hook #1 of a standard set: format-on-write, test-on-change, session-start orientation, and an append-only agent audit log (the "black box" for provable walk-back).

Pilot

Add a managed / policy settings tier

Guardrails that must be non-overridable — destructive-git gating, marketplace allowlist, model allowlist, org CLAUDE.md — belong in managed settings, not user files. This is the difference between a convention and a control.

Pilot

Ship a repo-bootstrap generator

One skill that scaffolds a new repo agent-first: CLAUDE.md, .claude/settings.json, the .mcp.json baseline, the staged-deletion path, and CI hooks — so "clone to productive agent" is under five minutes.

Ecosystem

Wire observability & FinOps for agents

OTEL export + the audit-log hook + cost-per-successful-task tracking. Prevents agent sprawl and gives the walk-back trail enterprise trust requires. You are stacking the reliability of every primitive — measure it.

Ecosystem

Current state of the estate

What exists today, verified against ~/.claude and this repo. The strengths are real; the gaps are all in distribution and specialization.

Component	Location	Status	Note
Permission model	~/.claude/settings.json	Strong	Allow-all Bash + ask-gated destructive git, global.
Deletion-staging hook	~/.claude/hooks/stage-deletions.sh	Strong	PreToolUse, verified firing. Blocks raw deletes → staging.
Slash-command library	~/.claude/commands/ · 33	Strong	Deep: audit, sprint, design, roadmap, onboarding.
Strategy corpus	~/.claude/*.md	Strong	Primitives, MAS scaling, deployment hurdles already written.
Custom subagents	~/.claude/agents/	Gap	Empty. No named, routable specialist agents.
Plugin marketplace	~/.claude/plugins/config.json	Gap	Empty `{}` — no ecosystem distribution path.
Model-routing policy	MODEL_CAPABILITY_MAP.md	Partial	Exists but stale (4.6-era); not bound to config.
MCP baseline	connected servers	Partial	Rich set connected; not standardized as a per-repo `.mcp.json`.
Managed / policy tier	managed-settings.json	Gap	Guardrails live in user files — overridable, not enforced.
Repo scaffolding	—	Gap	No agent-first bootstrap generator.

The four capability pillars

Each pillar: where GTCX is now → the standard to adopt → which settings layer owns it → how it reaches all 17 repos.

Spin up repos

scaffold agent-first

● Current

New repos are hand-configured. CLAUDE.md, settings, MCP, and hooks are copied ad hoc or missing — each repo starts cold for agents.

● Standard

A /bootstrap-repo skill that emits CLAUDE.md, .claude/settings.json, .mcp.json, the staged-deletion path, and CI lint/test/build hooks.
Inherit global guardrails automatically; only repo-specific overrides are written locally.
Bar: clone → productive agent in < 5 min.

● Owns

Global generator skillProject committed .claude/Policy org CLAUDE.md

● Rollout

Ship the generator via the marketplace; run it once per repo. Drift caught by a deploy:check-style CI gate that diffs each repo's .claude/ against the template.

Run commands

autonomy, gated

● Current

Strong: global allow-all Bash, ask-gated destructive git, and a firing deletion hook. This pillar is the most mature — the work is to formalize and enforce it.

● Standard

allow: ["Bash"] for autonomy; ask for reversible-risky (push, reset --hard); hook-deny for irreversible deletes.
Enable the OS sandbox (sandbox.enabled) with a network allowlist for untrusted execution.
CI parity: same lint + typecheck + test + build gate a PR must pass runs in-agent.

● Owns

Policy deny / ask rulesGlobal hooks + sandboxLocal personal allows

● Rollout

Move the destructive-action rules from user settings into managed-settings.json with allowManagedPermissionRulesOnly so they cannot be weakened per-repo.

Deploy subagents

the unit of autonomy

● Current

Every task runs through the generic agent. No named specialists, no defined tool-scoping, no orchestration beyond a single thread.

● Standard

A defined fleet in ~/.claude/agents/: scaffolder, reviewer, test-author, security-auditor, migrator, explorer — each with scoped tools + model.
Workflows for deterministic fan-out: pipeline() by default, parallel() only at real barriers; adversarial verify stages.
isolation: "worktree" when agents mutate files concurrently; background/remote for long jobs.

● Owns

Global agent defsMarketplace shared fleetProject repo-specific agents

● Rollout

Author agents once, bundle in the marketplace plugin. Supervision hierarchy: a lead thread spawns and merges; conflict detection via worktree isolation.

Route by model

right tier, right task

● Current

Routing lives in a stale doc, not in config. No fallbackModel, no per-role tiering, no effort-level policy.

● Standard

Bind role→tier in agent frontmatter and per phase in workflows (see tiers below).
fallbackModel for overload; effortLevel scaled to task hardness.
availableModels (policy) to bound cost centrally.

● Owns

Global defaults + fallbackAgent per-role modelPolicy availableModels

● Rollout

Refresh MODEL_CAPABILITY_MAP.md to current tiers, then encode it — the map becomes executable config, not advice.

Opus 4.8 [1m]

claude-opus-4-8

Deep reasoning, system architecture, multi-file refactors, adversarial review, judge panels, hard debugging.

Lead thread · plan · review

Sonnet 5

claude-sonnet-5

Best price/performance for implementation. Full-stack code, test authoring, most refactors — the fleet workhorse.

