strom-research 1.13.2-beta.3 → 1.13.2-beta.5

package/assets/lang/cs.json

@@ -1641,12 +1641,15 @@
   "ui.arg.year": "rok",
   "ui.error.arg.extra": "nečekaný argument „{arg}“",
   "ui.error.arg.extra.hint": "strom help {cmd} — hodnotu s mezerami dát do uvozovek",
+  "ui.error.arg.extra.fix.hint": "správně: {fix}",
+  "ui.error.arg.extra.several.hint": "více záznamů: {line}",
   "ui.error.option.is-arg": "{opt} se u {cmd} píše jako argument, ne jako volba",
   "ui.error.option.is-arg.hint": "{cmd} <{arg}>",
   "ui.error.option.unknown": "neznámá volba {opt} u {cmd}",
   "ui.error.option.unknown.hint": "volby: {options} · strom help {path}",
   "ui.error.option.unknown.none": "neznámá volba {opt} u {cmd}",
   "ui.error.option.unknown.none.hint": "{help}",
+  "ui.error.option.unknown.status.hint": "jistota (--status) patří jen k údaji (E…), nikdy ke jménu ani k rodině: strom cite E… S… --status probable|proven, strom event edit E… --status …",
   "ui.error.option.no-value": "{opt} se píše bez hodnoty",
   "ui.error.option.needs-value": "{opt} potřebuje hodnotu {value}",
   "ui.error.option.needs-value.hint": "strom help {path}",

package/assets/lang/de.json

@@ -1641,12 +1641,15 @@
   "ui.arg.year": "Jahr",
   "ui.error.arg.extra": "unerwartetes Argument „{arg}“",
   "ui.error.arg.extra.hint": "strom help {cmd} — Werte mit Leerzeichen in Anführungszeichen",
+  "ui.error.arg.extra.fix.hint": "richtig: {fix}",
+  "ui.error.arg.extra.several.hint": "mehrere Datensätze: {line}",
   "ui.error.option.is-arg": "{opt} ist bei {cmd} ein Argument, keine Option",
   "ui.error.option.is-arg.hint": "{cmd} <{arg}>",
   "ui.error.option.unknown": "unbekannte Option {opt} für {cmd}",
   "ui.error.option.unknown.hint": "Optionen: {options} · strom help {path}",
   "ui.error.option.unknown.none": "unbekannte Option {opt} für {cmd}",
   "ui.error.option.unknown.none.hint": "{help}",
+  "ui.error.option.unknown.status.hint": "eine Sicherheit (--status) gehört nur zu einer Angabe (E…), nie zu einem Namen oder einer Familie: strom cite E… S… --status probable|proven, strom event edit E… --status …",
   "ui.error.option.no-value": "{opt} hat keinen Wert",
   "ui.error.option.needs-value": "{opt} braucht einen Wert {value}",
   "ui.error.option.needs-value.hint": "strom help {path}",

package/assets/method/core.md

@@ -4,11 +4,14 @@ You work to the Genealogical Proof Standard: search thoroughly, cite every
 fact, analyse and correlate the evidence, resolve conflicts, and write the
 conclusion so the next researcher can follow it.
 
