romanprigodskii/indelible
A self-study tutor: diagnostics, closed-sheet sessions, cold rechecks, spaced mistake review and calendar planning. Works with Claude, ChatGPT Work and Codex; ordinary chats use manual records.
Changelog
[0.1.7] - 2026-11-13
- Add portable OpenAI plugin packaging, a repository marketplace and Codex skill metadata while retaining the Claude plugin.
- Adapt runtime selection to actual Python, storage and isolated-builder capabilities in Claude Code, ChatGPT Work and Codex. Resolve script paths from the installed skill rather than requiring Claude environment variables; preserve existing CLI and data schemas.
- Create
AGENTS.mdrouting beside the canonicalCLAUDE.mdin new workspaces and live/paused subjects. Addinstructionsto backfill older workspaces, preserve existing text, back up changed adapters and reject broken markers before writing. - Add standalone ChatGPT/manual instructions with a complete portable Markdown record, diagnostic-first sessions, cold rechecks and dated mistake reviews. Manual results remain
[unverified]; automatic pack/unpack, hidden generated keys without a builder and cross-chat storage are not claimed. - Document installation, host limitations and provider handling of uploaded records. Extend regression tests to cover adapters, packaging, versions and the privacy check's narrowly allowed manifest identity fields.
- Pin Linux CI to Ubuntu 24.04, which still provides Python 3.9, after the
ubuntu-latestmigration to 26.04 broke interpreter setup.
All notable changes to this project are recorded here. The format follows Keep a Changelog, and versions follow Semantic Versioning.
[0.1.6] - 2026-09-28
The night audit: rechecks that keep counting after the first, a path for a new topic that didn't land, marking that blames Claude's sheets rather than the learner, answer keys that open only as far as the evidence, and instructions that say each rule once.
Changed
- Level 4, upkeep and sentinels are scheduled. Mastery 4 needs a second cold pass a week after the first, and
plan.mdpromised a sentinel about 4 weeks after a mistake retires, but nothing ever listed either, so level 4 and 5 were out of reach and owned topics went unretrieved until the mocks.ind due --listnow has three more tiers after "oldest due": 5, the sentinel of each retired mistake 28 days after its last pass (or a week before the date), read from the archive too, until one is graded; 6, the level-4 recheck of a topic at 3, from 7 days after its first pass; 7, an upkeep recheck of a topic at 4 or 5, 21 days after its last pass (cold, mock or checkpoint, never one sat within 24 hours of the topic being seen), brought forward to 2 days before the date and never on or after it. "Needs repair" is tier 8 (8_needs_repairin--json). DUE counts them ("later rechecks ready now", "last check"), and the plain-words table insheets.md§12 gives both. Rule R5 again inmethod.md(a default graded [one-learner], since its week and 3 weeks are parameters; R5 and principle 3 carry the literature),session-open.md§5,plan.md§6,measure.md§8, the builder's SERVE lines, the README's method and the contract say so. - A recheck that leaves a topic below mastery 3 comes round again. After a topic's first 2-day recheck nothing served it cold again, so a topic that missed it stayed at 2 even after its mistake was fixed and passed four spaced returns, and topics built on it stayed gated. Now a topic whose last recheck left it below 3 gets a new 44–72 h window from its next warm exposure (the fix sheet, the logged feedback, drills), which books it as
session taughtdoes (a protected obligation to place, unplaced on demand; not while a wrong idea on it is unfixed, whose repair books it), andind due --listoffers it in tier 1, marked "again"; lint L7, the "Start by" line,plan checkand the late-recheck flags judge it by that window, as a first serve. Awordsrecheck, which levels don't read, never opens one:topics.jsonkeeps the type of the last cold serve (last_cold_type).grade recordnames each topic a recheck left below 3 and what opens its window, anderror repairrecomputes the levels before it moves a window, so a topic held at 2 only by that mistake is not treated as one.measure.md§8,session-open.md§5,session-grade.md,session-teach.md§1 and the contract (with its late-recheck,session exposeanderror repairlines) say so. - A fix comes at least 24 hours before its recheck, not 12. Method rule R6,
plan.md§4 and §5 and close check C5 asked for a repair at least 12 h before the recheck it serves, but a repair is an exposure, and the 24-hour rule takes its topic off any recheck less than 24 h later. All of them now say 24 h: C5 flags a recheck within 24 h that holds a topic with an unfixed mistake, and says to move the recheck or serve it without the topic. The contract givesrepair's due date as the next day, with the 24-hour rule setting the earliest serve.error repair, which promised a recheck "at least 12 h after the fix", now names the earliest time; the brief's BELIEFS DUE and OTHER DUE lines add "not now: " to a mistake that can't be served yet, asdue --listdoes; andsession-teach.md(after a repair and after a mock-miss review),measure.mdandtaxonomies.mdsay 24 hours. - Practice sheets say the learner may ask for a hint. The hint ladder and the worked-example fallback start only when the learner speaks up, but every drill sheet said "no AI" and nothing said they could ask, so a stuck novice sat in silence or wrote "I don't know" throughout. On theory, external, example, repair and drills sheets, the "I don't know" line now adds "Stuck on a question after a real try? Tell me its number: you get a small hint, never the answer", and drills say "no other AI". Mixed, review and measuring sheets are unchanged. The drills hand-over and the welcome card say the same, and the floundering timeout counts from the learner's first ask, which Claude can see. The sample workspace was rebuilt.
- Only the session-open brief counts an open. Every
ind brief <s>used to count one for the "not taken after 2 opens" rule and writedata/sheets.jsonl, so two status looks a few hours apart could bring "Sit them now, or drop that sheet?".ind briefnow writes nothing;ind brief <s> --open, run in session-open step 1, counts. The flag reads "not taken after 2 sessions". Status really changes nothing, as the README says. topics.jsonno longer storescold_passes. The list was filled by any sitting at 75% or more, including one-question, out-of-window and practice sittings that the level rules reject, and nothing read it. Cold passes are found inattempts.jsonlby the level rules, andlevel_basisnames the one that counts. An older workspace's list is left as it is and ignored. The sample workspace was rebuilt.- The exam date question states the published format instead of asking for it. The exam Q2 card asked about ten things, among them attempts and retake cost, which nothing records or uses, and a public exam's full format, which Claude can state. It now asks the date (or window), registration, start time and accommodations, then says "I'll plan to the published format: … [published]; say if yours differs"; with no public format, or one Claude isn't sure of, it asks for the format instead and never states one from memory. Q4 (materials) had only exam wording, so a language or code learner was asked about unseen official papers: it now has a language line (an app, a book, a course, a phrasebook) and a code or skill line (a book, course or tutorial to follow).
profiles.mdno longer mentions retake questions. - An on-demand learner's return is one short opener. After two weeks away, persona D got the re-entry session, whose last step re-plans a week an on-demand learner doesn't have, and the weekly review's 3-line summary was never wired to the open, and counted what was due ("Due now: one 2-day recheck and 3 reviews"), which re-entry bans. On demand, re-entry now skips the week re-plan (the close names the next window as usual), and its opener carries the 3 lines, which count nothing due and ask nothing; the first comes from the brief's LAST SESSIONS alone (sessions and questions, no recheck rate), since the open may not run
ind stats: the third is a statement (Say "change …" any time). The why line uses the learner's own words and is skipped when all there is is a one-word label. - "How am I doing" is a status question, and status has a shape. The trigger sent a casual question on the train to the weekly review: three check-in questions, a
reviews/file, compaction and proposals. Status had no answer shape, so replies varied and could show raw metric names, and nothing told Claude to keep the brief's private block to itself, so a status look before a recheck could name its topics. "How am I doing" now routes to status (SKILL.md,README.md,review.md); the review's triggers are "weekly review", "review my week" or a yes to the close offer. A newreview.md§9 gives status at most 5 plain lines (this week's sessions, the latest 2-day recheck, what's solid, the next session, one to-do), never reads out anything below the line, translates stats labels, writes nothing, and ends with "Full weekly review? (yes/no)". Learners with sessions of 30 minutes or less get the weekly review as 3 lines with that offer, as the week-1 "3-line check-in" inmeasure.md§11 promised; the check-in questions come only on a yes.close.md§8, read at every close, now makes that offer (or gives the 3 lines) at the first close of a new week; nothing made it before, so the review ran only when asked for by name. - The learner is told how photos reach the session. Paper plus phone photos is the main answer route for exams and courses, but nothing said how a photo gets from the phone into the terminal, and the P8 offer of a synced photos folder had no follow-through. The welcome card now has a photos line fitted to the learner's computer and phone (AirDrop or the inbox folder on a Mac; email the photos to yourself and save them in the inbox on Windows; attach them in Claude Code desktop; a paste works too, but keeps only a transcript), and the recheck, drills and diagnostic hand-overs point to it. When the learner says "sent",
session-grade.md§2 lists image and PDF files newer than the sheet in the inbox, Downloads and any recorded photos folder, names them, and files them only after a yes. P8 now asks for a folder just for study photos and records its path in the rootCLAUDE.md.PRIVACY.mdsays that those file names reach the conversation. - Marking names each question by what it asked, not its number alone. The verdict list, the account menu and every feedback line said "Question 4", even for a sheet sat days earlier (a solo block, marking carried over, a photo sent later), so a learner who does several sheets a day had to dig out the old sheet or gave no account.
session-grade.md§3 now names each question by its number, a gist of at most 12 words from the visible spec (never the key) and the learner's answer as filed: "4 (median of six waiting times; you wrote 11.2)". An official item, which has only a pointer in the spec, is named by its number and the learner's answer, never the official one. The account menu, feedback, the result card and diagnostic accounts use the same form; the Right line stays numbers only. A vocabulary probe quotes the sentence the word came from instead of "in question 6", and "Which questions?" before a recheck is marked adds "Topics in your own words are fine." - After a slip, Claude names the check instead of asking the learner for it. The quantitative taxonomy's careless treatment still said "ask which check would have caught it", a last quiz after the pointer that
session-grade.md§7, where Claude names the check form, had already dropped. It now reads: point to the line and the learner's value ("last line: your 0.35"), then name the check that would have caught it.session-grade.md§6 adds the value to the pointer when the miss is a number: the learner's value, never the right one, so the corrected line stays theirs. - The first-teach warning covers sittings and prompt suggestions. It asked the learner not to expand Claude's tool calls or thinking only while sheets were built, but a sheet's answers are also in the conversation while the learner works it and at marking, and the app's prompt suggestions are written from the conversation too. The one-time line in
teach.md§8 now covers working a sheet and asks the learner to turn prompt suggestions off in/config. - A mock is sat without check lines, as the exam is. Every mock, and every official question on a diagnostic or checkpoint, needed a written check beside each answer (lint L2 failed one without), while
measure.mdasks for exam conditions and the exam has no check column: at about 90 seconds a question, copying the supporting words took a large share of the clock, so a mock measured a slower task than the exam, and a checkpoint's pre-registered threshold, whoseif_belowis binding, was applied to it. L2 now passes amockwith no check lines, and an official question (origin: official:) on a diagnostic or checkpoint; a question Claude wrote there keeps its check. The rules box of a mock, or of a sheet of official questions only, says "Check your answers as you would in the exam: there is no check line on this sheet.", and a sheet where only some questions have one says "where a Check line is printed". Such questions are gradedcheck: n/a, which check coverage leaves out.builder.mdrule 4,sheets.md§1–§3 and §7,profiles.md,measure.md§3, §5, §6 and §12,session-grade.md§7, method rule R41 and §6, the README and the contract say so. - Feedback never tallies the learner's repeats. Law 10 banned praise, sarcasm and "simply", but not counts of the learner's repeats ("that's the fourth time", "every session this week"), a person-level judgement that names no error; in the method's development such tallies were frequent, several of the "habits" counted turned out to come from the sheets' own design, and they fed the learner's frustration. Law 10 (in
SKILL.mdand the contract, in the same words) now says never to tally the learner's past misses ("you did it again", not a bare "again", which check hints use), and thesession-grade.md§6 Don't table has the row; a miss on a mistake's own re-serve may still be named as a fact about its fix, with no count.taxonomies.md§7 has repeats checked first for a shared item design (the last sub-item, a printed formula, a multi-part chat question, an undefined word, a late block), which is logged as Claude's defect with a structural fix; otherwise one structural change is proposed at the weekly review, never a reminder.review.md§6 and method's Law 10 row point to it. - A label says what form counts, and marking holds to the exam's form. The key accepted every defensible form, but nothing told the learner so, and nothing covered the form an exam requires: a learner who believed an answer isn't finished until its brackets are opened spent the clock on algebra, while an answer in the learner's own name for a standard rule was marked right though the exam would reject it.
