/sdd:graph — Artifact Graph Skill

---

name: graph description: Build and query the SDD artifact graph. Use when the user wants to validate frontmatter edges, find impact/ancestors/chain for an ADR or spec, detect orphans or cycles, or backfill edges from prose. Currently supports validate / impact / ancestors / chain / orphans / cycles, with workspace-mode aggregation; backfill lands in Story 7. allowed-tools: Bash, Read, Glob, Grep, Task argument-hint: [] [--scope ] [--module ] [--table | --mermaid | --json]

/sdd:graph — Artifact Graph Skill

You are running /sdd:graph. This skill builds an in-memory directed graph of the project's ADRs, specs, and governed source files, then answers queries against it.

This skill differs from other SDD skills: instead of orchestrating Claude through markdown instructions, it delegates the deterministic build/validate/traverse work to a Python helper at lib/graph.py in this skill's directory. The output of the helper is the contract — JSON output (Story 6) is the stable shape any future MCP would consume. Markdown is reserved for orchestration and review UX (e.g., the backfill mode's accept/edit/reject in Story 7).

Process

0. Resolve artifact paths: Follow the Artifact Path Resolution pattern from references/shared-patterns.md to determine the ADR and spec directories. The resolved ADR directory is {adr-dir} and spec directory is {spec-dir}.

   In Story 2 the helper accepts a single root and a single ADR/spec dir per invocation. Workspace-mode aggregation (multiple modules, [module]/ID prefixes) lands in Story 5.

1. Parse the verb and flags from $ARGUMENTS.

   Currently supported: validate, impact <id>, ancestors <id>, chain <id>, orphans, cycles. Backfill (backfill) is recognized at the argparse layer and returns a clear "not yet implemented (planned for Story 7)" message with exit code 2.

   Traversal verbs require an artifact ID argument (e.g., ADR-0023, SPEC-0018, or [api]/SPEC-0001 in workspace aggregate mode). If the ID is unknown, the helper exits 1 and suggests closest matches.

   The --module <name> flag scopes a workspace-mode invocation to a single module: the helper builds only that module's graph with unprefixed IDs. Without --module in a workspace project, the helper aggregates all modules with [module]/ID prefixes. On single-module projects (no .gitmodules and no ### Workspace Modules table), --module is rejected with a clear error.

2. Locate the helper script.

   The helper lives at {skill-dir}/lib/graph.py, where {skill-dir} is the absolute path to this skill's directory (the Base directory for this skill line in the skill invocation header). The helper is invoked via python3 and reads from the project root passed via --root.

3. Invoke the helper.

   For validate:

   python3 {skill-dir}/lib/graph.py validate --root {project-root} [--adr-dir DIR] [--spec-dir DIR]
   

   For traversal verbs:

   python3 {skill-dir}/lib/graph.py impact {ARTIFACT-ID} --root {project-root}
   python3 {skill-dir}/lib/graph.py ancestors {ARTIFACT-ID} --root {project-root}
   python3 {skill-dir}/lib/graph.py chain {ARTIFACT-ID} --root {project-root}
   

   • {project-root} is the working directory (typically .).
   • {adr-dir} and {spec-dir} are passed only if Step 0 resolved a non-default location (e.g., a workspace module). For a single-module project, omit them and the helper defaults to docs/adrs/ and docs/openspec/specs/ under the root.
   • Traversal verbs refuse to run if validation has hard errors. Run validate first if the user reports unexpected output.

4. Present the helper's stdout to the user verbatim.

   The helper emits markdown directly. Do not reformat or summarize unless the user asks.

5. Surface the helper's exit code.

   • Exit 0: graph validates clean (no hard errors). Warnings, if any, are visible in the output.
   • Exit 1: hard errors. The graph is not queryable until they are fixed. Do not proceed to other verbs.
   • Exit 2: invocation error (bad arguments, missing root). This is a skill bug — surface it as such.

Workspace mode

Per SPEC-0014 § "Workspace Detection," the helper auto-detects workspace mode:

1. Look for .gitmodules in the project root (parses [submodule "name"] and path = entries).
2. Fall back to a ### Workspace Modules table in the project-root CLAUDE.md (Module / Root columns).
3. Otherwise treat the project as single-module.

