P3: Agent Usage Audit — assigned to agent-manager team

[rooftop-Owl]

Assignment 3: Agent Usage Audit (P3)

From: HQ (system health session, 2026-02-04)
To: Agent-manager team
Priority: Medium
Blocks: Deferred findings #3 (Local Architect) and #4 (Oracle Underutilization)

---

Background

Independent Sisyphus flagged that 29 agents is significant surface area. Each agent adds routing complexity, context cost (described in AGENTS.md, profiles), and maintenance burden (model mapping x 4 profiles). But we don't actually know which agents earn their keep. This audit answers that question.

This also unblocks your own deferred work: Oracle underutilization and Local Architect proposals were put on hold because we couldn't justify adding or merging agents without usage data. This audit provides that data.

Key Context: File Count vs Roster Count

There are 16 agent files in .claude/agents/ but the roster claims 29 agents. The gap:

• astraeus agent files (16): architect, build-error-resolver, code-reviewer, config-manager, doc-updater, e2e-runner, fidelity-auditor, planner, refactor-cleaner, security-reviewer, star-curator, suggest-model-swap, supervisor, system-ops, tdd-guide, + AGENTS.md
• oh-my-opencode built-ins (no files here): sisyphus, atlas, hephaestus, metis, prometheus, oracle, explore, librarian, frontend, ml-engineer, multimodal-looker, document-writer
• Converted to skill: journalist (no longer an agent)

The audit must distinguish between these categories. We can only prune/merge what we own (astraeus files). oh-my-opencode built-ins are upstream — we can only adjust their routing or stop listing them.

Methodology

1. Session scan: Use mcp_session_list and mcp_session_search to find delegation patterns across recent sessions. Search for agent names in delegation calls.
2. Invocation count: For each agent, estimate how often it has been invoked in the last 10-20 sessions. Rough numbers are fine — directional, not precise.
3. Overlap analysis: Flag agents with overlapping mandates (e.g., Doc Updater vs Document Writer, Fidelity Auditor vs Code Reviewer, Refactor Cleaner vs standard refactoring).
4. Context cost estimate: How many lines does each agent add to the injection chain (AGENTS.md roster entry, profile mapping, agent file)?

Deliverable

File report at .sisyphus/notepads/agent-usage-audit/report.md with these sections:

• Agent Inventory: Table with agent name, source (astraeus/omo/skill), file path, profile mapping
• Usage Analysis: Table with agent name, estimated invocations (last 10-20 sessions), usage tier (high/medium/low/never)
• Overlap Analysis: List agents with overlapping mandates. For each pair: keep both / merge / convert one to skill
• Recommendations: For each low/never-use agent, recommend ONE of: KEEP (with justification), MERGE into parent agent, CONVERT to skill (like Journalist was), REMOVE (if truly dead weight), or N/A (oh-my-opencode built-in)
• Oracle + Local Architect Re-evaluation: Based on usage data, revisit deferred findings

Constraints

• Report only — do NOT remove, merge, or modify any agents. HQ reviews first.
• Do NOT add new agents (including Local Architect — separate decision after audit)
• Do NOT change model routing
• Rough numbers are fine — directional insight, not precision metrics
• Focus on astraeus-owned agents — oh-my-opencode built-ins get a row but cannot be pruned

How to Deliver

1. File report at .sisyphus/notepads/agent-usage-audit/report.md
2. PR to HQ if any doc corrections needed alongside
3. Comment on this issue with findings or blockers
4. Close this issue when report is filed

[code-yeongyu]

This appears to be an internal audit task for a different project (astraeus), not an oh-my-opencode issue or feature request. Closing as it was filed in the wrong repository.