builder.mdrule 4 now has each label say "any equivalent form" or name the one form wanted, and on exam and course subjects the exam's required form in its own wording (sheets.md§5 has the Don't row).session-grade.md§9 says "defensible" means what the exam's marking accepts: a right answer in a form the exam won't take ishalf, in the subject's answer-form code, a slip, with the exam's form said in one line, and anitem_wordingdefect when the label never named it; the quantitative taxonomy's F row has an own-name example. The verbal taxonomy gains an F row (over the word limit, or a word where a letter was asked), so persona A's codes, and the sample workspace, have one; a language, code or essay subject marked to an exam's form adds its own letter, since the language table's F is grammatical form (taxonomies.md§1). No standing rules-box line was added: "form" means a verb form on a language sheet. - The Least-sure line asks about the idea, not a slip. Its label, "Least sure of (item numbers)", said nothing about which doubt it meant, so a careful learner named the items where an arithmetic slip was possible, which the method itself calls the wrong signal, and a named right answer then cost its level. The label is now "Least sure I chose the right idea (item numbers):", one wording for every subject, and every reference, the README and
SKILL.mduse it. The rules box's time line adds "Then fill in the Least-sure line: item numbers, or “none”.", so a line left blank no longer reads as "none"; the box stays at 8 lines. The sample workspace was rebuilt. - Naming a right answer on the Least-sure line costs a delay, not the level. Named questions never counted toward mastery, even when right, and
--shakyopened one mistake per question. So an item named on the only in-window 2-day recheck held its topic at 2 even after every +3-day re-serve came back right, with a burst of re-serves for one doubt, and leaving the line blank became the learner's best move.grade record --shakynow opens one shaky mistake per named item and topic, holding all its right named answers. When that mistake's first counted re-serve comes back right, those answers count toward the level at their own sitting, so the recheck can still make 3; a miss there leaves them out, and a right answer is never counted as wrong.grade recordthen names the recheck that now counts (only one whose named answers a re-serve on the sheet being graded confirmed, never that sheet's own pass), sets it as the topic's last recheck and closes its booking.method.md§4,measure.md§8,sheets.md§4 and the contract say so. - A floor the diagnostic called "Already yours" is enough to build on. The floor check asked for "mastery 3 or above", but a diagnostic can give at most 3p, which the CLI ranks below 3, so a learner told "Already yours: z-scores" was then told z-scores "isn't solid yet", and "the new topic next time" was wrong for any floor that needs a 2-day recheck first. The floor check in
session-teach.md§2 now asks for level 3p or above; a floor below it is served first, and the new topic waits until that floor's 2-day recheck has passed.measure.md§10 calls a floor owned at the same line,close.md§9 runs the floor check before building the next theory ahead (and builds the floor topic's sheets instead), and the weekly discrimination sheet for a confusable pair counts 3p too (session-teach.md§5,plan.md). - The hint ladder says where a pointer comes from. Rungs 3 and 4 ("look at the verb in the second sentence", "one step further") need the question, but nothing said how Claude gets it during a sitting: the build step said not to review a sheet's content, and working the question out would put its answer in Claude's thinking before anything is filed, which Law 1 forbids.
session-teach.md§3 now says: before anything is filed, hint only from what the learner has written or quoted; on drills, file that question's work alone (scan ingest --asks 5a, as a gate photo is filed), sokey openprints only its entry and the pointer comes from its solution; on theory, external, example and repair sheets, point to the sheet's own worked case, read from the visible sheet (never.indelible/). Pencil marking names the step from the visible theory sheet too.sheets.md§6 limits "don't review the sheet's content" to build time, andsheets.md§10,session-grade.md§2, the--askshelp and the contract name the hint filing. - A sentence question follows the numbers it is about. Builder rule 3 put a sentence or verbal item first on every drills sheet, "while the learner is fresh", and lint W2 warned otherwise; in the method's development, sentence questions set apart from their numbers were missed while the computed questions beside them were right, and the same content asked as numbers first, then one sentence inside the item, landed. On a drills sheet with both computed and sentence questions, each sentence question is now the last question of the item whose numbers it uses, and names them ("Using your answers to 3a–3b, …"). W2 warns on a sentence item with no computed item before it on its topic, and passes a sheet of sentence or reading items only. New method rule R51 ([one-learner]) and a
method.md§6 note give the reason;builder.md,session-teach.md§4,sheets.md§1 and §7 and the contract say so. - Worked cases carry a principle prompt. R12 cited the fading research, but the skill took only its fading half: its one self-explanation prompt was an open "why does this step follow?" on drill item 1, and theory and repair worked cases had none, though in that research prompts to name the principle behind each worked step are what carried the method to unfamiliar questions (Atkinson, Renkl & Merrill, 2003). Every worked case that gets marked (theory, repair, and drills item 1 at mastery 0–1) now gets one or two keyed choice questions on a named step, "Step 2 of the worked case is done because: (a) … (b) … (c) …", answered by letter, the options being the rule, the contrast's wrong idea and the warning's wrong turn, all on the sheet; on a verbal subject a focused "why not " may stand in on drills. They are marked like any pencil, count in
est_min, and never go on measuring sheets, nor on an example sheet, which is read and closed with nothing marked. R12 reads "…with a principle prompt on each worked case" and cites both studies;builder.mdrules 7 and 8,session-teach.md§1, §2 and its "Why this shape" line, andmethod.md§6 say so. The sample theory sheet has one. - On a new topic the failure gate stops every time. The gate after item 3 fired only on 2 failed checks, "I don't know"s or blanks, but a wrong idea passes its own check (the mean put back into the mean formula holds), so a novice could practise it through a whole block, often two, before any feedback; and in sessions of 30 minutes or less, marking of new drills could carry over to the open of the day that topic's recheck was due, just before it or not at all. A drills block with a topic below mastery 2 now carries
gate: "always"(lint L6 requires it; no level on file counts as 0), and its sheet prints "Stop here and send a photo of items 1–3. Go on once I've marked them.": Claude files those items, opens only their key, gives the verdicts, the account and the standard for each miss, and repairs a wrong idea before item 4. Such drills are marked in the session they are sat, andsheet issuerefuses them on a solo block, which has nobody to mark them, so a new topic is taught at a session with Claude. In short sessions, drill questions are cut rather than their marking; marking that still carries over happens at the open, before the recheck builder starts, with the learner's account for each miss, and logs areviewexposure so the topic comes off that day's recheck. R36 andmethod.md§6 cite Shute (2008);builder.mdrule 7,session-teach.md§4 and §6,session-open.md§2 and §4,close.md§6 and §8,plan.md,sheets.md§1, §2 and §7, the schema and the contract say so. The sample drills stop after item 4 for marking. The sample workspace was rebuilt. - A recheck is warned when it asks for an operation never shown, and every pencil names its operation. Lint L12 tied a drills sheet's operations to its theory, but a 2-day recheck could still ask for an operation no sheet of the topic had worked or practised, and a theory pencil with no
opleft the drills nothing to match. New warning W7 names acold:item whoseopno earlier theory, external, example, repair or drills sheet of its topic used (a topic with no teaching sheet is skipped, and a recheck is never blocked); builder rule 1 keeps recheck items to operations the topic's sheets used. L12 now also fails a pencil question with noopon a theory, external, example or repair sheet. No exemption was added for a fully worked item: a drills block holds one operation, so it could never change the result.builder.mdrules 1 and 8,sheets.md§7 and the contract say so. - Drills sat without Claude follow the same marking rule.
session-open.mdstep 2 andsession-teach.md§6 marked a solo block's photos after the day's recheck, while marking of new drills carried over to the next open now comes before it: drills on a topic whose first 2-day recheck is still ahead are now marked in the open, before the recheck is built, in both places, so the recheck is never sat on drills nobody has marked. - A learner who insists on a recheck too soon gets practice, and the recheck stays booked.
session-open.mdsaid the 24-hour rule and the recheck window could be overridden with the result labelled, but the only instrument it named, acoldsheet, can't pass lint L7, and every workaround (another type, another origin) is banned, whileplan.mddescribed a different route. Now one route: the booked recheck stays as it is, and the questions the learner asked for come as amixedpractice sheet whose items are allorigin: new(never acold:orerror:item), shown as "[practice] not counted (seen too recently)" and logged withsession override; grading logs a drill exposure, thenplan checksays whether the recheck must move.SKILL.md's ban,session-open.md§3 step 6 and §5,plan.md§4, the L7 row insheets.mdandbuilder.mdrule 12 say the same; the gap between measurements keeps its own route (measure.md§4). - A quick session keeps its recheck, and an on-demand session can teach. A quick session was "an unplanned drop-in, or one shorter than the learner's usual length" serving "nothing new and no measurement", which covered every on-demand session (they have no planned blocks), forbade the first teach
measure.md§11 gives an on-demand learner, made a 60-minute drop-in by a 120-minute learner a session with nothing new, and could be read as dropping the 2-day recheck, which is a measurement.session-open.md§4 now defines it for scheduled mode only: opened with no block and under half ofsession.length_min, it serves the 2-day recheck if its window is open, then the other due items in tier order; a new topic only as an override (session override), sized to the minutes left; no diagnostic, mock or checkpoint. A longer drop-in is an ordinary session, and on-demand sessions are never quick.sheets.md§6 says which sessions issue sheets with no--block. - Missed sessions are asked about in one message. The open asked about each missed block in its own message, then "what got in the way" after each skip, then the alarm's three choices, so two missed commutes could cost a 20-minute train session five exchanges, on the very "missed a couple of days, felt behind" that had made persona B quit before; it never used what the learner had already said, and it listed the misses before re-entry's promise not to.
session-open.mdstep 2 now asks about all of them at once, by day ("Monday's and Tuesday's sessions have no record…"), records what the learner already said without asking, asks "what got in the way" at most once per open and only after a skip with no reason, and at re-entry neither lists nor counts them (one question, only when a block was the learner's own). The three choices come in their own message, with no count ("IELTS hasn't been running lately"), and, for the subject being opened, only once two skips in a row are confirmed, and never at re-entry, which re-plans the week anyway. In a session of 30 minutes or less, or when the notes hold the learner's if-then plan for misses (persona B), they come in the close message instead of at the open, in place of a weekly review offer, so a 20-minute train session starts on its recheck.plan.md§2 and §7,session-teach.md§6 and question P6 say the same, andplan.mdno longer calls persona B's commute sessions soft blocks, which are never asked about. - A recheck whose window has passed no longer blocks the plan.
