How KRouter turns sessions into “what was decided” lines, how it lives next to another collector, and what moving the
work from the nightly run to the end of each session costs. Code: import_sessions.py, session_hook.py,
answer_cards.py. Measured on the author’s machine on 2026-10-30 unless said otherwise.
The 2026-10-28 checkup listed 8 cards whose answers held a number or path not found in the page, and 18 questions
too broad to tell which page they were about. All 8 were written on 2026-10-10, the day the whole-token check
(4d02772) went in, and all 8 had passed the model check (v: 1). Two causes:
answer_cards.py rewrote a page’s cards only when the page changed; a card
that a later, stricter check would refuse stayed in the file, blocked at read time, for good. Now every run holds
kept cards to the current checks and asks a page again once when one fails (new cards pass the same checks, so the
page is not asked again the next night).1 : 3.3 → 1:3.3, 6230 / N70 / N73 → 6230/N70/N73, 1.2MB / 498 行 → 1.2MB/498, 2026-08-23 / 24), or
quoted session_item_, which the token pattern cut to session_item. faithful() now also reads the page with
spaces around / and : removed, and a trailing _ is part of a token. The other 2 rewrote a date
(08-16 → 8/16) and are still refused; the prompt now says to copy values as written.recall_index.broad_question() is now
the one test both use: a new card with fewer than two content words is dropped, and the prompt asks for the page’s
own subject in the question (“Roadster 那期音效比人声低多少？”, not “音效该比人声低多少？”).Dry run on the author’s vault after the change: 16 pages re-asked once (about 35 000 characters), the 2 remaining
unfaithful cards and the broad ones among them. Tests: tests/test_recall_index.py.
collect_day.py)The host’s daily job (collect_day.py + distill_llm.py) writes one type: daily-log page per day under
05 时间日志/YYYY-MM/: a stub that lists every session as `claude:<id>`, then a distilled day summary that
cites sessions by an 8-character prefix. It reads session metadata (first question, turn count), not the turns:
its distilled logs say “索引未提供” where the content would be. KRouter’s pages are the turn-by-turn record of one
session-day. They are not the same content, so the plan is to link them and remove true duplicates, not to pick one.
type: daily-log page
covers a session it names (`agent:id`, or a bare 8+ hex prefix) on that page’s own date (a session that
ran past midnight is not covered for the next day), and covers its whole date when a collector wrote it
(source_kind in its frontmatter). On the author’s last 7 days: 41 session-day pages, 38 matched by id, 3 by day.--overlap link (default): write the page, add covered_by: "[[the log]]" and a line pointing to it.
--overlap skip: write nothing for a covered session-day. Saved with --consent.forked_from), and one wholly inside another gets no page (and its old page, if ours, is removed; in hook
mode the earlier session’s transcript is found by id and read back). Same words asked at another time are not a
fork. On the author’s last 7 days: 1 such duplicate.Not done here, on purpose: turning on --consent on the author’s machine. Suggested line, once the model lane works:
import_sessions.py --consent --distill --cli /opt/homebrew/bin/claude --overlap link --days 1.
Tests: tests/test_import_sessions.py.
session_hook.py end already imports the session that ended, in a detached process, and distills it when consent
included --distill. So “real time” needed no new hook; it needed three fixes before it is reasonable to leave on:
| Before | Now | |
|---|---|---|
A session that ends several times (window closed, /clear, resumed and ended again) |
the whole excerpt re-sent each time its turn count changed | only the turns since the last digest, plus 2 turns of context and the earlier lines for the change-of-mind pass |
| Two imports at once (two windows, hook and nightly) | both send the same session | one at a time (flock); the second finds the first one’s digest |
| Model lane | whatever the session’s shell held: llm.plan tries *_API_KEY first, so a DeepSeek key in the shell sent the excerpt to DeepSeek while the nightly job used claude |
the CLI named at --consent --cli goes first |
Cost. Last 7 days (2026-10-24 to 30): 27 sessions, 41 session-days, 316 turns, median 4 turns and 3 400 characters per page. Distilling the 32 session-days with at least 2 turns sends 444 000 characters (about 14 000 per session-day, 63 000 a day; for comparison the 2026-10-28 nightly sent 138 000 for topic dossiers and 54 000 for answer cards). With the incremental digest, moving this to session end costs about the same as doing it at night: the extra per additional end is the prompt templates, 2 context turns (at most about 1 500 characters each) and the earlier lines, so a few thousand characters at most, instead of the whole excerpt again. Each end runs one python process and, for a session that grew, 2 or more model calls in the background; nothing waits on it.
