typed-failures

name	typed-failures
description	Convert an app's untyped exceptions (ValueError, RuntimeError, bare Exception, ActivityError) to SDK-typed AppError subclasses so every classified failure carries category / code / audience / retryable on the wire. The Automation Engine attributes failures from those typed fields without parsing exception strings. Interactive and incremental: surveys the app first, asks the developer to confirm categories at ambiguous sites, applies edits one batch at a time, then reports remaining blind-spot (silent-swallow) sites as a phase-2 punch list for owner approval.
mandatory_triggers	["/typed-failures","add typed failures to this app","onboard typed errors","wire app to failure attribution"]
optional_triggers	["convert exceptions to AppError","add typed exceptions","make my errors typed"]
owner	connector-platform-team
last_updated	2026-05-11
staleness_days	90
inputs	[{"app_root":"auto-detected — the directory containing app/ and pyproject.toml"}]
outputs	["app/failures.py (typed classes + optional classifier)","edited raise sites across app/ (ValueError/RuntimeError → typed classes)","tests/unit/test_failures.py (wire-shape tests + classifier dispatch tests — match the app's existing test directory layout if different)","pyproject.toml (SDK pin bump to >= v3.7.0 if needed)","blind-spot report (printed, not committed — phase-2 punch list)"]
gates	["All 15 SDK-provided leaves come from application_sdk.errors. Apps may not invent new categories — only subclass leaves.","Codes follow <CATEGORY>_<SPECIFIC>, connector-agnostic. No vendor name in code.","No logic / control-flow changes unless explicitly approved. Silent-swallow paths report as phase-2 unless owner approves promotion."]

typed-failures

Onboard an app onto the SDK's typed AppError hierarchy so its failures surface as structured events through the Automation Engine. The end state: every existing raise site in the app raises a typed AppError subclass; bare exception handlers that re-raise classify-or-passthrough; the wire payload (ApplicationError.details) carries FailureDetails(category, code, audience, retryable, evidence); the AE parser surfaces those onto DAGNode.error_details and the emit pipeline lands them in otel_logs.service_logs as queryable typed attributes.

The canonical reference example is atlan-athena-app PR #57 — read it before running this skill on a new app to see the end-state shape and the conventions in practice.

---

Mission

For the app being onboarded:

1. Discover every existing exception raise / except / silent-swallow site.
2. Convert existing raises to typed AppError subclasses with the right (category, code, audience, retryable) — interactive Q&A only at ambiguous sites, auto-classify obvious ones.
3. Report the blind spots — silent-swallow paths and bare-passthrough sites that would benefit from typing but require logic changes the owner must approve.

Zero logic / control-flow changes by default. Only swap classes at existing raise sites. Promoting a silent-swallow to a raise is a logic change requiring owner sign-off — surface it in the phase-2 report instead.

---

When to use this skill

• Onboarding a new app onto failure attribution for the first time.
• Auditing an app that's already partially typed (e.g. some clients are typed but the activity layer isn't).
• After an SDK version bump that adds new typed leaves (refresh the app's failures.py if the new leaves cover existing raise sites better).

Not the right tool when:

• The owner has explicitly opted out of failure attribution for this app.
• The app is a one-off script with no Temporal workflows behind it.

---

SDK contract — the 15 leaves you may subclass

All leaves are defined in application_sdk.errors. You may not invent new categories. Apps subclass these and override code (per the convention below) and optionally audience (when the locus differs from the SDK base default).