plan checkfailed a recheck that was never placed and whose window had closed, and one placed after its window, though nothing done while planning can fix either: the late-recheck rule runs at the next session.plan.mdandcalendar.mdallow no preview and no calendar write until the check passes, so re-planning was blocked for days, and the pull was toward an earlyplan cancel, which also clears the brief's LATE RECHECK flag. Both are now a WARN,late_recheck, with the same fix, so the plan passes and the WARN becomes a preview line. A recheck placed outside a window that is still open still fails. The contract andplan.md§4 say so. - Moving the recheck asks for no prediction. A "skip the recheck" override moves the recheck, but
session-open.mdstill asked for a one-line prediction about specific questions of a sheet not yet built, which the learner had never seen, so any answer was a total in disguise. It now gets no prediction and nosession override: nothing is sat, and the move is the record. Law 9, the contract's copy and method rule R34 say so. - Re-entry asks for the learner's reason once when the notes have none. The re-entry session opens with the learner's own "why", from the learner notes the brief prints, and skipped the line when it was missing there, so a learner coming back after a week away, the moment the reason matters most, heard nothing of it. It now asks once ("in a line, why does this matter to you?"), lets the answer stand in for the line, and records it under "Learner notes" in the subject
CLAUDE.mdonce the recheck is handed over, so the next brief carries it. When the subject has notarget.why, which is where the weekly review reads the reason, it goes there too withind set, after a--dry-runshows the field empty, since the open never readssubject.json. It is never looked up insubject.jsonat the open (Law 6); "skip", or a one-word label, leaves the line out. - A learner who answers from a phone has a route the rules allow. Persona B reads sheets and types answers on a phone, but
profiles.mdandsheets.mdsaid typed answers go "in a file, never chat", the recheck hand-over asked for paper and a photo, marking suggested a typed file a phone can't make, and the HTML sheet itself said "Answer on paper". Withformat.answer_form"typed", the rules box in every format now asks for the answers typed, numbered as on the sheet with their checks, and sent in one message (or a text file) when the learner stops, with the start and stop times and the Least-sure line; a failure gate asks for "your answers to" its three items, not a photo. The message is filed verbatim withind scan ingest <s> <id> --typed -(evidence kindtyped, not a photo transcript). The recheck hand-over (whose typed form replaces both the paper and the photos lines), the welcome card, readback row 9,sheets.md§8 and §10,session-grade.md§2,session-teach.mdandprofiles.mdsay so. The route follows what the sheet asked for, not where the learner sits: every language subject's answer form is "typed", so one message is the normal route at a laptop too, and marking suggests a typed file only for any other answer form. The sheet is still handed over as an HTML or Markdown file; a theory sheet pasted into chat because the file won't open sends its drills to a new conversation or a later session (Law 2). - The language placement fits a 20-minute session. It was one
probeof 40 words in sentences, 5 can-do prompts and a listening item, billed as 10 minutes, but a probe holds at most 10 questions, an honest estimate at the default paces came to about 66 minutes, andsheet issuerefused it in persona B's 20-minute block. It is now two probes, in sessions 1 and 2, pooled as one measurement (the CLI already pools probes sat within 72 h): at most 8 high-frequency words in sentences, answered with a one-word gloss, then 3 can-do prompts answered in one typed line each, plus a listening item when there is audio. Both go in session 1 when its work minutes hold them.measure.md§1 and §11 (B's first teach moves to day 3, its recheck to day 6),profiles.mdand the persona B fixture say so. - The daily 15-minute slot for long sessions is a logged catch-up, not a review. Learners on 120-minute sessions were offered "a daily 15-minute review" with no content and no timing, and as worded it could send persona C, whose risk is cramming, to reread her notes each night, an unlogged exposure that voids the next morning's recheck. It is now a 15-minute catch-up, offered in scheduled mode only (
session-open.md§4 says an on-demand session is never a quick session and may teach, so a catch-up there would not keep to due items), in the preview as an if-then plan in the learner's words: a logged quick session (ind session open <s> --planned 15, no block) that serves due rechecks and mistakes only, never new material and never rereading notes. Taken up, it goes under Learner notes.teach.mdQ6,session-open.md§4, acatch-uprow inmethod.md(opt, [design]) and the persona C fixture say so. - The materials question has an interview wording. Q4 gave interviews, vivas and oral exams the exam wording, so a learner preparing for a panel was asked about unseen official past papers and question banks. It now asks what they are preparing from: the job advert or brief, their own work (a thesis, a portfolio, a project) and any list of likely questions, plus a book or course to follow; model answers are kept sealed until marking.
- A one-off request with an exam ahead gets its answer, not the onboarding interview. The skill description was 1,013 of its 1,024 characters, about a third of them on how the skill works, and it excluded "one-off questions with no ongoing goal" while setup step 3 sent anyone with "a lasting goal" to
teach: "explain Bayes, my final is in 3 weeks" fitted both. The description now names the triggers in 959 characters: every one it had, with marking (from photos or typed answers) named as a trigger in its own right, plus the learner's own revision timetable and a question asked in a workspace (which goes toask), and it excludes a single explanation, quiz or flashcard set with nothing to track, even before an exam, and marking other people's work. Setup step 3 decides by what was asked: a single answer, explanation or quiz is answered, with one line offering setup; a session to be run ("start maths") goes toteachwith "just start" offered; study to be set up, planned, or tested and tracked goes toteach.evals/trigger.jsongains near misses (a quiz now, flashcards, a child's spelling test, an explanation with a final ahead) and four more that should start the skill, among them a revision timetable and typed answers to mark. A newtests/test_skill_meta.pychecks the description (at most 1,024 characters) and compatibility (at most 500), and that every trigger entry has a query and a true/false label, no query twice and both labels; the README says the set itself needs a model to run. - Maths in chat is bracketed so it reads one way. The ban on LaTeX in chat asked for Unicode maths but said nothing about writing a fraction or a power on one line, so "x + 1/2n" could be read several ways.
SKILL.mdnow asks for brackets around any numerator, denominator or exponent of more than one symbol: (x + 1)/(2n), e^(−x²/2). - The builder's own rules live in
builder.md. The main conversation loadedsheets.mdin every session, and more than a third of it was for the builder: what the templates print (§2), the don'ts table (§5) and the checker's rules L1–L14 and W1–W8 (§7). They are now sections ofbuilder.md, which the builder reads, andsheets.mdkeeps a short stub under each old heading, so every "§N" pointer still lands. The theory and repair page orders insession-teach.md§1–§2 were a second copy of builder rules 7 and 8; the repair order moved into rule 8, andsession-teach.mdkeeps one line each. A test now checks that every pointer to a reference's section or step, in the docs and in the CLI's messages, names one that exists. - One table of check forms. The written check for each kind of question was set out four times, in
sheets.md§3,session-grade.md§7,session-teach.md§4 andprofiles.md§2, and the copies had drifted apart (for a proof, "Which step would you be pushed on? Do it another way." on the sheet, "Which step are you least sure of?" at marking; for language production, a back-translation "into your own language" on the sheet, "into the instruction language" in the profiles, which for a learner taught in the language they are tested in is the same language).sheets.md§3 now holds the only table, whose hint is also the line to say when coaching; the other three point to it and keep only what is theirs: when to coach, what a failed check means at marking, and which row a topic's layer takes.taxonomies.mdsends a careless slip's check there too. - The session open uses the budget the CLI prints.
session-open.md§4 spelled out the budget formula (open and close minutes, work fractions, the grading estimate) and three worked figures, and then said never to recompute them. It now keeps only whatind session openprints and how to use it, and the short, quick and measurement sittings. §5 lists the recheck's tiers in one paragraph, asind due --listprints them. - Evidence routes and materials entries have one home each. How to file each kind of evidence was a table in both
sheets.md§10 andsession-grade.md§2; the one table is now insession-grade.md§2, with the code, online-test and earlier-day rows it lacked, andsheets.md§10 points to it. The shapes ofmaterials.sourcesandmaterials.ration, and the rationing rules thatmeasure.md§12 already gave, sat insheets.md§11, which every session loads; they are now inmeasure.md§12, andsheets.md§11 keeps the labels. - A session reads each of its references once. The session route in SKILL.md said to load
session-grade.md"again for each later sheet", so a long session with a recheck, pencils and drills read the same file three or four times. It now sayssession-grade.mdis loaded at the recheck, or at the first sheet back in a session with none, and stays loaded for the later sheets (a test checks the row), and a reference is re-read only after a compaction;close.mdis still read at each close. - The questions asked later live with the onboarding questions. The table of follow-up questions (P1–P11, and the express questions put off) sat in
session-open.md§6, which every session loads, andteach.mdpointed to it. It is nowteach.md§10, beside the questions it defers.session-open.md§6 keeps the triggers that fire in a session and points to the words,close.md§8 points to Q4's own words, andreview.mdasks P1 and P3 at week 1 and P10 at week 2. - A session's instructions say each rule once, under a cap. SKILL.md and the five references a session reads (
session-open,session-grade,session-teach,sheets,close) had grown in this release to about 160,000 characters, much of it the same rule said in two or three files. With the moves above, each rule is now said once, in the file that owns it, with a pointer from the others: why the recheck is built at the open (session-open.md§2), the cut order (close.md§3), the looked-since question (session-grade.md§10), the solo-block rules (session-teach.md§6), the late-recheck steps and the alarm's choices (plan.md§7), the labels (sheets.md§11), the Least-sure marking (session-grade.md§3 and §8), and whatind due --listandind grade recordprint. Scheduling a tutor lesson is inplan.md§9 only. The "why" paragraphs at the ends ofsession-open.mdandsession-teach.mdare now links tomethod.md, which holds the reasons. The six files total under 110,000 characters (about 99,000 in 0.1.5, before this release's new rules), and a test keeps them there (README, "Contributing"; CONTRACT §1); another checks thatbuilder.md's checker table lists every lint rule in order, with its title.
Added
- Every sheet carries a code to match its photo. Titles are neutral by design, so every subject's recheck is called "2-day recheck", and nothing on a sheet or a notebook page tied a photo to its sheet. A page could be filed against the wrong sheet, which marks that sheet taken and opens its key.
sheet newnow gives each sheet a code, the subject id in capitals and a running number (IELTS-07, never a topic word;sheet buildsets one on an older row), printed at the end of the header line as "Sheet IELTS-07".sheet issue,sheet showand the "sheets out" line ofsession statusname it. A notebook page starts with the code and the start time, andsession-grade.md§2 matches the photo to its sheet before filing: with no code and more than one sheet out, or a different code, it asks which sheet it is. The recheck hand-over names the code, andsheets.md§12 says the code is the learner's label for the page, so the plain-vocabulary rule against showing ids doesn't cover it. The sample workspace was rebuilt. - A failed check is recorded, not reworked. A check that didn't hold, with the answer left as it was, could only be recorded as
filled, and nothing told the learner what to do with it, so a check could be rewritten until it agreed. The rules box's failed-check line, still one line, now says to mark it ✗ or "no" and leave the answer and the check as they are. A new check value,failed, records it. It counts toward check coverage, andgrade record,ind statsand the weekly review count failed checks, with those on right answers apart, since they point at the check, its tolerance or the key. For a rounded number, the check hint and the answers file'scheckstate the tolerance, never looser than the precision asked for, and marking holds to it (builder.mdrule 4,sheets.md§3,session-grade.md§3 and §7,measure.md§5). The sample workspace was rebuilt. - Lint L11 and L12 tie the drills to what the theory worked. The theory and its drills come from two builder runs, and nothing checked that a drill's operation had been shown, or even that a theory sheet had a worked case: one with only a rule section passed. L11 fails a theory or repair sheet with no
workedsection, or with the rule before it. L12 fails a new drills item whoseopno theory, external, example or repair sheet of its topic has shown, as a pencil question'sopor in a worked section's new optionalopslist; a topic with no such sheet (a tutor's, a migrated one) is skipped.builder.mdhas the drills builder read the paired theory and use its operation names, asks for one worked case per operation, never renames anopto pass, and fades the worked example in each new-operation block rather than only at the top of the sheet;session-teach.mdandsheets.md§7 say the same. The careless test counts none of those worked or part-worked items as the operation done unaided (session-grade.md§5,taxonomies.md); it still named only items 1–2. Since L12 matches operation names, a variant the theory never worked can still reach a sheet, so marking gains "an untaught case is my mistake":session-grade.md§5 gives such a miss nokindand logs a defect in the newuntaughtcategory, with a "builder:" fix that MY RULES carries to later builds. On a recheck the question is also left out ofgrades.json("Not counted (never taught)" on the result card), so a case Claude never taught can't fail the topic's recheck or hold it below mastery 3; on other sheets it keeps its verdict. ind session extend <s> --min Nrecords the one extension. Nothing recorded an extension, so during onesession statusprinted "closing time" at every check, a second extension could be offered, and after a context compaction the extension was forgotten. The command writes it to the lock, moves the close start, and refuses a second one or one over min(session.extension_max_min, a quarter of the planned minutes, the next fixed start less 15 minutes less the planned end). It finds that next start itself (another open block of any subject, atime.blockedspan, or 30 minutes before bedtime, counting only what starts once the session is under way) and names it with the cap; before another block it says to close or to move that block with the learner's yes.session statusshows "extension until HH:MM", and the overrun still counts the extension.close.md§2 runs it on a yes.- Session facts are written when they happen. A new
SKILL.mdfile-contract line: a promise goes toledger add owedas it is made, an extension tosession extend, a chat explanation tosession expose, and a discussed sealed question to its contamination defect, at once. After a compaction, Claude rebuilds the state fromsession statusand the open to-dos. The close checks the promises' to-dos rather than writing them from memory. - The brief raises the "hasn't run lately" alarm.
