@geml/geml 1.1.1 → 1.4.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/geml.js CHANGED
@@ -8,23 +8,32 @@
8
8
  // M2: inline parsing of flow blocks (§5 — emphasis/strong/strike, code, math,
9
9
  // media embeds, links, auto-references, footnotes) and build-time reference
10
10
  // validation (§8 — unique ids, resolvable internal/cross-document references).
11
- import { readFileSync, writeFileSync } from "node:fs";
12
- import { basename, dirname, join, resolve as resolvePath } from "node:path";
11
+ import { readFileSync, writeFileSync, realpathSync, statSync, existsSync } from "node:fs";
12
+ import { basename, dirname, isAbsolute, join, relative, resolve as resolvePath, sep } from "node:path";
13
13
  import { fileURLToPath } from "node:url";
14
14
  import { spawnSync } from "node:child_process";
15
15
  import { commit, restore, verify, listRevisions, resolveContent, firstChangedContent } from "./history.js";
16
- import { renderHtml } from "./render.js";
16
+ import { renderHtml } from "./render-html.js";
17
+ import { normalizeBlockId } from "./block-edit.js";
17
18
  import { coerce, parseAttrs } from "./attrs.js";
18
- import { parseInline } from "./inline.js";
19
+ import { META_REF_SRC, parseInline } from "./inline.js";
19
20
  import { parseTable } from "./table.js";
20
21
  import { buildChart } from "./chart.js";
21
22
  import { mdToGeml } from "./from-md.js";
22
23
  import { serialize } from "./serialize.js";
23
24
  import { gemlToMd } from "./to-md.js";
24
25
  export { mdToGeml } from "./from-md.js";
25
- export { renderHtml } from "./render.js";
26
+ export { renderHtml } from "./render-html.js";
26
27
  export { serialize } from "./serialize.js";
27
28
  export { gemlToMd } from "./to-md.js";
29
+ // A block id is any non-whitespace run (§4), so it may contain regex
30
+ // metacharacters. Every place that builds a RegExp from an id MUST run it
31
+ // through this first, or a crafted id (`#a(`, `#(x+x+)+y`) turns a labeled-close
32
+ // or reference match into an uncaught `SyntaxError` or a ReDoS on the main
33
+ // parse path (SEC: document-controlled RegExp injection).
34
+ function reLit(s) {
35
+ return s.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
36
+ }
28
37
  // Type registry: which body mode each typed block uses. Unknown types are a
29
38
  // warning and fall back to `raw` (forward compatibility, §3/§8).
30
39
  const REGISTRY = {
@@ -34,6 +43,7 @@ const REGISTRY = {
34
43
  table: "raw", // structured table parsing lands in M3
35
44
  output: "raw", // captured result of a code block (stored, never executed)
36
45
  note: "flow",
46
+ text: "flow", // addressable prose container: an id/attrs for a run of flow, no callout chrome
37
47
  meta: "data",
38
48
  };
39
49
  // §7: built-in diagram renderer registry. Unknown formats are a warning (the
@@ -45,6 +55,11 @@ const DIAGRAM_RENDERERS = new Set(["mermaid", "graphviz", "dot", "d2", "plantuml
45
55
  const FENCE_OPEN = /^(={3,})[ \t]+([A-Za-z][A-Za-z0-9_-]*)[ \t]*(\{.*\})?[ \t]*$/;
46
56
  const HEADING = /^(#{1,6})[ \t]+(.*?)[ \t]*(\{[^}]*\})?[ \t]*$/;
47
57
  const LIST_ITEM = /^[ \t]*(?:[-*]|\d+\.)[ \t]+(.*)$/;
58
+ // Maximum block/list nesting depth the recursive-descent scanner will build
59
+ // before emitting a diagnostic instead of recursing further. Guards parse()
60
+ // (scanBlocks / parseList) and, in step, the renderer against a deeply nested
61
+ // document overflowing the call stack (DoS). 256 is far past any real document.
62
+ const MAX_NESTING = 256;
48
63
  function isCloseFence(line, openLen) {
49
64
  const t = line.replace(/\s+$/, "");
50
65
  return /^=+$/.test(t) && t.length === openLen;
@@ -62,15 +77,67 @@ function slug(text) {
62
77
  // ---------------------------------------------------------------------------
63
78
  // §4: substitute `{{key}}` in flow text with the matching `=== meta` value.
64
79
  // An unknown key is a build error (single-source-of-truth, fail loudly).
80
+ // The scan mirrors the §5.3(1) verbatim atoms: a `{{key}}` inside a code span
81
+ // or inline math is left untouched (so GEML prose can document this very
82
+ // syntax), and a backslash-escaped character can neither open a span nor a
83
+ // `{{…}}` reference — `\{{key}}` renders as the literal text `{{key}}`.
84
+ const META_REF = new RegExp(META_REF_SRC, "y");
65
85
  function interpolate(text, line, ctx) {
66
86
  if (!text.includes("{{"))
67
87
  return text;
68
- return text.replace(/\{\{\s*([A-Za-z_][A-Za-z0-9_-]*)\s*\}\}/g, (full, key) => {
69
- if (ctx.meta.has(key))
70
- return ctx.meta.get(key);
71
- ctx.diags.push({ severity: "error", message: `unknown metadata reference \`{{${key}}}\``, line });
72
- return full;
73
- });
88
+ let out = "";
89
+ let i = 0;
90
+ while (i < text.length) {
91
+ const c = text[i];
92
+ if (c === "\\" && i + 1 < text.length) {
93
+ out += c + text[i + 1];
94
+ i += 2;
95
+ continue;
96
+ }
97
+ if (c === "`") {
98
+ let n = 0;
99
+ while (text[i + n] === "`")
100
+ n++;
101
+ const close = text.indexOf("`".repeat(n), i + n);
102
+ if (close >= 0) {
103
+ out += text.slice(i, close + n);
104
+ i = close + n;
105
+ continue;
106
+ }
107
+ out += text.slice(i, i + n); // unclosed run: literal, keep scanning
108
+ i += n;
109
+ continue;
110
+ }
111
+ if (c === "$") {
112
+ const close = text.indexOf("$", i + 1);
113
+ if (close > i + 1) {
114
+ out += text.slice(i, close + 1);
115
+ i = close + 1;
116
+ continue;
117
+ }
118
+ out += c;
119
+ i++;
120
+ continue;
121
+ }
122
+ if (c === "{" && text[i + 1] === "{") {
123
+ META_REF.lastIndex = i;
124
+ const m = META_REF.exec(text);
125
+ if (m) {
126
+ const key = m[1];
127
+ if (ctx.meta.has(key))
128
+ out += ctx.meta.get(key);
129
+ else {
130
+ ctx.diags.push({ severity: "error", message: `unknown metadata reference \`{{${key}}}\``, line });
131
+ out += m[0];
132
+ }
133
+ i = META_REF.lastIndex;
134
+ continue;
135
+ }
136
+ }
137
+ out += c;
138
+ i++;
139
+ }
140
+ return out;
74
141
  }
75
142
  // Register a block id, flagging duplicates as errors (§4: ids unique per doc).
76
143
  function registerId(ctx, id, line) {
@@ -122,6 +189,7 @@ function parseList(lines, i, base, ctx) {
122
189
  const root = mkList(matchMarker(lines[i]));
123
190
  const stack = [{ list: root, indent: matchMarker(lines[i]).indent }];
124
191
  let prevBlank = false;
192
+ let tooDeep = false;
125
193
  while (i < lines.length) {
126
194
  if (lines[i].trim() === "") {
127
195
  prevBlank = true;
@@ -139,9 +207,21 @@ function parseList(lines, i, base, ctx) {
139
207
  const parent = top.list.items[top.list.items.length - 1];
140
208
  if (!parent)
141
209
  break; // deeper indent with no parent item: defensive stop
142
- cur = mkList(mk);
143
- (parent.children ??= []).push(cur);
144
- stack.push({ list: cur, indent: mk.indent });
210
+ if (stack.length >= MAX_NESTING) {
211
+ // Refuse to nest deeper than the cap: keep the item at the current level
212
+ // rather than building a model that overflows the renderer (DoS). One
213
+ // diagnostic per over-deep list; content is preserved, just flattened.
214
+ if (!tooDeep) {
215
+ ctx.diags.push({ severity: "error", message: `list nesting too deep (max ${MAX_NESTING})`, line: base + i + 1 });
216
+ tooDeep = true;
217
+ }
218
+ cur = top.list;
219
+ }
220
+ else {
221
+ cur = mkList(mk);
222
+ (parent.children ??= []).push(cur);
223
+ stack.push({ list: cur, indent: mk.indent });
224
+ }
145
225
  }
146
226
  else {
147
227
  // §5: a change of marker type (bullet ↔ ordered) at the same level ends
@@ -159,7 +239,7 @@ function parseList(lines, i, base, ctx) {
159
239
  }
160
240
  return { block: root, next: i };
161
241
  }
162
- function scanBlocks(lines, base, ctx) {
242
+ function scanBlocks(lines, base, ctx, depth = 0) {
163
243
  const blocks = [];
164
244
  const diags = ctx.diags;
165
245
  let i = 0;
@@ -205,7 +285,7 @@ function scanBlocks(lines, base, ctx) {
205
285
  // of any length ≥ 3 followed by the block's id). The labeled close is a
206
286
  // *local* close: it can't be gotten wrong by miscounting `=`, so it is the
207
287
  // safe way to nest (§3).
208
- const labeled = attrs.id !== undefined ? new RegExp(`^={3,}[ \\t]+#${attrs.id}[ \\t]*$`) : null;
288
+ const labeled = attrs.id !== undefined ? new RegExp(`^={3,}[ \\t]+#${reLit(attrs.id)}[ \\t]*$`) : null;
209
289
  const body = [];
210
290
  let j = i + 1;
211
291
  let closed = false;
@@ -242,7 +322,16 @@ function scanBlocks(lines, base, ctx) {
242
322
  ctx.refs.push({ kind: "internal", anchor: of.slice(1), line: openLineNo });
243
323
  }
244
324
  if (mode === "flow") {
245
- block.children = scanBlocks(body, base + i + 1, ctx);
325
+ if (depth >= MAX_NESTING) {
326
+ // Refuse to recurse past the cap: emit a diagnostic and keep the body
327
+ // as raw so the parser returns cleanly instead of overflowing the
328
+ // call stack on a pathologically nested document (DoS).
329
+ diags.push({ severity: "error", message: `block nesting too deep (max ${MAX_NESTING}); body kept as raw`, line: openLineNo });
330
+ block.raw = body;
331
+ }
332
+ else {
333
+ block.children = scanBlocks(body, base + i + 1, ctx, depth + 1);
334
+ }
246
335
  }
247
336
  else if (mode === "data") {
248
337
  block.data = parseData(body);
@@ -453,8 +542,46 @@ export function parse(source, opts = {}) {
453
542
  }
454
543
  // The id that a fence/heading line defines, matching how scanBlocks derives it
455
544
  // (parseAttrs for the attribute object; heading text slug when no explicit id).
456
- function idOfHeading(braces, text) {
457
- return (braces ? parseAttrs(braces).id : undefined) ?? slug(text);
545
+ // The slug MUST come from the INTERPOLATED text — scanBlocks slugs after
546
+ // interpolate(), so `# {{title}} Setup` registers the substituted slug; slugging
547
+ // the raw text here would create a phantom id the parser never registered and
548
+ // make the real one unaddressable. `ctx` is an inert context carrying the
549
+ // document's meta (diagnostics are discarded — spans never report).
550
+ function idOfHeading(braces, text, line, ctx) {
551
+ return (braces ? parseAttrs(braces).id : undefined) ?? slug(interpolate(text, line, ctx));
552
+ }
553
+ // The matching close of the fence opened at lines[i] (equal-length run, or the
554
+ // labeled `=== #id` close when the block carries an id): the index just past
555
+ // the close line, and whether one was found — an unterminated block runs to
556
+ // the end of the scope.
