Token cost in a session
Where the tokens of an autopilot or review-pr run go once a session is running, what every session pays before it starts work, what was changed on 2026-10-25, and how to check the changes. The session and prompt-cache side (cold sessions, ticket boundaries, Foreman wakes) is in foreman-cache-economics.md; this document covers what that one does not.
Every figure comes from /bbs:token-audit, whose script
(.claude/skills/token-audit/scripts/token_audit.py) reads local Claude Code,
Codex and OMP session logs:
python3 .claude/skills/token-audit/scripts/token_audit.py usage --since 2026-10-15 --project babysit-worktrees --harness claude
python3 .claude/skills/token-audit/scripts/token_audit.py context --repo <repo>Figures are weighted tokens: input 1, 5-minute cache write 1.25, 1-hour cache write 2, cache read 0.1, output 5. Because output is included, the totals are higher than the input-equivalent totals in foreman-cache-economics.md (681.6M here against 476.9M there for the same 136 sessions). Shares are comparable within one report; they are not prices.
Baseline: worker sessions, 2026-10-15 to 2026-10-25
136 Claude Code worker sessions (babysit-worktrees projects), 681.6M weighted.
| Metric | Value |
|---|---|
| Main thread / subagents | 81% / 19% |
| Main thread by component | cache read 53%, 1h cache write 17%, output 11% (of the total) |
| Subagents by component | 1h cache write 10%, cache read 6%, output 2% |
| Context at the first request, median | 57.8k |
| Peak context, median | 265k |
| Turns per session, median | 152 |
| Compactions | 148 |
| Models, main thread | Opus 50%, Sonnet 31% |
| Models, subagents | Sonnet 19% (already set by CLAUDE_CODE_SUBAGENT_MODEL=sonnet) |
Main thread by active skill:
| Skill | Share |
|---|---|
| qa | 20% |
| review-pr | 18% |
| autopilot | 18% |
| no skill yet (Foreman phase integration and setup) | 11% |
| plan-draft | 6% |
| browse | 5% |
Subagents: review-pr's correctness angles A, B and C cost 7% between them (355k–459k per run); the cleanup, altitude and conventions finders cost about 2.5%; the verifiers about 4%.
Friction found in the same window:
| Signal | Count | Cause |
|---|---|---|
bbs autopilot verification begin/record rejected | ~230 | subject changed after review-pr --fix (44), "already owns" an open attempt (42), empty log (25), missing --attempt (22), wrong gate name (18), cwd outside the checkout (11), extra results keys (5) |
--help lookups | 644 | agents guessing flags the skill did not state |
| Orca heartbeats | 584 | removed in v1.107 |
| Reviews repeated on the same ticket with the full fan-out | 22 of 68 | no rule for what a second review should target; one ticket was reviewed 6× on its whole web/ diff |
What the baseline says
- The main thread is the cost. Cache reads of a ~165k context, over ~150 turns, are most of every run. A cheaper subagent model can only touch the 19%. Trimming skill text saves 1–2%.
- The one-hour subagent cache TTL cost ~4%.
subagentPromptCacheTtl: "1h"made each short-lived subagent write its prefix at 2× instead of 1.25×, and most subagents never live long enough to read it back. - Friction costs turns, and turns cost context. Every rejected verification
call and
--helplookup is a turn that re-reads the whole context. - Repeated reviews were full reviews. About 7.5% of cost went to review runs that re-read a whole ticket diff to check a small fix.
Background context: what every session loads
Measured with token_audit.py context in the kiem-lai checkout (the
2build-consuming repo) on 2026-10-25.
| Item | Before | After |
|---|---|---|
| Claude Code skill listing (user, project, plugins) | ≈26.3k chars ≈ 7.5k tokens | ≈10.4k chars ≈ 3k tokens |
claude.ai synced skills (anthropic-skills:*), almost all unused | 10.0k chars | 0.1k (skillOverrides) |
User skills (~/.claude/skills), 10 of 14 unused in 30 days | 5.9k chars | 3.2k (5 removed, 4 name-only) |
| Project ops/content skills | 4.0k chars | 0.4k (disable-model-invocation: true on 17 of 19) |
2build/.claude/skills relisted as 2build:* when work touches files under 2build/ | +6.5k chars | unchanged (see below) |
Stale skill-root copies without SKILL.md | 2 | 0 |
Auto-memory MEMORY.md | 25.1 KB / 174 lines (over the 24.4 KB read limit) | 16.0 KB / 137 lines |
The listing is part of the cached prefix, so it is re-read on every turn: 4.5k tokens less is about 2–3% of the main-thread cost of a 150-turn run, in every repo that loads these roots.
