ResearchZosho

Absorb a conversation the person had with another assistant (ChatGPT, Claude, any chat) as a starting point for research. Give a path (a saved transcript or an export such as conversations.json), a url, or the pasted text. The transcript is shelved as it is; the person's questions join the open questions; the claims the assistant made come back as a list to check and are filed as nothing, since an assistant's say-so is not a source. verify=true files one research run that checks those claims against sources. An export with many conversations takes the first `limit` (default 25) and reports how many remain. Paths from the keeper only (the terminal, the pages).

What became of an access request: pending, denied (with the reason), or approved, in which case the token, once, when the claim secret is right.

Read material into the library: a file, a folder (a mounted network share is a folder), or a url. mode=keep (default) copies the extracted text onto the shelves; the file itself is never copied. mode=link reads the files where they are and keeps only a pointer and a hash: nothing is copied, and if the drive is not mounted the library says so when it reads. mode=survey counts a folder and says what keeping it would take (files, text size, free disk) without shelving anything. For a folder, survey first, then ask keep or link, unless the person has already said. Paths are only accepted from the keeper (the terminal, the pages); a url from any patron with write access. No model runs; to ask a question of what was added, library_research with sources=shelves and the collection.

Ask The Librarian — the long-term research library. Returns the answer PACKAGE: full relevant entries (findings, shelf articles, captured documents) with states, claim types, sources and disputes, open frontier threads, and holds_nothing=true with NO entries when the library has nothing. Reviewed background: it never overrides direct evidence you gather yourself.

A browser's bookmarks as a starting point: the exported bookmarks file, Chrome's Bookmarks JSON, or lines of urls. Each page is fetched onto the shelves as a collection (folder=… takes one folder); one that cannot be read becomes a source request. watch=true has the housekeeping re-read the pages and keep a new copy when one changes. Paths from the keeper only.

Discovery by combination: two areas of the library that no source read together, joined by specific terms both areas' claims share (Swanson's fish oil and Raynaud's). op=run computes the candidate pairs from an area (or the hottest ones) and files up to a few proposals as open questions of type bridge; dry=true only shows the pairs. op=list the open proposals; op=accept files the research run that tests one (its verified claims are the test); op=dismiss drops it; op=settings reads or sets an area's dials (sources, reach, strict, toward, away, since, per_night); op=measure reads the ledger: proposed, kept, dismissed, corroborated.

Recall notices: everything that happened to findings and investigations after a cursor — added, state changes (retired, disputed, accepted), edits, supersessions. Re-check what you cited with one call. Pass the previous next_cursor as since; 0 = from the beginning.

Check the person's own draft — a memo, a chapter, notes (text, markdown, Word, PDF) — as a starting point. One voice: its definite statements come back as claims to check, its questions join the open questions, and the citations it carries (urls, DOIs, arXiv ids) are fetched onto the shelves; one that cannot be read becomes a source request. Nothing becomes a finding. verify=true files one research run that checks the claims against sources, the draft's own citations included. Paths from the keeper only.

The databases the library owner gave read access to (SQLite, PostgreSQL, MySQL or MariaDB, SQL Server, MongoDB, DuckDB and data files). op=list names them. op=schema with `database` gives tables, columns, row counts and sample rows: read it before writing a query. op=query runs ONE reading statement: `sql` (a single SELECT or WITH), or for MongoDB a `collection` with a `filter` or a `pipeline` as JSON. Rows are capped, so count and group in the query. The result is saved as a page; cite its locator. Only the owner may use this, and connections are added from the command line.

The DNA store beyond the shelf. op=status: how many documents and how many labelled, per kind, and whether the person shares. op=pack-export (the owner) with path writes a pack of the labelled documents (the shared pool's format: the works' layers and provenance, nothing of the person); op=pack-import (the owner) with path reads one in; op=pack-refresh fetches the pack researchzosho.org announces, when there is one. op=podcast-like with words: the podcast episodes like one, among the transcripts the library holds, the same show set aside; op=podcast-read (the owner) reads the transcripts for their DNA now.

Everything the library holds on a person or an organisation, as one document: the relations the graph has, the claims by the class of their source (public record, own publication, told, supplied, third-party), a timeline, what was told in conversation, the reports to dive into, the open questions, what was looked for and not found, and the subjects it links to. A view over the shelves, written fresh each time; a research run adds to it.