plan.md§7 relies on an alarm to stop a low-priority subject from being starved, but the brief never computed it and the read it needed was one the session open forbids, so it never fired for a subject nobody opened, and misses recorded across separate opens were never counted. Every brief now looks at every live subject in scheduled mode: the last 2 planned blocks both missed (a rebooked block's recorded miss included) with no session since, or no session for max(5 days, 2 × the median gap between its planned days). FLAGS gives it by the subject's title ("Spanish for a trip hasn't run lately: the last two planned sessions didn't happen"); ALARM below the line gives the block ids (a block missed, rebooked and missed again is named once, with both times) and the three choices. Cancelled blocks don't count as planned days, so dropping a subject shortens its limit rather than silencing it. A subject's own brief shows only the missed blocks, since a long gap there is the re-entry session; a paused subject, or one with an open "Ask again" to-do of its own (a to-do with no subject doesn't count), is silent.session-open.mdstep 2 is now the one script for missed blocks (plan.mdpoints to it) and offers the alarm's choices inplan.md's words, "Ask me again on " in place of "I know, ask me later". - The rules Claude sets itself after a mistake reach later conversations. A
rulefix ("a stated rule you follow from now on") and a "builder:" fix thatsession-grade.mdsaid to put "in the next build brief" lived only inledger.jsonland the chat, and the recheck is built at the next open, in a new conversation, so nothing read them again. The brief now ends, below the line, with MY RULES: the subject's open defects whose fix is aruleor starts "builder:", newest first, at most 5 (140 characters each, then "+N more"). A fix that proposes a change to the skill stays out.sheets.md§6 andbuilder.mdcopy the "builder:" lines into the builder's NOTES,session-grade.md§9 logs a key fix that way, and the weekly review's liveness check drops a rule nothing has used (ledger close). - The weekly review says whether the plan still fits the date. The feasibility verdict of
measure.md§9 ran atteachand on the results card, then never again, so a learner asking "are we going at the right speed?" got improvised numbers or general reassurance. The weekly review now re-runs it as one[mine]runway line when there is a date (with the learner's own step hours once two topics have a cold pass); a short verdict brings the three-cut question and is a re-baseline trigger inplan.md§8. "Am I on track?" and similar pace questions route to the review, which answers with that line and its figures, then offers the full review; with no date, it says what fits in the current 4-week cycle. ind doctorreports which language toolchains are on PATH, and code onboarding uses it. Stage 0 checked Python, renderers and the time zone but never the compiler, while the code diagnostic is a 15-minute capstone in the learner's editor, and on the express path the "have you ever run a program?" line of Q3 came only after that capstone was built.doctornow reportstoolchains(cargo,rustc,go,node,javac,gcc,python3,py): looked up on PATH, never run, so no toolchain proxy can start a download. The Found line names the learner's toolchain, and a missing one makes session 1 "install it, then the starter task". The express Q1 card for code carries Q3's code line, so "never" leads to the floor probe and the on-ramp rather than the capstone. The readback names the learner's code folder, outside the study folder, and the Learner notes keep its path forscan ingest --dir. The persona D fixture and the contract say so.- Moving the date has a recipe, and the CLI flags what the old date left behind. "The date gets fixed" was a re-baseline trigger and
measure.md§9 offers "move the date" as a cut, but nothing said what to change: the date also lives in the checkpoints, the rationed tests and each checkpoint's mirror decision, andset <s> target.datechanged only the date, so a checkpoint or final mock could stay after the exam withplan checkpassing.plan.md§11 now has "The date moves" (set the date, re-date or drop each checkpoint and rationed test, dropping one by writing its list again without it since neither has a dropped status, replace or close the mirror decisions, re-read the phase row and re-run the fit verdict, then move the blocks after the date).plan checkhas two new warnings:after_datefor a block after its subject's date (admin blocks aside) andcheckpoint_after_datefor an armed checkpoint or an assigned rationed test on or after it. Setting the date earlier also brings forward any open mistake whose stored due date the new date no longer allows (the date − 2 days, never before tomorrow), and prints each one. - Solo blocks get their sheets before the learner leaves. A block the learner works alone had its sheets "handed over in the close message before it" (
session-teach.md), whileclose.mdbuilt the next sheets only after that message, and nothing marked a block as solo: when the build ran late or stalled, the learner sat down to an empty slot, and the 2-day recheck the drills fed slipped too.plan add --solomarks such a block, andplan move <B> --solo(or--not-solo) marks one already planned without moving it. Only a block with practice sheets to issue ahead can be solo: a 2-day recheck, a block whose content names one, and oral, tutor-lesson, buffer and admin blocks are refused, since their card would tell a learner on their own to sit a recheck or to ask Claude.plan listand the brief show it ("on your own"), and its calendar card names the sheets folder instead of "open Claude".close.md§6 builds and issues a solo block's sheets before the close message, which gives their paths, and close check C8, information only until now, fails while the next block is solo with no sheet issued (--defermakes it a to-do). The contract,plan.md,sheets.md,calendar.mdandsession-open.mdsay the same. - A question Claude worded badly can be withdrawn. "I didn't understand the question" had no path: an "I don't know" on an ambiguous stem went through "no idea" to a wrong idea and a fix sheet on content the learner owned, and only an undefined word counted as Claude's mistake.
session-grade.md§9 now has an unclear-question check, which comes first in the taxonomy's check order: reread the stem, and if it can't be said in five words what it asks and in what form, or it has two readings, ask one reworded line in chat (nothing from the key, no teaching). Right: the question is withdrawn ("withdrawn, my wording. Not counted"), left out ofgrades.jsonwith nokind, reported on the result card, and logged as a defect in a newitem_wordingcategory, apart from a wrong key'scontent_error, so the weekly review counts wording defects on their own; its "builder:" fix reaches later builds. Wrong: it keeps its verdict and the reworded answer is the account. A question the learner asks about while a sheet is sealed is noted for this check, and the builder's audit now asks the five-word test of every question, not only verbal ones. - A check done in the head is recorded as one. A learner may decline the check lines for a sheet, as any override, but a check they said they ran in their head could only be recorded as
missing: marking coached every blank on a sheet whose lines they had declined, and check coverage said they hadn't checked, so the data that would test the written-check rule was never gathered. A new check value,head[self-report], records it. It stays in check coverage's denominator, since it is not a written check, andgrade record,ind statsand the weekly review report it apart, with the misses among head checks beside those among written ones.session-grade.md§7 doesn't coach the blanks of a sheet with that override and, after each miss's account, shows the one written line that would have caught it;session-open.mdstep 6 says the override holds for one sheet, since R41 is core;review.md§5 proposes the written lines again after two sheets where a head-checked miss had one that would have caught it. Method §4,measure.md§5 and the contract say so. - A blank Least-sure line is recorded as blank. The grades file took only a per-question
least_sure, so a line left empty looked the same as "sure of everything", and every wrong answer on that sheet counted as an unnamed one, which inflated the statistic that would test the written-check rule; the line was left blank on many sheets in the method's development. A new grades field,least_sure_line(named,noneorblank), is copied onto every attempt row:ind grade recordtakesnamedwhen a question is named, and otherwise refuses a sheet with the line until it saysnoneorblank, ornamedwhen every question named was left out of the file (withdrawn as unclear, say). The unnamed-wrong share leaves out sheets whose line was blank,ind statsadds "least-sure line filled on k of n sheets", and the result card says "Least-sure line: left blank" instead of the unnamed count.session-grade.md§3 says never to read a blank as "sure of everything", ask for it afterwards or count the blanks. - A second-language misread is probed before it becomes a wrong idea. A learner who misreads a known phrase (persona A reading "twice as many" as "two more") doesn't know they misread, so they can't answer "a word stopped me", and the miss was filed as a concept error with a fix sheet on a procedure they own. The plain-worded twin that separates word from concept ran only on diagnostics and early probes.
session-grade.md§5 now asks, whenlearner.l1differs fromlearner.instruction_langand a sentence answer is wrong on an operation the learner did right in symbols, for one probe line without the sentence frame (a formula, a number or their first language), never showing the key. Right: the miss is V, and its belief line names the reading (or, for a sense never defined for them, the undefined-word rule applies). Wrong: the content modes as usual. The taxonomy check order andmeasure.md§7 point to it. - Official tests marked online have a safe route into the session. Online practice tools show the correct answers and explanations on the same pages as the results, and nothing said how to report such a test, so the easy way (pasting the review pages) put the answers into the chat before the miss-review, where some chat apps offer pasted text back as you type.
