cost-report

name	cost-report
description	Weekly LLM cost breakdown by provider / gateway / skill, posted to private DM
when_to_use	["Scheduled weekly","User asks \"how much am I spending?\"","After a noticeable cost spike"]
toolsets	["terminal","file"]
parameters	{"window":{"type":"string","default":"7d"},"format":{"type":"string","enum":["markdown","json","csv"],"default":"markdown"}}
security	{"trust":"trusted","notes":"Read-only over local usage data. No untrusted input, no external writes.\n"}
model_hint	google/gemini-3.7-flash

cost-report — LLM Cost Breakdown

Generate a human-readable (or machine-readable) cost report from Hermes' own usage ledgers.

Procedure

1. Surface the numbers Hermes already keeps. There is no hermes logs export usage stream (that subcommand never existed) — use the built-in insights and logs views:

   hermes insights --days ${WINDOW}     # token + cost breakdown from session history
   hermes logs --since ${WINDOW} --component agent   # raw log lines for detail
   

2. Aggregate per-provider / per-skill splits. hermes insights rolls usage up globally. For the per-provider / per-skill / per-gateway tables below, pull from your tracing layer (self-hosted Langfuse via the observability/langfuse plugin — Part 20) or from a JSONL usage log you collect with the Part-20 log-tail pipeline. With such a JSONL file, DuckDB inspects it quickly:

   duckdb -c "
     CREATE TABLE logs AS SELECT * FROM read_json_auto('/tmp/hermes-logs.jsonl');
   
     -- By provider
     SELECT provider,
            SUM(cost_usd) AS cost,
            SUM(tokens_in) AS tok_in,
            SUM(tokens_out) AS tok_out,
            COUNT(*) AS calls
     FROM logs
     GROUP BY 1
     ORDER BY 2 DESC;
   "
   

3. Produce four tables:

   A. By provider

   Provider     Cost($)  Tokens-in   Tokens-out   Calls
   anthropic    18.44    2.1M        380K         412
   openai       6.20     1.2M        220K         187
   gemini       0.45     890K        140K         523
   

   B. By gateway

   Gateway     Cost($)  % of total
   telegram    14.22    56%
   cli         8.10     32%
   discord     2.77     11%
   cron        0.50     2%
   

   C. By active skill

   Skill                 Cost($)  Calls  Avg-cost
   claude-code           9.40     22     $0.43
   lightrag-query        4.11     189    $0.02
   pr-review             3.20     8      $0.40
   weekly-dep-audit      1.25     1      $1.25
   

   D. Daily trend (simple ASCII sparkline)

   Mon ▂
   Tue ▃
   Wed ▅█  ← weekly-dep-audit ran
   Thu ▃
   Fri ▄
   Sat ▂
   Sun ▁
   Total: $25.53
   

4. Flag anomalies. Use a 3x median-absolute-deviation rule on daily spend. Note any days or skills that exceed the threshold:

   ⚠ Wed spent $9.80, 4.5x typical. Driven by weekly-dep-audit.

5. Recommend savings. Pattern-match the data:

   • Any single skill > 30% of weekly cost → suggest a cheaper model for that skill (hermes model → "Configure auxiliary models")
   • Input tokens > 10x output tokens on any provider → prompt caching is already auto-on for Anthropic/OpenRouter/Portal; check that mid-session /model switches aren't resetting the cache prefix (Part 20)
   • Gemini/Flash calls doing work the primary model could absorb → check auxiliary.<task> is not pinned to an expensive model
   • Opus/GPT-5/Grok calls in cron or triage lanes → require explicit opt-in routing (model_aliases + cron job model pins)

6. Deliver. Post to the job's --deliver target (a private DM is the classic choice). Attach the raw JSON if format is json.

Cron wiring

Jobs live in ~/.hermes/cron/jobs.json, created via hermes cron create (the old cron.yaml list format was removed):

hermes cron create "0 9 * * 1" \
  "Run the cost-report skill: window=7d format=markdown" \
  --skill cost-report --name weekly-cost-report --deliver telegram

See also

• Part 20: Observability & Cost
• cost-routing playbook