What the shelves hold on a claim: accepted entries bearing on it, open disputes touching it, and a verdict computed from their states — established | disputed | not_established. No model composes it.

A READING AID, never a record. id + rung re-explains an entry for a reader at that rung: beginner (new to the field), familiar (knows the field but not this work) or written (the sources' own register). term (+ in, the entry it appears in) explains a term as it is used there. Shelves only: written from the entry, its findings and its captured sources; every paragraph is read back against what it cites and marked when the shelves do not support it. grounding is shelves | thin | none; when none, offer carries the library_research call that would fill the gap — pass it as it is (quick: true runs it now, ahead of the night's work), then call library_explain again. terms[] are words the reader may need next; each is a further library_explain. Needs a model drive the first time; cached after.

The queue of open questions the housekeeping's explorer researches a few of each night — op=list returns them in queue order with type, parked, position, tonight, the report that left each (report, report_title, report_fate: kept | waiting | disputed | retired | none), perspective, subjects, language, and similar (the head of a group that reads alike); list takes filters: type, show (queued | parked | all, default all), report (an id or its prefix), fate, who (perspective text), subject, language, q (words); hints=true adds answered {id, title, state} where a claim on the shelves already answers a question. op=add queues one attributed to the patron; op=next moves one to the head; op=later to the tail; op=park keeps one out of the explorer's reach, with why (the reason, listed back as parked_why with the date); op=unpark returns it; op=drop closes one without researching it; op=tidy removes duplicate lines (write access). A report's leftover questions are filed parked.

One entry in full by id: body, every source with locator and edition, supersedes / superseded_by, subjects, related entries, and the review record {round, reviewer, decision, stale}. An investigation also carries sections[] {heading, chars}, sources[] as rows {n, locator, title, edition, published, fetched, language, same_as}, claims[] (its findings with bodies) and open_questions[]. A write-up can run past what one call may carry: pass section (a heading; "answer" = the write-up without the harness's sections; "workers", "references", "evidence") to get one section, and offset / max_chars for a window of the body.

The library on ITSELF: what ResearchZosho is, which version is running, what it can and cannot do, how to do something in it, and why something does not work. Returns its live state (version, what this library holds, the model and whether it answers, the web search, the fields that are on, the chat's commands, the pages) and the sections of the program's own documentation that answer the question, with where to ask what neither answers. Use it for any question about the library or the program rather than answering from memory.

What the person's own lists hold: a Calibre library, an inventory, a reading list — every list shelved through library_items. An exact lookup: every word of the query must appear in an entry (title words, an author, a year). Ask it for each candidate to tell owned from not owned; a search snippet is a guess, this is not.

The claims waiting for the keeper's decision — drafts, and accepted claims whose review went stale — and the decisions. op=list returns items[] {id, title, state, stale, kind, tier, confidence, writer, date, subjects, language, sources, report, report_title}, oldest first, with filters report (an investigation id or its prefix), subject, kind (extraction | synthesis | interpretation | speculation), tier, confidence, writer, state (draft | stale), language, q (words in the title). op=accept puts claims into every answer from now on; op=dispute files why; op=retire keeps them on disk and out of every answer — each takes ids[] (or id), or report for every waiting claim of one investigation (write access).

A list of things — books, tools, places, an inventory of line items — as a starting point for research. Give a path (one item per line, a markdown table, or a CSV), a url, or the list as text. The list is shelved as it is so the library knows what the person has; each item is checked against the shelves (held or not); then, through the lens ("{item}: what it is, who made or wrote it, what it is for, and what is known about it" unless the person says what they want to know), as=frontier (default) files a question per item on the open questions, as=runs sends them out as research runs with a lane per item (batches of 8), as=none only reports what is held. Paths from the keeper only.

One job from the ledger ({job}: state queued|running|done|failed|stopped, elapsed_s, result when finished), or your jobs (active[], finished[], paused) when job_id is omitted. op=stop {job_id} stops one run: queued, it never starts; running, it ends at its next turn. op=pause holds the runner (queued runs wait, a running one holds at its next turn); op=resume lets it go (write access).

The GRAPH around a focus: a person, place, event, document, organisation, work or concept by name, or an entry id. Returns the neighbourhood — nodes {id, kind, label, also[], wikidata} and edges, where every edge IS a finding {from, predicate, to, finding, state, confidence, disputed} — plus the open questions touching it. This is how a family shelf and a computer-science shelf meet at one person.

