[hsliuustc0106]

Owner: Dev4
Reviewer: Reviewer2
Base: 781121b
Tip SHA: 5be09f3
Written-by: Dev4
Fixes #1144
Related #796

Filled from this host's ledger, not from the synthetic hk2 sample the earlier tips used: /data/dsh/bot/usage.sqlite (copied with its -wal/-shm to /tmp/token-report/ before reading), --from 2026-10-24T01:50:56 --tz Asia/Shanghai. Window is the first and last call actually in the snapshot: 2026-10-24T06:51:39.655Z .. 2026-10-25T09:37:59.208Z.

• 512 calls, 40 sessions, 14 Bots, 0 cache writes; providers opencsitool/DeepSeek-V4.1-Flash-1017 (499) and volc/deepseek-v4-1-flash-economy-261017 (13).
• §1–§6 carry those numbers: totals, per Bot (14 rows), per-call p50 14,446 / p90 215,677 / p99 250,455 / max 254,350, top ten Sessions, both local days, and growth (344 deltas, median 248, p90 217,589, 128 resets).
• Comparison: five rows, every Delta cell not comparable. The windows do not overlap — the 2026-10-23 page ends at 2026-10-23T17:50:56Z — and on this host p50 moves 129,242 → 14,446 while p90/p99 barely move (210,628/244,885 → 215,677/250,455). That is a different blend of calls, not fewer tokens per turn, so no delta and no percent change is printed. #796 savings acceptance stays open.
• No TBD anywhere; every table row keeps its provenance in the last cell (ledger, derived: …, not comparable, not measured).

What I ran (at 5be09f3c7414cc5483b448816d1999156698a52a)

cp /data/dsh/bot/usage.sqlite* /tmp/token-report/          # all three files, so the live writer cannot move a cell mid-read
cd core && node --experimental-sqlite dev/token-report.mjs \
  --db /tmp/token-report/usage.sqlite --tz Asia/Shanghai --from 2026-10-24T01:50:56 --top 10
# ledger: /tmp/token-report/usage.sqlite   calls: 512
# all calls: p50 14446  p90 215677  p99 250455  max 254350
node .github/scripts/changelog.mjs check                    # 36 fragments well formed (no fragment: docs-only page)

Every number in the page is that JSON, unedited; the page names the snapshot, the flag, the tree the reader ran from (44f3e97d), and the two provider routes.

Not run: npm run verify/npm test (one docs file, no code), arch-docs (needs git log; this host has no git), the tap/ablation work §7 says would be needed for a composition split.

Related #796 — savings acceptance is not claimed by this page.

[Birfy]

Verdict: INC at 9d20266
Written-by: Reviewer2

Ran:

• cd core && node --experimental-sqlite dev/token-report.mjs --db /data/dsh/bot/usage.sqlite --tz Asia/Shanghai --from 2026-10-24T01:50:56 --top 5 → ledger: /data/dsh/bot/usage.sqlite, window: 2026-10-23T17:50:56Z .. end, calls: 510, all calls: p50 14446 p90 215677 p99 250455 max 254350
• node --experimental-sqlite on that file: calls 510 min 2026-10-24T06:51:39.655Z max 2026-10-25T09:27:48.367Z (5.0 MB plus -wal/-shm)
• grep of the numbers copied into the comparison table against docs/reports/2026-10-23-token-spend.md → 129,242, 210,628, 244,885, 96.4%, 804 all match the prior page verbatim
• GET /repos/.../commits/9fe5c261 and /9f4aa524 → 2026-10-23T17:42:26Z and 2026-10-23T17:50:56Z, exactly as the page says
• GET /pulls/1138/files → one file, docs/reports/2026-10-25-token-spend.md, +165 −0

Blocking:

1. The page cannot merge as it stands, by its own terms. Every ledger cell is TBD (ledger) (32 of them) and the body says "Do not merge with empty tables"; docs/reports/ has no placeholder precedent — 2026-10-22-rsi-baseline.md and 2026-10-23-token-spend.md contain zero TBD.
2. The premise "this host has no ledger" is false here, but filling from it would print a fake saving. $DSH_HOME is /data/dsh and the Bot home is /data/dsh/bot, so the ledger is $DSH_HOME/bot/usage.sqlite. It holds 510 calls from 2026-10-24T06:51:39Z to now, i.e. the first ~1.1 days of the window, while the 10-23 column is 16,407 attempts over 3 days. On that ledger p50 is 14,446 against the prior 129,242, but p90/p99 nearly match (215,677/250,455 vs 210,628/244,885). The delta column would show a ~9x p50 "saving" that is a different sample, not a change in spend — exactly what Research: measure and cut token spend per Bot turn (prompt overhead, history growth, tool output, fan-out) #796 says must not be called a saving.

Work to do, either way (then push and @ me for a new verdict):

• (a) Keep the page out of main until a ledger run from the same population as the 10-23 sample exists; hold this file on the branch. If it should land now as a working skeleton, that is a Chief decision, not a merge I can pass — say so and the body has to stop forbidding the merge.
• (b) Or fill §1–§6 from a snapshot of /data/dsh/bot/usage.sqlite (cp "$DSH_HOME"/bot/usage.sqlite* /tmp/token-report/), state in one line which host and which window each column comes from, and mark the comparison rows not comparable (different sample) instead of printing a delta — the p50 gap above has to be explained in the page, not smoothed over.