557
+ function fenceClose(lines, i, open) {
558
+ const openLen = open[1].length;
559
+ const id = open[3] ? parseAttrs(open[3]).id : undefined;
560
+ const labeled = id !== undefined ? new RegExp(`^={3,}[ \\t]+#${reLit(id)}[ \\t]*$`) : null;
561
+ for (let j = i + 1; j < lines.length; j++) {
562
+ if (isCloseFence(lines[j], openLen) || (labeled && labeled.test(lines[j])))
563
+ return { end: j + 1, closed: true };
564
+ }
565
+ return { end: lines.length, closed: false };
566
+ }
567
+ // A heading's span covers its whole SECTION: the heading line through the line
568
+ // just before the next heading of same-or-higher level (fewer-or-equal `#`) in
569
+ // the current scope, or end-of-scope. Fenced blocks are skipped whole — a `#`
570
+ // line inside a `=== code` body is content, never a boundary.
571
+ function sectionEnd(lines, i, level) {
572
+ let j = i + 1;
573
+ while (j < lines.length) {
574
+ const open = FENCE_OPEN.exec(lines[j]);
575
+ if (open) {
576
+ j = fenceClose(lines, j, open).end;
577
+ continue;
578
+ }
579
+ const h = HEADING.exec(lines[j]);
580
+ if (h && h[1].length <= level)
581
+ return j;
582
+ j++;
583
+ }
584
+ return lines.length;
458
585
  }
459
586
  // Walk `lines` exactly as scanBlocks does — same fence close rules (equal-length
460
587
  // or labeled `=== #id`), same flow-only recursion via REGISTRY — recording the
@@ -462,7 +589,7 @@ function idOfHeading(braces, text) {
462
589
  // First definition wins, mirroring ctx.ids (a duplicate id is a build error, so
463
590
  // `get`/`set` operate on the one the parser actually registered). `base` is the
464
591
  // absolute line offset of this slice within the whole document.
465
- function collectSpans(lines, base, out) {
592
+ function collectSpans(lines, base, out, ctx, depth = 0) {
466
593
  const add = (id, start, end) => {
467
594
  if (!out.has(id))
468
595
  out.set(id, { start, end });
@@ -486,33 +613,27 @@ function collectSpans(lines, base, out) {
486
613
  } // hidden line: no id
487
614
  const open = FENCE_OPEN.exec(line);
488
615
  if (open) {
489
- const openLen = open[1].length;
490
616
  const type = open[2];
491
617
  const id = open[3] ? parseAttrs(open[3]).id : undefined;
492
- const labeled = id !== undefined ? new RegExp(`^={3,}[ \\t]+#${id}[ \\t]*$`) : null;
493
- let j = i + 1;
494
- let closed = false;
495
- for (; j < lines.length; j++) {
496
- if (isCloseFence(lines[j], openLen) || (labeled && labeled.test(lines[j]))) {
497
- closed = true;
498
- break;
499
- }
500
- }
501
- const end = closed ? j + 1 : j;
618
+ const { end, closed } = fenceClose(lines, i, open);
502
619
  if (id !== undefined)
503
620
  add(id, base + i, base + end);
504
621
  // Only a flow body is scanned for nested blocks (raw/data bodies are
505
622
  // opaque), so an id inside a `code` body is *not* addressable — exactly
506
623
  // the parser's contract.
507
- if ((REGISTRY[type] ?? "raw") === "flow") {
508
- collectSpans(lines.slice(i + 1, closed ? j : end), base + i + 1, out);
624
+ if ((REGISTRY[type] ?? "raw") === "flow" && depth < MAX_NESTING) {
625
+ collectSpans(lines.slice(i + 1, closed ? end - 1 : end), base + i + 1, out, ctx, depth + 1);
509
626
  }
510
627
  i = end;
511
628
  continue;
512
629
  }
513
630
  const h = HEADING.exec(line);
514
631
  if (h) {
515
- add(idOfHeading(h[3], h[2]), base + i, base + i + 1);
632
+ // Section span (heading through its prose and nested blocks). The walk
633
+ // still advances one line at a time so every nested id inside the
634
+ // section registers its own span — spans intentionally OVERLAP: #sec
635
+ // contains #code, and each remains addressable on its own.
636
+ add(idOfHeading(h[3], h[2], base + i + 1, ctx), base + i, base + sectionEnd(lines, i, h[1].length));
516
637
  i++;
517
638
  continue;
518
639
  }
@@ -523,29 +644,49 @@ function collectSpans(lines, base, out) {
523
644
  // with the physical lines produced by splitLines(source).
524
645
  export function blockSpans(source) {
525
646
  const out = new Map();
526
- collectSpans(source.replace(/\r\n?/g, "\n").split("\n"), 0, out);
647
+ const lines = source.replace(/\r\n?/g, "\n").split("\n");
648
+ // Inert context: heading auto-ids slug the interpolated text (parser parity);
649
+ // its diagnostics are discarded — the span scan never reports.
650
+ const ctx = { diags: [], ids: new Map(), refs: [], meta: collectMeta(lines) };
651
+ collectSpans(lines, 0, out, ctx);
527
652
  return out;
528
653
  }
529
654
  // Split into physical lines while *keeping* each line's terminator, so
530
655
  // join("") is byte-exact and slicing by span never rewrites line endings.
656
+ // A line ends at `\n` or at a LONE `\r` (old-Mac style) — the same boundaries
657
+ // the span scan's `\r\n?` -> `\n` normalization sees, so span indices always
658
+ // address the same lines the parser counted.
531
659
  function splitLines(source) {
532
- return source.split(/(?<=\n)/);
660
+ return source.split(/(?<=\n|\r(?!\n))/);
661
+ }
662
+ // `--head`: narrow any id's span to its HEAD line — the single declaring line
663
+ // (a heading's `# … {#id}` line, a typed block's opening fence, a footnote's
664
+ // `[^id]:` line). The head is by construction the FIRST line of the span, so
665
+ // the narrowing is parse-free and needs no type check. Main use: `set --head`
666
+ // edits a block's attributes (caption/compute/lang/…) without re-sending its
667
+ // body, or renames a heading without rewriting its section.
668
+ function narrowToHead(span) {
669
+ return { start: span.start, end: span.start + 1 };
533
670
  }
534
671
  // Depth-first search for the document-model node carrying `id`, descending into
535
672
  // flow-block children (and list-item children) so a nested id is found too.
536
- function findBlockById(blocks, id) {
537
- for (const b of blocks) {
673
+ // Returns the containing sibling array and index, not just the node: the model
674
+ // is FLAT a heading does not own its section; the section's prose and blocks
675
+ // are its FOLLOWING SIBLINGS — so a section consumer needs the array.
676
+ function findBlockSite(blocks, id) {
677
+ for (let i = 0; i < blocks.length; i++) {
678
+ const b = blocks[i];
538
679
  if ((b.kind === "heading" || b.kind === "block") && b.id === id)
539
- return b;
680
+ return { siblings: blocks, index: i };
540
681
  if (b.kind === "block" && b.children) {
541
- const hit = findBlockById(b.children, id);
682
+ const hit = findBlockSite(b.children, id);
542
683
  if (hit)
543
684
  return hit;
544
685
  }
545
686
  if (b.kind === "list") {
546
687
  for (const it of b.items) {
547
688
  if (it.children) {
548
- const hit = findBlockById(it.children, id);
689
+ const hit = findBlockSite(it.children, id);
549
690
  if (hit)
550
691
  return hit;
551
692
  }
@@ -554,6 +695,21 @@ function findBlockById(blocks, id) {
554
695
  }
555
696
  return undefined;
556
697
  }
698
+ // Model-side section boundary: within one sibling array, the section opened by
699
+ // the heading at index k runs to the next sibling heading of same-or-higher
700
+ // level, or the array end. This is the SAME rule sectionEnd() applies to raw
701
+ // source lines (where skipping fenced bodies makes "next heading" well-defined)
702
+ // — the two sides must stay in lockstep; the get-set suite pins their parity
703
+ // (ids covered by the raw slice == ids covered by the --json envelope).
704
+ function sectionEndIndex(siblings, k) {
705
+ const level = siblings[k].level;
706
+ for (let m = k + 1; m < siblings.length; m++) {
707
+ const b = siblings[m];
708
+ if (b.kind === "heading" && b.level <= level)
709
+ return m;
710
+ }
711
+ return siblings.length;
712
+ }
557
713
  // ---------------------------------------------------------------------------
558
714
  // CLI
559
715
  // ---------------------------------------------------------------------------
@@ -572,45 +728,69 @@ function parseStamp(s) {
572
728
  return new Date(Date.UTC(+y, +mo - 1, +d, +h, +mi, +se));
573
729
  }
574
730
  const VERSION = "1.0"; // GEML spec version this CLI targets
575
- const PARSER_VERSION = "1.1.1"; // reference implementation; keep in sync with package.json
576
- const USAGE = `geml — GEML reference CLI
577
-
578
- Usage:
579
- geml <file.geml|-> parse -> document-model JSON (stdout)
580
- geml get <file.geml|-> #id [--json] print ONE block by id (raw span, or --json node)
581
- geml set <file.geml|-> #id [--from f][-o f] replace ONE block by id (new content: --from/stdin)
582
- geml revert <file.geml> #id [--to <sel>] restore ONE block to a past revision (sel: -N|latest|id)
583
- geml check <file.geml|-> [--json] validate only: diagnostics + exit code
584
- geml render <file.geml|-> [-o out.html] render to one self-contained HTML file
585
- geml fmt <file.geml|-> [-o out.geml] re-serialize to canonical GEML
586
- geml convert <file.md|-> [-o out.geml] Markdown -> GEML
587
- geml export <file.geml|-> [-o out.md] GEML -> Markdown (lossy)
588
- geml history <commit|verify|show|restore|log> <file.geml> [...]
589
- geml codemap <build|verify|render|serve|refresh|mcp> [...] code-graph toolkit (geml codemap --help)
590
- geml --help | --version [--json]
591
-
592
- Use '-' as the file to read from stdin.
593
- Exit codes: 0 ok · 1 document/operation error · 2 usage error.`;
731
+ const PARSER_VERSION = "1.4.2"; // reference implementation; keep in sync with package.json
732
+ const USAGE = `geml — GEML reference CLI
733
+
734
+ Usage:
735
+ geml <file.geml|-> [--to <fmt>] [--from <fmt>] [-o out] transform a document (default: --to json)
736
+ <fmt>: json | html | md | geml
737
+ --to md -> Markdown (lossy)
738
+ --to html -> self-contained HTML
739
+ --to geml -> canonical re-format
740
+ --to json -> document-model JSON (default)
741
+ A Markdown input converts the other way:
742
+ geml notes.md -> GEML
743
+ --from overrides the input format (any input):
744
+ geml notes.txt --from md treat as Markdown
745
+ geml - --from md read Markdown on stdin
746
+ geml get <file.geml|-> [#id] [--json] [--head] with #id: print that block
747
+ (a heading id = its whole section; --head = head line;
748
+ --json = model node). Without #id: list all addressable
749
+ ids (--json = array).