A meeting transcript as a starting point: WebVTT (Zoom, Teams, Otter) or "Name: words" lines. The transcript goes on the shelves with its decisions; the questions raised join the open questions; the claims made come back to check, with who said them. Nothing becomes a finding. verify=true files one research run that checks the claims. Paths from the keeper only.

The models this library can run on, by name: list them; add one (a name, a host address or word — anthropic, openai, openrouter, gemini, nebius, huggingface — a model id, prices per million tokens as in/out/cache, an effort, a monthly cap in dollars); remove one; keep a host's key. A key is never returned. Write access.

The person's own shelf: their Calibre library (the calibre setting), the podcasts they follow, what they bought. op=profile: what they went for — the books they added one by one (a day of fifty or more additions is an import: holdings, not taste) and any they rated themselves — the subject mix among the liked books with its tilt against the whole shelf, the authors rated well more than once, decades, languages, sample titles, podcasts followed, shops bought from; an imported rating is set aside. op=have with title (and author): whether a book is already on the shelf, in the purchase ledger or on a holdings list. op=liked with words: the rated books on a subject or by an author, best first. op=dna with title: a shelf book's DNA — the text counted (length, sentences, dialogue, reading grade, the feeling's arc) and the model's reading of it over twelve windows (kind, themes, moods against an ordinary book of its kind, pace, darkness, scale, humour, style, drive, topics), made now when missing; without a title (the owner) the night's reading of the shelf runs now.

What the person owns: their purchases, imported from their own records (the zip Amazon's data request returns, a mailbox of order confirmations, a bank or card statement as CSV). op=status says what the ledger holds and where each file's records begin; op=find with words (and year, shop) looks something up: what was bought, when, where, for how much; op=import with path reads a file on this machine (the owner only); op=forget with file drops what one file added. The ledger is a reference, never a requirement.

Before researching: WHO studies this question and what would each of them insist on asking (STORM's move). Returns perspectives [{perspective, why, questions[]}] and sub_questions[] ready to pass to library_research. One or two searches and one judge call; needs a model drive.

The person's preferences file for books: what they want more and less of, each line with its source (drafted from the shelf, learned from the reactions, or said by the person). op=show (default) the lines; op=draft (the owner) writes one from the shelf; op=set with name and value changes a line (weight.subject 2 | roam 0.4 | popularity one|none|known | words "…" | exclude.author|subject|kind|series <value> | prefer.mood|subject|kind|pace|author <value> | avoid.<the same> <value>); op=unset drops one; op=export and op=import with path carry the file; op=learn reads the reactions into it. "Like X" and browsing read it. A program sets a line only when the person said so in this conversation.

A list of questions — a syllabus, an exam, what the person needs answered — onto the open questions in order (as=frontier, the default; the housekeeping's explorer works through them at night), or as research runs one each up to 10 (as=runs; the rest are filed as open questions). One question per line. Paths from the keeper only.

Writes the person's reaction to a book down: like, dislike, seen or own, by title (and author); about=<title> when the book was a row of a "like" list for that title. A disliked book leaves later lists, a seen or owned one is marked, and the preferences file learns from the likes and dislikes. The owner, and only when the person said so in this conversation.

The captured raw text behind a source locator — the verbatim path: read the evidence, not only the claim. Returns title, edition, captured_at and the text (truncated at max_chars).

A reading list as a starting point: BibTeX, RIS (Zotero, EndNote), a CSV export, or lines of DOIs, urls and titles. Every entry with a locator is fetched onto the shelves as a collection; one that cannot be read becomes a source request; an entry with a title only is listed so the person can give it a locator or research it by name (library_items). watch=true has the housekeeping re-read the pages. Then library_research with sources=shelves and the collection reads from them. Paths from the keeper only.

Take a report or a claim out of the library for good. id is a report (I-…) or a claim (F-…). For a report, what=all (default) removes it and the claims that are its alone, what=report the report only, what=claims its claims only; a claim another report also cites stays and is named. dry=true shows the plan and removes nothing. Files are deleted, the index and the changes log updated; what the run fetched from the web stays. There is no undo — retire is the gentler thing for a claim.

Ask to be let into this library. No access is needed to ask. Returns a request id and a claim secret (shown once); the owner approves or denies; collect the outcome with library_access.

