doc-lattice

A deterministic, offline traceability engine for design and production documentation.

doc-lattice tracks the dependencies between your markdown docs. When a downstream document derives from an upstream one (an integration guide built on an API design, an engineering design built on a product brief), it records that link in frontmatter. When the upstream changes, doc-lattice tells you exactly which downstream docs went stale, and a CI gate keeps stale work from shipping silently.

It is pure tooling: no network (except the optional linear command), stores no secrets, uses no LLM, and needs no database. The dependency graph is derived from your docs on demand, never committed.

The problem it solves

Docs drift apart. Someone changes the API contract, revises a requirement, or reverses an architecture decision, and the documents downstream of that decision keep citing the old version. Nothing breaks loudly; the docs just quietly disagree, and the drift surfaces as a bug, a re-do, or an argument weeks later.

doc-lattice makes those dependencies explicit and checkable. Each downstream doc declares what it derives from and records a hash of what it last saw. A change upstream that the downstream hasn't acknowledged is drift, and check fails CI on it until a human consciously reconciles the link.

Where it fits

doc-lattice is domain-agnostic: it needs nothing but markdown files with frontmatter. Three doc sets it fits naturally:

• Software product docs. Product briefs feed engineering designs, which feed runbooks and integration guides. When a requirement changes, impact lists every downstream doc that cited it, and check keeps the ones that never acknowledged the change from passing CI quietly.
• Game studio design docs (the project's original home). Art direction, economy tuning, and core-loop docs sit upstream of dozens of character, level, and systems specs. One retuned economy value can quietly invalidate a season of downstream work; drift detection surfaces that the day it happens instead of weeks later in a playtest.
• Policy and compliance doc sets. Procedures and checklists derive from a controls document or a policy. An unacknowledged upstream edit there is an audit finding waiting to happen; a CI gate turns it into a red build instead.

How it works

You annotate docs with two things:

• Stable ids. Every tracked file declares an id in its frontmatter. Sections are addressed by their heading's GitHub slug by default; an explicit {#anchor} tag on the heading provides a stable id independent of heading text. Section ids are file-scoped, so the same anchor in two files does not collide with file ids or each other.
• derives_from edges. A downstream doc lists the upstream ids it depends on. Each edge carries a seen hash: a fingerprint of the upstream content at the moment the dependency was last reconciled.

From those annotations doc-lattice builds a lattice: an id-indexed graph of nodes (your docs) and edges (the derives_from links). Every command reads from that one structure. The seen hash is the load-bearing trick: comparing it against the upstream's current content hash is what turns "these docs depend on each other" into "this dependency is out of date."

Drift states

check classifies every edge into one of four states:

The content hash is sha256 of a canonicalized copy of the text, truncated to 128 bits. Canonicalization normalizes line endings, strips trailing whitespace per line, and trims leading and trailing blank lines, so those cosmetic edits never trip drift. Internal whitespace is preserved, so rewrapping a paragraph (which moves its line breaks) does count as a change.

The hash input never includes frontmatter. A file ref hashes the canonicalized document body; a section ref hashes only the canonicalized text of the target section. Because reconcile writes only seen, and seen lives inside frontmatter, a reconcile write can never change any target's hash. That is what lets reconcile --all converge in one pass over a stable snapshot: acknowledging one edge cannot invalidate another. Convergence covers the reconcilable edges only, since reconcile skips BROKEN ones and they remain findings after the pass. The write scope and selector semantics restated here are owned by RECONCILE.md.

Broken refs and tool errors

A ref that points at nothing is a normal, reportable lattice state: check calls it BROKEN and exits 1. Invalid config or lattice frontmatter, unreadable or non-UTF-8 documents, containment failures, and incoherent ids are tool errors that exit 2. An index is incoherent when two files repeat a file id or two headings in one file resolve to the same file-scoped anchor. Equal anchors in different files, and a file id equal to another file's anchor, remain distinct TargetId(file_id, anchor) keys and do not collide. A write refused because the pipe's reader closed (for example, piping into head) exits 141, printing nothing.

A Markdown file without an opening --- fence is valid untracked prose. Once a file opens YAML frontmatter with ---, it must include a closing --- fence; otherwise every lattice-loading command names the file, asks for the missing close, and exits 2 instead of omitting the node.

Fenced frontmatter with no id is graded by what else it declares, since dropping a node also drops every edge it declares. A block carrying any of derives_from, authority, or tickets meant to be a node, so the missing id is a tool error that names the file and exits 2 rather than a silent omission. Any other id-less block is frontmatter this engine does not own, so the file is skipped with a warning on stderr naming it, and the exit status is unchanged. The two tiers are described in full under Frontmatter reference.

The authority ladder

Separately from drift, lint enforces a structural rule: authority only flows downhill. Docs can declare an authority of binding, derived, or exploratory. A derives_from edge from a more-authoritative doc to a less-authoritative one is an inversion (a binding spec should not derive from an exploratory sketch), and lint fails on it. lint is pure structure, independent of drift, and exits 1 on a violation just like check.

An edge whose source or target declares no authority cannot be ranked. lint reports it as unranked rather than as a violation, so it never fails the gate. Every human run ends with a coverage line counting violations and unranked edges, and --format json carries the same edges in a skipped array with a source-unannotated or target-unannotated reason.