750
+ geml set <file.geml|-> #id [--head|--body] [--in f[#src]|-] [-o f] replace ONE block by id
751
+ (--in F takes F's block #id, F#src takes #src, else stdin raw;
752
+ default = whole block · --head = head line · --body = body)
753
+ geml add <file.geml|-> (--append | --before #id | --after #id) [--in f[#src]|-] [-o f] insert a fragment
754
+ (1+ blocks and/or prose; content keeps its own ids, a clash is refused)
755
+ geml delete <file.geml|-> #id [#id2 …] [-o f] remove one or more blocks
756
+ (a missing id is skipped; a dangling reference is a warning, not a refusal)
757
+ geml rename <file.geml|-> #old #new [-o f] rename an id and every reference to it (id-boundary safe)
758
+ geml revert <file.geml> #id [--rev <sel>] [--head] undo one block to a past revision (splice / resurrect / remove)
759
+ (sel: -N | latest | id-prefix; default -1)
760
+ geml check <file.geml|-> [--root d] [--json] validate only: diagnostics + exit code
761
+ (--root widens cross-doc refs to dir d, e.g. the repo root)
762
+ geml history <commit|verify|show|restore|log> <file.geml> [...] .gemlhistory version sidecar
763
+ geml codemap <build|verify|render|serve|refresh|find|mcp> [...] code-graph toolkit (alias: codegraph)
764
+ geml --help | --version [--json]
765
+
766
+ Use '-' as the file to read from stdin.
767
+ Mutations (set/add/delete/rename) write the whole updated document in place for a
768
+ file, or to stdout for '-' input; -o redirects it (-o - = stdout).
769
+ Exit codes:
770
+ 0 ok
771
+ 1 document/operation error
772
+ 2 command usage error.
773
+ `;
594
774
  // One-line usage for each subcommand — the single source for both the error
595
775
  // shown on misuse and the `<cmd> --help` text.
596
776
  const SUBHELP = {
597
- get: "usage: geml get <file.geml|-> #id [--json]",
598
- set: "usage: geml set <file.geml|-> #id [--from FILE] [-o out.geml]",
599
- check: "usage: geml check <file.geml|-> [--json]",
600
- render: "usage: geml render <file.geml|-> [-o out.html]",
601
- convert: "usage: geml convert <file.md|-> [-o out.geml]",
602
- export: "usage: geml export <file.geml|-> [-o out.md]",
603
- fmt: "usage: geml fmt <file.geml|-> [-o out.geml]",
604
- revert: "usage: geml revert <file.geml> #id [--to <sel>] [--changed] [--dry-run] [-o out] (sel: -N | latest | id-prefix; default -1)",
777
+ get: "usage: geml get <file.geml|-> [#id] [--json] [--head] (with #id: that block, a heading id = its whole section, --head = its head line; without #id: list every addressable id, --json = array)",
778
+ set: "usage: geml set <file.geml|-> #id [--head|--body] [--in F | --in F#src | --in -] [-o out.geml] (content: --in F takes F's block #id, --in F#src takes #src, else stdin raw; default = whole block, --head = head line — both normalize the id to #id — --body = body; guarded splice, refused if it breaks the doc)",
779
+ add: "usage: geml add <file.geml|-> (--append | --before #id | --after #id) [--in F | --in F#src | --in -] [-o out.geml] (insert a GEML fragment — 1+ blocks and/or prose — at a position; --in F takes all of F, --in F#src takes #src, else stdin raw; content keeps its own ids, a collision is refused)",
780
+ delete: "usage: geml delete <file.geml|-> #id [#id2 …] [-o out.geml] (remove one or more blocks; a missing id is skipped with a note, not an error; a reference left dangling is a warning, not a refusal — delete never fails on a live reference)",
781
+ rename: "usage: geml rename <file.geml|-> #old #new [-o out.geml] (rewrite an id's declaration AND every reference — [[#id]], [text](#id), chart data=#id, footnote [^id] — id-boundary safe, skipping raw block bodies; #new must be free; refused if it breaks the doc)",
782
+ check: "usage: geml check <file.geml|-> [--root <dir>] [--json] (--root: resolve cross-doc refs within <dir> instead of the file's own directory)",
783
+ revert: "usage: geml revert <file.geml> #id [--rev <sel>] [--changed] [--append|--before #x|--after #x] [--head] [--dry-run] [-o out] (reconcile #id to a revision: splice / resurrect / remove; sel: -N | latest | id-prefix; default -1)",
605
784
  history: "usage: geml history <commit|verify|show|restore|log> <file.geml> [...]",
606
- codemap: `usage: geml codemap build --root <repo> # auto-detect languages, run the indexer(s), and merge into one codemap
607
- geml codemap build (--db <graph.db> | --adapter joern|scip --raw <in>)+ --root <repo> [--out .geml-code-graph] [--container module|dir|file] [--lang <JAVASRC|NEWC|…>] [--joern <path>] [--history [-m msg]]
608
- geml codemap verify [dir] geml check + profile reference checks
609
- geml codemap render [dir] every doc -> sibling .html (open index.html from disk)
610
- geml codemap serve [dir] [--port 8140] [--watch] [--background|--stop] live viewer: pages render from .geml on request; --watch re-runs the recipe when sources change
611
- geml codemap refresh [dir] [--force] [--commit] [--background|--hook] re-run the recorded build recipe (_index/refresh.json); --commit lands it as its own commit
612
- geml codemap mcp stdio MCP server (GEML_GRAPH_DIR or graph_dir arg)
613
- (<dir> for verify/render/serve/refresh defaults to ./.geml-code-graph)`,
785
+ codemap: `usage: geml codemap build [--root <repo>] # auto-detect languages, run the indexer(s), and merge into one codemap (--root defaults to the current directory)
786
+ geml codemap build (--db <graph.db> | --adapter joern|scip --raw <in>)+ [--root <repo>] [--out .geml-code-graph] [--container module|dir|file] [--lang <JAVASRC|NEWC|…>] [--joern <path>] [--history [-m msg]]
787
+ geml codemap verify [dir] geml check + profile reference checks
788
+ geml codemap render [dir] every doc -> sibling .html (open index.html from disk)
789
+ geml codemap serve [dir] [--port 8140] [--watch] [--background|--stop] live viewer: pages render from .geml on request; --watch re-runs the recipe when sources change
790
+ geml codemap refresh [dir] [--force] [--commit] [--background|--hook] re-run the recorded build recipe (_index/refresh.json); --commit lands it as its own commit
791
+ geml codemap find <name> [dir] locate a symbol by substring name -> doc#id + src (stdout, no browser)
792
+ geml codemap mcp stdio MCP server (GEML_GRAPH_DIR or graph_dir arg)
793
+ (<dir> for verify/render/serve/refresh/find defaults to ./.geml-code-graph; codegraph and code-graph are accepted as aliases of codemap)`,
614
794
  };
615
795
  // Set from argv at dispatch time; when true, errors are emitted as a JSON
616
796
  // envelope so an agent that standardizes on --json never has to parse text.
@@ -634,12 +814,66 @@ function readInput(file) {
634
814
  fail(file === "-" ? "cannot read stdin" : `cannot read ${file}`);
635
815
  }
636
816
  }
637
- // A cross-document resolver rooted at the input's directory (cwd for stdin).
638
- function resolverFor(file) {
639
- const baseDir = file === "-" ? "." : dirname(file);
817
+ // A cross-document resolver rooted at the input's directory (cwd for stdin),
818
+ // CONFINED to that directory's subtree. A reference that resolves outside the
819
+ // base via a `..` escape, an absolute path, or (on Windows) a different drive
820
+ // — is refused (returns null, i.e. an unresolvable ref) so a crafted document
821
+ // cannot turn `geml check`/parse into an arbitrary local-file read oracle. §8.
822
+ //
823
+ // A purely LEXICAL check is not enough: a symlink that sits lexically inside the
824
+ // subtree but points to `../../outside.geml` passes `path.relative` yet reads an
825
+ // external target. So after the cheap lexical gate we resolve BOTH the base and
826
+ // the target through `realpathSync` (following every symlink component) and
827
+ // re-check that the REAL target still lies within the REAL base subtree before
828
+ // reading. A target that does not exist makes `realpathSync` throw — handled as
829
+ // an ordinary unresolvable ref (null), never a crash.
830
+ //
831
+ // `root` (CLI `--root`, an explicit per-invocation user grant — never
832
+ // document-controlled) widens the confinement base from the input's own
833
+ // directory to an ancestor the user names, so repo-relative `../` references
834
+ // between sibling directories can be checked. It moves WHERE the boundary
835
+ // stands, never whether it is enforced: both gates below run against the
836
+ // widened base, so escapes past the root are refused exactly as above. The
837
+ // viewer/web surfaces never pass a root — their boundary is unchanged.
838
+ function resolverFor(file, root) {
839
+ const dirAbs = resolvePath(file === "-" ? "." : dirname(file));
840
+ const baseAbs = root === undefined ? dirAbs : resolvePath(root);
841
+ // Canonicalise the base once. If the base itself cannot be realpath'd, no
842
+ // cross-doc ref can be safely confined — resolve nothing.
843
+ let realBase = null;
844
+ try {
845
+ realBase = realpathSync(baseAbs);
846
+ }
847
+ catch {
848
+ realBase = null;
849
+ }
850
+ const outside = (from, to) => {
851
+ const rel = relative(from, to);
852
+ return rel === ".." || rel.startsWith(".." + sep) || isAbsolute(rel);
853
+ };
640
854
  return (d) => {
855
+ if (realBase === null)
856
+ return null;
857
+ // References resolve FROM the document's own directory; the gates below
858
+ // confine them to the (possibly widened) base.
859
+ const targetAbs = resolvePath(dirAbs, d);
860
+ // Cheap lexical gate: reject an obvious `..`/absolute/other-drive escape
861
+ // before touching the filesystem.
862
+ if (outside(baseAbs, targetAbs))
863
+ return null;
864
+ // Real (symlink-resolved) gate: a symlink pointing out of the subtree
865
+ // resolves to a real path outside `realBase` and is refused here.
866
+ let realTarget;
867
+ try {
868
+ realTarget = realpathSync(targetAbs);
869
+ }
870
+ catch {
871
+ return null;
872
+ }
873
+ if (outside(realBase, realTarget))
874
+ return null;
641
875
  try {
642
- return readFileSync(resolvePath(baseDir, d), "utf8");
876
+ return readFileSync(realTarget, "utf8");
643
877
  }
644
878
  catch {
645
879
  return null;
@@ -650,10 +884,22 @@ function resolverFor(file) {
650
884
  // dump (cheap for agents). `--json` prints the diagnostics array for machines.
651
885
  function runCheck(args) {
652
886
  const json = args.includes("--json");
653
- const file = args.find((a) => a === "-" || !a.startsWith("-"));
887
+ const root = flag(args, "--root");
888
+ const file = args.find((a) => a === "-" || (!a.startsWith("-") && a !== root));
654
889
  if (!file)
655
890
  fail(SUBHELP.check);
656
- const doc = parse(readInput(file), { resolveDoc: resolverFor(file) });
891
+ // A mistyped --root must be a usage error (exit 2), not a wall of misleading
892
+ // "cannot resolve document" errors from a resolver confined to nothing.
893
+ if (root !== undefined) {
894
+ let isDir = false;
895
+ try {
896
+ isDir = statSync(root).isDirectory();
897
+ }
898
+ catch { /* missing -> not a dir */ }
899
+ if (!isDir)
900
+ fail(`--root ${root} is not a directory`);
901
+ }
902
+ const doc = parse(readInput(file), { resolveDoc: resolverFor(file, root) });
657
903
  if (json) {
658
904
  console.log(JSON.stringify(doc.diagnostics, null, 2));
659
905
  }
@@ -737,39 +983,92 @@ function runHistory(args) {
737
983
  fail(historyError(e, file, historyPath));
738
984
  }
739
985
  }
740
- // `geml convert <file.md|-> [-o out.geml]` — Markdown -> GEML.
741
- function runConvert(args) {
742
- const file = args.find((a) => a === "-" || (!a.startsWith("-") && a !== flag(args, "-o")));
986
+ function runTransform(argv) {
987
+ const out = flag(argv, "-o") ?? flag(argv, "--out");
988
+ const fromRaw = flag(argv, "--from");
989
+ const toRaw = flag(argv, "--to");
990
+ const [file] = positionals(argv, ["-o", "--out", "--from", "--to"]);
743
991
  if (!file)
744
- fail(SUBHELP.convert);
745
- const { geml, notes } = mdToGeml(readInput(file));
746
- for (const n of notes)
747
- console.error(`note: ${n}`);
748
- const outPath = flag(args, "-o") ?? flag(args, "--out");
749
- if (outPath) {
750
- writeFileSync(outPath, geml);
751
- console.error(`wrote ${outPath}`);
992
+ fail("no input file (use '-' to read from stdin)", 2);
993
+ // A bare `--to`/`--from` (no following value) is a mistyped flag, not a
994
+ // silent fall-through to the default — flag() would return undefined and we
995
+ // must not quietly ignore it.