measure.md§5 now takes the platform's per-question right/wrong list as the evidence (typed, or a screenshot with the answer and explanation columns cropped out, never the review pages), and verdicts come from it. For the miss-review (§12), the learner switches off "show correct answer and explanation" and writes the why-line and the "now" answer on paper. Answers pasted anyway are never quoted or used: Claude asks which explanations were read rather than inferring it, and those questions skip the "now" answer.session-grade.md§10 logs acontaminationdefect only if Claude asked for the pages, andsheets.md§10 has the evidence row. - A tracked pre-push hook runs the privacy check. The repository's privacy rule was kept only when the maintainer remembered to run
dev/privacy_grep.pybefore a push.dev/hooks/pre-pushnow runs it and stops the push on any exit but 0 (a hit, or a missing denylist);git config core.hooksPath dev/hooksturns it on once per clone, as the README's contributing section says. CI still runs no privacy step: the denylist and the accepted git identity are private by design, so a check there would pass nothing real or fail every commit..gitattributeskeeps the hook's line endings LF on a Windows clone, where a CRLF checkout madeshfail every push, and the hook runspython3to test it before using it, so the Microsoft Store stub that stands in for a missingpython3falls through topy -3. Tests check both. - A test checks every CLI usage the docs give. The docs are how Claude learns the CLI, and nothing tied their 400 or so
ind …usages to it, so a renamed flag or a typo in a reference would surface as a usage error mid-session.tests/test_docs.pyreads every backtickedind …span and fencedindline inSKILL.md, the references andbuilder.md: the command must exist, every flag, optional[--…]groups included, must be one of its options spelt in full (argparse would take--budgetfor--budget-min), a literal value must be one of the option's choices, and a complete usage must parse. All pass today. - Every test run checks the version in all four places. CI compared only
plugin.jsonwithSKILL.md, so a release could ship with a stalelib.VERSION(which--versionand every.icsPRODID print) or no CHANGELOG entry, the drift 0.1.1 had to fix.tests/test_versions.pynow requiresplugin.json, theSKILL.mdmetadata,lib.VERSIONand the newest CHANGELOG release to agree, and--versionto print it, so the check runs locally too; the CI step it replaces is gone. - A theory sheet says what the object is and why. The template asked for a worked case and a rule, never what the object was or why the rule followed from it, and in the method's development procedures taught that way faded before their 2-day recheck, while the learner kept asking "what is it?". A new section kind,
meaning(printed as "What it is and why"), goes between the worked case and the rule: at most 5 lines, one everyday anchor or a picture in words, then why the rule follows; a pure convention says in one line that it is learned as given. New lint rule L13 fails a theory sheet on a procedural, conceptual or code topic without one, and W6 warns on other topics, and on a box over about 5 lines (80 words) or after the rule. When the learner asks "what is it?" or "why?" about the object, Claude points to the box; with none, it logs its owncontent_errorand puts the answer on the next sheet rather than teaching in chat above the drills. New method rule R52 [lit-mixed] cites Rittle-Johnson, Schneider & Star (2015);builder.mdrule 8,session-teach.md§2 and §3,sheets.md§1, §5 and §7, the schema and the contract say so. Untitled sections now get their kind's heading in Markdown too. The sample workspace was rebuilt. - A new topic's theory sheet opens with a guess. The first minutes of a theory sheet were passive reading, and nothing turned the diagnostic's miss on a topic into a question the reading answers, though asking before study helps learning of that material once the answers follow, even when most guesses are wrong (Pan & Carpenter, 2023). A new section kind,
prequestion, printed as "Before you read: have a guess" with a fixed line ("Write a guess on paper before reading on. Wrong guesses are expected and help. The answer is further down this sheet."), holds one or two questions that the worked case then answers. The builder adds one on a new topic at mastery 0–1, on the structure of the topic's diagnostic item (NOTES may name it: "prequestion from item "). It has no ask and no key, and nobody marks it. Lint L11 fails one on any sheet but a theory, or after the first worked section. New method rule R53 [lit-mixed], with amethod.md§6 note on its limits;builder.md,session-teach.md§2,sheets.md§1, §6 and §7, the schema and the contract say so. The sample theory sheet opens with one. - A step the learner is to write gets a printed place. The only printed structure per answer was the box and its check line, so a builder had to make every step its own graded question, which inflated the counts, or ask for the working in a sentence, which in the method's development got none: a printed blank inside the item got filled, a printed request got read past, and trace tables had nowhere to write a row. The sample's own drills asked for "steps 2 and 3" with nowhere to write step 2. An item may now carry
scaffold: working lines (["u =", "u′ =", "u′/u ="]) or an empty table ({"columns": [...], "rows": N}), drawn in all three formats between its text and its first box; it is never graded, keyed or counted, and lint L4 and L8 read its labels. New lint rule L14 refuses a scaffold on cold, mixed and measuring sheets, and on the last two items of a drills block, so the block shows whether the working now comes unprompted. On mastery 0–1 drills every item after the fully worked one but the last two carries the procedure's scaffold (on a repair's drill block, which has no worked item, every item but the last two), and the faded item 2 prints its remaining steps as lines. Marking never grades the lines, and points to a wrong or empty one as the step. New method rule R54 ([one-learner]);builder.md,session-teach.md§1 and §2,session-grade.md§7,sheets.md§2, §5 and §7, the schema and the contract say so. The sample drills print working lines on items 2–4, and their estimate rose to 10 minutes. The sample workspace was rebuilt. - A new topic that didn't land is taught again before its recheck. Nothing said what to do when a new topic's drills came back mostly wrong: "I don't know" opened no mistake, so the topic went straight to a cold 2-day recheck the learner had been told not to prepare for, a certain fail that measured nothing; "no idea" instead opened a "wrong idea" per question, each needing a fix sheet built on a wrong idea the learner never had.
session-teach.md§4 now has the path: drills on a topic taught this session or the last under half right [practice], "I don't know" and blanks counted as misses, get nokindfor those misses (abeliefonly for a wrong idea visible in the work, one per idea); the learner hears "This one hasn't landed yet … Next session we start it again from a different worked case, and the 2-day recheck moves to after that"; a to-do books the re-teach; the close doesn't place its recheck, and the next open leaves the topic off the recheck until the re-teach, whosesession taughtmoves the window, even one that has passed. While the to-do is open and the topic has had no teaching since,plan checkraises no unplaced-recheck FAIL or late WARN for it, and the brief's LATE RECHECK, RECHECK NOW anddueleave it out (due --listnames it "off until re-taught"), so a missed session never turns it into a late probe of a topic that never landed.grade recordnames such a topic and prints the to-do command, but only for a topic taught within the window's far end (72 h by default) before the drills, and never on a repair's drill block (a wrong idea awaiting repair on the topic), which gets a new repair page instead. The "please don't review it" request moved from "closed" to after the drills are marked, and only when they landed.session-grade.md§5 gives one mistake per wrong idea on a sheet (kindon its first question, the same account on the others, which a test checks), the "Show me this one" route gets the same carve-out, and new method rule R55 cites Rosenshine (2012).session-open.md§5,close.md§6,taxonomies.md§1 and the contract say so. - A question asked outside a session has a route. In a workspace, "what does 'median' mean again?" went to the default
session, which opens a full locked session, or was answered in chat under no rule that logged it, so a 2-day recheck the next morning could count a topic explained the evening before. A newaskcommand, with its steps inreferences/ask.md, outside what a session loads (and routing inSKILL.md: a content question with no session running goes toask), reads the brief, the sheets out (issued, or taken and not yet marked, assession statuscounts them) and the topic idssession exposeneeds, keeps a question on a sheet not yet marked sealed, asks before telling, and logs at once any explanation withsession expose --kind chatand any pointer back to a sheet, a warm look the next recheck can't see, with--kind review; a WARN about a booked recheck leads toplan checkand a move, told to the learner without naming its topics. Law 2 now says every explanation in chat is logged that way, in a session or not, and that a recheck's topics are never named before it is marked. A question on a sheet out and not yet marked is never discussed outside a session; in one, a practice sheet still gets the hint ladder and the gate's repair (session-teach.md§3–4), and a measuring sheet nothing before it is filed. The contract and the README's command table say the same. - A run of blocked days is one
time.blockedentry. A date entry covered a single day, so "sick till Monday" or a week away took oneind setcall per day. An entry may now carryto_date:{"date":"2026-10-14","to_date":"2026-10-18","what":"sick"}blocks both days and every day between, inplan check, its fixes and thesession extendcap.to_datewithoutdate, or before it, is refused. Theplan.md§11 sick recipe,ind schema indelibleand the contract use it. - Lint warns on a recheck word the learner has only read. L4 accepts
defined_on:<sheet>once that sheet defines the word, but a word printed once in a theory sheet's words box and never used in a question is not yet the learner's, and meeting it cold on a recheck is Claude's mistake, not theirs. New warning W8: on a cold, mixed or measuring sheet, a word resolveddefined_on:that appears in no question the learner answered (an item's text or label on a practice sheet sat or graded). The builder uses the word in a drills question first, or plain words on the recheck (builder.mdrule 5);sheets.md§7 and the contract list it. A warning only: a recheck is never blocked. - An official test marked online can be registered.
measure.md§5 took the platform's right/wrong list as the evidence for an online practice test, but the platform holds the answers, andsheet newneeded a key entry for every question of a mock or checkpoint, so there was no way to put such a test on the record at all.ind sheet new <s> <id> --spec <file> --marked-onlinenow registers adiagnostic,mockorcheckpointwhose items are allofficial:pointers with an empty key, and marks its sheet row"marked_online": true(named in the contract'ssheets.jsonlfields and inind schema sheet);key openprints it and says the verdicts come from the list filed as evidence, andgrade recordworks as usual. The builder is briefed with "marked online" in NOTES;measure.md§5,sheets.md§10,builder.mdand the contract say so. The hand-over asks for the stop time and the Least-sure line on paper before the test is submitted, since the platform shows the results then, so the line is recorded as the learner filled it, not as blank.
Fixed
- Lint times a words recheck. A
wordsrecheck carriescold:<topic>items and closes the topic's booking, but L7 skipped the type, so one sat inside the 24 hours after the words card, or on a topic with an unfixed mistake, passed lint and used up the recheck;profiles.mdasked Claude to check the window by hand. L7 andsheet issuenow judge a words recheck as a cold sheet: its first serve by the 44–72 h window, every serve by the 24-hour rule and untreated mistakes, and its re-served mistakes as on a cold sheet. Later rungs, placed by hand, have no window, and one word on a topic is fine, since words sheets don't feed levels. - The 7-day retention figure counts the rechecks that come a week or more on. It took cold questions 144–216 h after the last warm exposure, but a level-4 recheck comes 7 days after a pass that was itself 44 h or more after teaching, and upkeep later still, so the band almost never filled.
ind statsand the weekly review now count every cold question 144 h or more after the last warm exposure ("retention 7 d or more", "a week or more"; "a week or more after last seen" in the review file). The contract's metric table says so. - A recheck topic with one counted question neither passes, fails nor uses up its recheck. A cold pass needs 2 counted questions on the topic, but one wrong answer (or one wrong beside a right answer on the Least-sure line) dropped a mastery-4 topic to 2, a topic whose only cold sitting had one question read "no evidence yet", and grading closed the booking and set
last_coldfor a topic served once, or left with one after a right answer on the Least-sure line, so the topic dropped off the due list and could not reach mastery 3 until it was taught again. The level rules now need 2 counted questions to drop a topic too, andlevel_basisnames such a sitting ("cold 1/1 on … did not count") while the topic is below 3. Such a sitting now closes nothing, not even a recheck booked for it with other topics that were served (nor does a topic seen too recently), and leaveslast_coldalone, andgrade recordsays why and when the window closes;session-grade.mdsays to serve it again inside the window.session-open.md§5 gives every recheck topic 2 questions before any mistake goes on, and drops the topic whose window stays open longest when they don't all fit. Awordsrecheck, which levels don't read, still closes its booking.measure.md§8 and the contract say so. - DUE says how many mistakes can't be served yet. It counted every mistake due by date next to 2-day rechecks "ready now", so a mistake repaired last night, held by the 24-hour rule until this evening, made the morning recheck look bigger than it could be. The count now adds "(1 not ready yet)" (technical: "not servable now"), and
due --jsongiveserrors_not_now. - A mistake that came back needs two fresh passes to retire early. On a short runway a mistake retires after passes on 2 different days, but the count took in passes from before its latest miss, so a wrong idea that failed its re-serve was retired after one right answer the next day. Only passes after the last miss count now, for every kind and for a retired item that missed its sentinel serve.
plan.md§6 and the contract say so. - The drills failure gate can fire for a novice. It read "If your check failed on 2 of items 1–3, or you left 2 blank", but the same sheet says "I don't know" is always accepted, so two "I don't know"s, or one failed check and one blank, never met it. At mastery 0–1, items 1 and 2 are worked and faded, so it watched one independent item. The gate now reads "If 2 of items 1–3 have a failed check, an “I don't know” or an empty box: stop and send a photo of 1–3." A drills block may set
gate_after, which lint L6 checks (3 items up to it, 2 after): at mastery 0–1 a block of 6 or more puts the gate after item 4, over items 2–4.builder.md,session-teach.mdandsheets.mdsay so, the gate-photo steps file and reveal by the three items the gate names,method.mdR36 describes the gate as it now works, and the sample workspace was rebuilt with it. - A theory sheet allows time to read it. The builder's estimate counted only the pencil questions, so a first theory sheet of over 400 words in a second language could say "Stop after 4 minutes", and lint passed it.