A worked example

Two docs. The upstream owns a decision; the downstream depends on it.

docs/api-design.md, the upstream:

---
id: api-design
layer: design
authority: binding
---
# API Design

## Pagination {#pagination}
List endpoints use cursor pagination: pass the last item's cursor as `after`.

docs/billing-integration-guide.md, which derives from the pagination decision:

---
id: billing-integration-guide
layer: technical
authority: derived
derives_from:
  - ref: api-design#pagination
    seen: 647cc64481bee8d8541ef7d1733b5204
tickets: [ENG-412]
---
# Billing Integration Guide

Invoice listings page through results with the cursor scheme the API design defines.

The ref api-design#pagination resolves file-scoped: it points at the section in the api-design file whose heading carries the {#pagination} marker. Markers are optional; a heading with no marker is addressed by its GitHub slug instead, and an explicit marker pins the id so the ref survives a later rewording of the heading. The seen hash records the pagination text the guide was last built against.

Now someone switches the API to page-number pagination. The {#pagination} section's content hash no longer matches seen, so:

$ doc-lattice check
STALE         billing-integration-guide -> api-design#pagination
1 edge: 0 OK, 1 STALE, 0 UNRECONCILED, 0 BROKEN

$ doc-lattice impact api-design#pagination
billing-integration-guide  ('/work/acme-api/docs/billing-integration-guide.md')  tickets: ENG-412

check exits 1, so CI is now red. A human reviews the guide against the new pagination scheme, updates the body if needed, and then locks in the new hash:

$ doc-lattice reconcile billing-integration-guide
reconciled 'billing-integration-guide.md': api-design#pagination

$ doc-lattice check
1 edge: 1 OK, 0 STALE, 0 UNRECONCILED, 0 BROKEN

The listing is empty because there is nothing left to act on; the verdict line is what states the run was clean.

That edit → check → review → reconcile loop is the whole workflow. reconcile is the only command that writes to your docs, and it only ever rewrites seen values and the aliases that read them.

Quick start

Prerequisites

• Python 3.13+ (the floor is load bearing; see AD-24)
• uv (curl -LsSf https://astral.sh/uv/install.sh | sh)

Install and run

Run the released CLI without installing it globally:

uvx doc-lattice --help

Or install it into an isolated tool environment:

uv tool install doc-lattice
doc-lattice --help

pipx install doc-lattice provides the same isolated installation. A conventional python -m pip install doc-lattice is also supported when installing into an activated virtual environment.

Development

uv sync --group dev
uv run doc-lattice --help

Contributor commands, gates, and the full verification set live in CLAUDE.md.

Commands

check and lint gate by default, exiting 1 when they find drift or an authority inversion. impact, reconcile, graph, and init are informational and exit 0 on success, so wiring impact into a CI gate never turns the build red. linear also exits 0 by default; pass --exit-code to gate on any DANGER or BLOCKED finding, and add --warn-exit to gate on WARNING as well.

The lattice-loading commands check, lint, impact, reconcile, graph, and linear accept --config PATH (path to .doc-lattice.yml; defaults to the file in the current directory). init deliberately does not accept config or load the lattice. It keeps its current-directory behavior and does not require Git: it reads local Git state only to guess the workflow's default branch, and falls back when that read finds nothing. Run uv run doc-lattice <command> --help for the full flag list.

Pass --indent N with JSON output on check, lint, impact, or linear to pretty-print the JSON with N spaces per level. JSON output is selected uniformly by --format json; --indent without an effective --format json is a usage error.

Use the global --no-color option before the command to disable colored output explicitly, for example doc-lattice --no-color check. Rich also honors the NO_COLOR environment variable; --no-color is the command-line equivalent. doc-lattice intentionally extends the NO_COLOR baseline: the standard itself only asks implementers to drop color and leaves bold, underline, and italic styling in place, but either lever here means no styling at all, so no terminal escape sequence reaches the output under either one, even when a terminal-forcing variable is set. This covers every command's output, not just help and usage-error text.

check and lint also accept --format human|json|github. human is the default. github emits one escaped GitHub Actions ::error workflow command per drift finding or ladder violation, each with a repo-relative file path, so findings attach inline to the offending doc in the pull-request diff. Output selection never changes gate exit codes.

The reported path is relative to GITHUB_WORKSPACE when that variable is set and contains the document, which is the repository checkout root under GitHub Actions; annotations therefore stay attachable no matter which subdirectory the command runs in. Otherwise the base is the invocation working directory, and a document outside it is reported by absolute path rather than failing. Under a workspace the base is never the config file's project root: a config under packages/game still reports packages/game/docs/down.md, not docs/down.md. Only a run with no workspace set, from inside packages/game, reports docs/down.md, because there the base is that working directory.

When a document falls outside that base, its annotation carries an absolute path, which GitHub drops instead of attaching. That is reported: the run warns once on stderr, naming the base and every document outside it, so a gate that fails with nothing shown on the diff is diagnosable from the workflow log. Stdout stays exactly the workflow commands, so the warning never interferes with what GitHub parses.