996
+ if (argv.includes("--from") && fromRaw === undefined)
997
+ fail("--from needs a format (geml | md)", 2);
998
+ if (argv.includes("--to") && toRaw === undefined)
999
+ fail("--to needs a format (json | html | md | geml)", 2);
1000
+ // Input format: an explicit --from wins (for any input, file or stdin), else
1001
+ // the file extension, else GEML (covers .geml, unknown extensions, and stdin).
1002
+ let inFmt;
1003
+ if (fromRaw !== undefined) {
1004
+ if (fromRaw !== "geml" && fromRaw !== "md") {
1005
+ fail(`--from: unknown input format '${fromRaw}' (want geml | md)`, 2);
1006
+ }
1007
+ inFmt = fromRaw;
1008
+ }
1009
+ else if (/\.(md|markdown)$/i.test(file)) {
1010
+ inFmt = "md";
752
1011
  }
753
1012
  else {
754
- process.stdout.write(geml);
1013
+ inFmt = "geml";
755
1014
  }
756
- }
757
- // `geml export <file.geml|-> [-o out.md]` — GEML -> Markdown (lossy). Writes
758
- // the output even with diagnostics, prints any lossy-projection notes, and
759
- // exits non-zero on a parse error same contract as render.
760
- function runExport(args) {
761
- const out = flag(args, "-o") ?? flag(args, "--out");
762
- const file = args.find((a) => a === "-" || (!a.startsWith("-") && a !== out));
763
- if (!file)
764
- fail(SUBHELP.export);
765
- const doc = parse(readInput(file), { resolveDoc: resolverFor(file) });
766
- const { md, notes } = gemlToMd(doc);
767
- if (out) {
768
- writeFileSync(out, md);
769
- console.error(`wrote ${out}`);
1015
+ // Output format: an explicit --to wins, else md input -> geml, geml -> json.
1016
+ let outFmt;
1017
+ if (toRaw !== undefined) {
1018
+ if (toRaw !== "json" && toRaw !== "html" && toRaw !== "md" && toRaw !== "geml") {
1019
+ fail(`--to: unknown output format '${toRaw}' (want json | html | md | geml)`, 2);
1020
+ }
1021
+ outFmt = toRaw;
770
1022
  }
771
- else
772
- process.stdout.write(md);
1023
+ else {
1024
+ outFmt = inFmt === "md" ? "geml" : "json";
1025
+ }
1026
+ const src = readInput(file);
1027
+ // md -> geml is a direct projection, not a parse/serialize round-trip: emit
1028
+ // the converter's GEML verbatim (the old `convert`; no diagnostics to raise).
1029
+ if (inFmt === "md" && outFmt === "geml") {
1030
+ const { geml, notes } = mdToGeml(src);
1031
+ writeOut(geml, out);
1032
+ for (const n of notes)
1033
+ console.error(`note: ${n}`);
1034
+ return;
1035
+ }
1036
+ // Otherwise load a document — a md input is converted to GEML first — and
1037
+ // project it to the target.
1038
+ let notes = [];
1039
+ let doc;
1040
+ if (inFmt === "md") {
1041
+ const conv = mdToGeml(src);
1042
+ notes = conv.notes;
1043
+ doc = parse(conv.geml, { resolveDoc: resolverFor(file) });
1044
+ }
1045
+ else {
1046
+ doc = parse(src, { resolveDoc: resolverFor(file) });
1047
+ }
1048
+ let output;
1049
+ switch (outFmt) {
1050
+ case "json":
1051
+ output = JSON.stringify(doc, null, 2) + "\n"; // == the former bare parse
1052
+ break;
1053
+ case "geml":
1054
+ output = serialize(doc); // == the former `fmt`
1055
+ break;
1056
+ case "html":
1057
+ output = renderHtml(doc, {
1058
+ source: file === "-" ? "stdin" : basename(file),
1059
+ // geml-code-graph embeds load + parse sibling codemap docs on demand.
1060
+ loadDoc: resolverFor(file),
1061
+ parseDoc: (s) => parse(s),
1062
+ });
1063
+ break;
1064
+ case "md": {
1065
+ const r = gemlToMd(doc); // == the former `export`
1066
+ notes = notes.concat(r.notes);
1067
+ output = r.md;
1068
+ break;
1069
+ }
1070
+ }
1071
+ writeOut(output, out);
773
1072
  for (const n of notes)
774
1073
  console.error(`note: ${n}`);
775
1074
  for (const d of doc.diagnostics)
@@ -777,55 +1076,35 @@ function runExport(args) {
777
1076
  if (doc.diagnostics.some((d) => d.severity === "error"))
778
1077
  process.exit(1);
779
1078
  }
780
- // `geml render <file.geml> [-o out.html]` GEML -> one self-contained,
781
- // interactive HTML artifact (the P0 runtime). Writes the file even when there
782
- // are diagnostics (a viewer should still show what it can), but exits non-zero
783
- // on any error so CI and agents get a hard signal.
784
- function runRender(args) {
785
- const out = flag(args, "-o") ?? flag(args, "--out");
786
- const file = args.find((a) => a === "-" || (!a.startsWith("-") && a !== out));
787
- if (!file)
788
- fail(SUBHELP.render);
789
- const doc = parse(readInput(file), { resolveDoc: resolverFor(file) });
790
- const html = renderHtml(doc, {
791
- source: file === "-" ? "stdin" : basename(file),
792
- // geml-code-graph embeds load + parse sibling codemap documents on demand.
793
- loadDoc: resolverFor(file),
794
- parseDoc: (s) => parse(s),
795
- });
796
- if (out) {
797
- writeFileSync(out, html);
798
- console.error(`wrote ${out}`);
799
- }
800
- else
801
- process.stdout.write(html);
802
- for (const d of doc.diagnostics)
803
- console.error(`${d.severity}: ${d.message} (line ${d.line})`);
804
- if (doc.diagnostics.some((d) => d.severity === "error"))
805
- process.exit(1);
806
- }
807
- // `geml fmt <file.geml> [-o out.geml]` — re-serialize the document model into
808
- // canonical GEML. Because `serialize` is the inverse of `parse`, `fmt` is a
809
- // pretty-printer whose output parses back to the same model (round-trip stable).
810
- function runFmt(args) {
811
- const out = flag(args, "-o") ?? flag(args, "--out");
812
- const file = args.find((a) => a === "-" || (!a.startsWith("-") && a !== out));
813
- if (!file)
814
- fail(SUBHELP.fmt);
815
- const doc = parse(readInput(file), { resolveDoc: resolverFor(file) });
816
- const text = serialize(doc);
1079
+ // Write to `-o out` (with a `wrote` note on stderr) or to stdout.
1080
+ function writeOut(text, out) {
817
1081
  if (out) {
818
1082
  writeFileSync(out, text);
819
1083
  console.error(`wrote ${out}`);
820
1084
  }
821
1085
  else
822
1086
  process.stdout.write(text);
823
- // A broken document must not be reported as a clean format. Surface the
824
- // diagnostics and exit non-zero, matching parse/render/check.
825
- for (const d of doc.diagnostics)
826
- console.error(`${d.severity}: ${d.message} (line ${d.line})`);
827
- if (doc.diagnostics.some((d) => d.severity === "error"))
828
- process.exit(1);
1087
+ }
1088
+ // Output-target rule shared by the MUTATION verbs (set, and — soon — add,
1089
+ // delete, rename, revert): a real file input with no `-o` is edited IN PLACE
1090
+ // (it's the obvious target, and it's what lets an agent chain edits without
1091
+ // re-reading a path back out of stdout); stdin (`file === "-"`) has no such
1092
+ // target, so it falls back to stdout. `-o` always wins when given: `-o -`
1093
+ // explicitly requests stdout (even for a file input), `-o <path>` writes
1094
+ // there. Every write announces itself with `wrote <path>` on stderr; stdout
1095
+ // stays reserved for the document bytes so it's still pipeable.
1096
+ function resolveOutTarget(file, oFlag) {
1097
+ const toFile = (path) => ({
1098
+ write(text) { writeFileSync(path, text); console.error(`wrote ${path}`); },
1099
+ });
1100
+ const toStdout = { write(text) { process.stdout.write(text); } };
1101
+ if (oFlag === "-")
1102
+ return toStdout;
1103
+ if (oFlag !== undefined)
1104
+ return toFile(oFlag);
1105
+ if (file === "-")
1106
+ return toStdout;
1107
+ return toFile(file);
829
1108
  }
830
1109
  // Positional args (a file, an id) are the non-flag tokens that aren't the value
831
1110
  // of a value-taking flag. `-` (stdin) is a positional, not a flag. An id may be
@@ -849,69 +1128,503 @@ function positionals(args, valued) {
849
1128
  }
850
1129
  return out;
851
1130
  }
1131
+ // `geml get <file>` with no id: list every addressable id — the document's
1132
+ // table of contents. Default output is one id per line with its kind (and, for
1133
+ // a heading, its level and text); `--json` is a machine-readable array so an
1134
+ // agent can pick its next `get #id` target. Ids are listed in document order
1135
+ // (the registration order parse() records), covering the same set `get #id`
1136
+ // resolves against: typed blocks, headings, and footnote definitions.
1137
+ function listIds(source, file, json) {
1138
+ const doc = parse(source, { resolveDoc: resolverFor(file) });
1139
+ const rows = doc.ids.map((id) => {
1140
+ const site = findBlockSite(doc.children, id);
1141
+ const b = site?.siblings[site.index];
1142
+ if (b?.kind === "heading")
1143
+ return { id, kind: "heading", level: b.level, text: b.text };
1144
+ if (b?.kind === "block") {
1145
+ const row = { id, kind: b.type };
1146
+ if (b.classes.includes("footnote"))
1147
+ row.footnote = true; // §5.2 footnote definition
1148
+ return row;
1149
+ }
1150
+ return { id, kind: b?.kind ?? "unknown" };
1151
+ });
1152
+ if (json) {
1153
+ console.log(JSON.stringify(rows, null, 2));
1154
+ return;
1155
+ }
1156
+ if (rows.length === 0) {
1157
+ console.error(`no addressable ids in ${file === "-" ? "stdin" : file}`);
1158
+ return;
1159
+ }
1160
+ // Align the id and kind columns; append a heading's level+text or a footnote flag.
1161
+ const idW = Math.max(...rows.map((r) => r.id.length + 1));
1162
+ const kindW = Math.max(...rows.map((r) => r.kind.length));
1163
+ for (const r of rows) {
1164
+ let line = `#${r.id}`.padEnd(idW + 1) + " " + r.kind.padEnd(kindW);
1165
+ if (r.kind === "heading")
1166
+ line += ` h${r.level} ${r.text}`;
1167
+ else if (r.footnote)
1168
+ line += " footnote";
1169
+ console.log(line.replace(/\s+$/, ""));
1170
+ }
1171
+ }
852
1172
  // `geml get <file.geml|-> #id [--json]` — print ONE block, addressed by id,
853
1173
  // without loading the rest of the document into context. Default output is the
854
- // block's exact source bytes (its full `=== … ===` span, or the source line for
855
- // a heading/footnote); `--json` prints that block's document-model node.