A RESEARCH RUN: send a question to The Librarian's workers. The library's own runner plans sub-questions (yours, if you pass them), researches each in parallel with web search + fetch noting every fact with its source, has a critic decide on a second round, and writes the investigation in sections; the result enters the library as a draft investigation attributed to you and the review crew extracts findings from it. Returns a job_id to poll with library_job. Write access. Takes minutes to hours — do not wait inline. A field that acts only when asked for, such as family history, joins the run only when field names it; a question that looks like one of its questions is researched as ordinary research, and the result then carries suggestion {field, why, how}: what the field's mode does and how to ask for it.

Search The Librarian's catalog (hybrid BM25 + dense, CJK-aware). Returns hits {id, kind, title, snippet, score, state, subjects[]} and next_cursor; use library_get for one entry in full.

Seeks: the same question researched several times apart (library_research with seek=<group>, each on its own copy of the shelves, on any model), and one report merged from them. op=list the groups and their seeks; op=read a seek's report (group, job_id) or the merged one (group); op=merge files the merge run: the judge reads the seeks' claims in pairs, agreed claims are filed as findings, contradictions as disputed findings with the question on the frontier, what only one seek found is listed; op=drop removes a group's seeks and their shelves (the merged report stays).

The searches the housekeeping keeps running on a cadence and reports what is new from — op=list (each with parked and due); op=add {name, query, every_days} keeps one; op=every {name, every_days} changes its cadence; op=park {name} keeps it without running it, op=unpark {name} puts it back in the rotation; op=remove {name} stops keeping it (write access).

The library's settings (write access): op=list shows every setting with what it is for and its value, a key only as set or not; op=set with name and value writes one (name as the key, or without its RESEARCHZOSHO_ prefix; a blank value, or op=unset, unsets it). A program sets a key or an address only when the person gave it in this conversation.

The library reached from outside the house (write access): op=status says how it stands; op=tailscale, op=public (Tailscale Funnel: anyone with the link) or op=cloudflare (a quick tunnel; mail=… lets only listed addresses through) turns the relay on when its program is on the machine, only when the person asked for it in this conversation; op=off turns it off. A missing program is installed by the person on their own screen: researchzosho share, or /share in the chat.

SHARPEN a rough research question BEFORE running it: the library looks at what it already holds, asks who studies the question and what each would ask, and answers with the question rewritten, the assumptions it made (read them: a model sharpens by narrowing), a brief (scope, sub_questions, sources, deliverables), what is already held, at most three questions_for_you, a mode and a size (quick | full), and research_question — the text to pass to library_research as it is or after the person edits it. Proposes only; never runs research. Needs a model drive; two or three calls.

library_id, library_name, contract, counts by kind and state, last_updated.

The controlled vocabulary: {id, label, broader, narrower[], count}. Facets (the part before '--') are the broader terms. Use an id with library_search's subject filter.

Submit one finding as a DRAFT. It enters the librarian's review like any research run and never touches canon directly. A claim with no sources is refused (error no_sources). Requires a patron with write access.

Register a webhook: every change on this library's feed (retired, disputed, revised, supplied, …) is POSTed to your address, signed with a secret (X-ResearchZosho-Signature: sha256=hmac). The changes feed stays the source of truth. Needs a named patron.

What the library suggests the person do next, from the shelves' own events: a forecast whose date passed, a story the shelves have grown past, a question of theirs a claim now answers, a question asked again and again, a report that landed, a dossier not researched for months. Each is classed now, next or hold with the probability they take it up; their marks are the outcomes a Brier score is kept over. op=list (default): the open suggestions, now first. op=mark (write): id, outcome (taken, dismissed, later), note.

A thing the person already has, read as a starting point for research: a code repository (a folder, or a git url cloned when git is installed — it says so when git is missing), a paper (a PDF or document file, a DOI, an arXiv page), a website or product page (any other url), or an issue tracker (a GitHub issues page, or an export file). The kind is detected from the thing, or given. op=survey (the default) reads it, shelves the text, files one draft claim saying what it is, what it says or does, what it claims and what it rests on, and returns numbered directions: research questions the library could answer about it. Nothing runs at that point. op=pick with picks="1,3" files a research run for each direction named; op=do with question files the person's own direction and its run. pick and do take the thing by name (as surveyed).

What a person tells the library about a subject from their own knowledge: what it was like, who was there, an incident, a date. Filed as a claim of class told, in their words, dated, with them as the writer; the questions it raises that the public record could answer join the open questions, so an account can open a search. Answer with 'noted:' and what it raised.

Remove a webhook you registered.