Structured output is always selected with --format; the accepted values per command are in the table above, and init is deliberately excluded from structured-output selection. The 1.x silent --json alias was removed in 2.0; see CHANGELOG.md for the migration.

impact walks the full transitive closure by default. Pass --depth N (N >= 1) to bound the walk to N hops from TOKEN: --depth 1 lists only the docs that derive directly from it. Human output is unchanged, and each JSON entry gains a "depth" field carrying the minimum number of hops at which that doc is reached.

Human check output lists problem edges only. OK edges are classified, counted, and reported everywhere else, but they are not listed as rows, so the default invocation on a large lattice shows what needs acting on rather than thousands of OK lines. This matches lint, which has always printed violations only, so the two gating commands now ship one output philosophy instead of opposite ones. The verdict line below is what makes the omission safe rather than lossy: the totals and every per-state count stay visible, so a clean run is explicit rather than silent. --only overrides the default in both directions, and --format json is unaffected.

Every human check run ends with a one-line verdict counting the classified edges and breaking them down per state, for example 101 edges: 96 OK, 5 STALE, 0 UNRECONCILED, 0 BROKEN. Every state is listed, including the ones with a zero count, so truncated output such as check | tail still states the result rather than trailing off into whichever edges sort last. The line is one record on one line at any terminal width, so check | tail -1 always gets the whole verdict.

The line is present on a clean lattice too, where it is the entire output, and a lattice with no edges at all reports 0 edges: 0 OK, 0 STALE, 0 UNRECONCILED, 0 BROKEN. Because the state names are always printed, match on the exit code rather than grepping human output for a state name.

--format json carries the same counts in a summary object alongside edges, so a wrapper can answer "is the tree clean?" without folding every edge record. JSON is the complete structured record and still carries one entry per edge, OK edges included; the problem-only default is a human-output rule and does not reach it. --format github is likewise unchanged: it emits annotations for problems and stays silent on a clean tree.

check accepts a repeatable --only STATE to select which states are displayed (case insensitive, e.g. --only stale --only broken); an unrecognized state exits 2 and names the valid set. Supplying it replaces the defaults in both formats that list records: human output shows exactly the selected states, so check --only OK lists the OK edges the default omits, and JSON narrows edges to them. Filtering is display-only: the exit code and the summary counts always reflect every edge, so check --only OK on a drifting lattice still exits 1 and still reports the drift in its verdict. One consequence is deliberate: under --only, the summary counts do not sum to the number of records in edges.

reconcile

A normal reconcile needs either a downstream id or --all, and running it with neither is an error. reconcile DOWNSTREAM_ID clears every drifting edge of one node, --ref REF narrows that to a single upstream ref, and reconcile --all clears every STALE or UNRECONCILED edge in the lattice, with or without --ref. Add --dry-run to any of those to preview the plan read-only, or run reconcile --recover alone to finish an interrupted transaction without loading the lattice.

A real run applies the whole selected batch as one durable transaction under a nonblocking advisory lock: it re-reads each file at write time, detects a destination that changed under it as a conflict, and rolls the batch back if anything fails before the commit. Its temporary artifacts are a project-root journal plus staged before and after images, covered by the .gitignore block that doc-lattice init prints.

See RECONCILE.md for selector details, dry-run and JSON output, the durability contract, and recovery.

Frontmatter reference

id, title, each tickets entry, derives_from[].ref, and derives_from[].seen must be single-line text carrying no control character. A value that decodes to any C0 code point (U+0000 to U+001F, tab, newline, and carriage return included), DEL (U+007F), or any C1 code point (U+0080 to U+009F) fails the load with a FRONTMATTER_ERROR naming the key and the code point. Every other character is accepted, accented text, CJK, emoji, and a no-break space among them. The rule is what carries the --no-color promise above into a document's own text, as the quoted spelling of a path carries it into a filename: a value reaches your terminal, so it cannot carry bytes your terminal acts on. An unknown key that decodes to one is spelled in the error rejecting it for the same reason. A block that fails to parse at all never reaches this rule, and holds the promise a different way: the YAML parser's own message echoes what it choked on, so it is printed in the same quoted spelling a filename gets, as one line whose line breaks read as \n. That trades the parser's caret, which no longer points at anything once the message is one line, for coordinates you can still read and a diagnostic that cannot act on your terminal.

Three spellings reach this rule, and two of them are ones you can write without meaning to.

A literal tab is the one control byte YAML itself admits as a raw byte, inside a quoted or a block scalar, and nothing on screen distinguishes it from a run of spaces, so a value that looks fine can fail. An escape in a double-quoted scalar is how everything else in the refused set reaches a value, since YAML rejects those as raw bytes. "\u001b" is the spelling to picture, but it is far from the only one: \0, \a, \b, \t, \n, \v, \f, \r, \e, and \N each name a refused code point directly, and \xNN and \U00000000 reach the same points as \uNNNN does. A newline from a block scalar is the third, and it arrives in two ways that need different fixes. A block written | or > keeps a trailing break, which |- or >- chomps away. A literal block (| in any chomping mode) also keeps the breaks between its lines, and no chomping indicator touches those, so a |- spanning two lines still fails; only the folded styles join their lines with a space. Write a multi-line value as >-, which is the spelling that survives this rule, and keep |- for a value on one line. Folding has two limits worth knowing, because a >- that hits either keeps an interior break anyway: a blank line between the lines is a paragraph break, and a line indented past the block keeps its own break. Keep a folded value's lines adjacent and equally indented.