Changes made on 2026-10-25
| # | Change | Where | Targets |
|---|---|---|---|
| 1 | Removed subagentPromptCacheTtl (back to the 5-minute default) | user ~/.claude/settings.json | 1h subagent writes, ~4% |
| 2 | Tiered review subagents: cleanup finders (Reuse, Simplification, Efficiency, Conventions) on the light tier; correctness angles, Altitude, verifiers and sweep on the mid tier; never above the session's model; inherit when the harness has no model selector (Codex, OMP); a routed model that is unavailable reruns on the inherited one, so no finder or vote is lost | review-pr/SKILL.md §5 | subagent share |
| 3 | Delta re-review: each review gate snapshots its tree with a temporary index (GIT_INDEX_FILE, the real index is untouched) and stores it as review_tree. A later review on the same base, requirement and plan targets git diff <review_tree> <now> with the previous report as context. It falls back to the whole ticket diff when the base, requirement or plan changed, or when the delta is most of the diff | autopilot/SKILL.md § Current-session gates | repeated full reviews, ~7.5% |
| 4 | Verification playbook: exact gate names, data.id, one open attempt per ticket, the three-key results file, --fix failing the attempt as expected, "already owns" recovery | autopilot/references/verification.md | rejected calls, ~2% |
| 5 | /bbs:token-audit skill and script, so the next audit is one command and needs no re-derivation | .claude/skills/token-audit/, tests/test_token_audit.sh | measurement |
| 6 | Effort medium for the main model | user ~/.claude/settings.json | output and thinking |
| 7 | Background context trimmed (table above): ops skills manual-only, unused skills removed (copies in ~/Backups/agent-skills-2026-10-25/), synced and artifact-* skills off by skillOverrides | kiem-lai .claude/skills/*/SKILL.md, user settings | per-turn context |
Tests: tests/test_autopilot_review_delta.sh drives the delta-review protocol
against the real bbs binary; tests/test_autopilot_in_session_planning.sh
pins the review-pr §5 and autopilot contracts; tests/test_token_audit.sh
checks the script against synthetic Claude Code, Codex and OMP logs.
Checked and not done
| Lever | Estimate | Why not |
|---|---|---|
| Auto-compact window 300k → 200k | ~10% of main thread | 2.4× more compactions; it trades recall for cost, and forced compaction was removed on purpose. User decision. |
| Batch verifiers (several findings per verifier) | ~2% | breaks the tuned one-finding, one-vote design |
| Shorter skill text | 1–2% | not worth the risk to tuned prompts, which are sent word for word |
| Light tier for correctness angles | part of 7% | correctness is where findings come from; judge by findings, not tokens |
Hide the nested 2build:* relisting | 6.5k chars when touching 2build/ | no setting found that disables a nested project skill root; a per-name skillOverrides on 24 names would also hide bbs:* |
Re-check
After a week of runs on the new skills:
python3 .claude/skills/token-audit/scripts/token_audit.py usage --since <date> --project babysit-worktrees --harness claude| Metric | Baseline | Expect | Change it tests |
|---|---|---|---|
sub:write_1h share | 10% | ~0; 5-minute writes appear instead at a lower weight | 1 |
| Subagent share | 19% | lower; cleanup finders on the light model in by model | 2 |
| review-pr share of the main thread | 18% | lower; second reviews name a review_tree | 3 |
Errored bbs autopilot verification calls | ~230 | a handful | 4 |
--help lookups per session | 4.7 | lower | 4 |
| First-turn context, median | 57.8k | lower by the trimmed listing (~4k) | 7 |
| Output share of the main thread | 11% | lower | 6 |
Watch for the cost of each lever as well: findings a light-tier finder missed
(compare review reports before and after), delta reviews that missed a
regression outside the fix, and ops skills nobody remembers to call by
/name.