For each discovered module, the helper resolves ADR/spec paths via the Artifact Path Resolution pattern (reads the module's CLAUDE.md for declarations, falls back to docs/adrs/ and docs/openspec/specs/).

Aggregate mode (no --module in a workspace): the helper builds each module's graph independently, prefixes every node ID with [module]/, and merges into a single graph. Cross-module edges authored as requires: ["[shared]/SPEC-0001"] resolve against the merged graph; cycle detection and ID resolution operate over the full aggregate so cross-module cycles are caught.

Module-scoped mode (--module foo): the helper builds only that module's graph with unprefixed IDs. Cross-module references in that module's frontmatter become unresolved-ID hard errors — this is intentional (use aggregate mode if you want cross-module behavior).

Single-module mode (no workspace detected): the helper operates on the project root with unprefixed IDs, identical to pre-Story-5 behavior.

Verbs

validate

Builds the graph and reports diagnostics. No ID argument required.

impact <id>

Renders a top-down ASCII DAG: queried artifact at the top, dependents flowing below via derived inverse edges. Direct dependents (one hop) are immediate children; transitive dependents are nested. Each edge is labeled with its derived inverse type and a (derived) annotation; the connector uses a dashed arrow ─ ─►. Default edge types for the source/target kind pair are unlabeled to reduce visual noise (e.g., governs for ADR→spec, implements for spec→ADR).

ancestors <id>

Renders ancestor paths in a single contiguous diagram with the queried artifact at the bottom (per SPEC-0018 § Layout rules). Each enumerated leaf-to-target path is rendered as a top-down chain (most-distant ancestor first, edge labels and ┆ continuation glyphs flowing down), separated by a blank line from the next path, with the queried artifact appearing exactly once at the bottom. The vertical connector uses the dashed glyph ┆ because the visual flow is the inverse of the authored relationship; edge labels reflect the derived inverse type with the (derived) annotation.

The vertical-stack approximation of multi-parent fan-in (vs. a side-by-side merging Y) is a tractable ASCII-only rendering — multi-parent inputs read as a sequence of independent ancestor paths feeding into the shared queried node.

chain <id>

Single contiguous bidirectional diagram: ancestors above (rendered as top-down chains with the target's title suppressed at the bottom of each), the queried artifact in the middle (rendered once as <title> (queried)), and impact below (rendered as a top-down indented tree). The two regions are visually joined by a single │ continuation through the queried node — no markdown subheadings, no ▼ glyphs.

orphans

Surfaces three categories of orphan as flat markdown tables (default for flat results per SPEC-0018):

1. Source files without governing artifacts — non-markdown source files in the project tree that contain no Governing: comment block. Discovered by a dedicated walk so these files do not become graph nodes (they remain invisible to traversal queries) but DO surface here. The walk uses the same exclusions as the graph builder (.git, node_modules, vendor, build/cache dirs, docs/, skills/, references/, etc.). Markdown files are skipped — they participate via frontmatter (ADRs, specs) or are out of scope for v1 (READMEs, ad-hoc docs).
2. Specs with no implementing code — specs that no source file's governing comment references.
3. ADRs with no implementing spec — ADRs that no spec declares implements: against.

Optional --scope <subtree> restricts category 1 to source files under the given path. Categories 2 and 3 always cover the full graph.

Operator-facing framing. A spec is flagged whenever no Governing: comment in source code references it; comment-less code is invisible by design (per SPEC-0018 § "Files without governing comments"). For repos that haven't yet attached governing comments to source code, expect every spec and ADR to be flagged. The output includes a one-line preamble explaining this so first-time readers know the verb is working as intended, not reporting a real catastrophe.

cycles

Lists any cycles detected during validation. Note: traversal/diagnostic verbs only run after validation passes (the hard-error gate in main()), so this verb's output in v1 is always "No cycles detected." The verb exists for tooling that wants to confirm cycle-freeness without running full validation.

backfill

Parses prose in ## Related, ## More Information, ## Overview, ## Decision Outcome, and ## Consequences sections to propose frontmatter edges, then writes only what the user explicitly accepts.