To find all three, do not pattern-match the YAML. Load it. Every spelling above is a property of the value a loader constructs, not of the line it is written on, and the ways to write one value are open-ended: a tab need not share a line with its key, a derives_from edge writes - ref: behind a sequence dash, a block-form tickets entry is a bare - "GTX-1" item with no key at all, a block scalar puts its value on later lines behind an optional anchor or tag, and an explicit ? key pair puts the key on its own line. A scan that reads the fence the way the loader does and then asks the loader for the values is exact against all of them, and reports the same field name and code point the error would:

uv run python - <<'PY'
import pathlib
from ruamel.yaml import YAML

yaml = YAML(typ="safe", pure=True)

def refused(text):
    return sorted({f"U+{ord(c):04X}" for c in text
                   if ord(c) < 0x20 or ord(c) == 0x7F or 0x80 <= ord(c) <= 0x9F})

def fields(meta):
    yield from ((k, meta[k]) for k in ("id", "title") if isinstance(meta.get(k), str))
    yield from ((f"tickets[{i}]", t) for i, t in enumerate(meta.get("tickets") or [])
                if isinstance(t, str))
    yield from ((f"derives_from[{i}].{k}", e[k])
                for i, e in enumerate(meta.get("derives_from") or []) if isinstance(e, dict)
                for k in ("ref", "seen") if isinstance(e.get(k), str))

for path in sorted(pathlib.Path(".").rglob("*.md")):
    lines = path.read_text(errors="replace").lstrip("\ufeff").split("\n")
    if not lines or lines[0].strip() != "---":
        continue
    end = next((i for i, line in enumerate(lines[1:], 1) if line.strip() == "---"), 0)
    try:
        meta = yaml.load("\n".join(lines[1:end]) + "\n") if end else None
    except Exception as exc:
        print(f"{str(path)!r}: frontmatter did not parse: {str(exc).splitlines()[0]}")
        continue
    if isinstance(meta, dict):
        for name, value in fields(meta):
            if found := refused(value):
                print(f"{str(path)!r}: {name}: {' '.join(found)}")
PY

Run it where doc-lattice is installed, since it borrows that installation's ruamel.yaml and its pure parser, which is the one the loader itself uses. Everything it prints is a value that will not load. doc-lattice check is the authority after the upgrade, and reports the first document it finds, one run at a time.

It prints each path quoted, through the same repr(str(path)) spelling every command uses, for the reason the --no-color guarantee under Commands rests on: a filename is repo-controlled text too, and a scan that hunts terminal control characters is a poor place to print one.

A code point it names is a value that will not load. A block-scalar line is a prompt to look rather than a verdict, since whether a break survives depends on the lines beneath it, which is also why a correct >- is listed. doc-lattice check is the authority either way: it names the key and the code point for the first document it finds, one run at a time.

Files with no id

Only a file declaring an id joins the lattice. How the rest are treated depends on what they wrote, because silently dropping a file also silently drops every edge it declares:

Those three keys are the exact set that wires a file into the graph, so an id-less block carrying one is a node that lost its id (to a typo like idd:, or to an edit) rather than metadata belonging to another tool. title and layer are not in the set: they describe a document without wiring it into the graph, so a block carrying only those is a warning, not an error.

The warning exists because a docs root can legitimately hold frontmatter this engine does not own, such as a tool's own name/description block. To silence it for one file, prefer ignore_globs: it drops the file from discovery, so nothing else a run reports is affected.

It is also a Python warning, so the PYTHONWARNINGS filters apply, but neither available form targets it precisely. PYTHONWARNINGS=ignore silences every warning the run emits. PYTHONWARNINGS=ignore:skipping looks narrower and is not: that field is a literal message prefix rather than a regular expression, and the warning for a document symlinked outside the project root opens with the same skipping prefix, so filtering on it hides that one too. The report is identical whether the load was accelerated by the load cache or not.

Section ids are optional: a heading is addressed by its GitHub slug by default (e.g. ## Error Handling resolves to error-handling). An explicit marker must be the trailing heading token and match {#[A-Za-z0-9][A-Za-z0-9_-]*}; a whitespace-separated ATX closing sequence may follow it (e.g. ## Error Handling {#errors} ##). A valid marker supplies the stable anchor independent of heading text. Invalid or nontrailing marker-like text is ordinary heading content, so the heading falls back to its generated GitHub slug. Section refs are file-scoped (file#anchor), so the same anchor in two files does not collide.

Addressable sections intentionally use a narrow Markdown subset: column-zero ATX headings at levels 1 through 6, including empty headings and optional ATX closing sequences. CommonMark backtick and tilde fences suppress headings inside them. Setext headings, headings in block quotes or list items, and indented headings are not addressable. Inline Markdown remains part of the raw heading text used for slugging. Heading and fence recognition is pinned to markdown-it-py==4.2.0; generated slugs and document-order duplicate suffixes target github-slugger@2.0.0 under JavaScript Unicode 17.0. Generated lowercase patches and contextual casing-property tables bridge the minimum supported Python 3.13 Unicode 15.1 table to that target.