Leaf class	Category	Default audience	Default retryable	Typical trigger
CancelledError	CANCELLED	APP_OWNER	False	Caller / workflow cancellation
AppTimeoutError	TIMEOUT	APP_OWNER	True	Bounded wait elapsed (network read, activity start-to-close)
RateLimitedError	RATE_LIMITED	USER	True	429 from remote, per-key quota signal
AuthError	AUTH	USER	False	Credentials missing / expired / invalid
AppPermissionDeniedError	PERMISSION	USER	False	Authenticated but no permission for the action
NotFoundError	NOT_FOUND	USER	False	Entity doesn't exist
AlreadyExistsError	ALREADY_EXISTS	USER	False	Idempotent-create conflict
InvalidInputError	INVALID_INPUT	USER	False	Argument / payload malformed regardless of system state
PreconditionError	PRECONDITION	USER	False	Inputs valid but state forbids action (schema mismatch, version conflict)
DependencyUnavailableError	DEPENDENCY_UNAVAILABLE	PLATFORM	True	Required service down (Dapr, Temporal, object store)
SourceUnavailableError	SOURCE_UNAVAILABLE	USER	True	Customer-controlled source system unreachable (their DB, SaaS API, on-prem endpoint)
ResourceExhaustedError	RESOURCE_EXHAUSTED	PLATFORM	True	Local resource limit (OOM, disk full, file handles)
DataIntegrityError	DATA_INTEGRITY	APP_OWNER	False	Returned data corrupt / violates invariants
InternalError	INTERNAL	APP_OWNER	False	App / SDK bug, invariant broken
UnimplementedError	UNIMPLEMENTED	APP_OWNER	False	Known capability gap

Litmus tests for ambiguous categories:

• DEPENDENCY_UNAVAILABLE vs PRECONDITION: if retrying the same call without any state change is expected to succeed → DEPENDENCY_UNAVAILABLE. If explicit state must change first → PRECONDITION.
• RATE_LIMITED vs RESOURCE_EXHAUSTED: 429 from a remote endpoint → RATE_LIMITED. Local OOM / disk full / worker-slot exhaustion → RESOURCE_EXHAUSTED.
• ALREADY_EXISTS vs PRECONDITION: entity already exists on an idempotent-create path → ALREADY_EXISTS. Resource exists but in the wrong state for the operation → PRECONDITION.
• UNIMPLEMENTED vs INTERNAL: known capability gap (feature not built) → UNIMPLEMENTED. Unexpected invariant violation (bug) → INTERNAL.
• INVALID_INPUT vs PRECONDITION: payload itself is malformed (irrespective of system state) → INVALID_INPUT. Payload syntactically valid but state blocks → PRECONDITION.

---

Code naming convention — <CATEGORY>_<SPECIFIC>

Codes are app-owned strings. The grammar is closed; the vocabulary is open within the grammar.

Rules:

1. Always start with the category name (AUTH_, PERMISSION_, DEPENDENCY_UNAVAILABLE_, etc.) so the code is self-describing without joining against the category column.
2. No vendor / connector name in the code. Vendor identity rides in the AE DAG node app_name label (set by AE, single source of truth) and in evidence. Write AUTH_INVALID_CREDENTIALS, not ATHENA_AUTH_INVALID_CREDENTIALS.
3. Reuse subtypes across connectors. Every connector with bad credentials should emit AUTH_INVALID_CREDENTIALS — same code, different app_name. This is what lets dashboards group by code to surface cross-connector patterns.
4. Stay generic. PERMISSION_ASSUME_ROLE_FAILED (works for any cloud-IAM connector), not PERMISSION_STS_ASSUME_ROLE_FAILED (AWS-specific). NOT_FOUND_COMPUTE_RESOURCE, not NOT_FOUND_WORKGROUP (Athena-specific).

Examples drawn from the Athena reference:

Code	Why this shape
AUTH_INVALID_CREDENTIALS	category prefix; "invalid credentials" is the cross-connector subtype
PERMISSION_DENIED	category prefix; no further subtype needed (this IS the common case)
DEPENDENCY_UNAVAILABLE_NETWORK	category prefix; "network" subtype distinguishes from other unavailability modes
INVALID_INPUT_CONFIG	category prefix; "config" distinguishes from data / query-shape input errors
INTERNAL_QUERY_PREP_FAILED	category prefix; "query_prep_failed" is specific enough to be a runbook entry

---

Property usage matrix

Every typed class produces a FailureDetails wire envelope via the SDK's AppError.to_failure_details(). Here's what each property is for and whether to set it.

