chore: add role agent definitions with pinned models #248

New Issue

2026-08-22T22:53:47Z

egullickson commented

2026-08-22 22:53:47 +00:00

Problem

An audit of the agent configuration found that no model enforcement exists anymore: commit c239bb9 deleted the old role-agent library and its enforce-agent-model.sh hook, but the hook registration was left dangling in settings.local.json (firing a nonexistent command on every subagent spawn), and no .claude/agents/ definitions replaced the deleted roster. Subagents therefore inherit the session model (opus) instead of running on cheaper pinned models.

Change

Add a role-based agent roster under .claude/agents/, one file per role, each with a frontmatter model: pin (the durable mechanism, replacing shell-hook enforcement). Each agent's prompt directs it to read the relevant .claude/skills/mvp-*/SKILL.md files before acting, so the skill library remains the single source of domain knowledge.

Agent	Model	Role
developer	sonnet	Scoped implementation, self-gates lint/type-check/tests
debugger	sonnet	Root-cause analysis with evidence; no fixes unless asked
quality-reviewer	opus	RULE 0/1/2 merge-gate review, read-only toolset
operator	sonnet	Deploys, rollbacks, backups, CI workflow changes
tech-writer	sonnet	Doc-sync and new docs under house style

Also updates the .claude/CLAUDE.md index (Agents table) and the root CLAUDE.md .claude/ row per doc-sync discipline. The dangling hook removal in settings.local.json is local-only (gitignored) and not part of this diff.

Acceptance criteria

Five agent definitions exist in .claude/agents/ with valid frontmatter (name, description, model; tools restricted for quality-reviewer)
Model pins: sonnet for developer/debugger/operator/tech-writer, opus for quality-reviewer
Every skill path referenced in agent prompts resolves to an existing SKILL.md
.claude/CLAUDE.md and root CLAUDE.md indexes updated

## Problem An audit of the agent configuration found that no model enforcement exists anymore: commit c239bb9 deleted the old role-agent library and its `enforce-agent-model.sh` hook, but the hook registration was left dangling in `settings.local.json` (firing a nonexistent command on every subagent spawn), and no `.claude/agents/` definitions replaced the deleted roster. Subagents therefore inherit the session model (opus) instead of running on cheaper pinned models. ## Change Add a role-based agent roster under `.claude/agents/`, one file per role, each with a frontmatter `model:` pin (the durable mechanism, replacing shell-hook enforcement). Each agent's prompt directs it to read the relevant `.claude/skills/mvp-*/SKILL.md` files before acting, so the skill library remains the single source of domain knowledge. | Agent | Model | Role | |-------|-------|------| | developer | sonnet | Scoped implementation, self-gates lint/type-check/tests | | debugger | sonnet | Root-cause analysis with evidence; no fixes unless asked | | quality-reviewer | opus | RULE 0/1/2 merge-gate review, read-only toolset | | operator | sonnet | Deploys, rollbacks, backups, CI workflow changes | | tech-writer | sonnet | Doc-sync and new docs under house style | Also updates the `.claude/CLAUDE.md` index (Agents table) and the root `CLAUDE.md` `.claude/` row per doc-sync discipline. The dangling hook removal in `settings.local.json` is local-only (gitignored) and not part of this diff. ## Acceptance criteria - Five agent definitions exist in `.claude/agents/` with valid frontmatter (`name`, `description`, `model`; `tools` restricted for quality-reviewer) - Model pins: sonnet for developer/debugger/operator/tech-writer, opus for quality-reviewer - Every skill path referenced in agent prompts resolves to an existing SKILL.md - `.claude/CLAUDE.md` and root `CLAUDE.md` indexes updated

egullickson added the

labels 2026-08-22 22:53:47 +00:00

egullickson referenced this issue from a commit

2026-08-22 23:01:21 +00:00

chore: add role agent definitions with pinned models (refs #248)

egullickson referenced a pull request that will close this issue

2026-08-22 23:26:23 +00:00

chore: add role agent definitions with pinned models (#248) #249