Configuration

doc-lattice runs zero-config (defaulting to a docs/ root), or reads .doc-lattice.yml from the current directory:

# doc-lattice configuration
docs_roots:
  - docs                  # directories to scan, or individual .md files (default: ["docs"])
# ignore_globs:           # paths to skip within those roots
#   - "**/archive/**"
# cache_key: my-docs      # opt-in incremental load cache slot (see Load cache below)
# cache_trust_stat: false # opt-in stat fast tier for read-only commands (accepts the mtime caveat)
# linear_team: ENG        # the Linear team the `linear` query targets

The project root is the resolved parent of the selected config file, including an explicit --config PATH, or the resolved current directory in zero-config mode. Relative docs_roots entries are interpreted from that project root. Every root must resolve inside it; an entry that escapes via .., an absolute path, or a symlink is rejected before any read.

Each docs_roots entry names either a directory to scan recursively for .md files, or a single .md file to track directly. An entry that does not exist is tolerated and contributes no documents. An entry that exists but is neither a directory nor a regular .md file, such as a non-markdown file, a socket, or a device, fails config load with a config error and exit 2. ignore_globs patterns match a directory entry's files by their path relative to that root; a direct file entry has no root-relative path, so it is matched by its basename instead.

Discovered document symlinks are resolved separately. A symlink whose target stays inside the project root is allowed, while one targeting anything outside is skipped with a warning. If multiple roots or symlink aliases resolve to the same document, it is loaded once under the first unresolved path discovered. Reconcile re-resolves that identity path before writing so a retargeted symlink cannot escape the project root.

For 2.0, binding_layers is unsupported. Delete it from 1.x configs; there is no replacement. lint's fixed binding > derived > exploratory authority ladder is unchanged.

Load cache (opt-in)

Large doc sets (thousands of files) can skip re-parsing unchanged docs with an opt-in cache. Set cache_key to a single safe segment (^[A-Za-z0-9][A-Za-z0-9._-]{0,63}$); it names a slot under your user cache home at <cache_home>/doc-lattice/<cache_key>/load-cache.json, where <cache_home> is $XDG_CACHE_HOME (when absolute) or ~/.cache. The cache lives outside every checkout on purpose: because .doc-lattice.yml is committed, every clone and git worktree of the project shares one warm cache with no per-checkout setup, which an in-repo cache could not do.

By default the cache re-reads and re-hashes each file's bytes every run, so its output is always byte-identical to an uncached run under any cache state (cold, warm, stale, structurally corrupt, or wrong version); only timing differs. That covers stderr as well as stdout: an entry records why a file is not a node, not just that it is not one, so the skip warning above reproduces on a warm run that never re-reads the file. Because two checkouts can share a slot, an entry stores the reason rather than the rendered sentence, and each run names the path it discovered. A structurally corrupt cache (unreadable, non-JSON, wrong version, or schema-invalid) is discarded wholesale and rebuilt; the cache is a trusted single-writer file under your own cache home, so it is not hardened against hand-edited tampering that stays schema-valid. Setting cache_trust_stat: true adds a faster tier for read-only commands that trusts a file whose size and modification time are unchanged, accepting that the file is not opened at all: a rewrite that preserves both its size and its nanosecond mtime is served stale, and a file made unreadable (for example a permissions change, which does not alter size or mtime) is served from cache instead of erroring, each until the file is touched. reconcile ignores cache_trust_stat and always verifies content, so it can never write frontmatter from stale data. cache_trust_stat: true requires cache_key; otherwise config loading is a tool error and exits 2. Two projects sharing a cache_key stay correct (a content-hash hit implies identical bytes); the only cost is overwrite churn, so prefer distinct keys. Delete the cache directory to reset it; a tool-version bump discards it automatically.

Any cache read failure, including an unreadable, invalid, or stale cache file, silently falls back to rebuilding from documents. A cache write failure emits one stderr diagnostic and is otherwise ignored: it does not change command results or exit codes.

Adopting doc-lattice in your docs repo

Both setups below share one rule. For an initial adoption with no established baseline, annotate your documents, then run doc-lattice reconcile --all once before enabling the gates:

uvx --python 3.13 --from doc-lattice==5.0.0 doc-lattice reconcile --all

Commit the annotated input state and start from an otherwise clean working tree before running this command, so its reconcile-only diff can be reviewed and reverted with git. That undo is separate from reconcile's own failure rollback, which restores an interrupted run but does not reverse a successful baseline you later decide was wrong; see RECONCILE.md for what that rollback does cover.

This acknowledges the current state of every STALE and UNRECONCILED edge, so the gates start from a known baseline instead of reporting the whole backlog on their first run. Do this only on a first adoption: init is rerunnable against an existing config, and on an established lattice reconcile --all would acknowledge legitimate drift you have not reviewed. It also does not by itself make CI green, because reconcile skips BROKEN edges and those remain findings. See RECONCILE.md for the selector semantics.