builder.mdnow adds reading time on theory, example and repair sheets (their words over 120 a minute, 90 in a second language; about 3 minutes a page on external sheets). A new lint warning, W5, flags an estimate that leaves no time to read even at 150 words a minute. Those sheets and external pages now say "Allow about N minutes, and read it all even if it takes longer" in place of the hard stop; drills and measuring sheets keep "Stop after N minutes". The sample workspace was rebuilt. - A theory sheet no longer says to close it before its pencils are marked. It ended "Close this sheet now. The drills come separately.", right under the pencil questions, while marking points back to a step of the worked case, and nothing asked the learner to send the pencils or to say "closed". It now ends "Send me your pencil answers and keep this sheet open until I've marked them. Then put it away and tell me “closed”. The drills come on their own sheet.", and its rules box says to do the pencil questions at the end with the sheet open; an example sheet keeps its own lines.
session-teach.md§2 adds the theory hand-over, lists the pencils where the sheet prints them (last, on a fresh case), and on textbook pages points to the book's page and example at marking. The sample workspace was rebuilt. - The word rule on every sheet no longer contradicts marking. It said "A word used and not defined on this sheet is my error: mark it V". On drills and rechecks, whose words are defined on the theory sheet, a learner was told a forgotten word was Claude's mistake, then told the opposite at marking, and "V" was never explained. The rules box now says "If a word here was never explained to you, on this sheet or an earlier one, write it beside that answer: that's my mistake, not yours."
session-grade.md§3 looks a written word up before the account: a word never defined is Claude's mistake, and one defined earlier is the account "a word stopped me", not asked for again. The sample workspace was rebuilt. - A course with a formula sheet lets the learner use it.
format.reference_sheetwas collected at onboarding and then dropped: every drill, recheck and mock said "Closed book" with no formula sheet, so a statistics novice was tested on recalling formulas the exam prints. With the flag set,sheet buildnow adds to the rules box of every closed-book sheet "You may use a clean copy of the exam's formula sheet, with nothing written on it", and the tools line names it; theory, external, example and repair sheets are unchanged.builder.mdrule 13 never asks the learner to recall a formula that sheet prints, only to choose and use it;profiles.mdandmeasure.mdsay the same. - A recheck topic gets at least two questions. A cold pass counts toward a level only with 2 counted questions on the topic, but lint passed a recheck with one, and
builder.md's own example served a topic once: the recheck was used up and the topic could not move. Lint L7 now fails acoldsheet where acold:topic has fewer than 2 questions (itserror:andsentinel:items count, as at marking).builder.mdasks for 2 per recheck topic and cuts whole topics when over budget, and its example serves each topic twice;measure.md§8 and the contract's level table state the minimum. - Symbols and code built-ins are words too. L4 never read code spans or fenced code, so
len,.appendor a slice reached drills untaught even when the subject listed them, and nothing said that a symbol (x̄, σ, ln, λ) needs defining like a word. L4 now reads code for the subject's ownlexiconandsense_listentries (never the seed list, whose words are often identifiers).builder.mdrule 5 andsession-teach.mdput every symbol, built-in and piece of syntax in the words box with how to say it aloud;profiles.md§9 seeds a maths, statistics or code subject'slexiconwith the syllabus's symbols and built-ins; and an undefined symbol is Claude'sundefined_termdefect, as a word is (session-grade.md,taxonomies.md). - A theory sheet can't call the word it teaches "everyday". L4 accepted
everydayon any sheet unless the word was in the lexicon ortheory.words, and its own theory FAIL message suggested it, so a theory sheet could use its central word a dozen times and never define it. On theory, example and repair sheets,everydaynow fails for a word the sheet teaches: one in the title, a section title, the topic's name or a pencil question, or used 3 times or more. One plain use in a box still passes. The theory message no longer points ateveryday, and on a sheet with notheory.words(drills, a recheck) theeverydayFAIL names the resolutions that sheet can use, not "define it in theory.words". - The builder is told every term resolution lint accepts, and the glossary gets filled.
builder.mdrule 5 andsheets.mdlisted three of the five resolutions, then said "use plain words", so a vocabulary probe had no documented way to test a lexicon word (measured_here), andanswer_in_passagewas never mentioned, so an exam's "NO MORE THAN TWO WORDS" question had to be paraphrased away. Both now list all five with their limits and documentanswer_in_passage. Nothing ever ranind glossary add, so theglossaryresolution always failed:session-grade.md§8 now adds each "use" word from a triage sheet and each word right on a words recheck. L4's "used without a resolution" line names the five resolutions. defined_on:<sheet>has to name a sheet that defines the word, and check hints are read. L4 accepteddefined_on:with any well-formed id, even one never built, so a learner could meet a word nobody defined and have the V miss logged as theirs. Lint now looks the sheet up (any sheet of the subject that is not void) and passes only if it lists the word intheory.wordsor resolves itdefined_here;sheet issueholds the sheet back until that sheet is issued. L4 also scans check hints, which are printed under every question and often quote the key word.builder.md,close.md(build a theory before its drills) andsheets.mdsay so.- Sheets on one block are sized together. L5 and
sheet issuemeasured each sheet alone against 0.8 × its block, so two 35-minute sheets were issued into one 60-minute block. Now the est_min of every other sheet already issued on the block (sat and graded ones too) is taken off, and the basis names them ("less 35 min on ielts-x-01"). When the open session runs on that block, the work minutessession openprinted (for its planned minutes with the extension) count instead, longer or shorter than the block, since a slot split into a recheck block and a session block is one session, and a session opened short has less time: a 45-minute sheet is refused in a 60-minute session's 36.7 work minutes. A recheck linted against the session's remaining minutes could be refused at issue against its own 15-minute block:session-open.md,plan.mdandsheets.mdnow issue every sheet of a session, the recheck included, against the session block (grading still finds the recheck block by topic and window), andbuilder.mdno longer says BLOCK is "for the budget only". - A measuring sheet keeps to its own minutes. L5 passed every diagnostic, mock and checkpoint ("sized by the exam clock") and ignored even an explicit
--budget-min, andsheet issueskipped them too, so a 90-minute diagnostic could be issued into a 60-minute block, turning unreached questions into apparent gaps. Both now hold a measurement to--budget-min, else its block's minutes less the 10 kept for recording (measure.md§4), else, for a mock or checkpoint, the exam's own minutes. Over budget, the message says to split a part into sittings or book a longer block, never to cut questions; the builder returns FAILED rather than cutting. The pace floor applies to measurements too, to the questions Claude wrote: an official paper's questions keep the exam's clock, so an official listening test marked 30 minutes is never told to stop after 51. - Lint no longer takes the sheet's estimate on trust. L5 compared
est_minwith the budget, butest_minis written by the builder whose sizing the rule exists to check, so a sheet of 6 reading questions marked "1 min" passed a 2-minute budget. L5 now works out the builder's own formula again (each question'space_s[layer], over 60, plus 1 minute, from the subject's pace) and fails a lower estimate, even when no budget is known. Triage sheets are exempt.builder.mdsays so, and adds that a question needing several written results gives each its own box, so the estimate counts the work, not the items. - "Sick till Monday" no longer dead-ends on a recheck. When sick days covered a recheck's whole window,
plan checksaid to move it to another day,plan moverefused (outside its window), and the recipe said not to cancel. For a recheck,plan checknow says "move it inside its window (…)", or, when no time is left in it, to cancel it, add a to-do and apply the late-recheck rule at the first session back.plan.md§11 gives those steps, a to-do that names the block but never its topic, the preview line, and the re-baseline when slack drops below zero. - A 2-day recheck that lapses is no longer lost. A recheck skipped in a session (an allowed override), or an on-demand learner's recheck whose window passed before the next session, vanished: the brief,
due --listand the close said nothing, so the late-recheck rule never ran and the topic stalled below mastery 3. The brief now flags it without naming its topic ("a 2-day recheck's window has passed"; LATE RECHECK below the line names it),due --listlists it as tier 0, andplan checkcalls the late recheck[measured], asplan.mddoes, not[practice]. A new close check, C9, fails when a recheck booked in the session's time was not sat while its window is still open, and prints theplan movethat fixes it: the first quarter hour after now, outside the sleep window and the half hour before bedtime, thatplan moveaccepts, or INFO when no time is left in the window. The close no longer marks a recheck block done, since only grading closes one, andclose.mdsays so. A topic served cold inside its window is not flagged late, and grading such a sitting closes the booking it served.session-open.mdmoves a skipped recheck inside its window, andplan.md§7 logs the feedback given at marking as a review exposure (a measuring sheet logs none). - Drills on a later day move the 2-day recheck window. The window counts from the last warm exposure (the contract, lint L7,
ind dueand the level rules all say so), but the windowsession taughtstored kept counting from the teach. With drills a day later,plan checkasked for a time thatplan placeandplan moverefused.session exposeanderror repairnow move the open recheck of a topic still waiting for its first recheck to 44–72 h after the exposure, print the new window, and name a placed recheck left outside it.grade recorddoes the same from the latest warm exposure, since a graded practice sheet logs a drill exposure at its sitting, and warns about a placed recheck left outside the window or within 24 hours of the sitting.session-teach.mdand the contract say so, in place of "24 hours clear of the drills", and after a repair. The sample workspace was rebuilt. - A recheck sent the next morning keeps its real sitting time. With the Start and Stop lines blank, or the photo filed on a later day without
--date, the filing day and the marking time stood in for the sitting, so a recheck sat inside its window could be recorded outside it, lose its level 3 and still use up the first recheck. Without--date,scan ingestandsheet satnow refuse a recheck, or a sheet with mistakes re-served, that was issued on an earlier day (any other sheet keeps today, with a note).grade recordno longer guesses the time of such a sheet unless it was issued today and is marked within max(3 h, 3 × its minutes); it asks fordateandstart. The warning fromsheet satuses the same fallback as grading. A sitting time more than 5 minutes before the sheet was issued is refused (a wrong date, or "9:00" for 21:00), and a guessed time is never before the issue. - A recheck near the end of its window says so. RECHECK NOW rounded 71.5 h up to "72 h" and gave no closing time, so a recheck built at the open and started half an hour later could fall outside its window: the pass then counted for nothing, silently, and still used up the first recheck. RECHECK NOW and
due --listnow cut the hours (71.5 h reads 71 h) and give the time the window closes, withCLOSINGinside 30 minutes (closes_atindue --json).sheet issueprints "Start by " for each topic of a recheck, andgrade recordprints "Not counted toward level 3 … treat it as a late recheck" when a first recheck was started outside its window.session-open.mdbuilds a closing recheck first. - Practice no longer uses up a 2-day recheck. A
cold:<topic>item on a mixed or other practice sheet closed the topic's recheck at marking and set it as served cold, so the real recheck was never offered again. Now only a measuring sheet closes a booking (acoldsheet, awordsrecheck, or the late-recheckprobeofplan.md§7); on practice, a note says the recheck stays open. - A mistake re-served on a warm topic makes no ladder move on any sheet. On a mixed or measuring sheet, a mistake whose topic was seen in the 24 hours before the sitting now gets the "not counted" note, as on a cold sheet. The rows still count as practice or measurement toward levels.
- Lint L7 checks mixed sheets too. Their
error:andsentinel:items must be due (forerror:) and not seen in the last 24 hours, and acold:item on a mixed sheet fails: a recheck in its window is acoldsheet. The builder now links each sheet to the block it will be sat in, so a mixed sheet built at the close is judged at the next block's start, not at build time, andsheet issuechecks a mixed sheet's L7 again, as it does a recheck's. A sheet built for a block and not issued yet no longer takes the block's minutes at issue, so linking the practice sheets early never makes the recheck the sheet refused. - With technical vocabulary, the brief no longer names the recheck topics above the line. DUE gave the topic ids of the 2-day rechecks ready now, and NOW/NEXT the topics of a recheck to book, in the part that may be read aloud. Both now give a count or a time only, as in plain words; the topics stay in RECHECK NOW below the line.