1174
+ // block's exact source bytes: a typed block's full `=== … ===` span, a
1175
+ // footnote's line, or — for a heading its whole SECTION (heading line through
1176
+ // the line before the next same-or-higher heading). `--json` covers the same
1177
+ // content: a block/footnote id prints its document-model node; a heading id
1178
+ // prints a section envelope `{kind:"section", id, level, blocks:[heading,
1179
+ // …siblings up to the boundary]}`.
856
1180
  function runGet(args) {
857
1181
  const json = args.includes("--json");
1182
+ const headOnly = args.includes("--head");
858
1183
  const [file, rawId] = positionals(args, []);
859
- if (!file || !rawId)
1184
+ if (!file)
860
1185
  fail(SUBHELP.get);
1186
+ // No id: list every addressable id — the document's "table of contents", so
1187
+ // an agent can discover what `get #id` can target without pulling the model.
1188
+ if (!rawId) {
1189
+ listIds(readInput(file), file, json);
1190
+ return;
1191
+ }
861
1192
  const id = rawId.replace(/^#/, "");
862
1193
  const source = readInput(file);
863
1194
  if (json) {
864
- // The model node — same shape `geml <file>` emits for it. Parsing is needed
865
- // to resolve the tree (and nested-block ids), but only the one node prints.
1195
+ // The model node(s) — same shapes `geml <file>` emits. Parsing is needed
1196
+ // to resolve the tree (and nested-block ids), but only the target prints.
866
1197
  const doc = parse(source, { resolveDoc: resolverFor(file) });
867
- const block = findBlockById(doc.children, id);
868
- if (!block)
1198
+ const site = findBlockSite(doc.children, id);
1199
+ if (!site)
869
1200
  fail(`no block with id \`${id}\``, 1);
1201
+ const block = site.siblings[site.index];
1202
+ // `--head` on a heading suppresses the section envelope (the lone heading
1203
+ // node IS the head). On a block/footnote id there is nothing finer than
1204
+ // the single node — the model has no sub-node for "just the fence line" —
1205
+ // so --head refines only the RAW output there.
1206
+ if (block.kind === "heading" && !headOnly) {
1207
+ // A heading id addresses its SECTION, so `--json` covers the same
1208
+ // content as the raw span: a self-describing envelope whose blocks[0]
1209
+ // is the heading node followed by its siblings up to the boundary.
1210
+ // `kind: "section"` lets a consumer branch — a block/footnote id still
1211
+ // yields the single model node (the model itself stays flat).
1212
+ const end = sectionEndIndex(site.siblings, site.index);
1213
+ console.log(JSON.stringify({ kind: "section", id: block.id, level: block.level, blocks: site.siblings.slice(site.index, end) }, null, 2));
1214
+ return;
1215
+ }
870
1216
  console.log(JSON.stringify(block, null, 2));
871
1217
  return;
872
1218
  }
873
1219
  // Raw: slice the source span byte-for-byte. No parse required, so `get` still
874
1220
  // returns the exact bytes even if the document has diagnostics elsewhere.
875
- const span = blockSpans(source).get(id);
876
- if (!span)
1221
+ const found = blockSpans(source).get(id);
1222
+ if (!found)
877
1223
  fail(`no block with id \`${id}\``, 1);
1224
+ const span = headOnly ? narrowToHead(found) : found;
878
1225
  process.stdout.write(splitLines(source).slice(span.start, span.end).join(""));
879
1226
  }
880
- // `geml set <file.geml|-> #id [--from FILE] [-o out]` replace ONLY that
881
- // block's source span with new content (from --from or stdin), preserving every
882
- // other byte. Prints the full updated document, or writes in place with -o. The
883
- // splice is re-parsed and rejected if it broke the doc: `set` never writes a
884
- // corrupt file.
1227
+ const NO_CONTENT = "no replacement content (use --in FILE or pipe it on stdin)";
1228
+ // `geml set <file.geml|-> #id [--head|--body] [--in F|F#src|-] [-o out]`
1229
+ // replace ONE existing block, addressed by #id, with new content, preserving
1230
+ // every other byte. Two content CHANNELS × three MODES:
1231
+ //
1232
+ // channels · `--in F[#src]` extracts a BLOCK from GEML file F (F is always
1233
+ // read as GEML — extension ignored, no md conversion): `--in F`
1234
+ // takes the block whose id == the target #id; `--in F#src` takes
1235
+ // #src. stdin (default, or `--in -`) is raw bytes.
1236
+ // modes · default replaces the WHOLE block, `--head` only the head line,
1237
+ // `--body` only the body. Default and `--head` NORMALIZE the
1238
+ // content's id to #id (its source id is irrelevant); `--body`
1239
+ // keeps the target's head verbatim, so #id is preserved naturally.
1240
+ //
1241
+ // Output follows resolveOutTarget (file -> in place, stdin -> stdout, `-o`/`-o -`
1242
+ // override) and every splice is guarded — re-parsed and rejected if it broke
1243
+ // the doc, so `set` never writes a corrupt file.
885
1244
  function runSet(args) {
886
1245
  const out = flag(args, "-o") ?? flag(args, "--out");
887
- const from = flag(args, "--from");
888
- const [file, rawId] = positionals(args, ["-o", "--out", "--from"]);
889
- if (!file || !rawId)
1246
+ const from = flag(args, "--in");
1247
+ const headOnly = args.includes("--head");
1248
+ const bodyOnly = args.includes("--body");
1249
+ if (headOnly && bodyOnly)
1250
+ fail("--head and --body are mutually exclusive", 2);
1251
+ const [file, rawId] = positionals(args, ["-o", "--out", "--in"]);
1252
+ if (!file)
890
1253
  fail(SUBHELP.set);
1254
+ // No id: there is no block to replace. Point the way to discovery, not a bare
1255
+ // usage line — `geml get <file>` lists every id `set` can target.
1256
+ if (!rawId)
1257
+ fail(`no #id given — run 'geml get ${file === "-" ? "<file>" : file}' to list addressable ids`, 2);
891
1258
  const id = rawId.replace(/^#/, "");
892
- // Both the document and the replacement can't come from stdin. Reject that up
893
- // front before consuming stdin so the document read below is unambiguous.
894
- if (file === "-" && from === undefined) {
895
- fail("reading the document from stdin needs --from for the new content", 2);
1259
+ // The raw channel is stdin `--in` omitted or `--in -`; anything else sources
1260
+ // a block from a file. Document and content can't BOTH be stdin: reject that
1261
+ // up front, before consuming stdin, so the document read below is unambiguous.
1262
+ const rawChannel = from === undefined || from === "-";
1263
+ if (file === "-" && rawChannel) {
1264
+ fail("reading the document from stdin needs --in for the new content", 2);
896
1265
  }
897
1266
  const source = readInput(file);
898
- // New content: an explicit --from file, else stdin.
899
- let replacement;
900
- if (from !== undefined) {
901
- replacement = readInput(from);
1267
+ if (bodyOnly) {
1268
+ runSetBody(source, id, from, rawChannel, file, out);
1269
+ return;
1270
+ }
1271
+ // default / --head: content is a whole block (default) or a bare head line.
1272
+ let content;
1273
+ if (rawChannel) {
1274
+ content = readInput("-");
1275
+ if (content === "")
1276
+ fail(NO_CONTENT, 1);
1277
+ // Default mode wants exactly ONE block. Pure prose has no head to carry the
1278
+ // id (steer to --body); multiple blocks are `add`'s job. --head takes a
1279
+ // lone head line, so it skips the whole-block shape check.
1280
+ if (!headOnly) {
1281
+ const shape = contentShape(content);
1282
+ if (shape === "empty")
1283
+ fail(NO_CONTENT, 1);
1284
+ if (shape === "prose")
1285
+ fail(`content is prose, not a block — use --body to set the body of #${id}`, 1);
1286
+ if (shape === "multi")
1287
+ fail("set replaces ONE block, but the content has multiple blocks (use add)", 1);
1288
+ }
902
1289
  }
903
1290
  else {
904
- replacement = readInput("-");
905
- if (replacement === "")
906
- fail("no replacement content (use --from FILE or pipe it on stdin)", 1);
1291
+ content = extractBlock(from, id, headOnly ? "head" : "whole");
907
1292
  }
908
- const updated = spliceBlock(source, id, replacement, file);
909
- if (out) {
910
- writeFileSync(out, updated);
911
- console.error(`wrote ${out}`);
1293
+ const normalized = normalizeBlockId(content, id);
1294
+ const updated = spliceBlock(source, id, normalized, file, headOnly);
1295
+ resolveOutTarget(file, out).write(updated);
1296
+ }
1297
+ // `--body`: swap ONLY the target block's body, keeping its head (and #id) and,
1298
+ // for a typed block, its close fence. Assembles head + new body + close and
1299
+ // reuses the guarded spliceBlock — the head carries #id, so the id survives
1300
+ // with no normalization needed.
1301
+ function runSetBody(source, id, from, rawChannel, file, out) {
1302
+ const found = blockSpans(source).get(id);
1303
+ if (!found)
1304
+ fail(`no block with id \`${id}\``, 1);
1305
+ const lines = splitLines(source);
1306
+ const headLine = lines[found.start] ?? "";
1307
+ const headText = stripEol(headLine);
1308
+ // A typed block keeps its closing fence; a heading section has none.
1309
+ let closeLine = null;
1310
+ const open = FENCE_OPEN.exec(headText);
1311
+ if (open) {
1312
+ const lastText = stripEol(lines[found.end - 1] ?? "").replace(/[ \t]+$/, "");
1313
+ const bid = open[3] ? parseAttrs(open[3]).id : undefined;
1314
+ const labeled = bid !== undefined && new RegExp(`^={3,}[ \\t]+#${reLit(bid)}[ \\t]*$`).test(lastText);
1315
+ if (isCloseFence(lastText, open[1].length) || labeled)
1316
+ closeLine = lines[found.end - 1] ?? "";
1317
+ }
1318
+ let body;
1319
+ if (rawChannel) {
1320
+ body = readInput("-");
1321
+ if (body === "")
1322
+ fail(NO_CONTENT, 1);
1323
+ }
1324
+ else {
1325
+ body = extractBlock(from, id, "body");
912
1326
  }
1327
+ let head = headLine;
1328
+ if (head !== "" && !/(\r\n|\r|\n)$/.test(head))
1329
+ head += "\n";
1330
+ let b = body.replace(/\r\n?/g, "\n");
1331
+ if (closeLine !== null && b !== "" && !b.endsWith("\n"))
1332
+ b += "\n";
1333
+ const replacement = closeLine !== null ? head + b + closeLine : head + b;
1334
+ // A typed block (closeLine !== null) must stay ONE block: enforce the
1335
+ // block-count invariant so a `===` fence in the raw body can't close it early
1336
+ // and inject siblings (SEC F2). A heading section body has no close fence and
1337
+ // may legitimately contain blocks, so it is not count-guarded.
1338
+ const updated = spliceBlock(source, id, replacement, file, false, closeLine !== null);
1339
+ resolveOutTarget(file, out).write(updated);
1340
+ }
1341
+ // `geml add <file|-> (--append | --before #x | --after #x) [--in F|F#src|-] [-o]`
1342
+ // — insert a GEML fragment (1+ blocks and/or prose) at a position. Unlike `set`,
1343
+ // `add` names no target id, so content keeps its OWN ids (no normalization); an
1344
+ // id colliding with the document (or duplicated within the fragment) makes the
1345
+ // re-parse fail and nothing is written. Bare prose is a valid fragment.
