Content
# specdrift
Deterministic spec lint for manifest-driven golden paths.
A golden path lives or dies by one promise: **the manifest is the single source of truth.**
Schemas validate the manifest's shape and static analyzers guard the code — but nothing
watches the space *between* artifacts, where "the manifest says X, the repo does Y" rots
silently. That gap is exactly where AI-generated changes decay. specdrift is the lint layer
for that gap, and it is deliberately boring: **it never calls an LLM — LLMs call it.**
```bash
dotnet tool install -g specdrift
specdrift validate .platform/manifest.yaml --schema manifest.schema.json --rules rules.yaml
```
## What it does (v1)
- **`validate`** — JSON-Schema validation of a YAML/JSON manifest **plus cross-field
invariants** you declare as data (`rules.yaml`): "when this path contains that value,
this other path must exist." Every finding has a rule id, a severity, and a message that
teaches the fix. Output: human text or `--format json`; exit code is the contract
(0 clean · 1 findings · 2 usage/IO).
Schema evaluation is an **in-tree, dependency-free evaluator** for the pragmatic 2020-12
subset manifests actually use (`type · enum · const · required · properties ·
additionalProperties · unevaluatedProperties · oneOf/anyOf/allOf/not · if/then/else ·
$ref(#/…) · pattern · min/maxLength · min/maximum · items · uniqueItems · min/maxItems`).
An unknown assertion keyword is a **hard fail** — the engine never guesses. Reports are
noise-free by design: a failed `if` merely deselects its branch, and a failed `oneOf`
yields ONE finding at the decision point, not one per alternative. (Why in-tree: every
candidate library failed this project's own gates — one ships a maintenance-fee EULA in
its NuGet binaries, one compiles code at runtime, one silently skips `if/then`.)
- **`drift`** — manifest ↔ repository reality, driven by a drift profile
(`.specdrift/drift.yaml`): a feature enabled with its package or wiring call absent
(each direction has its own finding — including "referenced but not enabled" dead
weight), committed API documents diverging semantically from built ones, and manifest
schema-version skew. Detection is TEXTUAL by design and says so; reports, never
auto-fixes.
```yaml
# .specdrift/drift.yaml
version: 1
manifest: .platform/manifest.yaml
schemaVersion: 1
wiring:
- feature: features.outbox
package: Platform.Messaging
call: AddPlatformOutbox
- feature: providers.auth # value-gated wiring
equals: openid
package: Platform.Auth
call: AddPlatformAuth
openapi:
- committed: specs/openapi.json
built: artifacts/openapi.json
```
- **`mcp`** — the same two verbs served over stdio MCP (`spec_validate`, `spec_drift`), so
coding agents ask the engine instead of guessing. Register it like any stdio server:
```json
{ "mcpServers": { "specdrift": { "command": "specdrift", "args": ["mcp"] } } }
```
## Profiles are data, the engine is generic
specdrift knows nothing about any particular platform. The schema, the invariant rules and
(coming with `drift`) the wiring tables are **profile data** that each golden path ships in
its own repository:
```yaml
# rules.yaml
version: 1
rules:
- id: SPEC0101
description: L2 caching needs a redis connection name
when: { path: features.distributedCaching.levels, op: contains, value: l2 }
require: { path: features.distributedCaching.redis.connectionName }
severity: error
message: "features.distributedCaching.levels includes 'l2' but redis.connectionName is missing - name the connection the cache should use."
```
Supported `when` ops: `exists`, `equals`, `contains` (array membership or substring).
Constraints: `require` (path must exist and be non-empty) and `forbid` (path must be absent).
A rule without `when` applies unconditionally.
## Determinism contract
- Same inputs → byte-identical report. No network, no clocks in output, no LLM.
- Unknown manifest schema versions are a **hard fail** — the engine never guesses forward.
## License
MIT.