- A failure-gate photo opens only its own questions. The photo of items 1–3 sent at a drills block's failure gate was filed like the finished sheet: it marked the sheet taken, and
key openthen printed the whole key into the conversation, answers to the questions still being worked included.scan ingest --asks 1a,2a,3anow files a gate photo: the sheet stays issued,key openprints only those questions (and says so) until the finished sheet is filed, andgrade recordwaits for the whole sheet, as does close check C1. Each line of.indelible/keys/opened.jsonlnow names the questions it printed.session-teach.md,session-grade.md,sheets.md§9 and Law 1, with its copy in the contract, say so (a test checks the three), and the sample workspace was rebuilt. - A second typed file no longer replaces the first.
scan ingest --typedalways wroteanswers/<id>.txt, so the finished sheet's file overwrote the gate photo's, and both evidence entries pointed at the same file. Each typed filing now gets its own file (answers/<id>-2.txt, and so on). - A second chat no longer walks into a running session blind. A learner who opened a new conversation mid-session got only "Carry on": the new chat didn't know which sheets were out, and could issue a second recheck on a topic already out.
session statusnow adds a "sheets out" line (issued or taken, not graded yet: id, type and issue time, never a recheck's topics). A read-then-close sheet already taken is left out, since it needs no grading.sheet issuerefuses a cold or mixed sheet that serves a recheck or mistake another such sheet, not graded yet, already serves, and names that sheet. Insession-open.mdstep 4, Claude reads the status and the open to-dos before building anything, and asks the learner to send photos to this chat from now on. cal icswrites onlyplan/ics/<name>.ics. Before, any path inside the workspace was accepted, so a wrong path could overwritedata/attempts.jsonlor a sealed key with no backup. Any other place, or a name without.ics, is now refused (exit 2) and nothing is written, as the README says.- A subject the learner names comes before the current folder.
SKILL.mdrouting took the subject from the current folder first, so a learner in the IELTS folder who said "let's do some spanish" got an IELTS session, against the CLI (ws.resolve_subject) and the contract, which put a named subject first. The order is now: a subject the message names (its id, its title or an obvious short name), then the current folder when it is inside a subject folder, then the block on now or next, then numbered options. - Numbered procedures carry the steps other files require.
session-grade.md§1 said "Never reorder these steps" but left out the looked-since question that §10 andsession-open.mdask before a 2-day recheck is marked; it is now step 2, before the key is opened, so the learner can't name questions after seeing the verdicts.SKILL.mdsetup step 5 loadsclose.mdbefore a late close. Law 6 said to read onlyind briefat the open whilesession-open.md§2 also readsind due --listandind sheet show: it now allows CLI output only (the brief, the listings that section names and the output of the commands the open runs, such asind doctorat step 8), never the raw data or views, and the contract, the workspaceCLAUDE.mdtemplate and the sample say the same. Re-entry no longer falls back to readingsubject.jsonfor the learner's why. Two cross-references pointed at the wrong setup step (close.md§10,migrate.md§2). - A question with no check line is never coached for one.
session-grade.md§7 said every answer on a drill or measuring sheet has a written check, but probes and words sheets (both measured) print none, so a late recheck run as a probe could be recordedmissingand the learner coached for a check the sheet never asked for, lowering check coverage. §7 now lists the sheet types with check lines, asprofiles.mdand lint L2 do, and says a question printed without one ischeck: n/aand never coached.grade recordrefuses any other value for such a question. - The references give the real reason a recheck is built at the open.
session-open.md,sheets.mdandsession-teach.mdsaid lint L7 checks the 2-day timing at the moment it runs, so a recheck couldn't be built ahead, while the CLI judges a sheet linked to a block at that block's start. The real reasons are thatind due --listcan only work out what is due now, and that a rendered recheck waiting in the learner's sheets folder could be looked at before it is sat. The references now say so, the docstrings ofcmd_sheet.pyandlint.pycall the block-start judging a safety net for a sheet issued before its block, and the contract listssheet new --blockandsheet lint --block --at --json. Nothing about when rechecks are built changes. - Session timers are set, kept and removed. The timing rules assumed Claude sees time pass, but it acts only when a message or a timer arrives, and
session-open.mdstep 8 named CronCreate with no parameters (it repeats by default), no prompt and no clean-up: an early "gotta go" or a distress stop left the warning to fire anyway, andclose.mdcounted on a planned-end timer and "about a minute" that nothing produced. Step 8 now loads CronCreate, CronDelete and CronList, sets one-shot jobs (recurring: false, pinned to the date, in local time) at T−10, the close start, a "no-answer" moment 2 minutes later (a minute later again on :00 or :30, where a one-shot job can fire 90 seconds early; one job with the planned end in sessions of 30 minutes or less), the planned end and each break, each with a prompt that checksind session statusfirst and stays silent when no session is open or the study frame is stopped.close.mdreplaces the no-answer job after an extension, closes on it only onceind session statusshows 2 minutes since the question, and deletes every job still pending at the close, at an abrupt exit and at a Law 13 stop, the one set after an extension included; parking a session deletes its jobs too. With no timer tool, the learner gets the warning and closing times at the first hand-over. The floundering timeout is checked againstind session statusat the learner's next message, andcalendar.mdpoints to these timers instead of an undefined "session timer". - A later
.icsexport sends only the new blocks.calendar.mdgave the export without--ops create, so every later sync re-sent every block already imported: the file no longer matched the preview ("Add 1" with 7 events in it), it added a recheck eventplan.md§1 keeps out of the calendar, and apps that duplicate re-imported events filled the calendar with copies. The export now uses--ops create, incalendar.mdand at setup (teach.md§7 step 9, which a second subject's setup runs again after the first subject's blocks are imported), the ack covers only the blocks in the file, and the delete-and-reimport route makes a full file without it. The contract lists--ops. - Code evidence keeps its project layout.
profiles.md,sheets.mdandsession-teach.mdfiled code as loose files or one text file, which dropped the file names and folders (a Rustpub mod parse;no longer resolved), so the record could not be rebuilt to settle a disputed mark, and marking ran on the live project.scan ingest --dir, which copies a project with its layout toanswers/<id>/, was built and tested but never mentioned. The references now file code in one call,--dir <project folder> --typed <output file>, mark from a temp copy of that snapshot, treat a build failure over a file the snapshot lacks as a filing gap (file it too, or mark that question from what was filed), and list the snapshot in the dispute checklist. The ingest note now names what--dirleft out (hidden files and folders, build output, links, files over 1 MB), where it gave only a count that left hidden files out. The contract lists--dir. - Times are written as local time, and a copied offset is caught. The CLI already reads a time with no offset as the workspace's local time, clock changes included, but every concrete example Claude saw (
teach.md,migrate.md, the--helptext and error messages) carried a fixed+01:00, and the--duehint for a date alone used today's offset for a date after the clock change. Lisbon changes on 25 Oct and New York on 1 Nov, inside A's and C's runways, so a copied offset put a Saturday block or a to-do an hour early with no warning.plan.md§1 now says to write--start,--dueand window times with no offset, the examples and hints carry none, andplan add|place|moveandledger add owedprint a one-line note on stderr when an explicit offset is not the zone's at that instant ("+01:00 is not the offset of Europe/Lisbon on Tue 27 Oct (+00:00), so this is 06:00 local time"). The time is still accepted, so a UTC time from a connector works. The contract says so. - A second subject settles where its hours come from, and
plan check's fixes respect priority and free time. Adding a subject never raised the weekly target or ceiling, which were derived from the first subject alone, so the week went over the ceiling. The ceiling fix then named the largest unprotected block of any subject, which in the sample was the priority-1 subject's Saturday practice the learner had asked for, while claiming to follow the drop order; the outside-window fix suggested the window's start even when another block held it. The second subject's Q6 card now asks "raise your total, or share the hours you have?", with the writes for each answer (teach.md§9). The ceiling fix prints the drop order (adding the lowest-priority subject's step only when the learner's order doesn't already end with it) and names a buffer block first, then the smallest unprotected block of the lowest-priority subject that covers the minutes over. The outside-window fix names the first free start in that day's windows, or says there is none. - A missed page no longer turns answered questions into blanks without anyone noticing. Only a photo pasted into chat was transcribed; photos and PDFs given by path were marked straight from the images, so a page read past became a row of
skips, which open no mistake and can send a topic to a probe. Blanks never appeared on the verdict list, and nothing can amend a verdict after the record. Photos and PDFs are now transcribed too, one line for every question number ([blank],[unreadable]or[not found]), and filed with the pages in one call (scan ingest … <paths> --transcript -). A transcript is now saved asscans/<date>-<id>-transcript.txt(then-transcript-2.txt), never under a page's name (it was-answers.txt, and filed with pages it took the next-pNname, so it read as one more page). A[not found]line on a finished sheet brings one neutral question before filing ("left blank, or on a page I didn't get?"). A failure-gate photo's transcript has lines only for the questions it shows, so the learner is never asked mid-drill about questions still being worked;session-teach.mdandsheets.mdfile it with--transcript -as well. The verdict list gets aBlank:line, andgrade recordprints a note naming any question of the sheet with no entry in the grades file. - Three rule citations now test what their rules claim. R1 ("plan from measurement, not self-report") cited only Simonsmeier et al. (2022), a meta-analysis of prior knowledge and learning gains that never compares measurement with self-report; it now leads with Zell & Krizan (2014), whose metasynthesis of 22 meta-analyses found self-evaluations of ability correlate only moderately with performance. R39 (first-language glosses) cited Abedi & Lord (2001), which tested plainer wording of test items, not glosses; it now cites Yanagisawa, Webb & Uchihara (2020), and
method.md§6 says why it stays [lit-mixed]: that evidence measures words learned, not content understood in a second language. Abedi & Lord stays on R9. R12 was graded [lit-strong] as a whole on Renkl & Atkinson (2003), a study of worked examples and fading, though half of it was a concrete case before the definition, which that study never tested. R12 now keeps the worked-example half and its grade under the same ID, so a saved override still names it, and anR12 orderrow grades concrete first [lit-mixed] on Fyfe et al. (2014);method.md§6 gives Kaminski, Sloutsky & Heckler (2008), where generic examples transferred better, andsession-teach.mdcalls concrete first a default instead of "always". Lint L11 still checks it, but accepts the rule first when the subject'soverridesholds a locked override ofR12 order(or ofR12, saved before the split), so the override can take effect; a sheet with no worked case still fails (builder.mdrule 8 and its checker table, and the contract). The reference list also gains Rohrer & Taylor (2007) and Dunlosky et al. (2013), which the README cites. - The build contract matches the code again.
dev/CONTRACT.mdsays it wins when the code disagrees, but it had fallen behind. It had noglossary add|listorsubject list, left out over twenty flags (among theminit --timezone,topic add --scope,error add --ask,doctor --quickandstats --since/--until) and the obligation form ofplan add, and missed exemptions the code keeps: repair sheets in L2, a single-topic recheck in L3, option labels andanswer_in_passagein L8, and the read-then-close sheets in C2. C6 gave half its promise words, and the brief's part for Claude showed 5 of its 16 sections. All now match the code. A newtests/test_contract.pychecks that every command and option the CLI takes has a synopsis in §7, that no synopsis names an option its command lacks, that the checker's rule ids match, and that C6 gives the code's regex. - The sample rebuild test compares contents, not only file names.
