Dependency Direction Analysis Skill

This skill runs the dependency direction detector locally and proposes sound architectural fixes for any violations found. It enforces two rules:

1. ddtrace.internal and ddtrace.contrib must not depend on product code. They are shared foundation layers; every product depends on them, so a dependency running the other way creates hidden coupling and risks circular imports (see the circular-import-analysis skill).
2. Products must not depend on each other directly. Tracing, AppSec, AI Guard, LLM Observability, Profiling, Dynamic Instrumentation, CI Visibility, Error Tracking, OpenFeature, OpenTelemetry, and Runtime metrics are each isolated: none of them are mandatory for a given dd-trace-py install, so one product can't assume another is present.

The guiding principle is the same as circular-import analysis: Separation of Concerns. Fixes must restructure ownership or add a decoupling layer, not paper over the problem with deferred imports.

When to Use This Skill

• The detect_layering_violations CI job reports new violations on your PR.
• You are adding a new module, or an import, that crosses from ddtrace/internal, ddtrace/contrib, or one product package into another product package.
• You are adding a brand new top-level ddtrace/<x> package or module and the CI job reports it as an uncovered/uncategorized top-level module.
• You are refactoring and want to verify you haven't introduced a new violation.

Running the Analysis

uv run --script scripts/import-analysis/layers.py analyze violations.json

This writes the results to violations.json and prints a summary to stdout. Requires uv on PATH (brew install uv or pip install uv). The output has two top-level keys:

{
  "violations": [ ... ],
  "uncovered": [ "ddtrace.newthing" ]
}

violations entries look like:

{
  "from": "ddtrace.internal.tracemethods",
  "to": "ddtrace.trace",
  "from_zone": "internal-core",
  "to_zone": "product:tracing",
  "score": 139,
  "in_tangle": true
}

• from / to — the two modules the violating import connects.
• from_zone / to_zone — which side of the rule they fall on (internal-core, contrib, or product:<name>).
• score — how bad this specific edge is (see "Severity scoring" below).
• in_tangle — the imported module is also part of a strongly connected component larger than one module, i.e. this violation is compounding an existing circular-import problem, not just crossing a boundary once.

uncovered lists direct children of the ddtrace package root (packages or .py modules) that are neither a key in layers.json's zones map nor listed in foundation.top_level. This is what catches a new top-level submodule that was added without anyone deciding which zone it belongs to — without it, a new package like ddtrace/newproduct/ would silently be treated as exempt foundation code and get zero dependency-direction enforcement. Unlike violations, a new entry here always fails CI on compare (see below), regardless of severity — it represents a config gap, not a graded issue.

To compare against the base branch the way CI does (new vs. pre-existing vs. worsened vs. removed, for both violations and uncovered modules):

uv run --script scripts/import-analysis/layers.py compare violations-base.json violations-pr.json

Clean up afterwards:

rm violations.json violations-base.json violations-pr.json

Zone Configuration

Zones are defined in scripts/import-analysis/layers.json, keyed by module prefix (longest match wins), so a product's own ddtrace.internal.<product> subpackage (e.g. ddtrace.internal.appsec) is carved out of the ddtrace.internal catch-all and treated as part of that product, not as foundation code. Modules with no matching prefix (e.g. ddtrace.ext, ddtrace.propagation, ddtrace.vendor) are unclassified "foundation" code and are exempt from every rule, both as importer and as imported module.

layers.json also has an exceptions list of zone-pairs that are deliberately exempt from the rules — this is how we record a considered decision without touching detection logic. For example, ddtrace/contrib/* modules are tracer integrations by design, so contrib -> product:tracing is listed as an exception rather than flagged on every run.

Only add an exception when the dependency is intentional and durable — not as a shortcut to make CI pass. If you're unsure whether an edge should be an exception or a bug, ask; this is a business/architecture decision, not something to infer from the code.

Fixing a new "uncovered top-level module" finding

When the CI job (or analyze) reports a new entry under uncovered, someone added a new direct child of ddtrace/ (a package or a .py module) that layers.json doesn't know about yet. Resolve it by editing scripts/import-analysis/layers.json:

• If it's a new product (mandatory-or-not feature area, isolated from other products), add it to zones as "ddtrace.<name>": "product:<name>", and add its ddtrace.internal.<name> counterpart too if one exists.
• If it's shared foundation code that everything may depend on and that itself has no restrictions (like ddtrace.ext or ddtrace.propagation), add it to foundation.top_level.
• If it's a carve-out of an existing product (e.g. a new ddtrace.internal.<product> subpackage), map it to that product's zone rather than leaving it to fall through to internal-core.