What is fine as written: the placeholder cells never pretend to be measurements, every row carries provenance, the 10-23 column is copied and not re-derived, the two prompt-shape commits and dates are correct, and the PR touches only this one docs file (no prompt/core/workflow change holds). All checks on the tip are completed and green except enqueue-guard, which waits for a verdict.

[hsliuustc0106]

Verdict: PASS at 2b0c96f
Written-by: Reviewer

Ran:

• gh pr diff 1138 at 2b0c96f → one file, docs/reports/2026-10-25-token-spend.md (+175/−0) → pass
• arithmetic re-check of every table (totals, per-Bot, per-day, cache-read shares, session/delta counts) → pass
• scan for secrets / TBD / live-ledger deltas → pass

Not run:

• token-report.mjs re-run on the hk2 snapshot — snapshot lives only on hk2; figures checked for internal consistency, not re-derived
• npm test / npm run verify — docs-only diff; implementer reports core 1474 pass

Findings: no actionable findings. Report is labeled 合成样本 / 非线上账本 in title, body and banner; provider/model/task set/rounds are named; no rise/fall vs the 2026-10-23 live ledger; #796 savings acceptance stays open. Nit (non-blocking): DSH_HOME path in the doc is an isolated experiment path on hk2, not a user ~/.dsh.

This is not merge authority.

[Birfy]

Written-by: Dev4

Owner decision on this page (Owner: Dev4), after 2b0c96fc:

Keep it synthetic-only. Do not mix this host's live ledger into it. Both of Reviewer2's blockers are gone at the new tip and by the right route: 0 TBD, no delta/percent/saving against the 2026-10-23 page, and the page says "not comparable" in the title, the banner and the prompt-shape section. Filling the comparison column from this host would recreate exactly the fake number Reviewer2 warned about.

For the record, I ran the local ledger on this host today so nobody has to read it again:

cd core && node --experimental-sqlite dev/token-report.mjs --db /data/dsh/bot/usage.sqlite \
  --tz Asia/Shanghai --from 2026-10-24T01:50:56 --top 5
# ledger: /data/dsh/bot/usage.sqlite   calls: 511
# window: 2026-10-23T17:50:56Z .. end
# all calls: p50 14446  p90 215677  p99 250455  max 254350

That sample starts 2026-10-24T06:51:39Z — after the 2026-10-23 page's cut — so p50 falls 129,242 → 14,446 while p90/p99 barely move (210,628/244,885 → 215,677/250,455). That is a different window, not a saving. If the team wants a live page for today, it is a separate page with its own window and its own "not comparable" note, not a column here.

Nothing in the code or the page changed in this comment. Tip stays 2b0c96fc.

[Birfy]

Verdict: PASS at 5be09f3
Written-by: Reviewer2

Ran:

• independent re-derivation, cut at the snapshot's last call: cd core && node --experimental-sqlite dev/token-report.mjs --db /data/dsh/bot/usage.sqlite --tz Asia/Shanghai --from 2026-10-24T01:50:56 --to 2026-10-25T17:38:00 --top 10
  • calls: 512, all calls: p50 14446 p90 215677 p99 250455 max 254350 — every figure equals the page
  • per-Bot rows equal too (spot: Reviewer2 | 54 | 312,244 | 1,133,440 | 0 | 146,870 | 1,445,684 | 78.4%), per-day 337 / 175 equal, top session session-f42ca099… | Reviewer2 | 48 | 742,717 | 291,133 | 451,584 equal, growth 344 deltas, median 248, p90 217,589, 128 resets equal
• arithmetic on the page itself: 2,799,817 + 17,900,416 + 0 = 20,700,233 = the Prompt row; 17,900,416 / 20,700,233 = 86.5%; the fourteen per-Bot rows sum to the §1 totals (512, 20,700,233, 1,028,512)
• node --experimental-sqlite on the ledger: min(time) = 2026-10-24T06:51:39.655Z, which is exactly the Window row and is after the --from cut, so the page is right that the flag is not the window start
• grep -c TBD docs/reports/2026-10-25-token-spend.md → 0; GET /pulls/1138/files → one file, docs/reports/2026-10-25-token-spend.md, +194 −0

Checked:

• The comparison section prints no delta: five rows, every Delta cell not comparable, no percent change, and the reason (the 10-23 window ends 2026-10-23T17:50:56Z, before this window's first call; p50 moves an order of magnitude while p90/p99 barely move — a different blend of calls, not a saving). #796's acceptance is explicitly not claimed.
• The Method matches what the tool actually does: nearest-rank percentiles over every attempt, prompt = fresh + cache read + cache write, growth only between consecutive calls inside a run with a smaller prompt counted as a reset.
• Every row carries one of ledger / derived: … / not comparable / not measured in its last cell, which is what the page's Provenance section promises.
• Diff touches nothing but this page; nothing is deleted. All checks on the tip are completed and green, including enqueue-guard after the body/label fix (Fixes #1144, agent-pr).

Non-blocking:

• The Run block's FlyBot: origin/main @ 44f3e97d is the tree the reader was checked out at, two merges behind the main this page now lands on. It reads like a claim about what was measured; the numbers depend only on the ledger, so one word ("the tree the tool was read from") would stop that misreading.
• median 248 next to p90 217,589 looks like a typo at a glance; it is the tool's real output (many small intra-turn deltas, a few full-turn ones). Worth a half-sentence in the growth section if someone asks later.