Default build agent

Haiku 4.5

claude-haiku-4-5

Fast + cheap at volume: search fan-out, formatting, lint fixes, structured extraction, hook evaluations.

Parallel · mechanical

Fable 5

claude-fable-5

High-throughput generative surface. Route per the refreshed capability map where speed × volume dominates.

Throughput · generative

Cross-cutting foundations

The plumbing that sits under all four pillars. These are the primitives that make reliability a property of engineering, not of model intelligence.

Foundation	What it does for the org	Config surface	Layer
Skills & commands	Reusable, invocable expertise; the "clone-to-productive" bar.	.claude/skills · commands · marketplace	Global + MP
MCP baseline	Standard connected tools (GitHub, Figma, Vercel, filesystem, memory) with a safe auto-approve set.	.mcp.json · enableAllProjectMcpServers · allowedMcpServers	Project + Policy
Hooks catalog	Event-driven guarantees: staging, format, test-on-change, orientation, audit.	hooks.{PreToolUse,PostToolUse,SessionStart,Stop}	Global
Memory & context	Persistent facts + plans; context hygiene across a multi-repo estate.	autoMemory · CLAUDE.md · plansDirectory	Global + Project
Governance	Non-overridable org guardrails, attribution, marketplace + model allowlists.	managed-settings.json · strictKnownMarketplaces · claudeMd	Policy
Observability	Audit trail + FinOps: provable walk-back and cost-per-successful-task.	OTEL env · audit-log hook	Global + Policy

RULE

Reliability compounds multiplicatively — five primitives at 97% is 86% end-to-end, not 97%. Standardize the primitives so every repo inherits the same 99%, rather than each re-deriving a weaker stack.

Phased adoption roadmap

A real sequence — each phase produces artifacts the next depends on. Phase 0 is structural and unblocks everything; do not parallelize past it.

PHASE 0

Institutionalize now

proven · low-risk

Adopt

▹

gtcx-agentic marketplace repo — a git repo bundling skills + agents + hooks; registered via extraKnownMarketplaces.

▹

Six-agent fleet — scaffolder, reviewer, test-author, security-auditor, migrator, explorer in ~/.claude/agents/.

▹

Refreshed MODEL_CAPABILITY_MAP — current tiers, bound to agent frontmatter + fallbackModel.

▹

Guardrails already live — permission model + deletion hook (done this session).

PHASE 1

Pilot in one repo

validate before spread

Pilot

▹

Full hooks catalog — format-on-write, test-on-change, session-start orientation, append-only audit log.

▹

Managed-settings tier — destructive rules + marketplace/model allowlists made non-overridable.

▹

MCP baseline as .mcp.json — committed, with the safe auto-approve set defined.

▹

Proof gate — pilot in gtcx-infrastructure; confirm the harness verifies (Level-2 verification) before rollout.

PHASE 2

Ecosystem-wide

all 17 repos

Scale

▹

Bootstrap generator — /bootstrap-repo applied to every repo; new repos start compliant.

▹

Drift detection in CI — each repo's .claude/ diffed against the template; violations flagged.

▹

Observability + FinOps — OTEL export, cost-per-successful-task dashboards, agent-sprawl watch.

Appendix — copy-paste config

Drop-in snippets for the Phase-0 artifacts. Paths assume ~/.claude (global) unless noted. Verify against your Claude Code version's settings schema before committing to policy.

~/.claude/agents/reviewer.md — a routable specialist subagent

---
name: reviewer
description: Adversarial code reviewer. Verifies correctness
  bugs and reuse/simplification on a diff. Reports,
  does not edit.
tools: Read, Grep, Glob, Bash
model: claude-opus-4-8
---
You are a staff engineer reviewing the current diff.
Rank findings most-severe first; verify each before
reporting. Prefer fewer, high-confidence findings.

~/.claude/settings.json — model routing + fallback

{
  "model": "claude-opus-4-8",
  "fallbackModel": ["claude-sonnet-5", "default"],
  "effortLevel": "high"
}

.claude/settings.json (committed) — register the internal marketplace

{
  "extraKnownMarketplaces": {
    "gtcx-agentic": {
      "source": {
        "source": "github",
        "repo": "gtcx/gtcx-agentic",
        "path": ".claude-plugin/marketplace.json"
      }
    }
  },
  "enabledPlugins": { "agentic-core@gtcx-agentic": true }
}

managed-settings.json (policy) — non-overridable guardrails

{
  "permissions": {
    "ask": ["Bash(git push *)", "Bash(rm -rf *)"],
    "disableBypassPermissionsMode": "disable"
  },
  "allowManagedPermissionRulesOnly": false,
  "availableModels": ["opus", "sonnet", "haiku"],
  "strictKnownMarketplaces": [
    { "source": "github", "repo": "gtcx/gtcx-agentic" }
  ]
}

workflow phase — route model per stage (Sonnet finds, Opus verifies)

pipeline(files,
  f => agent(reviewPrompt(f),
        { phase: 'Review', model: 'claude-sonnet-5' }),
  r => agent(verifyPrompt(r),
        { phase: 'Verify', model: 'claude-opus-4-8',
          effort: 'high' })
)

NOTE

These are scaffolds, not a merge-ready patch. Marketplace repo slugs, the plugin manifest, and whether allowManagedPermissionRulesOnly should be true are org decisions to confirm during the Phase-1 pilot.