Implementation Team

Implementation Team

This repo's own .claude/ directory is the SpecRoute implementation team — the agents, skills, commands, and hooks that build SpecRoute itself. It is not the consumer template (those live in runtimes/.claude/).

For per-agent definitions, see .claude/agents/README.md.

What lives in .claude/

Path	What's there
.claude/agents/	12 implementation agents
.agents/skills/	The same skills in vendor-neutral shape — the repo-level skills root Codex and other non-Claude CLIs read
.claude/skills/	6 contributor skills (scaffold-artifact, add-vendor, example-walkthrough, frontmatter-lint, all-hands, doc-currency-check)
.claude/commands/	4 slash commands (/audit, /parity, /sanitize, /status)
.claude/hooks/	SessionStart status banner, PreToolUse sanitization gate, PostToolUse frontmatter check
.claude/agent-memory/	Per-agent persistent context
.claude/settings.json	Tracked, project-wide settings
.claude/settings.local.json	Gitignored — per-user permissions, MCP enables
.claude/.forbidden-strings.txt	Gitignored — per-installation sanitization wordlist

The 12 implementation agents

Verified against disk on 2026-07-27 (9 opus, 3 sonnet):

Agent	Owns	Model
prd-author	PRDs (full + lightweight + SRS templates and examples)	opus
spec-author	Spec triplet, lightweight feature spec, technical spec, ADRs	opus
agent-roster-architect	Agent roster template, archetypes, examples	opus
skill-author	Skill template, folder-per-skill convention	opus
command-author	Per-vendor command templates and examples	sonnet
hooks-author	Per-vendor hook templates, scripts, event matrix	sonnet
prompt-engineer	Global / phase / task prompt templates and per-vendor sets	opus
framework-docs-author	agentic-docs/, workflows, rules	opus
runtime-architect	runtimes/.<vendor>/, MCP single source, sync tools	opus
sanitization-auditor	/sanitize, /audit's sanitization layer, wordlist guidance	opus
template-quality-reviewer	Production-grade-and-immediately-usable bar	sonnet
docs-currency-auditor	Vendor facts, version anchors, transition dates, cross-mirror consistency	opus

Regenerate rather than hand-editing the Model column:

grep -h '^model:' .claude/agents/*.md | sort | uniq -c

Each agent has a .md definition with:

• Frontmatter. name and description (with trigger phrases) are the only required fields; model and color are optional, and an absent model inherits the parent session's. These 12 set all four for legibility.
• Operating principles.
• Owns (what files the agent governs).
• Don't use for (boundaries; hands off to neighboring agents).
• Optional memory to read/write persistent context.

See Agents for the frontmatter contract and the semantic-tier mapping table, and Frontmatter Contracts for the full field reference.

The 6 contributor skills

Interactive workflows for SpecRoute contributors:

• scaffold-artifact — bootstrap a new artifact (PRD / spec / agent / skill / command / hook / prompt / runtime) with the right frontmatter and target path.
• add-vendor — interactive walkthrough for adding a new agent CLI to the matrix.
• example-walkthrough — guided end-to-end build of examples/sample-project/.
• frontmatter-lint — interactive frontmatter validation across agents / skills / commands with offered fixes.
• all-hands — convenes the relevant implementation agents to plan, implement, and review a piece of work end to end over this repo's own roster.
• doc-currency-check — audits vendor facts, version anchors, dates, counts, and the four synchronized vendor-matrix copies.

See Skills for the folder-per-skill convention.

The 4 slash commands

Deterministic pre-commit checks:

• /audit — comprehensive sweep (sanitization + frontmatter + vendor-matrix consistency + broken links + TODO health).
• /parity — body-aware skill parity across all six runtime layouts plus MCP source-of-truth coverage; agent formats remain vendor-native.
• /sanitize — scans working, staged, and historical publication content for forbidden identifiers, machine paths, secrets, and private-source overlap.
• /status — SpecRoute skeleton state report (which top-level dirs exist, which artifacts have been drafted, what's outstanding).

See Commands and Sanitization.

The 3 hooks

Always-on automation:

• session-start-status.sh (SessionStart) — prints the SpecRoute skeleton status banner so Claude orients without re-grepping.
• pre-bash-sanitize.sh (PreToolUse) — blocks publish commands when forbidden terms appear in tracked files, untracked additions, the staged index, or outgoing history.
• post-edit-frontmatter.sh (PostToolUse on Write/Edit) — validates frontmatter on agent / skill / command file writes; warns on stderr.

See Hooks and Sanitization.

Why .claude/ and runtimes/.claude/ exist separately

Path	Who reads it	Contents
.claude/	This repo's contributors (Claude Code)	The implementation team that builds SpecRoute
runtimes/.claude/	Consumers who drop it into their own repo	The template that uses SpecRoute

Don't confuse them. The consumer drops runtimes/.claude/ into their repo; they do not copy .claude/. The implementation-team agents draft PRDs, specs, and agent definitions for the consumer template — they aren't themselves the template.

Per-agent persistent context

Agents with memory: project in their frontmatter read/write .claude/agent-memory/<agent-name>/<topic>.md. Reference implementations:

• sanitization-auditor/checklist.md — canonical wordlist pointer + audit protocol.
• runtime-architect/vendor-matrix-progress.md — vendor-by-vendor build-out state.
• framework-docs-author/docs-status.md — drafting progress for agentic-docs/.

See Agent Memory.

Per-user vs tracked settings

• .claude/settings.json — tracked, project-wide.
• .claude/settings.local.json — gitignored, per-user overrides (permissions, MCP enables).
• .claude/.forbidden-strings.txt — gitignored, per-installation sanitization wordlist.
• .claude/.provenance-sources.txt and .claude/.provenance-allowlist.txt — gitignored, local inputs for private-source overlap auditing.

See Sanitization for how the wordlist is populated and refreshed.

See also

• Agents · Skills · Commands · Hooks — the four primitives
• Automation Decision Framework — when to reach for which
• Agent Memory — per-agent persistent context
• Sanitization — the three-layer sanitization mechanism
• Frontmatter Contracts — required-field reference