examples/README.mdpromised that a rebuild gives the same files, but the test compared their names, so the committed sample went on saying 0.1.1 in its.icsPRODID through three releases and nothing failed. It now compares every file's contents, reading CRLF as LF (a Windows checkout may convert them), and skips that part where the sample's time zone can't be loaded, since the brief then adds a TIME ZONE line. The.icsPRODID is also checked against the current version on every runner. - A recheck in a split slot is judged when it is sat. A slot split into a recheck block and a session block opens its session on the later session block, and every sheet, the recheck included, is issued against it, so lint and
sheet issuejudged the recheck at the session block's start, up to the recheck block's length after the real sitting. A window closing in those minutes refused a recheck that was valid when the learner sat it. Once the subject's session is open on the block, L7 now judges at now; before the session opens, a sheet built ahead is still judged at the block's start. - A record file that can't be read at all is named, not read as empty. Bad JSONL lines are kept aside in the quarantine and skipped, so a file whose every line was bad (a calendar file written over
data/attempts.jsonl, say) read as empty:statssaid "No graded questions yet" and the brief only counted unreadable lines. The brief now names such a file (plain words: "a whole record file can't be read", and its lines are no longer also counted as lines "kept aside"; for Claude, anUNREADABLE FILE:line: no grading, planning or stats until it is back). It names the copy the learner restores it from, never Claude, since only the CLI writes data files: its.bakwhen that has a readable line, elsegit checkout HEAD --when the workspace's last commit holds a readable copy, else "No copy here", never implying one exists. The line also says what to tell the learner and, with no copy, to ask about a backup of their own or going on without that part of the record;session-open.md§3 step 1 points to it and keeps "nothing is lost" for lines kept aside.statsnames each graded-question file with unreadable lines, says "No graded questions can be read." when none are left, and lists them underunreadablein--json, with the same restore step ascopy. - The 30 minutes before bed are no longer graded as well studied. R29 ("nothing in the sleep window or within 30 minutes of bedtime") was graded [lit-strong] on Mazza et al. (2016), a single study of 40 people that found sleep between two study sessions helps; it never tested a gap before bed. R29 now protects the sleep window alone, on two meta-analyses (Berres & Erdfelder, 2021: more is recalled after sleep than after as long awake; Newbury et al., 2021: a night without sleep impairs memory), and an
R29 bufferrow grades the 30 minutes before bedtime [design], a wind-down choice.method.md§2 lists the 30 minutes as a parameter, and §6 says what Mazza et al. found.plan checkis unchanged. - The break interval is graded as a guess. The breaks row ("one 10-minute break per 75 minutes") was graded [lit-mixed] on Biwer et al. (2023), a study of set breaks (6 minutes after every 24 of study, or 3 after every 12) against breaks taken at will, which never tested 75 minutes or a 10-minute break. The
breaksrow now says breaks come at set times, [lit-mixed] on that study, and abreaks intervalrow grades 10 minutes every 75 [one-learner].method.md§6 says what the study found, and themigrate.mdconflict table grades the break the same way. The default is unchanged. - Python scripts keep LF line endings on a Windows clone. Nothing kept them LF, so a clone with
core.autocrlf=truegave every script a CRLF#!/usr/bin/env python3line, and runningdev/privacy_grep.pyorscripts/indelible.pyby its own name (under WSL or Git Bash) failed with "env: 'python3\r': No such file or directory"..gitattributesnow checks*.pyout LF, as it does the pre-push hook; the contract's layout says so, and a test checks it.
[0.1.5] - 2026-09-28
The learner is never sent to find their own mistake.
Changed
- Marking points to the error, and the learner makes the fix. The account asked for each miss is now how the learner got their answer (
1) I did it this way: ___ 2) a word stopped me 3) I guessed 4) I can see my slip: ___), not "what happened?". On a topic met that day, "what happened?" asked them to find an error they couldn't see. Claude then names the question and the step, and asks once for the corrected line. A slip whose fix doesn't come is treated as a wrong idea, and a wrong idea gets its fix sheet. Law 11 and a new ban inSKILL.mdsay so, and "one of these is wrong: find it" is gone for good. Method rule R42 andmethod.md§5 give the reasons (Große & Renkl, 2007; Baars et al., 2014). - Careless needs more than a named rule. Naming the right method counts as a slip account only when the learner writes the corrected line unaided once the step is pointed to. The same operation must also be done right, unaided, elsewhere (never the worked items 1–2 of a first drill). A guess not forced by time becomes the wrong idea "no method yet for …", and its fix sheet gives a second worked case in place of the contrast.
- A failed check is a flag, not a hunt. The rules box on sheets with check lines now says: "If a check fails and you can't see why within a minute, keep your answer, put its number on the Least-sure line and go on: I'll show you where at marking."
- Checks on a new topic are ones the learner can run. A theory or fix sheet's worked case now ends with its check, worked as a step. So does item 1 of a first drill block, which covers textbook pages. The drills' check hints name that check. "Another way" and "the weakest step" wait until a topic reaches mastery 3, because a new learner has one method and no sense yet of where they are weak. Rechecks give only the form of a check, never a topic's own method, so the check doesn't label the question.
- "Which line is the first wrong one?" and "find the error" items appear only on topics the learner owns, or where finding errors is the exam's own question form. Below that, the wrong working sits beside the right one. A mock's miss-review asks why the learner chose their answer, and asks for the answer now only after Claude has pointed to the line or step.
- The sample workspace was rebuilt: the new rules-box line, and the
.icsPRODID now carries the current version.
Added
- Sheet checker rules. L10 (fails): a check hint that asks for a search ("find your mistake", "check your work for mistakes"), a re-solve ("redo", "double-check") or a confidence rating, or says only "check your answer". Subject words such as "the standard error", "error bars" or "the error message" pass. W3 (warns): on a topic below mastery 3, a check line with no hint, or one that needs a second method or the weakest step. W4 (warns): a theory or repair sheet whose worked case has no "Check:" step.
[0.1.4] - 2026-09-26
Changed
- Every environment variable is read by its literal name. The workspace lookup no longer keeps a reference to the whole environment, and the Windows browser search names
PROGRAMFILES,PROGRAMFILES(X86)andLOCALAPPDATAone by one.
[0.1.3] - 2026-09-26
Changes from the second directory validation.
Changed
- Child processes started by
examples/build_sample.pyand the tests get only a short, explicit list of environment variables (such asPATH, andSYSTEMROOTon Windows), never a copy of the whole environment. - A new plugin icon: a check mark written in ink, with an ink dot.
[0.1.2] - 2026-09-26
Changes from the first directory validation.
Changed
- The skill no longer declares
allowed-tools, so it pre-approves no commands. Claude Code asks before each script call until the learner adds the permission rule the README describes.
Added
- A plugin icon (
.claude-plugin/icon.svg).
[0.1.1] - 2026-09-25
Documentation and packaging for the plugin directory, and the fixes found while preparing it.
Added
- Docs for the directory submission. The README now covers what the plugin runs, writes and sends; example prompts; a synthetic sample workspace to try (
examples/sample-workspace/, with its fixed clock); where it works; troubleshooting; and support and security. PRIVACY.md, a privacy policy: what data exists, where it lives, what is sent where, how long it is kept, how to delete it, and the protective defaults that apply if a learner says they are under 18.SECURITY.md, how to report a vulnerability privately, what counts, supported versions and what to expect.- A guard for claude.ai chat and the mobile app in
SKILL.md. Before any script runs, the skill checks for a folder that lasts between conversations. Without one, it says that v0.1 can't keep the record there and offers a limited manual mode. Cowork with a shared folder is described as untested. displayNameinplugin.json, and a listing description that says v0.1 is built for Claude Code.- Tests (
tests/test_examples.py) for the sample workspace, for the browser's offline flags, and for the scripts staying inside the workspace.
Changed
- A headless browser that prints a PDF is kept off the network. Chrome, Chromium and Edge used to contact their makers' services in the background as soon as they started (updates, safe-browsing lists, DNS over HTTPS), even though the page printed is a local file. Both launches, the sheet printer and
doctor's test print, now switch that traffic off and send anything left to a proxy address that doesn't exist, with no host name resolving.doctor's fallback test print also runs without extensions. - The scripts write and delete only inside the workspace.
cal icsrefuses an output path outside the workspace.sheet newdeletes the builder's answers file only when it is inside the subject's.indelible/tmp/; anywhere else, it leaves the file and says so. allowed-toolsno longer listsBash(python *indelible.py *), which the skill never uses. The README andSECURITY.mdnow say plainly that the remaining patterns are text patterns, broader than the one script, and give a stricter rule.- The one-time notice before the first script call now says the script keeps the record as files on the learner's computer and sends nothing over the internet.
- The onboarding asks "Are you 18 or over?"; indelible is intended for adults, and the under-18 defaults are a safety net.
- The README describes the
CLAUDE.mdfiles the workspace holds, the commands Claude runs for programming subjects and formigrate, and what deleting one subject leaves behind in v0.1. - For programming subjects, Claude deletes its temporary test copy of the learner's code once the marks are recorded, and says once that the learner's build tool may download declared dependencies.
- The README labels every v0.2 item as planned and not yet available.
Fixed
- The skill and the command-line tool report version 0.1.1, matching
plugin.json(indelible.py --versionand the.icsPRODID said 0.1.0, and CI's version check failed). The sample workspace was rebuilt to match. - Tests skip the clock-change cases when there is no time zone database (Windows without
tzdata) instead of failing. CI installstzdataon one Windows job, so those cases still run there.
[0.1.0] - 2026-09-25
The first release. Built for Claude Code (terminal or desktop).
Added
- The skill.
SKILL.mdwith 14 laws, the absolute bans, a router, a command table (teach,session,close,status,diagnose/mock,plan/reschedule,review,syncandmigrate) and a manual mode for machines without Python. Each command has a reference file, exceptstatus, which changes nothing in the plan or records. - The method, as structure. Written backwards checks beside every answer and one closing Least sure of: ___ line per sheet, in place of per-answer confidence flags. A 2-day recheck (44–72 hours) booked for everything taught. Mistakes on a 1-day, 3-day, 1-week and 3-week ladder, with wrong ideas fixed before they are re-tested cold. Mastery levels 0–5 computed from the record: same-day practice reaches at most level 2, and anything higher needs a cold recheck or a measurement.
indelible.py, a command-line tool (Python 3.9+, standard library only) that is the only writer of the learner's data:- setup:
doctor,init,subject add,set,schema; - session start:
brief(at most 4,500 characters) anddue; - the session lock, its time budget and the close checklist:
session open|status|expose|taught|override|close; - sheets: specs, the sheet checker (
sheet lint), rendering to PDF, HTML or Markdown, issuing, and evidence filing (scan ingest); - answer keys sealed apart, opened only by
key openonce an attempt is filed; - grading records (
grade record), mistakes (error ...) and topics (topic ...); - plan blocks and their checks (
plan add|place|move|cancel|done|miss|list|week|check); - calendar operations (
plan diff,cal ack) and.icsexport (cal ics); - the ledger of to-dos, decisions and Claude's own mistakes,
note append,stats,review weekandcompact.
- setup:
- Sheet building by a builder subagent, so answers never enter the main conversation.
- A learner-owned workspace of plain JSON, JSONL and Markdown files, with generated views and a
.gitignorethat keeps keys, typed answers, photos and the inbox out of version control. - Tests (unittest) on macOS, Linux and Windows with Python 3.9 and 3.13, and CI.
- Evals: four synthetic learners (
evals/fixtures/persona-a…d.json) and a trigger set (evals/trigger.json). - Maintainer tooling:
dev/CONTRACT.md(the build contract) anddev/privacy_grep.py(a pre-push privacy check: a local denylist, the handle allowlist, commit and tag identities, and optional text fingerprints).
Deferred to 0.2
- Hooks declared in the skill's frontmatter.
- A claude.ai and mobile-app mode (a state file carried between chats).
- Two-way calendar sync. In 0.1, Claude proposes blocks,
plan checkvalidates them, and calendar changes go one way after a preview and a yes. - A scheduling solver.
- Scheduled tasks, plugin wrapper skills, and pin/unpin.
- LaTeX sheet templates and KaTeX, a dashboard, and migration tooling in the command-line tool.
forget: a command that previews exactly what it would delete and deletes only after a yes. In 0.1 the learner deletes a subject folder or the workspace themselves.