Property	Layer	Use it?	How
category	ClassVar	✅ always	Inherited from each leaf's SDK base class
code	ClassVar	✅ always	Override per leaf following <CATEGORY>_<SPECIFIC>
audience	ClassVar	✅ override only when needed	Inherit from SDK base; override per leaf only when locus differs
default_retryable	ClassVar	⚪ rarely override	Inherit from SDK base; override only for unusual retry semantics
message	instance	✅ always	Set at every raise site with concrete human-readable text
cause	instance	✅ where applicable	Set via from exc chain at classifier sites; not needed when there is no upstream exception
evidence fields (dataclass attrs you add)	instance	✅ when useful	Add fields specific to the leaf (e.g. field on InvalidInput-style errors, db_name/schema_name on a query failure). These land in FailureDetails.evidence automatically.
app_name	instance	❌ deliberately NOT set	AE attributes via DAG node label — single source of truth. Setting it on every leaf creates a second source of truth and confuses Grafana cuts.
run_id	instance	❌ deliberately NOT set	AE attaches from Temporal context at ingest. The producer doesn't carry it.
suggested_action	instance	⚠️ optional, at raise site	Free to set at the raise site when there's specific guidance to convey ("regrant Glue read access"). Don't pre-default per leaf — canned text is rarely better than the activity-specific context the call site has.

Document the three deliberate non-uses in the app's failures.py module docstring (see the Athena reference for the exact wording).

---

Audience routing — who acts on a failure

The audience field is what AE uses to route alerts and SLA cuts. Closed three-value enum:

• USER — the customer must act. Credentials, permissions, source config, source-side network, customer-controlled resources.
• PLATFORM — infra ops must act. A shared Atlan service is unavailable (Dapr, Temporal, object store, pod OOM).
• APP_OWNER — the app team that wrote the failing code must act. Connector bugs, SDK bugs, unimplemented capabilities, data-integrity failures.

The deliberate override pattern: when an SDK base class's default audience doesn't fit your leaf's actual locus, override it. Example from Athena:

@dataclass(kw_only=True)
class AthenaSourceUnreachable(DependencyUnavailableError):
    """Athena endpoint not reachable.

    Audience overridden to USER (the SDK base defaults to PLATFORM) because
    source-side reachability is the customer's network / firewall, not the
    platform's.
    """
    code: ClassVar[str] = "DEPENDENCY_UNAVAILABLE_NETWORK"
    audience: ClassVar[Audience] = Audience.USER

Note: new code for an unreachable source raises SourceUnavailableError directly (already audience=USER), so no override is needed. The override pattern above still applies where no leaf fits the locus — e.g. a source-side timeout is an AppTimeoutError whose APP_OWNER default you override to USER.

If you're tempted to override audience more than once or twice per app, re-check the SDK base default — most apps don't need to override at all.

---

Phase 1 — Discovery (no edits)

When invoked, do this first. Make zero changes during discovery. End discovery with a structured report and explicit user confirmation to proceed.

Step 1.1 — Verify SDK version

The typed-errors infrastructure (application_sdk.errors) requires SDK ≥ v3.7.0. Check pyproject.toml [tool.uv.sources] block.

If the pin is older, propose a bump:

SDK pin is currently v3.0.1. Typed errors require >= v3.7.0. The bump pulls in temporalio 1.27 and OpenTelemetry 1.41. OK to bump?

If the user declines, abort — the skill cannot proceed without the typed errors package.