1346
+ function runAdd(args) {
1347
+ const out = flag(args, "-o") ?? flag(args, "--out");
1348
+ const from = flag(args, "--in");
1349
+ const before = flag(args, "--before");
1350
+ const after = flag(args, "--after");
1351
+ const append = args.includes("--append");
1352
+ const posCount = (append ? 1 : 0) + (before !== undefined ? 1 : 0) + (after !== undefined ? 1 : 0);
1353
+ if (posCount !== 1)
1354
+ fail("add needs exactly one position: --append | --before #id | --after #id", 2);
1355
+ const [file] = positionals(args, ["-o", "--out", "--in", "--before", "--after"]);
1356
+ if (!file)
1357
+ fail(SUBHELP.add);
1358
+ const rawChannel = from === undefined || from === "-";
1359
+ if (file === "-" && rawChannel)
1360
+ fail("reading the document from stdin needs --in for the new content", 2);
1361
+ const source = readInput(file);
1362
+ // Content: --in F#src -> block #src; --in F -> all of F (a multi-block
1363
+ // fragment is fine here); stdin -> raw. No id-normalization: add keeps ids.
1364
+ let content;
1365
+ if (rawChannel)
1366
+ content = readInput("-");
1367
+ else if (from.includes("#"))
1368
+ content = extractBlock(from, "", "whole");
913
1369
  else
914
- process.stdout.write(updated);
1370
+ content = readInput(from);
1371
+ if (content.trim() === "")
1372
+ fail("no content to add (use --in FILE or pipe it on stdin)", 1);
1373
+ // Resolve the physical-line insertion point.
1374
+ const lines = splitLines(source);
1375
+ let at;
1376
+ if (append) {
1377
+ at = lines.length;
1378
+ }
1379
+ else {
1380
+ const anchorId = (before ?? after).replace(/^#/, "");
1381
+ const span = blockSpans(source).get(anchorId);
1382
+ if (!span)
1383
+ fail(`no block with id \`${anchorId}\` in ${file === "-" ? "stdin" : file}`, 1);
1384
+ at = before !== undefined ? span.start : span.end;
1385
+ }
1386
+ const updated = insertFragment(source, lines, at, content, file);
1387
+ resolveOutTarget(file, out).write(updated);
1388
+ }
1389
+ // Splice `fragment` into `source` at physical-line index `at` (splitLines
1390
+ // coords), separating it from adjacent content with a single blank line so
1391
+ // blocks don't fuse, then GUARD: the re-parse must be error-free (a colliding
1392
+ // or duplicate id surfaces as an error diagnostic) and no pre-existing id may
1393
+ // vanish. Returns the updated text; on any violation fail()s and writes nothing.
1394
+ function insertFragment(source, lines, at, fragment, file) {
1395
+ const beforeIds = parse(source, { resolveDoc: resolverFor(file) }).ids;
1396
+ const before = lines.slice(0, at);
1397
+ const after = lines.slice(at);
1398
+ // The preceding line must end in a newline so the fragment starts on its own.
1399
+ if (before.length && !/(\r\n|\r|\n)$/.test(before[before.length - 1])) {
1400
+ before[before.length - 1] += "\n";
1401
+ }
1402
+ let frag = fragment.replace(/\r\n?/g, "\n");
1403
+ if (!frag.endsWith("\n"))
1404
+ frag += "\n";
1405
+ // A single blank separator on each side that has adjacent content and isn't
1406
+ // already blank — keeps a following head / preceding block from fusing.
1407
+ const blank = (s) => stripEol(s).trim() === "";
1408
+ const sepBefore = before.length && !blank(before[before.length - 1]) ? "\n" : "";
1409
+ const sepAfter = after.length && !blank(after[0]) ? "\n" : "";
1410
+ const updated = before.join("") + sepBefore + frag + sepAfter + after.join("");
1411
+ const reparsed = parse(updated, { resolveDoc: resolverFor(file) });
1412
+ const errs = reparsed.diagnostics.filter((d) => d.severity === "error");
1413
+ if (errs.length) {
1414
+ const first = errs[0];
1415
+ fail(`adding the content would break the document: ${first.message} (line ${first.line}); not written`, 1);
1416
+ }
1417
+ const now = new Set(reparsed.ids);
1418
+ const dropped = beforeIds.find((x) => !now.has(x));
1419
+ if (dropped !== undefined)
1420
+ fail(`adding the content would drop block \`#${dropped}\`; not written`, 1);
1421
+ return updated;
1422
+ }
1423
+ // `geml delete <file|-> #id [#id2 …] [-o]` — remove one or more blocks. A
1424
+ // missing id is SKIPPED with a note (declarative "ensure absent", not an
1425
+ // error). Unlike set/add, delete's write is LENIENT: removing a complete block
1426
+ // can't break the parse structurally, but it may leave a reference dangling —
1427
+ // that is a WARNING, never a refusal (delete is reversible via revert + history,
1428
+ // and `geml check` still flags the dangling ref afterward). Contained/overlapping
1429
+ // spans (a nested block inside a deleted heading section) are handled by deleting
1430
+ // the UNION of target lines, so a line is never spliced twice.
1431
+ function runDelete(args) {
1432
+ const out = flag(args, "-o") ?? flag(args, "--out");
1433
+ const pos = positionals(args, ["-o", "--out"]);
1434
+ const file = pos[0];
1435
+ if (!file)
1436
+ fail(SUBHELP.delete);
1437
+ const ids = pos.slice(1).map((s) => s.replace(/^#/, ""));
1438
+ if (ids.length === 0)
1439
+ fail("delete needs at least one #id (run 'geml get <file>' to list ids)", 2);
1440
+ const source = readInput(file);
1441
+ const spans = blockSpans(source);
1442
+ const toDelete = new Set();
1443
+ let found = 0;
1444
+ for (const id of ids) {
1445
+ const span = spans.get(id);
1446
+ if (!span) {
1447
+ console.error(`skipped #${id}: no such block`);
1448
+ continue;
1449
+ }
1450
+ found++;
1451
+ for (let i = span.start; i < span.end; i++)
1452
+ toDelete.add(i);
1453
+ }
1454
+ if (found === 0) {
1455
+ resolveOutTarget(file, out).write(source);
1456
+ return;
1457
+ } // nothing to remove
1458
+ const updated = splitLines(source).filter((_, i) => !toDelete.has(i)).join("");
1459
+ // Lenient guard: surface any resulting error diagnostic (a reference now
1460
+ // dangling) as a WARNING, but write regardless.
1461
+ const reparsed = parse(updated, { resolveDoc: resolverFor(file) });
1462
+ for (const d of reparsed.diagnostics.filter((x) => x.severity === "error")) {
1463
+ console.error(`warning: ${d.message} (line ${d.line}) — left dangling by delete; run 'geml check' to see it as an error`);
1464
+ }
1465
+ resolveOutTarget(file, out).write(updated);
1466
+ }
1467
+ // `geml rename <file|-> #old #new [-o]` — the one verb that reaches OUTSIDE a
1468
+ // block: it rewrites #old's declaration AND every reference to it. #new must be
1469
+ // free; the guarded re-parse refuses anything that would break the doc.
1470
+ function runRename(args) {
1471
+ const out = flag(args, "-o") ?? flag(args, "--out");
1472
+ const [file, rawOld, rawNew] = positionals(args, ["-o", "--out"]);
1473
+ if (!file || !rawOld || !rawNew)
1474
+ fail(SUBHELP.rename);
1475
+ const oldId = rawOld.replace(/^#/, "");
1476
+ const newId = rawNew.replace(/^#/, "");
1477
+ if (oldId === newId)
1478
+ fail("#old and #new are the same id — nothing to rename", 2);
1479
+ const source = readInput(file);
1480
+ const before = parse(source, { resolveDoc: resolverFor(file) });
1481
+ if (!before.ids.includes(oldId))
1482
+ fail(`no block with id \`${oldId}\``, 1);
1483
+ if (before.ids.includes(newId))
1484
+ fail(`id \`${newId}\` already exists; not written`, 1);
1485
+ // Renaming an id that has recorded history breaks the revert-lineage for it
1486
+ // (revert keys by id and can't follow #old -> #new across the boundary). Warn
1487
+ // so the user knows a later `revert #new` won't reach pre-rename revisions.
1488
+ if (file !== "-") {
1489
+ const hp = historyPathFor(file);
1490
+ if (existsSync(hp)) {
1491
+ try {
1492
+ if (blockSpans(resolveContent(hp, "latest").text).has(oldId)) {
1493
+ console.error(`warning: #${oldId} has history; revert across this rename is not tracked — see docs`);
1494
+ }
1495
+ }
1496
+ catch { /* unreadable/empty history: no warning */ }
1497
+ }
1498
+ }
1499
+ const updated = rewriteId(source, oldId, newId, file);
1500
+ const reparsed = parse(updated, { resolveDoc: resolverFor(file) });
1501
+ const errs = reparsed.diagnostics.filter((d) => d.severity === "error");
1502
+ if (errs.length) {
1503
+ const e = errs[0];
1504
+ fail(`rename would break the document: ${e.message} (line ${e.line}); not written`, 1);
1505
+ }
1506
+ if (!reparsed.ids.includes(newId))
1507
+ fail(`rename did not produce #${newId}; not written`, 1);
1508
+ if (reparsed.ids.includes(oldId))
1509
+ fail(`#${oldId} still present after rename; not written`, 1);
1510
+ // Every OTHER id must be untouched. The `#old` match boundary treats a char
1511
+ // outside [A-Za-z0-9_-] as an id terminator, but ids may contain e.g. `.`
1512
+ // (`#foo.bar`), so renaming `#foo` could silently rewrite the *different* id
1513
+ // `#foo.bar` -> `#baz.bar`. Reject when the set of ids other than the rename
1514
+ // pair changed at all (SEC/correctness: collateral id corruption).
1515
+ const othersBefore = before.ids.filter((id) => id !== oldId).sort().join("\n");
1516
+ const othersAfter = reparsed.ids.filter((id) => id !== newId).sort().join("\n");
1517
+ if (othersBefore !== othersAfter) {
1518
+ fail(`rename would also change other ids sharing the \`${oldId}\` prefix (e.g. \`#${oldId}…\`); not written`, 1);
1519
+ }
1520
+ resolveOutTarget(file, out).write(updated);
1521
+ }
1522
+ // Rewrite id `old` -> `new` everywhere it is a declaration or reference, id-
1523
+ // boundary-safe: `#old` is replaced only when NOT followed by an id char, so a
1524
+ // longer id like `#old2` / `#old-x` is untouched. Covers the declaration
1525
+ // (`{#old …}`, labeled close `=== #old`), block references (`[[#old]]`,
1526
+ // `[t](#old)`, chart `data=#old`) and footnotes (`[^old]`). RAW / data block
1527
+ // BODIES (code/diagram/math/table/meta) are skipped — a `#old` there is literal
1528
+ // text, not a reference. (Known residual: id-less raw bodies and inline
1529
+ // code/math spans in flow content — see design §8.)
1530
+ function rewriteId(source, oldId, newId, file) {
1531
+ const doc = parse(source, { resolveDoc: resolverFor(file) });
1532
+ const spans = blockSpans(source);
1533
+ const protectedLines = new Set();
1534
+ for (const b of doc.children) {
1535
+ if (b.kind === "block" && (b.mode === "raw" || b.mode === "data") && b.id) {
1536
+ const span = spans.get(b.id);
1537
+ if (span) {
1538
+ const br = bodyRange(source, span);
1539
+ for (let i = br.start; i < br.end; i++)
1540
+ protectedLines.add(i);
1541
+ }
1542
+ }
1543
+ }
1544
+ const esc = reLit(oldId);
1545
+ const hashRe = new RegExp(`#${esc}(?![A-Za-z0-9_-])`, "g");
1546
+ const fnRe = new RegExp(`(\\[\\^)${esc}(?![A-Za-z0-9_-])`, "g");
1547
+ const lines = splitLines(source);
1548
+ for (let i = 0; i < lines.length; i++) {
1549
+ if (protectedLines.has(i))
1550
+ continue;
1551
+ lines[i] = lines[i].replace(hashRe, `#${newId}`).replace(fnRe, `$1${newId}`);
1552
+ }
1553
+ return lines.join("");
1554
+ }
1555
+ // Extract one block from a GEML file for `--in`. `spec` is `F` (block whose id
1556
+ // == the target) or `F#src` (block #src) — the last `#` splits path from id, so
1557
+ // a `#` inside the path is tolerated; F is read as GEML regardless of extension
1558
+ // (blockSpans + splitLines, no parse — same slice `geml get` prints). `part`
1559
+ // selects the whole span, its head line, or its body. A missing file or absent
1560
+ // id is an operation error (exit 1); the caller writes nothing.