+<!-- for link verify enrich intake locate narrate -->
 - **Certainty is explicit.** `proven` — you read the original record yourself
   and it states the fact directly. `probable` — good evidence, but indirect or a
   single reading of hard handwriting. `possible` — weak evidence. `lead` — a
   family story, a family tree, an index or a guess: something to check, never a
   fact. Family trees and memories are leads, always.
+<!-- end -->
+<!-- for link verify enrich intake locate -->
 - **A namesake is not your person.** Identify by year, place, house, occupation,
   spouse and parents — the mother's maiden name decides most cases. Two entries
   that give one name different parents (another mother) are **two people** —
@@ -20,25 +23,36 @@ conclusion so the next researcher can follow it.
 - **Negative results are results.** Every search is recorded (`strom search add
   … --result negative`), with exactly what was covered. The most expensive
   mistake is to search the same book twice.
+<!-- end -->
 - **Check the premise first.** Before opening anything, read the brief's
   *Already known* — the searches of the task's books, people and surnames;
   `strom searched <where>` only for a book, place or years it does not list.
+  Work that answers another open task as well closes it with yours:
+  `strom session close --done "…" --next "…" --also-done T…`.
+<!-- for link verify enrich intake -->
 - **Extract everything the first time** you open a record: names, ages, house
   numbers, occupations, godparents, witnesses, midwife, remarks in the margin.
   Opening the same page again later costs more than writing it down now.
+<!-- end -->
+<!-- for link verify enrich intake locate -->
 - **Nothing is deleted.** A wrong fact is retracted with a reason; a changed
   fact is edited with a reason. Contradictions become conflicts, competing
   explanations become hypotheses — never silently pick one.
+<!-- end -->
 - **Write as you go.** Record each finding as soon as you have it. Your context
   can be cut or summarised at any moment, and a session working alone is
   stopped at its time limit (the brief says when; near the end strom's output
   counts down): what is only in the conversation is lost, what is in strom is
-  not. Going through many images, record each stretch when it is done (every
+  not.
+<!-- for link verify enrich intake -->
+  Going through many images, record each stretch when it is done (every
   ten images, each district or year): the finds, and `strom search add …
   --pages` for what was searched in vain — never all of it at the end.
   A fact of a person goes on the person (`event add`); a note or the diary is
   for what is not a fact yet. What you held back while who it was stayed a
   hypothesis goes on the person when you decide it.
+<!-- end -->
+<!-- for link verify enrich intake locate -->
 - **Lessons belong where they apply.** A quirk of a register (its calibration,
   two years per page, a hand that writes 7 like 1) goes to `strom lesson add
   --on B…`, so whoever opens that book next sees it.
@@ -50,6 +64,8 @@ conclusion so the next researcher can follow it.
   queue holds back every other task about people off the tree
   (`strom task list --off-tree`); they come back by themselves once the link
   is recorded.
+<!-- end -->
+<!-- for link verify enrich intake -->
 - **Say what a hypothesis would connect.** A hypothesis about a connection —
   whose child, the same person, a couple, siblings — gets each variant's links
   when you record or argue it: `strom hypothesis link H… B --child P… --of F…`
@@ -63,11 +79,15 @@ conclusion so the next researcher can follow it.
   (`strom hypothesis variant H… "<claim>"`). Deciding one whose variants have
   links names the variant it is decided for:
   `strom hypothesis decide H… --variant B --decision "…"`.
+<!-- end -->
 - **Stay within the task.** New questions become new tasks (`strom task add`,
-  with where and done-when), not detours. A new task whose images are not here
+  with where and done-when), not detours.
+<!-- for link verify enrich intake locate -->
+  A new task whose images are not here
   gets them now: through the archive's connector (built first when it has
   none), or — where the archive allows only that — it waits at once for what
   the user should download (`strom task wait T… --images B…:<numbers> --on
   "…"`); a later session would only find that out.
+<!-- end -->
 - **Hand over cleanly.** Close with a summary of what was proven, what was
   searched in vain, and the next cheapest step.

package/assets/method/reading.md

@@ -13,7 +13,11 @@ about 80 views write down what you found before you open more.
 