Privacy. What is sent does not change: redacted questions (up to 600 characters) and closing answers (up to 900), no tool output, no thinking. What changes:
--revoke or delete a transcript; at session end
it is sent within a minute. redact() removes key-shaped strings, not personal data: on 2026-10-29 the author’s
sessions held a third person’s name and date of birth, which either path would send.--cli).claude -p used by the distill itself) end too; their transcripts have no human turns and are
skipped, after waiting for the lock.Keeping a session out, built since: #不记 (or #nocapture) in any question of a session keeps it out of import
and distill, and removes a page already written for it; --exclude-root <dir> (repeatable, saved with --consent)
does the same for every session under that project root.
A desktop window often stays open for days and is archived long after the work in it is done, so SessionEnd fires
late or not at all; it is a top-up, not the trigger. What runs the work on a schedule instead:
extras/host-daily-evolution/intraday_capture.py, from an hourly timer
(launchd.intraday.example.plist: StartInterval 3600), runs import_sessions.py --intraday when a slot is due.
Slots are wall-clock times in one zone (--slots 12:10,18:10 --tz Asia/Shanghai by default), so the machine’s
daylight-saving change does not move them. A machine asleep at a slot runs it once on the first tick after it
wakes (the latest missed slot only); a failed slot is tried once more on the next tick, then left to the nightly
job. --intraday needs consent, uses the lane and distill choice recorded then, and covers today and yesterday,
so a session that crossed midnight is seen; with the incremental digest a page that did not grow sends nothing.
State and a one-line summary per run: ~/.local/state/krouter/intraday.json, intraday.log.session_hook.py prompt (UserPromptSubmit) appends a typed correction to
05 时间日志/会话摘录/YYYY-MM-DD｜纠正候选.md (status: provisional, time, project, claude:<id8>, the cue word,
“待下次蒸馏判定”). Pattern match only: correction cues (改成/改为/换成/以后都/一律/锁死/不要再/别再/记住/纠正/搞错/说错,
from now on, instead of, that's wrong, …), minus a sentence that ends as a rhetorical question (不是吗/对不对/
是不是/难道不是, isn't it). Skipped: no consent, a prompt over 1500 characters, a session marked #不记, and any
prompt under a KRouter model call (llm.cli_env() sets KROUTER_IN_DISTILL, so the distill’s own claude -p
never records its excerpt as your correction). Redacted, cut to 300 characters, nothing sent, never fails the
session, prints nothing into it. The page is a clue (time-log tier in recall); the next distill of that session
settles each candidate on its own (corrections.py): a digest line sharing enough of its content words makes it
成立 (the verdict names that line and page), none makes it 不成立. No extra model call.Capture writes memory without a yes: a yes/no gate on every digest line was one more chore that did not get done, so nothing ever took effect. In its place every run that changes something is visible and reversible:
[krouter-run <id>]. Other changes in the vault are left out of it (git commit --only). Hook-written
candidates are committed as they were just before they are settled, so undoing a run puts them back to pending.90 系统文件/自动化/记忆变更记录.md (status: generated, so recall never cites
it): each page that gained lines and its session, each added line (a change of mind shown as 取代：原来 A → 改成
B), each candidate’s verdict, each page removed. The entry is committed separately, so undoing keeps the record.import_sessions.py --undo <id> reverts that run’s commit and notes the undo in the ledger.KROUTER_INJECT=confirmed
restores the earlier confirmed-pages-only default. A run that changes nothing writes no commit and no entry.bench/decision_capture/)Set 6 (bench/real_questions/capture_bench.py) measures capture on the author’s real sessions with model judges; it
is private and its numbers depend on the judge. This bench is the small public counterpart: 14 written sessions (10
Chinese, 4 English) with 14 labeled decisions, 3 of them changes of mind, and 4 sessions where nothing was decided
(discussion, a one-off, steps only, “I’ll think about it”). Scored by rule: each decision lists the words a line must
hold; forbid lists steps and discussion. Reported: recall, change recall (a line with “→”), false-extraction rate,
steps-or-discussion lines, lines on the no-decision sessions.
python3 bench/decision_capture/run.py --oracle # scorer check: recall 1.0, false 0.0
python3 bench/decision_capture/run.py --cli /opt/homebrew/bin/claude --save lines.jsonl --show
python3 bench/decision_capture/run.py --replay lines.jsonl
No baseline yet: on 2026-10-30 the claude CLI answered “OAuth session expired and could not be refreshed”.
Lexical scoring undercounts a correct line worded with none of the listed alternatives; --show prints misses and
false lines to read. Tests: tests/test_decision_capture_bench.py.