1561
+ function extractBlock(spec, targetId, part) {
1562
+ const hash = spec.lastIndexOf("#");
1563
+ const fragFile = hash >= 0 ? spec.slice(0, hash) : spec;
1564
+ const fragId = hash >= 0 ? spec.slice(hash + 1).replace(/^#/, "") : targetId;
1565
+ let text;
1566
+ try {
1567
+ text = readFileSync(fragFile, "utf8");
1568
+ }
1569
+ catch {
1570
+ fail(`cannot read ${fragFile}`, 1);
1571
+ }
1572
+ const span = blockSpans(text).get(fragId);
1573
+ if (!span)
1574
+ fail(`no block with id \`${fragId}\` in ${fragFile}`, 1);
1575
+ const lines = splitLines(text);
1576
+ if (part === "head")
1577
+ return lines.slice(span.start, span.start + 1).join("");
1578
+ if (part === "body") {
1579
+ const b = bodyRange(text, span);
1580
+ return lines.slice(b.start, b.end).join("");
1581
+ }
1582
+ return lines.slice(span.start, span.end).join("");
1583
+ }
1584
+ // Strip a single trailing terminator (`\r\n`, `\r`, or `\n`) from one line.
1585
+ function stripEol(line) {
1586
+ return line.replace(/(\r\n|\r|\n)$/, "");
1587
+ }
1588
+ // The body sub-range of a block span: [head+1, close) for a closed typed block,
1589
+ // otherwise [head+1, end) — a heading section (no close fence) or an
1590
+ // unterminated block whose span already runs to end-of-scope.
1591
+ function bodyRange(text, span) {
1592
+ const lines = splitLines(text);
1593
+ const open = FENCE_OPEN.exec(stripEol(lines[span.start] ?? ""));
1594
+ if (open) {
1595
+ const lastText = stripEol(lines[span.end - 1] ?? "").replace(/[ \t]+$/, "");
1596
+ const bid = open[3] ? parseAttrs(open[3]).id : undefined;
1597
+ const labeled = bid !== undefined && new RegExp(`^={3,}[ \\t]+#${reLit(bid)}[ \\t]*$`).test(lastText);
1598
+ const closed = isCloseFence(lastText, open[1].length) || labeled;
1599
+ return { start: span.start + 1, end: closed ? span.end - 1 : span.end };
1600
+ }
1601
+ return { start: span.start + 1, end: span.end };
1602
+ }
1603
+ // The shape of default-mode stdin content, section-aware: a heading OWNS its
1604
+ // section (`# H …blocks…` is ONE unit, not many), matching sectionEnd/blockSpans.
1605
+ // Used to reject pure prose (-> --body) and multi-block content (-> add) before
1606
+ // the splice — extraction via --in is inherently one block and skips this.
1607
+ function contentShape(content) {
1608
+ const bs = parse(content).children;
1609
+ let blockUnits = 0, proseUnits = 0, i = 0;
1610
+ while (i < bs.length) {
1611
+ const b = bs[i];
1612
+ if (b.kind === "heading") {
1613
+ i = sectionEndIndex(bs, i);
1614
+ blockUnits++;
1615
+ }
1616
+ else if (b.kind === "block") {
1617
+ i++;
1618
+ blockUnits++;
1619
+ }
1620
+ else {
1621
+ i++;
1622
+ proseUnits++;
1623
+ }
1624
+ }
1625
+ if (blockUnits === 0)
1626
+ return proseUnits === 0 ? "empty" : "prose";
1627
+ return blockUnits + proseUnits === 1 ? "single" : "multi";
915
1628
  }
916
1629
  // Replace block #id's source span in `source` with `replacement`, preserving
917
1630
  // every other byte, and GUARD the result: the re-parse must be error-free, #id
@@ -919,15 +1632,22 @@ function runSet(args) {
919
1632
  // can silently swallow a neighbour). Returns the updated document text; on any
920
1633
  // violation it calls fail() and never returns a corrupt document. Shared by
921
1634
  // `set` and `revert`.
922
- function spliceBlock(source, id, replacement, file) {
923
- const span = blockSpans(source).get(id);
924
- if (!span)
1635
+ function spliceBlock(source, id, replacement, file, headOnly = false, guardCount = false) {
1636
+ const found = blockSpans(source).get(id);
1637
+ if (!found)
925
1638
  fail(`no block with id \`${id}\``, 1);
926
- const beforeIds = parse(source, { resolveDoc: resolverFor(file) }).ids;
1639
+ const beforeDoc = parse(source, { resolveDoc: resolverFor(file) });
1640
+ const beforeIds = beforeDoc.ids;
927
1641
  // Keep the bytes before and after the target span exactly; give the new block
928
1642
  // a single trailing newline so the following block still starts on its own
929
1643
  // line (unless it is the file's last line, which may legitimately lack one).
930
1644
  const orig = splitLines(source);
1645
+ // `--head`: splice only the id's head line; everything below stays
1646
+ // byte-identical. The guard below still applies — the replacement must
1647
+ // re-declare `{#id}` and, for a typed block, keep the fence pairing intact
1648
+ // (an opening line that no longer matches the untouched close fence breaks
1649
+ // the re-parse), or the splice is refused.
1650
+ const span = headOnly ? narrowToHead(found) : found;
931
1651
  const before = orig.slice(0, span.start);
932
1652
  const after = orig.slice(span.end);
933
1653
  let inject = replacement.replace(/\r\n?/g, "\n");
@@ -952,9 +1672,20 @@ function spliceBlock(source, id, replacement, file) {
952
1672
  if (dropped !== undefined) {
953
1673
  fail(`replacement would drop block \`#${dropped}\` (malformed content?); not written`, 1);
954
1674
  }
1675
+ // For a typed block with a close fence, the body is opaque and swapping it
1676
+ // keeps exactly ONE block. A raw `--body` can embed a `===` fence of the
1677
+ // block's length that closes the target early and turns the remainder — plus
1678
+ // the close line we re-appended — into NEW sibling blocks, including an id-less
1679
+ // `=== meta` that redefines document metadata (the dropped-id check above
1680
+ // cannot see an id-less injection). Guarded callers refuse any count change.
1681
+ // (Not enforced for heading sections / whole-block set, whose replacement may
1682
+ // legitimately span several top-level blocks.)
1683
+ if (guardCount && reparsed.children.length !== beforeDoc.children.length) {
1684
+ fail(`replacement changes the block count (a fence in the body closed #${id} early and injected sibling block(s)?); not written`, 1);
1685
+ }
955
1686
  return updated;
956
1687
  }
957
- // `geml revert <file.geml> #id [--to <sel>] [--changed] [--dry-run] [-o out] [--history PATH]`
1688
+ // `geml revert <file.geml> #id [--rev <sel>] [--changed] [--dry-run] [-o out] [--history PATH]`
958
1689
  // Restore ONE block to a past revision's version — a targeted, guarded splice
959
1690
  // that leaves the rest of the document untouched. <sel> (default `-1`): `-N` (N
960
1691
  // revisions back from current), `latest`, or an id prefix/suffix. `--changed`
@@ -964,9 +1695,16 @@ function spliceBlock(source, id, replacement, file) {
964
1695
  function runRevert(args) {
965
1696
  const changed = args.includes("--changed");
966
1697
  const dryRun = args.includes("--dry-run");
1698
+ const headOnly = args.includes("--head");
967
1699
  const out = flag(args, "-o") ?? flag(args, "--out");
968
- const to = flag(args, "--to") ?? "-1";
969
- const [file, rawId] = positionals(args, ["--to", "--history", "-o", "--out"]);
1700
+ const to = flag(args, "--rev") ?? "-1";
1701
+ const before = flag(args, "--before");
1702
+ const after = flag(args, "--after");
1703
+ const append = args.includes("--append");
1704
+ if ((append ? 1 : 0) + (before !== undefined ? 1 : 0) + (after !== undefined ? 1 : 0) > 1) {
1705
+ fail("revert takes at most one position: --append | --before #id | --after #id", 2);
1706
+ }
1707
+ const [file, rawId] = positionals(args, ["--rev", "--history", "-o", "--out", "--before", "--after"]);
970
1708
  if (!file || !rawId)
971
1709
  fail(SUBHELP.revert);
972
1710
  if (file === "-")
@@ -974,21 +1712,25 @@ function runRevert(args) {
974
1712
  const id = rawId.replace(/^#/, "");
975
1713
  const historyPath = flag(args, "--history") ?? historyPathFor(file);
976
1714
  const source = readInput(file);
977
- const curSpan = blockSpans(source).get(id);
978
- if (!curSpan)
979
- fail(`no block with id \`${id}\` in ${file}`, 1);
980
- const curBlock = splitLines(source).slice(curSpan.start, curSpan.end).join("");
981
- // Extract block #id's source from a reconstructed revision (undefined if the
982
- // block did not exist there).
1715
+ const curFull = blockSpans(source).get(id); // undefined => absent now
1716
+ const curBlock = curFull === undefined ? undefined : (() => {
1717
+ const span = headOnly ? narrowToHead(curFull) : curFull;
1718
+ return splitLines(source).slice(span.start, span.end).join("");
1719
+ })();
1720
+ // Extract #id's block from a reconstructed revision (undefined => absent
1721
+ // there). Under `--head`, extract only the head line.
983
1722
  const pick = (text) => {
984
1723
  const s = blockSpans(text).get(id);
985
- return s ? splitLines(text).slice(s.start, s.end).join("") : undefined;
1724
+ if (!s)
1725
+ return undefined;
1726
+ const span = headOnly ? narrowToHead(s) : s;
1727
+ return splitLines(text).slice(span.start, span.end).join("");
986
1728
  };
987
1729
  // Resolve the source revision, formatting any history-layer error cleanly.
988
1730
  const target = (() => {
989
1731
  try {
990
1732
  if (changed) {
991
- const found = firstChangedContent(historyPath, curBlock, pick);
1733
+ const found = firstChangedContent(historyPath, curBlock ?? "", pick);
992
1734
  if (!found)
993
1735
  fail(`no earlier revision changes \`${id}\``, 1);
994
1736
  return found;
@@ -999,22 +1741,135 @@ function runRevert(args) {
999
1741
  fail(historyError(e, file, historyPath), 1);
1000
1742
  }
1001
1743
  })();
1002
- const oldBlock = pick(target.text);
1003
- if (oldBlock === undefined)
1004
- fail(`block \`${id}\` does not exist at revision ${target.id}`, 1);
1005
- if (oldBlock === curBlock) {
1006
- console.error(`#${id} is unchanged at ${target.id}; nothing to revert${changed ? "" : " (try --to -2, or --changed)"}`);
1744
+ const oldBlock = pick(target.text); // undefined => absent at R
1745
+ // Common write path (bespoke message; -o path redirects; -o - -> stdout).
1746
+ const emit = (updated, verb) => {
1747
+ const dest = out ?? file;
1748
+ if (dest === "-")
1749
+ process.stdout.write(updated);
1750
+ else
1751
+ writeFileSync(dest, updated);
1752
+ console.error(`${verb}${dest === file ? "" : dest === "-" ? " -> stdout" : ` -> ${dest}`}`);
1753
+ };
1754
+ // Reconcile #id between now and revision R across the four presence cells.