 - **Look only through views**: `strom media view B0001:57` (record set and
   image number) makes a file in `.strom/views/` — open that file. Whole images
-  come reduced: good for "is our surname on this page?", not for reading.
+  (and `--grid`) come reduced: for finding the entry, never for reading it.
+  **Transcribe only from a crop** at full resolution (`--crop`, or `--half`
+  where the entry fills the page). A whole view unclear, or the entry not where
+  expected: `--half both` before calling it not found — a negative search only
+  after the pages were seen at full resolution.
 - **The views of a scan, or of a batch, in one call**, then open all the files
   it lists together — not one call and one look per view:
   `strom media view B0001:57-60 --half both` (the two pages of each double
@@ -26,11 +30,13 @@ about 80 views write down what you found before you open more.
   width, height as fractions) or `--half left|right` for one page of a spread.
   A crop comes at full resolution (small ones enlarged). `--contrast` for faded
   ink, `--rotate 90` for sideways pages.
+<!-- for link verify enrich part -->
 - **Too small to read?** When a crop says it is enlarged, the scan has no more
   detail there. A connector that can fetch a part of an image sharper says so
   in that line: `strom fetch <connector> --recordset B0001 --images 57 --crop
   …` (one request). Then view the same crop again: it comes from the sharper
   part by itself.
+<!-- end -->
 - **Browsing a book is a reader's job**: `strom read B0001 --images 40-69
   --question "…"` — readers in batches of ten, each with the full question.
   Their reports stay in notes/readings/; `strom read` ends with what they
@@ -44,14 +50,14 @@ about 80 views write down what you found before you open more.
   It waits for its readers (often ten minutes or more): run it in the
   foreground and let it finish — your session ends with your turn, and
   whatever is left running in the background ends with it.
-  (Your own subagents can read too — on your own model, never a faster one — in batches of at most twelve — but what they
-  report is lost with your session unless you write it down: record each
-  report as it comes back, before you send the next readers.) Only the entries
+  (A scan-reader subagent where your agent has one, strom-scan-reader, or your
+  own subagents — on your own model, never a faster one — read too, in batches
+  of at most twelve with the whole question; what they report is lost with
+  your session unless you write it down: record each report as it comes back,
+  before you send the next readers.) Only the entries
   that will be cited need your own eyes, at full resolution.
 - Old handwriting is decoded, not copied: never a weaker model for handwriting,
   never a guess. "Illegible" is a valid and valuable answer.
-- **Extract everything the first time**: names, ages, house numbers,
-  occupations, godparents, witnesses, midwife, remarks in the margin.
 - **Report image by image, as you go**: the image and page, what was found (or
   nothing), what was illegible and where, the hand, how sure each name is.
 - **Cite the image, and where the entry is on it**: `strom source add …
@@ -61,12 +67,16 @@ about 80 views write down what you found before you open more.
   numbering jumps) were not searched: say so in `--note` (strom read puts the
   gaps its readers saw there), or leave them out of the range
   (`--pages 40-52,55-69`).
+<!-- for verify enrich intake -->
 - Page ↔ image: `strom recordset calibrate B0001 --point 57=112` (measured on
   the image, never guessed); then `strom media view B0001 --page 112` works.
-- **No images here yet:** a connector for that archive
-  (`strom connector list`) fetches them through strom:
-  `strom fetch <connector> <book> --images 40-69 --recordset B0001` — only the
-  images you need, never a whole book "just in case".
+<!-- end -->
+<!-- for link verify enrich connector -->
+- **No images here yet:** the book's connector fetches them through strom
+  (the brief's record set gives the exact `strom fetch`) — only the images you
+  need, never a whole book "just in case".
+<!-- end -->
+<!-- for link verify enrich no-connector -->
 - **The archive has no connector yet: build one — now, you.** The user does
   not know connectors exist and will not ask for one. Tell them in a sentence
   ("for this archive I am preparing a downloader; it will fetch only the images
@@ -74,10 +84,14 @@ about 80 views write down what you found before you open more.
   follow its DISCOVERY.md (the portal's terms and robots.txt first, official
   exports preferred), test it with `strom connector test`, then fetch. It is
   built once per archive and serves every later task.
+<!-- end -->
 - You never download from an archive yourself (curl, a script, your browser
-  tools) — only through a connector, paced by strom. The user saves images by
+  tools) — only through a connector, paced by strom.
+<!-- for link verify enrich no-connector by-hand -->
+  The user saves images by
   hand only where the archive does not allow automation (the connector then
-  finds books and gives links only) or where a check stops it. Ask with
+  finds books and gives links only), where its connector only finds books, or
+  where a check stops it. Ask with
   `strom task wait T… --images B0001:40-69 --on "…"`: strom makes the folder
   of the inbox they go into, shows it to the user with the book's link, and
   checks the numbers of what arrives. Write `--on` for the user, in their
@@ -88,3 +102,4 @@ about 80 views write down what you found before you open more.
   the entry and save that view too (40a.jpg); a full-resolution scan can be
   ordered from the archive. Then take the next task; when the images are
   registered the task comes back by itself.
+<!-- end -->

package/assets/method/recording.md

@@ -2,8 +2,11 @@
 
 Record a found entry in one batch: write the lines to a file in notes/ and run
 `strom batch --file notes/<file>` — all or nothing: when a line fails nothing is
-written and every failing line is said at once; fix them, run it again.
-`#name` labels what a line creates, `@name` uses it later.
+written and every failing line is said at once; fix them in the file with your
+file-editing tool, not with sed or another shell command (the tree's rules let
+the shell run strom, not edit files), run it again. `#name` labels what a line
+creates, `@name` uses it later; a found search comes after the source it found
+(`--found @s`).
 
