skywatch-fixing-osprey-rules

name	skywatch-fixing-osprey-rules
description	Use when fixing Osprey SML validation errors or reviewer-identified issues. Contains error categories, fix patterns, and debugging workflow. Not triggered on general coding tasks — only when resolving specific SML errors.
polytoken	{"tags":["skywatch"]}

Fixing Osprey Rules

This skill provides error categories, diagnosis patterns, and fix workflows for resolving Osprey SML validation errors and reviewer-identified issues.

Input Expectations

You receive a structured issue report from the reviewer. Each issue has:

• File path and line number
• Error description
• Severity (Critical, Important, Minor)
• Layer and check reference

Fix ALL issues in one pass. Do not fix one, re-review, fix another.

1. Understanding osprey-cli Error Output

osprey-cli reports errors in this format:

Error at <file>:<line>:<column>: <message>

Key elements:

• File path is relative to the rules project root
• Line and column numbers are 1-indexed
• Error messages describe the validation failure

Exit codes:

• 0: validation passed
• Non-zero: one or more errors found

The --dry-run flag validates without pushing. Always use it during development.

2. Type Mismatches

Error patterns:

• incompatible types in assignment
• found multiple different types in list literal

Root cause: Mixing RuleT and bool in when_all lists.

Type rules:

• RegexMatch(...), comparisons (>, <, ==, !=), or/and on bools → bool
• Rule(...) produces RuleT
• RuleT or RuleT is also RuleT
• not Rule(...) produces RuleT; not bool_val produces bool
• All items in when_all must be the SAME type

Fix option (a): If most items are RuleT, wrap booleans in Rule():

# Before (mixed types - wrong):
Rule(when_all=[RuleA, SomeBool > 5, RuleB])

# After (all RuleT):
_ThresholdRule = Rule(when_all=[SomeBool > 5])
Rule(when_all=[RuleA, _ThresholdRule, RuleB])

Fix option (b): If most items are bool, keep everything as bool:

# Before (mixed types - wrong):
Rule(when_all=[RuleA, SomeBool, AnotherBool])

# After (all bool - if RuleA can be decomposed):
Rule(when_all=[SomeBool, AnotherBool, SomeOtherBool])

3. Import Cycles

Error: import cycle detected here

Root cause: File A imports file B, which imports file A (directly or transitively).

Fix option (a): Extract shared definitions to a new file imported by both:

# Before: A imports B, B imports A
# After:
# shared.sml ← common definitions
# A imports shared.sml
# B imports shared.sml

Fix option (b): Move the definition causing the cycle to the file that needs it.

Fix option (c): Re-evaluate whether the import is actually needed.

4. Undefined Variables

Error patterns:

• unknown identifier '<name>'
• unknown local variable '<name>'

Root cause (a): Missing import — the file defining the variable is not imported. Fix: Add the missing Import(rules=['path/to/file.sml']).

Root cause (b): Underscore-prefixed variable used cross-file — _PascalCase variables are file-local and cannot be imported. Fix: Rename without underscore prefix if it needs to be imported, or duplicate the definition in the consuming file.

Root cause (c): Typo in variable name. Fix: Correct the spelling to match the definition.

5. Function Call Errors

Error patterns:

• unknown function '<name>'
• unknown keyword argument '<name>'
• missing keyword argument(s) '<name>'
• invalid argument type for '<name>'

Fixes:

• Unknown function: Check UDF signatures (from investigator report or reference). The function may not exist or may have a different name.
• Unknown keyword: Check the function's accepted parameters. Remove or rename.
• Missing keyword: Add the required parameter.
• Invalid type: Check the expected type and convert.

6. Duplicate Definitions

Error: features must be unique across all rule files

Root cause: Same variable name defined in multiple files that are both imported.

