token-usage-audit

Token Usage Audit

This skill owns the local token usage audit cron. It is disabled by default and is operational plumbing for the resident's agent, not a chat skill the agent should mention by name.

What It Does

The script at scripts/audit_token_usage.py reads only tenant-local operational metadata:

• Hermes dashboard session summaries from http://127.0.0.1:9119/api/sessions?limit=500
• local cron metadata from cron/jobs.json
• metadata-like local JSON/JSONL files only as a fallback when dashboard usage is unavailable

It does not call an LLM. It does not emit raw message content, prompts, raw session ids, private hosts, secrets, env values, or profile/memory prose. Session ids are used only for in-process grouping and become ordinal refs like session_001.

Cron Contract

When explicitly enabled with --token-usage-audit-cron "<expr>" or TOKEN_USAGE_AUDIT_CRON="<expr>", the AgentVillage installer creates one Hermes cron:

• name: Edge — token usage audit
• schedule: the configured cron expression
• delivery: Telegram only after the script wake gate returns wakeAgent:true
• skill: token-usage-audit
• script: skills/token-usage-audit/scripts/audit_token_usage.py

The script prints a final JSON line. When no action is needed, the line is:

{"wakeAgent":false}

When action is needed, the script prints a short sanitized agent prompt plus a final JSON line with wakeAgent:true. The agent should use only the facts in that prompt, keep any user-facing message brief, name a likely cron only when confidence is high or medium, and suggest pausing/reporting the driver. If attribution is unknown or ambiguous, say that plainly instead of guessing.

Alert Policy

Defaults:

• lookback: 24 hours
• cooldown: 72 hours per driver
• state file: memory/token-usage-audit.json
• total usage threshold: 100,000 tokens before spend-driver alerts

The script wakes only for actionable drivers, such as:

• cron-source usage is at least 60% of meaningful total usage
• one known cron is at least 40% of usage and at least 50,000 tokens
• one session is at least 250,000 tokens
• unknown scheduled-work bucket is very large, but emitted as unknown rather than guessed
• low-budget metadata is present and indicates a near-exhausted token budget