Don't add it to foundation.top_level just to silence the check — that defeats the point of the coverage check. Ask if it's unclear which zone fits.

Severity Scoring

Each violation's score combines three structural signals (no git history involved):

• Rule weight — internal-core/contrib violations start higher (3) than product-vs-product violations (1), because foundation code reaching upward is a worse inversion than two peers leaking into each other.
• Afferent coupling of the target (ca from betsy's ModuleMetrics) — how many other modules already depend on the module being imported. A violation that reaches into a heavily-relied-upon module has a bigger blast radius to eventually unwind.
• Cycle bonus (+5) — added when the imported module's nccd (from betsy) is greater than 1.0, i.e. it's already part of an import tangle. Fixing the layering violation first often makes the tangle easier to break too.

Use the score to prioritize: fix the highest-scoring violations first, especially any marked in_tangle.

Architectural Patterns for Fixing Violations

Never use deferred imports (import x inside a function body) as a fix. They hide the structural problem and impose a runtime cost on every call.

Understand the edge first

# What exactly does <from> import from <to>?
grep -n "^import ddtrace\|^from ddtrace" <path/to/from/module>.py

Identify the exact names crossing the boundary before choosing a fix — often only a small fraction of the target module is actually needed.

Pattern 1 — Core event bus (for contrib -> product violations)

When to use: A contrib integration wants to notify or be observed by a product (this is the most common shape for contrib -> product:X violations). This is the documented pattern in .cursor/rules/isolated-responsibility.mdc.

The contrib patch dispatches an event; it does not import the product:

from ddtrace.internal import core

core.dispatch(f"{event}.before", (kwargs,), allow_raise=True)
resp = func(*args, **kwargs)
core.dispatch(f"{event}.after", (kwargs, resp), allow_raise=True)

The product registers a listener, guarded by its own enable flag, inside its own package — not inside contrib:

from ddtrace.internal import core

def load_my_product():
    core.on("some.integration.before", _before_handler)

Neither side imports the other; ddtrace.internal.core is foundation code both may depend on.

Pattern 2 — Dependency inversion (for internal-core -> product violations)

When to use: ddtrace.internal needs to call into a product, but the product also needs to be the one driving behavior (e.g. registering a hook, supplying a callback).

Define a Protocol or abstract base inside ddtrace.internal (or a small neutral module); the product implements it and registers itself explicitly. ddtrace.internal depends on the abstraction, never on the concrete product package.

Pattern 3 — Extract shared types into a third, unclassified module

When to use: Two zones share a data type, constant, or protocol that both legitimately need, but neither should own.

Create a thin module outside both zones' prefixes (so it's unclassified foundation code, e.g. ddtrace._types or similar) containing only the shared contract. Both sides import from it; neither imports from the other.

Pattern 4 — Move the code to the zone that owns it

When to use: The violation exists because a function/class ended up in the wrong package. This is the simplest and often best fix.

If ddtrace.internal.tracemethods calls something that conceptually belongs to the tracing product, move it into ddtrace.trace/ddtrace._trace so the dependency direction reverses: the product depends on internal-core (allowed), not the other way round.

Pattern 5 — Question whether the target should be foundation code

When to use: A product-to-product violation involves a genuinely general-purpose utility that happens to live inside a product package (e.g. a formatting helper under ddtrace.trace that other products also want).

Move the utility down into ddtrace.internal (or an unclassified module) so every product can depend on it without depending on each other. Don't do this for anything that's conceptually part of the product's public contract (e.g. Tracer, Span) — those stay put, and the dependency on them should go through Pattern 1 or 2 instead.

Decision checklist before proposing a fix

1. Identify the exact cross-boundary names — grep the violating file.
2. Classify the relationship:
   • Contrib notifying/observing a product → Pattern 1 (core event bus)
   • internal-core needs product behavior → Pattern 2 (dependency inversion)
   • Shared data type/constant → Pattern 3 (extract)
   • Wrong home for the code → Pattern 4 (move)
   • Misplaced general-purpose utility → Pattern 5 (relocate to foundation)
3. Consider whether this is actually an intentional, durable dependency — if so, propose adding it to layers.json's exceptions list instead of restructuring code, but say so explicitly and explain why; this is a call for the humans reviewing the PR, not something to decide unilaterally.
4. Verify by re-running uv run --script scripts/import-analysis/layers.py analyze violations.json after the change and confirming the violation is gone (or, if compared against a saved base snapshot, that it doesn't appear as new).