     source add "Baptism of Jan Novák 1885" --kind baptism --recordset B0001 --clip B0001:57@0.05,0.40,0.45,0.18 --locator "pag. 112, entry 2" --language la --information primary --transcript @notes/entry.txt #s
     event add P0001 CHR --date "25 JUN 1885" --place "Týnec" --house 13 --cite @s --quote "baptizatus est" --with "godparent:Marie Dvořáková" --with "midwife:Anna Nová" --with "officiant:P. Josef Kříž" --status proven
@@ -22,7 +25,8 @@ written and every failing line is said at once; fix them, run it again.
 - `--transcript` is the entry's words as they stand in it — its own language,
   spelling and abbreviations, line by line, unread letters `[?]`; the Strom app
   shows them next to the entry's image. Every entry read gets them.
-- A fact already in the tree gets the citation (`cite E…`), not a second fact —
+- A fact already in the tree gets the citation (`cite E…`, `--status` is a
+  fact's only — a name or a family cited takes none), not a second fact —
   and what the record adds to it: `event edit E… --age husband:27 --age wife:17
   --house 21 --with "witness:…" --with "officiant:…"` (filling in needs no reason).
 - The names a record gives go on the person: `name add` (a maiden name completes

package/dist/agents/files.js

@@ -16,6 +16,7 @@ import { agentAppUrl, appSite } from "../core/stromapp.js";
 import { configDir } from "../core/paths.js";
 import { agentBrowser } from "../core/connector.js";
 import { CHROME_ALLOW, CHROME_DENY, chromeDomain } from "../core/browser.js";
+import { claudeScanReader, opencodeScanReader, SCAN_READER } from "./scanreader.js";
 export const MARKER = "<!-- strom: generated above this line (strom agents sync); your own notes below are kept -->";
 function agentsMd(tree) {
     const lang = langName(tree.config.lang);
@@ -313,6 +314,8 @@ export function opencodeConfig(tree) {
                 description: "strom run: the agent working alone, nobody to ask — what the rules do not allow is refused",
                 permission: refused(permission),
             },
+            // reading scans: the shell for strom and the file reader, nothing else
+            [SCAN_READER]: opencodeScanReader(tree, permission, refused),
         },
     };
 }
@@ -343,6 +346,7 @@ function dropAgentFiles(tree) {
         ["AGENTS.md", agentsMd(tree)],
         ["CLAUDE.md", claudeMd(tree)],
         [path.join(".claude", "settings.json"), JSON.stringify(claudeSettings(tree), null, 2) + "\n"],
+        [CLAUDE_SCAN_READER, claudeScanReader(tree)],
         ["opencode.json", JSON.stringify(opencodeConfig(tree), null, 2) + "\n"],
         [path.join(".grok", "config.toml"), grokConfig(tree)],
         [path.join(".grok", "rules", "strom.md"), GROK_RULES],
@@ -370,7 +374,9 @@ function dropAgentFiles(tree) {
     }
     return gone;
 }
-export const AGENT_FILES = ["AGENTS.md", "CLAUDE.md", path.join(".claude", "settings.json"), "opencode.json", path.join(".grok", "config.toml"), path.join(".grok", "rules", "strom.md")];
+/** Claude Code's subagent for reading scans, in the tree. */
+export const CLAUDE_SCAN_READER = path.join(".claude", "agents", `${SCAN_READER}.md`);
+export const AGENT_FILES = ["AGENTS.md", "CLAUDE.md", path.join(".claude", "settings.json"), CLAUDE_SCAN_READER, "opencode.json", path.join(".grok", "config.toml"), path.join(".grok", "rules", "strom.md")];
 export function syncAgentFiles(tree) {
     // an archive: no agent works on it, and nothing of one shows in its folder (Milan's decision, 2026-10-03) — strom's own
     // files of the agents go (what a person wrote in them stays); research switched on writes them again
@@ -396,6 +402,7 @@ export function syncAgentFiles(tree) {
     write("AGENTS.md", agentsMd(tree));
     write("CLAUDE.md", claudeMd(tree));
     write(path.join(".claude", "settings.json"), JSON.stringify(claudeSettings(tree), null, 2) + "\n");
+    write(CLAUDE_SCAN_READER, claudeScanReader(tree));
     write("opencode.json", JSON.stringify(opencodeConfig(tree), null, 2) + "\n");
     write(path.join(".grok", "config.toml"), grokConfig(tree));
     write(path.join(".grok", "rules", "strom.md"), GROK_RULES);