1755
+ if (curBlock === undefined && oldBlock === undefined) {
1756
+ fail(`\`${id}\` exists in neither the document nor ${target.id} (try --changed)`, 1);
1757
+ }
1758
+ // both present -> SPLICE (undo set)
1759
+ if (curBlock !== undefined && oldBlock !== undefined) {
1760
+ if (oldBlock === curBlock) {
1761
+ console.error(`#${id} is unchanged at ${target.id}; nothing to revert${changed ? "" : " (try --rev -2, or --changed)"}`);
1762
+ return;
1763
+ }
1764
+ if (dryRun) {
1765
+ console.error(`would revert #${id} to ${target.id}:`);
1766
+ process.stdout.write(oldBlock.endsWith("\n") ? oldBlock : oldBlock + "\n");
1767
+ return;
1768
+ }
1769
+ emit(spliceBlock(source, id, oldBlock, file, headOnly), `reverted #${id} to ${target.id}`);
1007
1770
  return;
1008
1771
  }
1772
+ // --head is only meaningful for the splice cell (it can't resurrect or remove).
1773
+ if (headOnly) {
1774
+ fail("--head only applies when the block exists in both the document and the target revision", 2);
1775
+ }
1776
+ // absent now, present at R -> RESURRECT (undo delete)
1777
+ if (curBlock === undefined && oldBlock !== undefined) {
1778
+ // Guard: if the block we'd resurrect is the same (modulo id) as one already
1779
+ // present under a different id, #id was likely renamed away — resurrecting
1780
+ // would duplicate it. Point at `rename` instead of writing.
1781
+ const cmpKey = normalizeBlockId(oldBlock, "__cmp__");
1782
+ for (const [cid, cs] of blockSpans(source)) {
1783
+ if (cid === id)
1784
+ continue;
1785
+ const csrc = splitLines(source).slice(cs.start, cs.end).join("");
1786
+ if (normalizeBlockId(csrc, "__cmp__") === cmpKey) {
1787
+ fail(`#${id} looks renamed to #${cid}; use 'rename #${cid} #${id}' to undo the rename`, 1);
1788
+ }
1789
+ }
1790
+ const { at, where, warn } = resurrectPosition(source, target.text, id, before, after, append, file);
1791
+ if (dryRun) {
1792
+ console.error(`would resurrect #${id} from ${target.id} at ${where}:`);
1793
+ process.stdout.write(oldBlock.endsWith("\n") ? oldBlock : oldBlock + "\n");
1794
+ return;
1795
+ }
1796
+ if (warn)
1797
+ console.error(`warning: anchors for #${id} are gone; appended at end`);
1798
+ emit(insertFragment(source, splitLines(source), at, oldBlock, file), `resurrected #${id} from ${target.id} at ${where}`);
1799
+ return;
1800
+ }
1801
+ // present now, absent at R -> REMOVE (undo add)
1802
+ // Guard: if the block we'd remove is the same (modulo id) as one present at R
1803
+ // under a different id, #id was likely renamed IN — removing would delete a
1804
+ // renamed block. Point at `rename` instead (the dangerous direction).
1805
+ {
1806
+ const cmpKey = normalizeBlockId(curBlock, "__cmp__");
1807
+ for (const [rid, rs] of blockSpans(target.text)) {
1808
+ if (rid === id)
1809
+ continue;
1810
+ const rsrc = splitLines(target.text).slice(rs.start, rs.end).join("");
1811
+ if (normalizeBlockId(rsrc, "__cmp__") === cmpKey) {
1812
+ fail(`#${id} looks renamed from #${rid}; revert would delete it — use 'rename #${id} #${rid}'`, 1);
1813
+ }
1814
+ }
1815
+ }
1009
1816
  if (dryRun) {
1010
- console.error(`would revert #${id} to ${target.id}:`);
1011
- process.stdout.write(oldBlock.endsWith("\n") ? oldBlock : oldBlock + "\n");
1817
+ console.error(`would remove #${id} (absent at ${target.id})`);
1012
1818
  return;
1013
1819
  }
1014
- const updated = spliceBlock(source, id, oldBlock, file);
1015
- const dest = out ?? file;
1016
- writeFileSync(dest, updated);
1017
- console.error(`reverted #${id} to ${target.id}${dest === file ? "" : ` -> ${dest}`}`);
1820
+ const span = curFull;
1821
+ const beforeIds = parse(source, { resolveDoc: resolverFor(file) }).ids;
1822
+ const updated = splitLines(source).filter((_, i) => i < span.start || i >= span.end).join("");
1823
+ const reparsed = parse(updated, { resolveDoc: resolverFor(file) });
1824
+ const errs = reparsed.diagnostics.filter((d) => d.severity === "error");
1825
+ if (errs.length) {
1826
+ const first = errs[0];
1827
+ fail(`removing #${id} would break the document: ${first.message} (line ${first.line}); not written`, 1);
1828
+ }
1829
+ const now = new Set(reparsed.ids);
1830
+ const dropped = beforeIds.find((x) => x !== id && !now.has(x));
1831
+ if (dropped !== undefined)
1832
+ fail(`removing #${id} would drop block \`#${dropped}\`; not written`, 1);
1833
+ emit(updated, `removed #${id} (absent at ${target.id})`);
1834
+ }
1835
+ // Choose the physical-line insertion point for a resurrected block. Explicit
1836
+ // --append/--before/--after win; otherwise infer from the block's neighbours in
1837
+ // revision R: the nearest id BEFORE it that still exists now (insert after it),
1838
+ // else the nearest id AFTER it that still exists (insert before it), else append
1839
+ // at end (warn=true). The deleted block's own former descendants are absent now
1840
+ // too, so they are naturally skipped as anchors.
1841
+ function resurrectPosition(source, revText, id, before, after, append, file) {
1842
+ const lines = splitLines(source);
1843
+ const here = blockSpans(source);
1844
+ if (append)
1845
+ return { at: lines.length, where: "end", warn: false };
1846
+ if (before !== undefined) {
1847
+ const a = before.replace(/^#/, "");
1848
+ const s = here.get(a);
1849
+ if (!s)
1850
+ fail(`no block with id \`${a}\` in ${file}`, 1);
1851
+ return { at: s.start, where: `before #${a}`, warn: false };
1852
+ }
1853
+ if (after !== undefined) {
1854
+ const a = after.replace(/^#/, "");
1855
+ const s = here.get(a);
1856
+ if (!s)
1857
+ fail(`no block with id \`${a}\` in ${file}`, 1);
1858
+ return { at: s.end, where: `after #${a}`, warn: false };
1859
+ }
1860
+ const revIds = [...blockSpans(revText).keys()];
1861
+ const idx = revIds.indexOf(id);
1862
+ for (let i = idx - 1; i >= 0; i--) {
1863
+ const s = here.get(revIds[i]);
1864
+ if (s)
1865
+ return { at: s.end, where: `after #${revIds[i]}`, warn: false };
1866
+ }
1867
+ for (let i = idx + 1; i < revIds.length; i++) {
1868
+ const s = here.get(revIds[i]);
1869
+ if (s)
1870
+ return { at: s.start, where: `before #${revIds[i]}`, warn: false };
1871
+ }
1872
+ return { at: lines.length, where: "end", warn: true };
1018
1873
  }
1019
1874
  // geml codemap <sub>: the code-graph toolkit ships as plain scripts in the
1020
1875
  // package's codemap/ directory (they are argv-driven programs, some
@@ -1027,6 +1882,7 @@ function runCodemap(args) {
1027
1882
  render: "render-all.mjs",
1028
1883
  serve: "serve.mjs",
1029
1884
  refresh: "refresh.mjs",
1885
+ find: "find.mjs",
1030
1886
  mcp: "mcp-server.mjs",
1031
1887
  };
1032
1888
  const sub = args[0] ?? "";
@@ -1037,10 +1893,28 @@ function runCodemap(args) {
1037
1893
  const r = spawnSync(process.execPath, [mod, ...args.slice(1)], { stdio: "inherit" });
1038
1894
  process.exit(r.status ?? 1);
1039
1895
  }
1040
- const entry = process.argv[1] ?? "";
1041
- if (entry.endsWith("geml.js") || entry.endsWith("geml.ts")) {
1896
+ // npm's unix bin shim is a symlink named plain `geml`, so detect "run as a
1897
+ // CLI" by resolving argv[1] to its real path, not by its spelling.
1898
+ const entry = (() => {
1899
+ const argv1 = process.argv[1];
1900
+ if (!argv1)
1901
+ return "";
1902
+ try {
1903
+ return realpathSync(argv1);
1904
+ }
1905
+ catch {
1906
+ return argv1;
1907
+ }
1908
+ })();
1909
+ // `entry` must be non-empty: in a browser bundle both sides degenerate to ""
1910
+ // (esbuild defines process.argv=[] and import.meta.url="", and the node-stub's
1911
+ // fileURLToPath is String()), which would run the CLI at import time and crash
1912
+ // the page. A real CLI invocation always has argv[1].
1913
+ if (entry && (entry === fileURLToPath(import.meta.url) || entry.endsWith("geml.ts"))) {
1042
1914
  const argv = process.argv.slice(2);
1043
- const cmd = argv[0];
1915
+ // The on-disk artifact is `.geml-code-graph/`, so people reconstruct the
1916
+ // command from the directory name — accept those spellings as `codemap`.
1917
+ const cmd = argv[0] === "codegraph" || argv[0] === "code-graph" ? "codemap" : argv[0];
1044
1918
  jsonMode = argv.includes("--json");
1045
1919
  const rest = argv.slice(1);
1046
1920
  if (cmd === "--help" || cmd === "-h") {
@@ -1067,24 +1941,21 @@ if (entry.endsWith("geml.js") || entry.endsWith("geml.ts")) {
1067
1941
  else if (cmd === "set") {
1068
1942
  runSet(argv.slice(1));
1069
1943
  }
1944
+ else if (cmd === "add") {
1945
+ runAdd(argv.slice(1));
1946
+ }
1947
+ else if (cmd === "delete") {
1948
+ runDelete(argv.slice(1));
1949
+ }
1950
+ else if (cmd === "rename") {
1951
+ runRename(argv.slice(1));
1952
+ }
1070
1953
  else if (cmd === "revert") {
1071
1954
  runRevert(argv.slice(1));
1072
1955
  }
1073
1956
  else if (cmd === "history") {
1074
1957
  runHistory(argv.slice(1));
1075
1958
  }
1076
- else if (cmd === "convert") {
1077
- runConvert(argv.slice(1));
1078
- }
1079
- else if (cmd === "export") {
1080
- runExport(argv.slice(1));
1081
- }
1082
- else if (cmd === "render") {
1083
- runRender(argv.slice(1));
1084
- }
1085
- else if (cmd === "fmt") {
1086
- runFmt(argv.slice(1));
1087
- }
1088
1959
  else if (cmd === "check") {
1089
1960
  runCheck(argv.slice(1));
1090
1961
  }
@@ -1093,14 +1964,13 @@ if (entry.endsWith("geml.js") || entry.endsWith("geml.ts")) {
1093
1964
  }
1094
1965
  else if (cmd !== "-" && !/[.\/\\]/.test(cmd)) {
1095
1966
  // A bare word that is neither a known command nor a path is almost always
1096
- // a mistyped command — say so, don't try to read it as a file.
1967
+ // a mistyped command — say so, don't try to read it as a file. (The
1968
+ // reclaimed verbs render/export/fmt/convert land here too.)
1097
1969
  fail(`unknown command '${cmd}'. Run 'geml --help'.`);
1098
1970
  }
1099
1971
  else {
1100
- // Default: parse a file (or stdin via '-') to the document-model JSON.
1101
- const doc = parse(readInput(cmd), { resolveDoc: resolverFor(cmd) });
1102
- console.log(JSON.stringify(doc, null, 2));
1103
- if (doc.diagnostics.some((d) => d.severity === "error"))
1104
- process.exit(1);
1972
+ // A file (or stdin via '-') is the transform entry: `--to`/`--from`/`-o`,
1973
+ // default `--to json`. The single door for every format conversion.
1974
+ runTransform(argv);
1105
1975
  }
1106
1976
  }