ID:ADR-034Status:proposed
Relationships
Related 5
Branches 3
claude/nifty-ptolemy-80epot current proposed
claude/fix-duplicate-plan-ids proposedclaude/data-rune-custom-queries-h8qg72 proposedmain proposedContext
Several in-flight features need to answer "which lines of which file produced this node?", and nothing in the pipeline can. The renderable tree knows what a node is and has no idea where it came from.
The demand is concrete, not speculative:
- SPEC-135 promises a line it cannot produce. Its acceptance criterion reads "returns structured findings — file, line, severity, error id, message", and
PipelineWarning(packages/types/src/pipeline.ts) carriesseverity,phase,pluginName,url?,message— no file, no line. Half of that promise has another route (Markdoc.validate()returnsValidateError, which already carrieslines/location); the cross-page half — broken{% ref %}, missing entities, nav slug failures, raised fromregister/aggregate/postProcess— has none. - The agent read→write loop stops one step short. SPEC-121 gives a machine
/page.json, SPEC-136'stouchingquery tells it which page documents the file being changed. Neither says where in the file, so the edit step is a text search — the fragile link in a chain otherwise made of addresses. - SPEC-136 D8's referrer-side region scoping requires it. Scoping an edge to the paragraph around a reference is what stops a typo three sections away from zeroing it; without location that phase re-parses and re-locates by hand.
- Shipped errors advertise a line they do not have.
read-file.ts:37documents its context parameter as producing "Referenced from: docs/getting-started.md:42". Both call sites pass a bare page path (snippet-pipeline.ts:178,file-ref-resolve.ts:213).site/content/runes/code.mdcarries 17lang-map.tsinvocations; a sandbox error names the file and leaves the reader to find which of the 17. - Node-level provenance (git blame per rune instance) is the original motivation and the weakest of the five — it is the only one with no consumer already waiting.
What is already true, verified against the pinned Markdoc
Measured rather than assumed, because two widely-held assumptions about this are wrong:
| Reality | |
|---|---|
node.lines | number[], always populated |
node.location | { file?, start: {line}, end: {line} } — populated by default |
Markdoc.parse(src, { location: true }) | Changes nothing observable. Location is not opt-in |
location.file | The one absent field. It comes from ParserArgs.file, which no call site passes |
| Line numbering | 0-indexed — # Title on file line 1 reports start.line: 0 |
escapeFenceTags | Line-preserving — maps line→line, substitutes within a line only |
parseFrontmatter | Strips frontmatter before parse, so body coordinates are offset by the stripped line count |
serialize() | A 4-field allowlist — $$mdtype, name, attributes, children |
Tag | Has no location field of any kind |
So location already exists, at every AST node, on every parse. What is missing is the single field that makes it interpretable.
Why a bare line number is not safe here
At least five build-path parse sites produce AST from a source that is not the page being rendered, and splice it into that page's tree:
| Site | Source |
|---|---|
site.ts:240,243 | partials — another file |
expand-pipeline.ts:348 | an expanded entity's body — another page |
collection-resolve.ts:115 | a collection body |
collection-helpers.ts:41 | a template body, parsed as '\n' + bodySource — a deliberate +1 shift, documented in place |
deferred-body.ts:60 | a stashed body |
include-pipeline.ts:183-184 already carries both fields across the clone:
if (node.lines) copy.lines = node.lines; if (node.location) copy.location = node.location;
That code is correct. The node genuinely originated at those lines of the partial. The defect is that without file, a spliced partial's coordinates are indistinguishable from the host page's own — the right data, uninterpretable. Anything consuming lines as "lines in this page" is wrong for included content today, silently, which is the failure class this repository keeps declining to tolerate (SPEC-131, SPEC-135 D8).
Where it dies
The AST → Renderable boundary, pipeline stage 2→3. createComponentRenderable (packages/runes/src/lib/component.ts:38) builds a Tag from cursors and never sees the originating Node. One seam.
Decision
Carry source location through the schema transform as a file-qualified range, sourced from Markdoc's own location, terminating at the renderable.
1. The unit is (file, start, end) — never a bare line number
Every location value names its own source. A consumer never has to know which of five parse sites produced the node it is holding, and a spliced partial reports the partial.
This is the decision the other five follow from. It is also why option 1 is rejected despite being free: the cost of this feature is almost entirely the cost of making the file correct, and a design that omits it is not cheaper, it is wrong.
2. Markdoc's location is the carrier; the work is passing file
No parallel mechanism. All nine Markdoc.parse() call sites in the build path (packages/content, packages/runes) pass { file } naming the real source of that text — the page for page bodies, the partial for partials, the originating entity's page for an expanded body.
include-pipeline.ts's existing carry-across then does the rest: because the cloned node keeps location, and location now names its file, a partial spliced into three pages reports the partial in all three. The splice problem solves itself, which is the strongest argument for using Markdoc's field rather than inventing one beside it.
