[zolotarev]

Friction — token cost is the main complaint of the first project's owner, but dpcli could not say where tokens go: skills/harness-retro/cost.py measures sessions, not modules or tasks, so process changes (v18 per-wave coordinator, v20/v22 Edit instead of Bash edits) could not be verified by numbers.

Change — stats --tokens reads Claude Code's subagent logs of the project (~/.claude/projects/<project>/*/subagents/*.jsonl); an agent is attributed by its first message, which is a dpcli brief (<ID> <type>: …, Review <ID>: …, Module <m>: …; the Russian protocol of older journals too). Prints context tokens of executors / reviewers / coordinators per module and the per-task median. Other agents are not counted. Opt-in: reading the logs takes ~3 s on ~300 agent logs.

Real numbers from the first project (anonymized): per module, one coordinator processed 76–148M context tokens — about as much as all executors of the module together; per-task median 4–16M. That is the evidence behind v18's "one coordinator per wave".

Test: tests/test_stats_tokens.py (fake logs: executors, reviewer, coordinators in both languages, a non-dpcli agent).

[github-actions]

VERDICT: accept

• Real friction with evidence in the PR body: process changes (v18, v20/v22) could not be checked by numbers. The feature is opt-in, stdlib-only and read-only.
• Correct on reading: the brief regex covers the English and Russian protocols. A non-dpcli agent is skipped. A missing log directory gives empty output. Journals are untouched. I did not run the tests myself, so CI must be green.
• Test tests/test_stats_tokens.py covers executors, reviewer, coordinators in both languages and a foreign agent. It would fail without the change.
• Process is complete: VERSION 24, CHANGELOG line, docs/guide.md row. No prompt or template growth, so tools/budgets.json is unaffected.
• Nit: agents whose first message looks like a brief but whose task ID is not in the journal (deleted or renamed task) are dropped without a word. CLAUDE.md says never truncate silently, so consider printing unattributed: N agents as a one-line hint.
• Nit: task_module maps ID to module with the last one winning. This is fine while task codes are unique per module, but note the assumption.
• Nit: the "context tokens" figure sums input and cache reads per turn, so it counts re-reads. The label says "context", which is fine. Say in the docs row that output tokens are excluded.