Ordinary offline setup

Bootstrap config and the drift and authority-ladder gates for a repo whose docs you want to track:

uvx --python 3.13 --from doc-lattice==5.0.0 doc-lattice init

This writes .doc-lattice.yml (only if absent) and always prints the reconcile-artifact .gitignore block (see RECONCILE.md), pre-commit hooks, and a GitHub Actions workflow that run doc-lattice check (drift) and doc-lattice lint (authority ladder) as your gates. Paste each where the output says. Pasting the pre-commit block installs no Git hook, so those two hooks stay inert until you enable the gates, which is a separate step with an ordering constraint on a first adoption. init only prints .gitignore guidance and never modifies that file. Pass --docs-root (repeatable) or --linear-team to bake those values into the generated config. The generated gates remain fully offline: they run only check and lint and do not require or receive LINEAR_API_KEY.

Branch resolution reaches init in 5.0, the same release boundary MANAGED_CI.md records for --default-branch and its own recipe. Before that release the printed workflow hard-wires its trigger: both the push and the pull_request filter render as branches: [main], init never inspects origin/HEAD, the run names no branch on stderr, and passing --default-branch exits 2 with No such option: --default-branch. On any such release, compare both filters against the branch you actually gate on before saving the file and edit both when it is not main. That hand edit is the only correction available there, and a repository on master, trunk, or develop that skips it installs a workflow that never runs.

Starting in 5.0, and through the two paragraphs after this one, the printed workflow triggers on one branch that init resolves by a fixed precedence: an explicit --default-branch wins; otherwise the local origin/HEAD remote-tracking ref is read; otherwise it falls back to main. The run always names the branch it used and where that came from on stderr, for example workflow triggers on branch trunk (origin/HEAD), so a repository on master, trunk, or develop does not silently install a workflow that never runs.

Treat the probe as a hint rather than an authority. origin/HEAD is cached local state: it is often absent in fresh or shallow clones, and after an upstream default-branch rename it can still name the old branch. init detects a target that no longer exists and falls back, but a stale target that is still present locally is indistinguishable from a current one without network access. Reading the reported source line is how you catch that, and --default-branch is how you fix it. Prefer passing it explicitly whenever you want a reproducible result, such as in an upgrade you intend to repeat. A missing remote, a missing git, or a directory outside a worktree all fall back quietly; ordinary init has no Git requirement. The probe runs git only from an absolute path outside the directory being scaffolded, so a checkout carrying its own git falls back rather than running it, as does a git reachable only through a relative PATH entry. A branch name that is supplied or detected but is not a supported literal name is a different case and is rejected with an error: names are limited to ASCII letters, digits, ., _, and - in /-separated parts, because a GitHub branches: filter is a glob pattern rather than a literal and *, ?, [, ], and ! would be matched as patterns. Git's own structural exclusions are rejected too, including .. anywhere in the name, a leading or trailing . on any part, a .lock suffix, and the reserved name HEAD: no branch can carry such a name, so a filter built from one would never match. Only that exact spelling of HEAD is reserved, so head, release/HEAD, and similar names are ordinary branch names and are accepted.

--default-branch applies only to this printed workflow. The Linear workflow the recipe in MANAGED_CI.md publishes is pinned to the exact main branch as a security control, so nothing there takes a branch either.

To test an unreleased commit, replace the PyPI requirement with a Git source such as --from git+https://github.com/Guardantix/doc-lattice@<commit>; released configurations should keep the exact PyPI version pin.

Enabling the gates

Pasting the pre-commit block installs no Git hook. It adds two hook definitions to .pre-commit-config.yaml, and nothing reads that file on commit until pre-commit has written .git/hooks/pre-commit in your clone. Until it has, the gates exist in CI only, and every local commit succeeds regardless of drift, including the one that introduces it. Enabling them is a separate, explicit act, and nothing earlier in this setup path performs it or provides the pre-commit runner it needs, because the path requires only uv:

uv tool install pre-commit
uv tool run pre-commit install

uv tool install puts the runner in a persistent environment. uv tool run then invokes it without needing uv's tool bin directory on PATH, which in a fresh shell it often is not; uv warns when that applies. If the machine already has a durable pre-commit from another installer, plain pre-commit install does the same job, and there is no reason to force a uv tool installation over it.

Prefer that pair over uvx pre-commit install. Pre-commit records the absolute path of the interpreter that installed the hook and runs it first, falling back to a pre-commit on PATH only when that path is gone. Installed through uvx, the recorded interpreter lives in uv's disposable cache, which uv is free to reclaim. On a machine whose only Python tooling is uv there is then no fallback either, and the hook fails closed: it exits 1 with a pre-commit not found error and blocks every commit until it is reinstalled. The hook records a path rather than resolving one each time, so re-run the install command after anything that moves or rebuilds the environment behind it, including renaming a parent directory.

.git/hooks/ is neither tracked nor cloned, so activation belongs to a clone rather than to the repository. Committing .pre-commit-config.yaml enables nothing for anybody else: each contributor runs the install command once in their own clone, and again after re-cloning. A fresh clone of an already-gated repository commits drift locally without complaint, and the offline workflow is what catches it on the pull request. Such a clone is the established case below, so it activates immediately and has no ordering to observe.