3. Normalize at capture — once, not per consumer
Two corrections compose, and both are silent when missed:
- 0-indexed → 1-indexed.
# Titleon file line 1 reportsstart.line: 0. - Frontmatter offset.
parseFrontmatterreturns the body only, so every coordinate is short by the stripped line count. Verified: a heading on real line 5 of a file with three frontmatter lines reportsstart.line: 1.
Both are corrected where the parse happens, so what enters the tree is a real file coordinate. A consumer that has to remember to add four is a consumer that will eventually forget, and an off-by-four line link is wrong in the way that looks right.
collection-helpers.ts:41's deliberate '\n' + prefix is a third such offset and is corrected the same way, at its own site.
4. It stops at the renderable; serialize() stays a 4-field allowlist
Location is carried in-process — through the schema transform and available to the pipeline hooks, the CLI, and /page.json. It does not cross the server→client boundary by default.
serialize() being an allowlist is what makes this free: adding a field to Tag cannot leak to the client, because serialize emits four named fields and would have to be changed to carry a fifth. That change is available as an opt-in for the dev server, the editor, and SPEC-121's structured representation — and stays off for ordinary builds, so SPEC-082's payload work is not undone by a feature most sites do not use.
5. Synthesized nodes inherit, and say so
The identity transform constructs nodes that were never authored — 31 makeTag( sites in engine.ts alone, injecting headers, icons, badges and wrappers.
Such a node takes the location of the authored node it was derived from, plus a flag marking it derived. Not null, which would strip location from exactly the structural elements a click-to-source consumer is most likely to land on; and not silent inheritance, which would claim an icon was authored at lines nobody wrote an icon on. The flag is what keeps "inherited" from becoming "false".
6. The plumbing lives in the two factories — rune files do not change
createContentModelSchema builds transform: (node, config) => { … } (packages/runes/src/lib/index.ts:761) and calls the author's transform from inside that closure, with node in scope. So the factory captures node.location and stamps it onto what createComponentRenderable returns.
No individual rune is edited. All ~114 core and plugin runes inherit location by virtue of the factory they already use. A rune that hand-rolls a Schema (icon.ts, badge.ts export a bare Schema) simply has no location until it opts in — degradation, not breakage.
Rationale
The decisive finding is that the mechanism already exists and is already being maintained: location is populated on every node of every parse, and two preprocess pipelines already treat carrying it across a clone as load-bearing enough to comment on. What has never been supplied is file — one argument, at nine call sites.
That reframes the work. This is not "add source tracking to the transform", which would be a large and speculative change to the renderable contract. It is "finish a mechanism that is three-quarters present, and make its output interpretable." The expensive-looking part, provenance through includes and expands, is already handled by code written for other reasons.
Option 1 is rejected precisely because it looks like the cheap version of this decision and is not: it ships the ambiguity rather than resolving it, on a codebase with five splice sites. Options 2 and 3 both work; each trades against something this repository has explicitly invested in — tree rebuild freedom, and SPEC-082's channel reduction.
Terminating at the renderable (4) is what keeps the cost proportionate. The consumers that justify the work — diagnostics, the CLI, /page.json, an agent's edit address — are all build-time. The one that would need serialization, click-to-source in a rendered page, is the one nobody has asked for.
Consequences
Taggains an optional location field, andserialize()does not. The allowlist is the boundary, and it is already written. Any future decision to serialize location is a deliberate edit to one function, not a drift.- All nine build-path parse sites must pass the right
file, and "right" is the whole point. A site that passes the host page for spliced content re-creates the ambiguity this ADR exists to remove, silently. Each site needs a test asserting the file it reports, and the partial/expand sites need one asserting they report the source rather than the host. - Three offset corrections are load-bearing: 0→1 indexing, the frontmatter strip, and
collection-helpers''\n'prefix. Each is invisible when wrong. A fixture whose expected line numbers are read off a real file — not computed from the parse — is the only test that catches them. - SPEC-135's MCP acceptance criterion becomes satisfiable for cross-page findings.
PipelineWarninggains optionalfile/line, which is a type change for every plugin that constructs one — additive, since both are optional. - SPEC-136 phase 4 stops needing its own locator. D8's referrer-side scoping reads the location that is already on the node.
- Memory grows per node, in-process only. A
{ file, start, end }per node across a 237-page corpus is the cost; it never reaches a client bundle under decision 4. Worth measuring on the dogfooded sites before phase 2 of anything that depends on it, not worth pre-optimizing. - Derived nodes need the flag honoured by consumers. A consumer that ignores it and offers "edit this" on an injected header sends an author to lines that do not contain what they clicked. The flag exists; using it is on each consumer, and the blame/provenance case is the one most likely to get it wrong.
- Runes with a hand-rolled
Schemaget nothing until they move to the factory or opt in explicitly.iconandbadgeare the known cases. - Nothing here produces a source map to rendered HTML. Post-identity-transform the tree contains synthesized nodes and consumed metadata; mapping a DOM element back to source is a separate problem this ADR deliberately does not solve.