Step 1.2 — Survey raise sites

Grep / read every *.py under app/. Categorise every raise statement:

• Bare-builtin raises (raise ValueError("..."), raise RuntimeError(...), raise Exception(...)) — in-scope, must be converted.
• SDK ActivityError / legacy errors — in-scope, convert.
• Already-typed AppError subclasses — out of scope (already done).
• Re-raises in except blocks (except X: raise, except X as e: raise e, except X: raise e from X) — in-scope for classify-or-passthrough wrap if the underlying exception comes from a known source (boto, requests, db driver, file I/O).

For each in-scope site, capture:

• file:line
• 5 lines of context (the code before/after the raise)
• The current exception class + message
• The control-flow position (inside __init__, inside an activity body, inside a helper function, inside a finally block, etc.)

Step 1.3 — Survey except blocks

Categorise every except block:

• Re-raises (terminates with raise / raise e / raise X(...) from e) → in-scope for classifier wrap.
• Logs and continues (except X: logger.warning(...) with no raise) → silent-swallow blind spot. Capture for phase-2 report. Do NOT touch in phase 1.
• Returns / sets-default and continues → also silent-swallow. Capture.
• Per-record swallows (for row in batch: try: transform(row) except: skip) → intentional best-effort, low severity. Capture but mark as "likely fine".

Step 1.4 — Map the app's activity surface

Read the app's entry points to understand the architecture:

• app/<connector>_app.py or app/connector.py — the main App class.
• app/handlers/*.py — HTTP-side preflight / test_auth.
• app/clients/*.py — connection / query layer (likeliest source of typed raises).
• app/transformers/*.py — usually per-record swallows; rarely raise.
• Any @activity.defn or @task decorated functions — activity boundaries.

Note which raise sites are inside @task-decorated bodies (these reach AE directly via ApplicationError.details) vs inside HTTP handlers (these never reach AE — they surface as HTTP responses; typed classification still helps preflight UX).

Step 1.5 — Report and confirm

Print the discovery summary as a single message with three sections:

## Discovery summary

App: <name>
SDK pin: <current> → <target if bump needed>

### Existing raises (will be converted)
- N bare ValueError raises (likely InvalidInputError for missing-config patterns)
- M bare RuntimeError raises
- K legacy SDK error raises
- (sites listed by file:line with one-line context each)

### Except blocks that re-raise (candidates for classifier)
- Site list with the underlying exception source (boto / requests / db driver)

### Silent-swallow blind spots (phase-2 — owner approval needed)
- High severity: <count> — silent swallows in activity-level code that may
  mask production failures (activity returns success despite real failure)
- Medium: <count> — bare exception handlers that bubble untyped
- Low: <count> — per-record swallows, likely intentional best-effort

Ready to proceed with phase 2? (y/n)

Wait for explicit user confirmation before any edit.

---

Phase 2 — Conversion (interactive, batched)

Apply edits in this order: failures module first, then raise sites, then tests, then SDK pin bump, then verification.

Step 2.1 — Build app/failures.py

If the file doesn't exist: create it. Start with the module docstring explaining the convention and the deliberate non-uses. Use the Athena reference's docstring as a template — adapt the connector name.

If the file already exists (partial-typing case): read it first. For each class you're about to add, check whether a class with the same code or with semantically overlapping triggers already exists.

• Same code already defined → reuse the existing class, don't duplicate. If the existing audience / retryable / evidence shape doesn't match what you'd write, surface the mismatch to the developer and ask before changing it (existing callers may depend on the current shape).
• Same trigger but different code → ask the developer which code is canonical; pick one and remove the duplicate. Two codes for the same root cause splits dashboard cuts.
• Existing classifier function present → extend the marker tuples rather than building a parallel classify_<app>_v2.

Never overwrite an existing typed class silently. The file is owned by the app team; preserve their work and additively extend.