Sub-commands (selected by mutually-exclusive flags):

Flag	Behavior
(none)	Print proposals as a markdown report. Read-only — nothing is written.
--apply <ID> [<ID>...]	Write the proposed edges for the listed artifacts to their frontmatter. Existing edge fields are merged (deduped) rather than overwritten.
--reject <ID> [<ID>...]	Record rejections in .sdd-graph-backfill-skip. Skipped proposals stay hidden until --reset.
--reset	Delete the skip file so previously-rejected proposals re-appear in propose.

Heuristics for inferring edge types from prose:

Prose pattern	Inferred field	Notes
"supersedes X" / "replaces X"	supersedes	hard replacement
"extends X" / "modifies X" / "builds on X"	extends	builds-without-replacing
"enables X" / "unblocks X"	enables	downstream-unlocking
Bare ADR ref in ## More Information / ## Related (no qualifying verb)	related	weak association
ADR mentions SPEC in ## Decision Outcome / ## Consequences	governs	code path the decision shapes
Spec's ## Overview mentions ADR	implements	spec realizes the decision
Spec's ## Overview mentions another spec	requires	capability dependency

Rejection memory. .sdd-graph-backfill-skip (line-delimited node_id|field|target_id records) is operational state, not configuration — it MAY be .gitignored if you don't want rejection decisions versioned. The file is ASCII-safe and conflict-resilient (no merge churn within a single sorted record set).

Orchestration. When the user runs /sdd:graph backfill, the skill should:

1. Run python3 {skill-dir}/lib/graph.py backfill (read-only) to enumerate proposals.
2. For each artifact in the report, present the proposal to the user via AskUserQuestion with options Accept / Edit / Reject.
3. On Accept: run python3 {skill-dir}/lib/graph.py backfill --apply <ID>.
4. On Edit: use Read + Edit to modify the artifact's frontmatter directly. Do NOT use --apply for the edited proposal — write exactly what the user agreed to.
5. On Reject: run python3 {skill-dir}/lib/graph.py backfill --reject <ID>.
6. After the loop, run validate to confirm the resulting graph still passes.

The helper enforces SPEC-0018 REQ "Backfill Mode" guarantees: nothing is written without --apply, and --reject is the only path that mutates the skip file.

Output formats

Per SPEC-0018 § Output Formats, the helper supports four output formats. Defaults are shape-aware: hierarchical results (chain, impact, ancestors) default to ASCII DAG; flat results (orphans, cycles) default to markdown table. Format flags are mutually exclusive.

Flag	Format	Use case
(none)	Shape-default	Terminal viewing — ASCII DAG for hierarchical, markdown table for flat
--table	Markdown table	Force tabular output on hierarchical verbs (columns: ID, Type, Edge, Authored)
--mermaid	Mermaid flowchart	Visual format for embedding in docs / PR descriptions. Authored edges = -->; derived edges = -.->
--json	Versioned JSON	Machine consumption — the contract for any future MCP, IDE plugin, or dashboard

JSON schema

Every JSON response includes a top-level schema_version field. The current version is "1". Breaking changes require a new schema version and a versioned addendum to SPEC-0018 § Output Formats.

Traversal verbs (impact, ancestors, chain):

Each result entry's edges[] describes the relationship FROM the result TO the traversal subgraph (the queried artifact and any other visited node) — not the reverse. A result with both an authored implements edge and a derived governed-by edge to the queried artifact would emit both edges; a result with only one direction emits only that one.

{
  "schema_version": "1",
  "query": {"verb": "impact", "id": "ADR-0023"},
  "results": [
    {
      "id": "SPEC-0018",
      "type": "spec",
      "module": null,
      "title": "SPEC-0018: Artifact Graph",
      "edges": [
        {"type": "implements", "target": "ADR-0023", "derived": false}
      ]
    }
  ]
}

Error envelope (any JSON-mode failure):

{
  "schema_version": "1",
  "query": {"verb": "impact", "id": "ADR-9999"},
  "error": {
    "code": "unknown-artifact",
    "message": "unknown artifact `ADR-9999`",
    "suggestions": ["ADR-0023", "ADR-0022", "ADR-0021"]
  }
}

Distinguishable from success responses by the presence of a top-level error field instead of results. Other error codes: graph-has-errors (validation failed; see validate --json for details).