On an initial adoption, enable them after the reconcile baseline, not while pasting the blocks. check exits 1 on unreconciled edges as well as stale ones, and an initial adoption commits exactly that state, because the baseline above has you commit the annotated input before reconcile --all acknowledges it. Gates enabled during setup therefore refuse the very commit the baseline depends on. The order is:

1. Paste the three blocks and commit them with the annotated input. The gates are inert, so this commit is not gated.
2. Run doc-lattice reconcile --all.
3. Run check and lint, and resolve whatever they still report. The baseline clears STALE and UNRECONCILED edges and nothing else: reconcile skips BROKEN edges by design, check exits 1 on them all the same, and neither command touches a lint finding. Anything left standing here is what the gate refuses in step 5.
4. Enable the gates with the two commands above.
5. Stage and commit the reconcile-only diff. Both hooks run on it and pass, which is also how you confirm activation worked.

An established installation enables them immediately. A conversion, or any repository that already has a baseline, skips the reconcile step and with it the ordering constraint, so run the two commands as soon as you find the gates off.

Confirm activation with a commit that stages at least one Markdown file. Both generated hook entries carry files: \.md$, so an empty or non-Markdown commit runs pre-commit and reports both doc-lattice hooks as Skipped. That looks like a working gate and proves nothing about it.

The hook entries run uvx --python 3.13 --from doc-lattice==5.0.0, so the pinned release has to resolve on every gated commit, out of uv's cache once it is warm and from PyPI when it is not. These gates are offline in the sense that matters for secrets, meaning they never require or receive LINEAR_API_KEY. That is not the same as running without a network.

Protected Linear setup in CI

To run linear in CI without exposing its API key to every workflow in the repository, install the hand-installable recipe in MANAGED_CI.md. You add two workflows you own, plus a GitHub environment whose deployment allow list is exactly main and which holds one dedicated secret. That secret is mapped onto LINEAR_API_KEY only on the final step of the trusted job. The environment is the boundary; the workflows are ordinary files in your repository.

The recipe replaced a managed setup that generated and maintained four committed artifacts for the same boundary, removed in 5.0. The same document carries the procedure for converting an installation left over from it, which changes no remote state.

See MANAGED_CI.md for the recipe, what it does without relative to the managed setup, requirements, conversion, and the security model.

Upgrading

Upgrades are hand-applied: nothing in your repository updates itself. init prints its blocks from the version of doc-lattice actually running, so every command below has to name the release you are moving to, and an old binary prints the old blocks. Substitute the target release for NEW_VERSION throughout. Read that release's section in CHANGELOG.md first: a release that changes generated output carries a ### Migration subsection with the steps specific to it.

Which paths apply depends on how you installed. The pre-commit snippet is yours either way, the ordinary workflow is yours, and a recipe installation's Linear workflow is yours too. A managed installation left over from 4.1.0 converts rather than upgrades; see Managed installs.

The pre-commit snippet (every install)

The pre-commit block is printed guidance rather than a generated file, so no command updates your .pre-commit-config.yaml for you. Print the block from the target release and compare it against the one you checked in:

uvx --python 3.13 --from doc-lattice==NEW_VERSION doc-lattice init

Replace your whole block with the printed one instead of hand-editing the pinned version in its two entry: lines. The block carries generated structure beyond those two commands, so bumping only the pins silently keeps an outdated hook shape.

Replacing the block does not need reactivation, because the installed hook reads .pre-commit-config.yaml on every commit rather than baking it in. An installation that never activated in the first place still has to; see Enabling the gates.

Run this from the directory that holds your existing .doc-lattice.yml, normally the repository root. Plain init resolves that file against the current directory rather than the Git root, so a run from a subdirectory does not see your config and scaffolds a second, nested one carrying default settings. From the right directory the run is safe against an existing install: init writes .doc-lattice.yml only when it is absent, and otherwise reports that the config already exists, leaves it untouched, and prints.

Ordinary installs

The same init run also prints the GitHub Actions workflow. Diff it against your checked-in .github/workflows/doc-lattice.yml, replace the file with the printed version, and commit it together with the refreshed pre-commit block.

The printed workflow's trigger branch is resolved per run, so pass --default-branch with the branch you actually gate on to make the upgrade reproducible rather than dependent on the local origin/HEAD of whichever checkout you ran it in. Either way, check the reported branch on stderr against your workflow before committing the replacement.

Recipe installs

Take the pre-commit block and the ordinary workflow from the plain init run above, passing --default-branch main as the recipe does at install time, since a recipe installation is pinned to main throughout and an omitted flag resolves the trigger against whatever origin/HEAD the checkout you ran it in has cached. Then replace your Linear workflow whole from the target release's MANAGED_CI.md step 2. Do not bump only its doc-lattice== pin: its structure and action pins can change between releases independently of the version it installs, which is the same reason the ordinary workflow is replaced whole.

Managed installs

A managed installation is one generated by the removed init --github before 5.0. It has no upgrade path, because its offline workflow invokes ci audit, a command 5.0 removed, so pinning it forward fails outright. It does not break on its own either: its workflows pin the exact release that generated them and go on running it, which is why nothing prompts the conversion.

Convert it instead. The procedure is in MANAGED_CI.md, it changes no remote state, and it is a local change of file ownership from the tool to you. Once converted, follow Recipe installs above for every later upgrade.