For each typed class you're going to need (new or added), define it in failures.py with:

• A docstring describing what triggers it and who acts
• code: ClassVar[str] following the convention
• audience: ClassVar[Audience] only if overriding the SDK base
• Evidence fields as dataclass attributes (kw_only, with None or sensible defaults)

If the app has marker-string-based exception classification (common in apps wrapping boto, requests, db drivers), define marker tuples and a classify_<app>_exception(exc) -> AppError | None function at the bottom of failures.py. Pattern from Athena:

AUTH_MARKERS: tuple[str, ...] = (...)
PERMISSION_MARKERS: tuple[str, ...] = (...)
NETWORK_MARKERS: tuple[str, ...] = (...)

def classify_<app>_exception(exc: BaseException) -> AppError | None:
    """Return a typed failure for exc, or None if no marker matches.

    Returns None when exc is already an AppError — don't reclassify.
    Caller pattern:
        classified = classify_<app>_exception(e)
        if classified is not None:
            raise classified from e
        raise
    """
    if isinstance(exc, AppError):
        return None
    msg = str(exc)
    if any(m in msg for m in AUTH_MARKERS):
        return <App>AuthFailed(message=msg, cause=exc)
    # ...
    return None

Why cause=exc AND raise … from e? They are two different fields serving two different consumers: cause=exc populates the SDK's wire envelope (FailureDetails.cause_repr — sanitised + length-capped string visible to dashboards and AE), while from e sets Python's __cause__ attribute (preserves the in-process traceback for local debugging and logger exc_info). Both are correct and both are needed; remove either and you lose one consumer.

Step 2.2 — Convert raise sites (auto-classify obvious, ask on ambiguous)

Auto-classify (no question needed):

• raise ValueError("X is required") / raise ValueError("X must be provided") → <App>InvalidConfig(message=..., field="X") — INVALID_INPUT / USER
• raise ValueError("Engine not initialized") / raise ValueError("X must be called first") → <App>EngineNotInitialized(message=...) / similar → INTERNAL / APP_OWNER
• raise RuntimeError("Failed to prepare ...") / template / formatting errors → <App>QueryPrepFailed(message=...) / similar → INTERNAL / APP_OWNER
• raise ActivityError(ATLAN_UPLOAD_ERROR ...) → AtlanUploadFailed(message=..., failed_count=N, total_count=M) → DEPENDENCY_UNAVAILABLE / PLATFORM
• except <ConnectionError|...>: raise after boto / requests / db calls → wrap with classify_<app>_exception(e) + passthrough

Ask the developer when:

• The same message could be one of two categories. Example: "Cursor is not supported" could be InternalError (programmer error) or UnimplementedError (known feature gap). Ask which.
• A new class name doesn't match an existing one. Show suggested name, ask to confirm or rename.
• The audience would differ from the SDK base default. Show the default, ask if override is intended.
• Evidence field naming. Show suggested fields, ask for additions.

Question format — keep it tight. One question per ambiguous site, with context and a default:

Site: app/clients/__init__.py:196 — raise ValueError("Cursor is not supported")

Context: SQLAlchemy returned a cursor we can't iterate. This is either: a) InternalError (default — PyAthena version mismatch / programmer error) b) UnimplementedError (we know this cursor type exists but haven't implemented it)

Suggested class: <App>CursorUnsupported(InternalError) Suggested code: INTERNAL_CURSOR_UNSUPPORTED Audience: APP_OWNER

a / b / customize?

Batch similar sites when they share the same pattern. Don't ask the same question 4 times for missing-config raises in the same file.

Step 2.3 — Apply edits

Use Edit for each site. Preserve the original message string verbatim (append it as the message= kwarg). Add evidence kwargs where the call site has obvious context.

Update imports — every modified file needs from app.failures import <classes> at the top.