Diagnostic verb orphans:

{
  "schema_version": "1",
  "query": {"verb": "orphans"},
  "results": {
    "code_files_without_governing": [...],
    "specs_without_implementing_code": [...],
    "adrs_without_implementing_spec": [...]
  }
}

Diagnostic verb cycles:

{
  "schema_version": "1",
  "query": {"verb": "cycles"},
  "results": []
}

validate:

{
  "schema_version": "1",
  "query": {"verb": "validate"},
  "results": {
    "nodes": 41,
    "authored_edges": 2,
    "derived_edges": 2,
    "diagnostics": []
  }
}

JSON output uses sort_keys=True and indent=2 for byte-identical reproducibility. Each entry is fully self-describing — consumers do not need to re-read markdown to interpret the result.

What validate reports

Per SPEC-0018 REQ "Graph Validation", three classes of finding:

Finding	Severity	Trigger
Unresolved ID	error	An edge references an artifact ID that has no node in the graph
Cycle	error	A cycle exists in any acyclic edge type (supersedes, extends, enables, governs, implements, requires)
Status inconsistency	warning	An ADR/spec is supersedes-targeted but the target's status is not superseded (ADR) or deprecated (ADR/spec)

The helper also reports authoring-time mistakes that are not in the canonical spec but are useful in practice:

Finding	Severity	Trigger
Authored derived field	warning	Frontmatter declares a derived field (governed-by, implemented-by, etc.) that should be computed, not authored
Schema misuse	warning	A spec declares an ADR-only field, or an ADR declares a spec-only field
Malformed edge list	error	An edge field is not a list, or contains a non-string entry
Duplicate ID	error	Two files declare the same ADR-XXXX or SPEC-XXXX

Architecture

+---------------------+      +---------------------+
| docs/adrs/*.md      |      | docs/openspec/...   |
| (frontmatter edges) |      |    /spec.md         |
+----------+----------+      +----------+----------+
           |                            |
           +----------+    +------------+
                      |    |
                      v    v
          +-----------------------------+
          | lib/graph.py                |
          |  - parse frontmatter (YAML  |
          |    subset; stdlib only)     |
          |  - parse governing comments |
          |  - build directed graph     |
          |  - validate (3 checks)      |
          |  - derive inverse edges     |
          +--------------+--------------+
                         |
              +----------+----------+
              | stdout (markdown    |
              | by default)         |
              +---------------------+

The helper is stdlib-only — no PyYAML dependency. The frontmatter parser is intentionally narrow: it handles scalars and inline-bracket lists, which is all the SPEC-0018 schema declares. Extending it to nested mappings or block-list YAML is future work that should not be undertaken speculatively.

Python version. Requires Python 3.10+ (uses PEP 604 union syntax str | None and Path.is_relative_to).

Cross-references

• Schema: references/shared-patterns.md § "Graph Edge Resolution"
• Author surfaces: skills/adr/SKILL.md § "Graph Edge Frontmatter", skills/spec/SKILL.md § "Graph Edge Frontmatter"
• Code-edge format: references/shared-patterns.md § "Governing Comment Format"
• Canonical spec: docs/openspec/specs/artifact-graph/spec.md (SPEC-0018)
• Canonical decision: docs/adrs/ADR-0023-frontmatter-dag-and-graph-skill.md

Rules

• Verb output is the helper's stdout — do not paraphrase it.
• Hard errors (exit 1) MUST be surfaced before any other action. Do not run other verbs against a graph that failed validation.
• The helper is the source of truth. If the helper output disagrees with this SKILL.md, fix the SKILL.md or the helper, not the user's expectations.
• Do not bypass the helper to "do graph work in markdown." Determinism and byte-identity require the script.
• When extending the helper (Stories 3-7), preserve the existing Python module structure: discovery → parsing → models → construction → validation. New verbs add new entry points but do not rewrite the core.