package/dist/agents/images.js

@@ -10,6 +10,11 @@
 export const IMAGE_MAX = 1568;
 /** What the newer models take in whole, and Claude Code, Codex and OpenCode pass on unshrunk. */
 export const IMAGE_MAX_LARGE = 2000;
+/**
+ * The long side of a whole image shown to find an entry on it (strom media view without --half, --crop or --split):
+ * the entry is found as surely as at 2000 px for half the tokens; it is read from a crop or a half at the model's size.
+ */
+export const OVERVIEW_MAX = 1400;
 /** A Claude model name or alias ("opus", "claude-opus-4-7", "claude-sonnet-5-5[1m]", "us.anthropic.claude-…"): its family and version. */
 export function claudeModel(model) {
     const m = /(opus|sonnet|haiku|fable|mythos)(?:[-.]?(\d{1,2})(?:[-.](\d{1,2})(?!\d))?(?!\d))?/.exec(model.toLowerCase());

package/dist/agents/profiles.js

@@ -35,6 +35,10 @@ was), then open the next. Every image and view stays in your context to the end:
 views, write down what you found and go on. A name or place the next search depends on:
 read it again on a crop at full resolution (\`strom media view … --crop\`)
 before you record it.
+OpenCode: this tree's subagent strom-scan-reader (strom and the file reader
+only) reads on your model, or on the one the user set for handwriting: hand
+it a batch with the whole question, and record what it returns at once (a
+negative result as a search by reader).
 `;
 /**
  * The agents strom can start working alone without the user's personal add-ons (agent.addons): Claude Code
@@ -63,6 +67,9 @@ comes back to you. Choose the model by the kind of work:
 | mechanical: downloads, renaming, counting, a plain grep | \`${m.cheap ?? "haiku"}\` |
 | judgement: identity, conflicts, which task next, writing to strom | nobody — you |
 
+Scans are read by this tree's subagent type \`strom-scan-reader\` (strom and the
+file reader only, so it starts small); the model above still holds.
+
 ${DELEGATION_RULES}
 - Send independent batches in ONE message so they run in parallel.
 - Do not read a subagent's transcript or output file — its images would come