For classifier-wrapped except blocks, change the pattern:

# before:
except Exception as e:
    logger.error("...", e)
    raise e

# after:
except Exception as e:
    logger.error("...", e)
    classified = classify_<app>_exception(e)
    if classified is not None:
        raise classified from e
    raise

Step 2.4 — Write tests

Create or update tests/unit/test_failures.py with two test classes:

TestTypedClasses — one test per leaf, asserting the wire shape:

def test_<app>_auth_failed_payload(self):
    fd = <App>AuthFailed(message="bad creds").to_failure_details()
    assert fd.category is FailureCategory.AUTH
    assert fd.code == "AUTH_INVALID_CREDENTIALS"
    assert fd.audience is Audience.USER
    assert fd.retryable is False

Cover audience overrides explicitly (a test that asserts the override took effect, e.g. AthenaSourceUnreachable.audience is Audience.USER not PLATFORM).

TestClassifier — parametrised tests over every marker in each tuple, asserting dispatch to the right typed class. Plus:

• test_unknown_message_returns_none — unrecognised exception → None
• test_does_not_reclassify_existing_app_error — AppError → None
• test_classified_carries_original_as_cause — from exc chain preserved

Update any existing tests that asserted bare ValueError / RuntimeError to assert the new typed class + check evidence fields.

Step 2.5 — Bump SDK pin (if needed)

If pyproject.toml [tool.uv.sources] pins the SDK to < v3.7.0, bump:

atlan-application-sdk = { git = "https://github.com/atlanhq/application-sdk", tag = "v3.7.0" }

Run uv lock and uv sync --all-extras --quiet. The lock-file diff will be large but mechanical.

Step 2.6 — Verify

Run the test suite:

uv run pytest tests/unit/ -q --no-header -p no:cacheprovider

All tests must pass (excluding pre-existing collection errors unrelated to this change). If a previously-asserting-ValueError test fails, update it — the test was right, the assertion just needs to be the new typed class.

For each new typed class, run the wire-payload check:

uv run python -c "
from app.failures import <App>AuthFailed
print(<App>AuthFailed(message='test').to_failure_details().model_dump())
"

Verify category / code / audience / retryable match the documented spec.

Step 2.7 — Run pre-commit (mandatory before any commit)

The SDK's CLAUDE.md mandates uv run pre-commit run --files <file> before any commit. The raise-site rewrites in Step 2.3 frequently touch ruff/pyright gates (unused imports from the old exception types, type-hint mismatches when constructors gain field= / evidence kwargs). Skipping this step ships edits that pass pytest locally but fail CI on lint / type.

Run on the modified files:

uv run pre-commit run --files \
  app/failures.py \
  tests/unit/test_failures.py \
  <every-modified-source-file>

Or on the full diff if simpler:

uv run pre-commit run --all-files

Fix any issues pre-commit surfaces (unused imports left from the old raises, formatter normalisations, type annotations on new evidence fields) and re-run until pre-commit passes. Do not commit until it does.

---

Phase 3 — Blind-spot report (the phase-2 punch list)

After all edits land and tests pass, produce the final report. This is the deliverable the app owner uses to decide what to address in the next PR.

Format — one section per severity, sites listed with file:line + one-line description + the typed class that would apply if the swallow were promoted to a raise:

## Blind-spot report — phase 2 punch list

These sites are currently silent-swallows or bare-passthrough. Closing them
requires adding try/except wrappers (logic change). They are intentionally
out of scope for this PR. The list is ranked by customer impact, not by
code change size.

### High severity — activity reports success despite real failure

- `app/<connector>_app.py:<line>` — extract loop catches all exceptions and
  logs; activity returns success with 0 records on auth failures. Would
  classify as <App>AuthFailed / <App>PermissionDenied / <App>SourceUnreachable
  depending on which marker matches. **This is the highest-impact gap.**

- (other high-severity sites)

### Medium — bare exception handlers bubble untyped

- `app/<connector>_app.py:<line>` — Glue API call has no error handling; bare
  botocore ClientError bubbles to AE as untyped ApplicationError.
- (other medium sites)

### Low — intentional per-record swallows

- `app/transformers/*.py:<line>` — per-record transform errors logged and
  skipped. Likely correct as-is; only worth promoting if every record fails
  (would classify as DataIntegrityError).