Linear integration

doc-lattice linear is the only network-touching command. It builds a trigger map from the loaded lattice, then fetches live ticket status over the Linear GraphQL API to report tickets that shipped against a spec that has since drifted. It reads LINEAR_API_KEY from the environment (export it before running; the error points you to impact for the offline view), and the client is https-only, redirect-refusing, size-capped, and SSRF-hardened. A transient HTTP 429 or 5xx gets two retries, for three total attempts. Without a usable Retry-After, retries wait 1 second and then 2 seconds. A non-negative integer Retry-After is honored up to the 30-second cap; negative, date-form, and invalid values use the fallback delay.

Security note: If linear is used in CI, use the protected GitHub setup. The command processes repository-controlled tickets and linear_team while LINEAR_API_KEY is present. Untrusted pull-request workflows should use only offline commands.

Canonical ticket ids are uppercase ASCII TEAM-NUMBER: TEAM starts with an uppercase letter and continues with uppercase letters or digits, while NUMBER is 0 or a decimal with no leading zeros. One linear run accepts at most 500 distinct ticket refs after its positional or --from scope is applied. Set the team the query targets with linear_team in .doc-lattice.yml, or pass --linear-team to init. Every other command runs fully offline.

Exit codes

Troubleshooting

LINEAR_API_KEY is not set. Only the linear command needs a key. Export a Linear API key (export LINEAR_API_KEY=lin_api_...) before running linear, or, when live Linear status is unnecessary, run impact instead: impact is the fully offline view of the same downstream reach and needs no key.

Linear returns HTTP 429 or 5xx. These are transient. The client makes at most three attempts, using the 1- and 2-second fallback delays or a capped, non-negative integer Retry-After. If it still fails, the error tells you to wait and re-run; impact stays available offline in the meantime.

A linear finding is BLOCKED not-found. A ticket the Linear filter does not return is treated as absence, not an error: it grades as a BLOCKED not-found finding rather than crashing the command. Confirm the ticket id exists and that linear_team targets the right team.

unclosed YAML frontmatter ... exits 2. A file beginning with --- must add another --- line after its YAML metadata. The message names the malformed file; a file with no opening fence remains ordinary untracked Markdown.

frontmatter in ... declares 'derives_from' but has no 'id' key exits 2. The file wired itself into the graph but never named itself, so it and every edge it declares would leave the lattice together. Almost always a typo in the id key (idd:, Id:) or an edit that dropped it: add the id back, or remove the lattice keys if the file is not meant to be tracked. The same message lists authority and tickets when those are what the block declared.

must not contain a control character; found U+000A at index ... exits 2. A frontmatter value decoded to a byte a terminal acts on, so the document is refused rather than loaded and printed. The fix depends on which code point the message names, and the three common ones do not share one. U+000A is a line break a block scalar kept, and chomping removes it only when it is trailing: change | to |-, or > to >-. A break between the lines of a value survives that edit, so if the error repeats, fold the value with >- and keep its lines adjacent, or put it on one line. U+0009 is a tab, which no chomping indicator touches: it is a literal tab or a \t escape in the value, and it has to come out. U+000D reaches a value only as a \r escape, since YAML folds a literal carriage return to a space, so it too has to come out rather than be re-chomped. Anything in the U+001B neighborhood is an escape sequence written into the value on purpose and should come out as well. The message names the key and the offending code point rather than repeating the value, so reading it never puts the byte back on your terminal. See the Frontmatter reference for the exact accepted set and for a scan that finds every spelling that reaches this rule, including the ones nothing on screen shows you.

skipping ... its frontmatter declares no 'id' on stderr. Not an error: a file with fenced frontmatter that declares no id and no lattice keys is left out of the lattice, and the exit status is unchanged. Expected when a docs root holds frontmatter belonging to another tool. Exclude the file with ignore_globs to silence it precisely; see Files with no id for why the PYTHONWARNINGS alternatives are blunter than they look.

duplicate id ... exits 2. A duplicate id makes the index incoherent, so loading the lattice fails with exit 2 (a tool error, distinct from the exit 1 that check and lint use for drift). The message names both registration sites so you can find the clash: either two files share an id, or two headings in one file resolve to the same anchor through equal markers or a marker/slug collision. Equal anchors in different files do not collide.

Documentation

Project structure

doc-lattice/
├── src/doc_lattice/         # the engine: a pure graph/report core behind a thin impure shell
│   ├── markdown_compat.py      # pinned heading and GitHub-slug compatibility adapter
│   ├── _github_slugger_data.py # generated slug and Unicode compatibility data
│   ├── persistence.py          # shared durable single-path filesystem primitives
│   ├── reconcile_transaction.py # reconcile lock, journal, commit, rollback, and recovery
│   ├── cli/                    # per-invocation runtime and one adapter per command
│   └── cache/               # phase-separated incremental load cache
│       ├── schema.py        # filesystem-free models and codec
│       ├── state.py         # filesystem-free run-local state
│       ├── lookup.py        # document reads and stats for cache-tier selection
│       └── store.py         # cache-file reads and atomic writes
├── tests/                   # test suite (mirrors sources; property-based hashing invariants)
├── scripts/                 # slug generation, section benchmark, repository and release tools
└── pyproject.toml           # project configuration

See ARCHITECTURE.md for module boundaries and their rationale.

License

MIT. See LICENSE.