Fix:

1. Identify which file is the canonical source for the variable
2. Remove the duplicate definition from the other file
3. Import the canonical source instead

7. Rule Constraint Errors

Error patterns:

• rules must be stored in non-local features — Rule assigned to _ prefixed variable
• variable interpolation in non-format string — description missing f prefix

Fixes:

• Non-local features: Remove _ prefix from Rule variable names. Rules must be importable (non-local) so they can be used in WhenRules and when_all.
• Non-format string: Add f prefix to Rule description: description=f'...'

8. Debugging Workflow

When you receive issues to fix, follow this process:

1. Categorise issues by root cause. Multiple issues may share the same root cause. Fix the root cause once rather than fixing each symptom individually.

2. Fix in dependency order. If fixing issue A would also fix issue B (cascading error), fix A first. Cascading errors are common with import cycles and type mismatches.

3. Re-validate after EVERY fix. Run uv run osprey-cli push-rules <path> --dry-run after each change to confirm the fix worked and didn't introduce new errors. This is a self-check, not the formal gate.

4. Handle cascading errors. When osprey-cli reports many errors, often only 1-3 are root causes and the rest are cascading effects. Fix the earliest/deepest error first and re-validate — many other errors may disappear.

5. If a fix introduces new errors, revert and try a different approach.

9. Quick Reference Table

Error	Likely Cause	Fix
incompatible types	Mixed bool/RuleT in when_all	Wrap bools in Rule() or decompose RuleT
import cycle	Circular imports	Extract shared defs to new file
unknown identifier	Missing import or _ prefix	Add import or rename variable
unknown function	Nonexistent UDF	Check UDF reference
features must be unique	Duplicate variable names	Remove dupe, import canonical
non-local features	Rule in _ prefixed var	Remove _ prefix
non-format string	Missing f prefix	Add f to description string

10. Examples — End-to-End Debugging

Example A: Type Mismatch

Error:

Error at rules/record/post/spam.sml:15:3: found multiple different types in list literal

Diagnosis:

1. Open the file, look at line 15
2. Find the when_all list
3. Identify which items are bool and which are RuleT
4. Choose fix option (a) or (b) based on majority type

Fix: Wrap the boolean in a Rule:

_IsNewAccount = Rule(when_all=[AccountAge < Day * 7])
Rule(when_all=[_IsNewAccount, SpamPatternRule])

Example B: Import Cycle

Error:

Error at rules/record/post/index.sml:3:1: import cycle detected here

Diagnosis:

1. Trace the import chain from the error file
2. Find where the cycle closes
3. Identify the shared definition causing the cycle

Fix: Extract shared definition:

# New file: models/shared_post_features.sml
Import(rules=['models/base.sml'])
PostText: str = JsonData(path='$.record.text', required=False)

Example C: Multiple Cascading Errors

Errors:

Error at models/record/post.sml:5:1: unknown identifier 'BaseModel'
Error at rules/record/post/spam.sml:8:1: unknown identifier 'PostText'
Error at rules/record/post/spam.sml:12:3: incompatible types in assignment

Diagnosis:

1. First error is the root cause — missing import in model file
2. Second error cascades — PostText can't resolve because model failed
3. Third error cascades — type inference fails because PostText is unknown

Fix: Fix only the first error (add missing import), re-validate. The other two errors will likely disappear.

Rationalizations to Block

Rationalization	Reality	Action
"I'll fix them one at a time"	Multiple issues may share a root cause. Fixing individually wastes cycles.	Categorise by root cause, fix all in one pass.
"This error doesn't matter"	All errors block the gate. There is no "doesn't matter."	Fix every issue you received.
"I'll skip the self-check"	Self-checking confirms your fix worked before the formal review.	Always run osprey-cli after fixing.
"I'll fix extra issues I found"	You fix what you're given. The reviewer identifies issues, not you.	Only fix the issues in your report. Don't expand scope.
"The cascading errors are separate issues"	They're almost always symptoms of one root cause.	Fix the deepest error first, re-validate, then check if others resolved.

---

Output: Report which issues were fixed and how. The orchestrator will dispatch the reviewer to formally re-verify.