## Recommendation

Address the high-severity gaps first in a phase-2 PR. The pattern is small
per site (≤5 lines) but does change control flow, so each gap is an owner
decision: "do we want this activity to fail loudly, or silently report
green-with-zero?"

Do not commit this report. Print it as the final skill output so the developer can paste it into the PR description or share with the owner for phase-2 sign-off.

---

Anti-patterns to avoid

1. Inventing new categories. The SDK enum is closed. If the existing category doesn't fit, the right move is to refine the leaf's audience or evidence — not to add a new category. If you genuinely believe a new category is needed, that's a Chris/SDK-team conversation, not a skill output.

2. Putting the vendor name in code. ATHENA_AUTH_INVALID_CREDENTIALS is wrong. AUTH_INVALID_CREDENTIALS is right. Vendor identity rides on the DAG node label.

3. Pre-populating app_name / suggested_action / run_id on the class definition. All three have documented reasons to stay None by default. app_name and run_id are deliberately never set by the producer (AE provides both at consumption). suggested_action may be set at a raise site when the call site has specific guidance worth shipping ("regrant Glue read access on the IAM role"), but do not default it on the class itself and do not set it when the existing message already contains the action ("aws_role_arn is required" doesn't need a separate suggested_action saying "provide aws_role_arn"). Read the property matrix above and the Athena failures.py module docstring.

4. Promoting silent-swallow paths to raises in this skill's output. That's a logic change. It goes in phase 2 after owner approval. Even if you're confident the swallow is wrong (e.g. activity reports green-with-zero on auth failure), it goes in the report, not the diff.

5. Adding granular evidence for every leaf. Evidence is for context that helps debugging or routes downstream. field on an InvalidInput-style class is useful (tells customer which field to fix). created_at on every class is noise. Match the Athena reference's restraint.

6. Mixing classifier markers across categories. AUTH markers and PERMISSION markers are distinct on the wire (different categories). If a marker is genuinely ambiguous (AccessDeniedException could be auth or permission), pick the more specific one (PERMISSION — authenticated but not authorised) and document the choice.

7. Skipping the discovery report. Always show the user what was found before making edits. Different apps have different exception landscapes — the discovery report is what makes this skill non-blind.

8. Auto-overriding audience without asking. The SDK base default is usually right. If you're tempted to override on more than one or two leaves per app, double-check — you're probably misreading the locus.

---

Reference: the canonical example

atlan-athena-app PR #57 (branch feat/typed-failure-classes-develop, base develop) is the canonical reference. It includes:

• app/failures.py — 7 typed classes + classifier with marker tuples
• Raise sites converted across clients/__init__.py, clients/glue_api_client.py, handlers/athena_handler.py, <connector>_app.py
• tests/unit/test_failures.py — 35 tests covering wire shape + dispatch
• Handler refactor: _classify_and_raise delegates to classify_athena_exception and handles both bare and already-typed inputs
• PR description with the convention, the wire-flow trace, the blind-spot punch list, and the open questions for the owner

Read it before running the skill on a new app to anchor on the end-state shape and the level of restraint expected (e.g. how many leaves to create, how rich the evidence fields should be, how to phrase docstrings).

---

Final output checklist

When the skill completes, you should have produced:

• app/failures.py — typed classes + (optional) classifier
• Edited raise sites — every in-scope raise swapped for typed
• Updated imports across modified files
• tests/unit/test_failures.py — wire-shape + dispatch tests, all passing
• Updated existing tests that asserted ValueError / RuntimeError
• pyproject.toml SDK pin >= v3.7.0 (if needed)
• All unit tests pass (uv run pytest tests/unit/ -q)
• Blind-spot report printed for the developer
• PR description draft (template the Athena PR's structure)

If any item is incomplete, say so explicitly to the developer. Do not declare the skill done if anything in the checklist failed.