Root cause by layer

Root-Cause-by-Layer — inner loop (∂L/∂code), the gradient computer

The Metaphor

The pathologist at the autopsy table. The visible symptom — the bleeding, the bruise, the collapsed lung — is never the disease. The disease lives in a failing organ, a contract between systems, a subtle pressure gradient two layers deep. A surgeon who sutures the bleeding without opening up dies a different patient next week. Five layers, always. Stop at layer 2 and the bug re-emerges in a different guise.

Overview

In Gradient Descent for Agents, this skill is how you compute ∂L/∂code. Without the 5-layer ladder, any proposed edit is a random direction in parameter space — on average it raises generalization loss even when it happens to lower training loss. A bug is a symptom. The disease almost always lives at a layer boundary, a contract violation, or a conceptual mismatch — rarely on the line that threw the exception. Walk the 5-layer ladder before proposing any fix.

Core principle: A change that turns the symptom green without a causal story at layers 3–5 is a suppression, not a cure. It re-emerges in a different guise within a few iterations.

Violating the letter of the layered analysis is violating the spirit. "I know the cause already, I'll skip writing it out" is the most common skip — write it down anyway. The discipline is the externalization.

When to Use

Any of these triggers means invoke this skill before proposing code:

• A test is red or flaky and you want to ship a fix
• An exception, type error, assertion failure, or unexpected output
• A CI gate is rejecting and you're considering widening the rule
• An LLM/tool returned a malformed result and you're considering a retry/regex
• You catch yourself typing try:, hasattr(, @pytest.mark.skip, # TODO clean up, or a retry loop as the primary fix
• Deadline / sunk-cost pressure is telling you "just get it green"

Do not use for: cosmetic edits, rename refactors, adding new features with no failing signal. This skill is for diagnosing, not building.

The Five Layers

Apply in order. You do not have a root cause until you can answer both 4 and 5 out loud.

1. Symptom — The exact error, failed assertion, rejected output, wrong number. Copy the raw text. Do not paraphrase.
2. Mechanism — Which code path produced the symptom? Name the function, branch, and state that led to it.
3. Contract — Which input/output shape, type, invariant, or documented promise did the mechanism violate? Whose contract — the caller's or the callee's? Is the contract explicit (signature/docstring/schema) or only implicit?
4. Structural origin — Which architectural layer is this in (domain, integration, transport, persistence, UI, config, …)? Is the bug inside that layer, or is it a leak across a boundary?
5. Conceptual origin — Which design concept is being misapplied: separation of concerns, ownership, single source of truth, explicit-over-implicit, data-flow direction, invariant location?

Only once 4 and 5 are named is a fix admissible. The admissible fix is at the layer you named — not above, not below.

Red Flags — STOP, you are about to symptom-patch

These phrases, appearing in your own thinking or answer, mean you skipped layers 3–5:

• "Just make it tolerant to both shapes"
• "Add a try/except so it doesn't blow up"
• "I'll clean up later / mark as tech debt / file a ticket"
• "It's a smaller change than fixing it properly"
• "The other callers still rely on the old shape" (have you grepped?)
• "Under time pressure this is good enough"
• "Widen the regex / threshold / whitelist"
• "Retry loop will handle the flakiness"
• "@pytest.mark.skip / xfail for now"
• "Hardcode the value that works"
• "Cache so we don't hit the broken path"
• "LLM was probably bad, just retry"

When one fires: stop, re-enter at layer 2, walk up.

Anti-Pattern Catalog

Anti-pattern	Why wrong	Do instead
try/except around the failing call	Hides contract violation; corrupted state flows downstream	Find the caller that violates the input contract; fix at that boundary
hasattr-or-subscript / duck-typing shim	Normalizes two implicit contracts into one ambiguous one	Pick ONE explicit type; convert at the boundary
Widening regex / threshold / whitelist until test passes	Treats the test as the spec; real spec is the contract	Ask what the rule intends; honor that
Retry loop around a flaky step	Normalizes non-determinism	Trace non-determinism to its source, remove it there
Hardcoding a "works for now" value	Freezes one case, breaks for the next	Derive the value from the contract that should produce it
@pytest.mark.skip / xfail to unblock	Converts a loss signal into silence	Fix root cause; if the test is wrong, fix the test with written justification
Patching log/UI so the error doesn't show	Error is still there, just invisible	Surface is a read-only view; fix reality
Adding a compat shim "for existing callers"	Usually there are NO existing callers	Grep for callers first; fix the single real caller
"LLM / network / filesystem was flaky"	Abdicates diagnosis	Capture the exact payload; diagnose which contract its output violates

How to apply — checklist

1. Read the symptom verbatim and paste it into your answer.
2. For each of layers 2–5, write one sentence. If you can't, read more code — do not guess.
3. If a Red Flag fired during step 2, delete your draft fix and restart at layer 2.
4. State the admissible fix at the layer you named in 4/5. Resist the urge to fix one layer higher (too shallow) or one lower (too invasive).
5. Add a regression test at the layer of the root cause, not at the layer of the symptom. A diagnosed-as-contract bug gets a contract-level test, not another end-to-end retry.
6. If time pressure is real, the minimal structural fix is almost always smaller than the "quick" symptom patch plus the follow-up cleanup it promises. Ship the structural fix.

Common Rationalizations

Excuse	Reality
"Deadline — clean up later"	Tech debt from symptom patches compounds faster than any deadline saves. "Clean up later" almost never happens, and when it does it's 3× the original cost.
"Make it tolerant so both shapes work"	Tolerance = two implicit contracts living together. Pick one explicit contract.
"Other callers might rely on old shape"	Did you grep? 90% of the time there are none — you invented them.
"Retry handles the flakiness"	Flakiness has a cause. Retry masks, it does not fix. You will see the same bug next week.
"Smaller diff is safer"	A 1-line symptom patch is usually a 3–5 line structural fix. Measure diffs in future bugs prevented, not in lines changed today.
"The test is wrong"	Sometimes true — but then write the correct test before you touch the code. Never delete a failing test without a replacement.

Measured failure mode

On the fixture tests/fixtures/root-cause-by-layer/ (a contract-boundary TypeError under 15-minute deadline pressure), a subagent without this skill ships:

email = user.email if hasattr(user, "email") else user["email"]

…labeled "tolerant," tagged as an "interim fix to be cleaned up later." Two implicit contracts, one masked boundary, one deferred cleanup that never happens — the canonical symptom patch.

With this skill loaded, the same subagent names layer 4 (domain ↔ integration boundary leak) and layer 5 (implicit-contract violation / separation-of-concerns), then fixes at the boundary by making the transport-layer signature explicit (email: str) and moving the extraction to the caller. No shim, no try/except, no deferred cleanup.

Δloss ≈ 6 rubric violations removed. Full baseline and GREEN artifacts in the fixture directory.