moshcode ps says which agent is blocked. moshcode cost says what the herd is spending — read out of the agent CLIs' own session logs, so nothing has to be instrumented, wrapped or proxied.

  session  engine  model          in    out   cache  cost    age
  api      claude  claude-opus-5  1.2k  27k   10.5M  $9.91~  42m
  audit    codex   gpt-5.6-sol    400   200   600    —       12m

  total  $9.91~  1.6k in · 27k out · 10.5M cached
  ~ estimated from published rates; unmarked figures are the engine's own.
⚠ no rate for gpt-5.6-sol — tokens counted, cost omitted.

Where the numbers come from

Decisions worth reviewing

• A price the engine computed is never overwritten. opencode and aider price their own messages; those are reported untouched. Claude Code and Codex record tokens only, so those are multiplied by a rate card and marked ~ — on a subscription that figure is "what this would cost on the API", not a bill. The distinction is carried into --json as costSource: "engine" | "rates" | "mixed".
• Unpriced models show tokens and no cost, never $0. Only Anthropic rates ship in the table; Codex/Gemini/Kimi/Qwen model ids (gpt-5.6-sol) do not appear on a public price list, so inventing one was the wrong call. ~/.moshcode/pricing.json takes your own rates and wins over the shipped table.
• Cache tokens get their own column. On a long agent session they are most of the traffic and a tenth of the price; folding them into in makes a $3 session look like a $60 one.
• Attribution is a heuristic and says so. A run is matched to the herd session with the same engine and directory that started before it. Right for one agent per directory, wrong when two share one — --json carries the run list so it stays checkable, and runs that match nothing are reported as their own line rather than absorbed into a neighbouring total.
• Engines that log nothing readable report no cost, not zero cost (gemini, kimi, qwen, deepseek, openagents), and the report names them.
• Claude Code's <synthetic> turns are skipped — they are messages Claude Code wrote itself (API errors, cancellations), not requests, and would otherwise show up as an unpriced model.
• Repeated messages are counted once, deduped on Claude's own message.id + requestId pair, so a resumed session or a folded-in subagent transcript does not double-bill.

Surface

moshcode cost                    # per herd session, last 24h
moshcode cost api                # one session, with its engine runs
moshcode cost --all --since 7d   # every engine session on the box, herd or not
moshcode cost --watch            # re-read every 10s — the slope is the point
moshcode cost --json             # machine-readable

Also moshcode herd cost, /cost in the pit, and cost(...) in moshscript. --all works with no herd at all, which covers foreground moshcode agents claude runs.

Performance

Bounded by mtime and, where the layout allows it, by directory and date before any file is opened; Codex rollouts are read head-and-tail rather than whole. On a box with ~600 Claude transcripts, cost --all --since 6h returns in ~0.7s.

Tests

39 new tests in test/cost.test.mjs — pricing and rate matching (dated snapshots, provider prefixes, user overrides), each reader against a throwaway $HOME, dedupe, <synthetic>, window and cwd filtering, malformed lines, aider parsing, attribution edge cases, and the CLI end-to-end via --json. Full suite green: 2084 tests, 0 failures. The README command table was regenerated from the schema.

🤖 Generated with Claude Code