package/dist/agents/scanreader.js

@@ -0,0 +1,96 @@
+// The scan reader: a subagent of its own kind for reading scans, in the agents that let a tree define one (Claude Code:
+// .claude/agents/ and --agents; OpenCode: an agent of mode subagent in opencode.json). It has the shell (for strom) and
+// the file reader (for its views) and nothing else — no subagents of its own, no web, no editing — and a short prompt,
+// so it starts small (a subagent of the agent's general kind starts with every tool and instruction the agent has).
+// Its model is the user's model.vision where they set one, else the model of the agent that starts it: never a
+// cheaper one picked by strom. Codex, Antigravity and Grok Build get none: Codex's agent roles cannot narrow its tools
+// and Antigravity has no such setting; Grok Build reads scans itself (profiles.ts), as the tree tells it.
+import path from "node:path";
+import { Settings } from "../core/config.js";
+import { writeFileAtomic } from "../core/json.js";
+export const SCAN_READER = "strom-scan-reader";
+export const SCAN_READER_DESCRIPTION = "Reads scans of old records for this family research (registers, indexes, land books, any old handwriting): give it the images (B0001:57-66, M0012) and the whole question — what counts as a find, which years and names. It opens the views through strom, reads them and returns what it read, image by image. It never writes to the research.";
+export const SCAN_READER_PROMPT = `You read scans of old records for a family research kept with strom. You are given images
+(B0001:57-66, M0012…) and the whole question: what counts as a find, which years and names. You read
+and report; you never write to the research and decide nothing.
+
+Views: run strom by its command — \`strom media view B0001:57-60 --half both\` (at most 12 images and
+24 views a call); an entry to read closely: \`--grid\` to see where it is, then \`--crop x,y,w,h\` at
+full resolution; \`--contrast\` for faded ink. Open ALL the views a call lists in ONE message — every
+Read call together, never one view at a time — then read them, then the next call. Every view stays
+in your context: after about 80 views, stop and report what you have. \`strom read …\` hands a range
+to strom's own readers: only when the question asks for it (in the foreground; let it finish).
+
+Reading: transcribe as written — the record's language, spelling and abbreviations, not modernised;
+[?] for each uncertain letter or word; never complete a name or date from what the question expects.
+"Illegible" beats a guess; "nothing" is a valid result.
+
+Return, image by image (image and page): found / nothing / unclear; each entry that answers the
+question word for word, where it stands and its columns (date, house, names, parents, godparents,
+witnesses, remarks); what was illegible and where; the hand; how sure you are of each name. Say
+plainly which images you did not open or read whole — never report an image you did not read as
+searched.
+`;
+/** The model the user chose for reading handwriting (model.vision: flag, env, tree, user config) — none: the starting agent's. */
+export function scanReaderModel(tree, agent) {
+    const r = new Settings(tree.env, {}).resolve("model.vision", tree.config, agent);
+    return r ? String(r.value) : undefined;
+}
+/** Claude Code's tools for it: the shell (on Windows its PowerShell too, where the tree's rules allow strom as well) and the file reader. */
+export function scanReaderTools(platform = process.platform) {
+    return platform === "win32" ? ["Bash", "PowerShell", "Read"] : ["Bash", "Read"];
+}
+/** .claude/agents/strom-scan-reader.md: a conversation (the terminal, the desktop app) finds it in the tree. */
+export function claudeScanReader(tree) {
+    return `---
+name: ${SCAN_READER}
+description: ${JSON.stringify(SCAN_READER_DESCRIPTION)}
+tools: ${scanReaderTools().join(", ")}
+model: ${JSON.stringify(scanReaderModel(tree, "claude") ?? "inherit")}
+---
+
+${SCAN_READER_PROMPT}`;
+}
+/** The same for --agents: a run is given it on the command line, whatever Claude Code makes of the folder's files. */
+export function claudeAgentsJson(tree) {
+    return {
+        [SCAN_READER]: {
+            description: SCAN_READER_DESCRIPTION,
+            prompt: SCAN_READER_PROMPT,
+            tools: scanReaderTools(),
+            model: scanReaderModel(tree, "claude") ?? "inherit",
+        },
+    };
+}
+/** The file a headless run of Claude Code is given (--agents): in .strom/, written afresh for each run. */
+export function claudeAgentsFile(tree) {
+    const file = path.join(tree.root, ".strom", "claude-agents.json");
+    writeFileAtomic(file, JSON.stringify(claudeAgentsJson(tree), null, 2) + "\n");
+    return file;
+}
+/**
+ * OpenCode's agent of mode subagent: the tree's rules with everything but reading and strom refused — no editing, no
+ * web, no subagents, no questions (nothing that would ask: a run nobody watches would end at it).
+ */
+export function opencodeScanReader(tree, permission, refused) {
+    const model = scanReaderModel(tree, "opencode");
+    const bash = permission.bash;
+    const only = {
+        ...permission,
+        bash: { ...bash, "*": "deny" },
+        edit: "deny",
+        webfetch: "deny",
+        websearch: "deny",
+        task: "deny",
+        question: "deny",
+        skill: "deny",
+        todowrite: "deny",
+    };
+    return {
+        mode: "subagent",
+        description: SCAN_READER_DESCRIPTION,
+        prompt: SCAN_READER_PROMPT,
+        ...(model ? { model } : {}),
+        permission: refused(only),
+    };
+}