ZeroClaw starts assigning model costs to individual conversations

ZeroClaw merged conversation-level attribution into its cost ledger on September 8. Previously, one random session identifier described a daemon's entire lifetime, leaving simultaneous chats mixed together under the same apparent session. A companion proposal tackles a different accounting gap: cache-creation tokens priced like ordinary input despite a provider's configured write premium. Together they expose two prerequisites for useful agent cost controls—knowing which conversation spent the tokens and which rate actually applies.

The facts

- Merged PR 10718 adds an optional conversation_id sourced from the tool-loop session key to each cost record. - The existing session_id remains a daemon-lifetime identifier, and old JSONL ledger rows read back without a conversation value. - The maintainer explicitly limits the merged work to ledger attribution; trace correlation remains open. - Open PR 10716 adds an optional cache_write_per_mtok rate and separately records cache_creation_input_tokens. - Without a configured write rate, the companion proposal preserves historical cost arithmetic rather than silently inventing a premium. - Neither source adds per-conversation budgets, aggregation screens or a new budgeting user interface.

Why it matters

A dashboard can display precise-looking money totals while answering the wrong question. Process-level grouping cannot explain whether one chat or several agents produced the bill, and missing cache-write pricing can make repeated misses appear cheaper than they were. Attribution is therefore a foundation for future controls, not a control itself. Keeping the old process identifier also matters: operators can still correlate a daemon lifetime without pretending that it identifies a user's conversation.

Current

Inspected on 2026-09-09. The ZeroClaw stable-release baseline is v0.8.5 published 2026-09-05T07:31:19Z. The main source was merged 2026-09-08T22:43:31Z when captured. Release metadata is a version boundary; an open proposal, a merged mainline change and a packaged release remain different states.

Evidence

The primary evidence is zeroclaw-labs/zeroclaw PR #10718 (https://github.com/zeroclaw-labs/zeroclaw/pull/10718). Supporting context comes from zeroclaw-labs/zeroclaw PR #10716 (https://github.com/zeroclaw-labs/zeroclaw/pull/10716); ZeroClaw — current official README (https://raw.githubusercontent.com/zeroclaw-labs/zeroclaw/master/README.md). These pages were inspected directly. Related project records explain the source cluster; they are not independent replications of contributor test claims.

Operator take

After the attribution change reaches the deployed build, inspect two concurrent chat sessions and confirm distinct conversation fields alongside the shared process field. For cache pricing, compare the configured rate sheet with the actual provider contract before enabling a write rate. The proposal deliberately leaves estimated-cost preflight arithmetic unchanged, so recording more accurate retrospective costs does not imply more accurate admission decisions. Retain the raw provider usage and invoice for reconciliation; the local ledger is an operational estimate whose identity and pricing dimensions now need separate tests.

Caveat

The attribution source reports unit-level persistence and legacy-row checks, not live multi-session daemon validation. The cache-write proposal remains open and did not exercise live Anthropic billing. These changes do not establish universal invoice accuracy, per-chat limits or automatic savings; the stable release baseline predates the merged attribution.