Skywatch Investigating Osprey Rules

Investigating Osprey Rules

Overview

You are investigating an Osprey SML rules project to produce a structured report. Your caller has provided a rules project path and an osprey-for-atproto repo path.

Output rules:

• Return ALL findings as text in your response. NEVER write files.
• Include exact file paths and line numbers for every finding.
• If something is missing or unexpected, report it explicitly — do not silently skip.
• Structure your report with the section headings defined below.

Investigation order:

1. Project Structure Inventory (this section)
2. UDF Discovery (Section 2)
3. Execution Graph Mapping (Section 3)

Produce each section in order. If a section cannot be completed (e.g., path inaccessible), report what's missing and continue to the next section.

---

Section 1: Project Structure Inventory

Validate the project directory structure, catalogue all SML files, extract the labels table from config/labels.yaml, and list model files with their key variable definitions.

Step 1.1: Validate Required Structure

Use glob and file_read to check for the required project components. Report each as present or absent.

Required components (check in this order):

| Component | Check | Required | |-----------|-------|----------| | main.sml | File exists at project root | Yes | | config/ | Directory exists | Yes | | config/labels.yaml | File exists | Yes | | models/ | Directory exists | Yes | | rules/ | Directory exists | Yes | | rules/index.sml | File exists | Expected but not fatal if absent |

Report format:

## 1. Project Structure Inventory

### 1.1 Structure Validation

Project path: /path/to/rules-project

  ✓ main.sml
  ✓ config/
  ✓ config/labels.yaml
  ✓ models/
  ✓ rules/
  ✓ rules/index.sml

If ANY required component is missing:

  ✗ models/ — MISSING (required)
  ✓ rules/

Report ALL missing components — do not stop at the first failure. Continue investigation with whatever IS present.

Step 1.2: List All SML Files

Use glob to find all .sml files in the project.

`glob` pattern: **/*.sml

Report format:

### 1.2 SML File Inventory

Found N .sml files:

  main.sml
  models/base.sml
  models/record/post.sml
  models/record/follow.sml
  models/label_guards.sml
  rules/index.sml
  rules/record/index.sml
  rules/record/post/index.sml
  rules/record/post/spam_detection.sml
  ...

List paths relative to the project root. Sort alphabetically.

Step 1.3: Extract Labels Table

Read config/labels.yaml and produce a table of all defined labels.

What to extract for each label:

• Label name (the YAML key)
• valid_for values (entity types: UserId, AtUri, PdsHost, Handle, etc.)
• connotation value (neutral, positive, or negative)

Report format:

### 1.3 Labels

Source: config/labels.yaml

  Label Name              | Valid For              | Connotation
  ----------------------- | ====================== | -----------
  alt-gov                 | UserId                 | neutral
  alt-tech                | AtUri                  | neutral
  amplifier               | UserId                 | neutral
  spam-post               | UserId, AtUri          | negative
  ...

Total: N labels defined

If config/labels.yaml does not exist, report:

### 1.3 Labels

  ✗ config/labels.yaml not found — cannot extract labels table

Step 1.4: Catalogue Model Files

Read each file in models/ (and subdirectories). For each file, extract the key variable definitions.

Variable types to identify:

| Variable Pattern | Type | Purpose | |-----------------|------|---------| | EntityJson(type='X', path='...') | Entity definition | Label target — used in LabelAdd/LabelRemove entity= | | JsonData(path='...') | Primitive extraction | Data values used in rule conditions | | Second = 1, Minute = 60, etc. | Time constant | Duration values for windows and deltas | | HasAtprotoLabel(entity=X, label='Y') | Label guard | Pre-computed check preventing re-labeling |

For each model file, report:

### 1.4 Models

#### models/base.sml

  Line  | Variable            | Type           | Details
  ----- | =================== | ============== | -------
  3     | UserId              | EntityJson     | type='UserId', path='$.did'
  5     | Handle              | EntityJson     | type='Handle', path='$.handle'
  8     | ActionName          | JsonData       | path='$.action', type=str
  12    | Second              | Time constant  | = 1
  13    | Minute              | Time constant  | = 60
  14    | Hour                | Time constant  | = 3600
  15    | Day                 | Time constant  | = 86400
  16    | Week                | Time constant  | = 604800

#### models/record/post.sml

  Line  | Variable            | Type           | Details
  ----- | =================== | ============== | -------
  5     | PostText            | JsonData       | path='$.record.text', type=str
  7     | PostUri             | EntityJson     | type='AtUri', path='$.uri'
  ...

#### models/label_guards.sml

  Line  | Variable            | Type           | Details
  ----- | =================== | ============== | -------
  4     | HasSpamLabel        | Label guard    | entity=UserId, label='spam'
  5     | HasBotLabel         | Label guard    | entity=UserId, label='bot'
  ...

Include line numbers for every variable definition. If models/ does not exist, report it as missing and skip this step.

---

Section 2: UDF Discovery

Discover all available User-Defined Functions (UDFs) that can be called from SML rules. Use dynamic discovery (reading Python source) when the osprey-for-atproto repo is accessible, falling back to static discovery (reading the reference file) when it is not.

Step 2.1: Attempt Dynamic Discovery

Try to locate register_plugins.py in the osprey-for-atproto repo path provided by the caller.

Expected location: {osprey-for-atproto-path}/plugins_atproto/src/register_plugins.py

Use file_read to check if this file exists.

If the file exists: Proceed with dynamic discovery (Steps 2.2-2.4).

If the file does NOT exist: Skip to Step 2.5 (Static Fallback). Report:

### 2.1 UDF Discovery Mode

  ✗ register_plugins.py not found at expected location
  → Falling back to static reference (may be outdated)
  Discovery mode: STATIC (low confidence)

Step 2.2: Extract UDF Class Names

Read register_plugins.py and find the register_udfs() function. Extract every class name returned in the list.

What to look for:

• A function decorated with @hookimpl_osprey named register_udfs
• It returns a list of class references (e.g., TextContains, IncrementWindow)
• The imports at the top of the file show where each class is imported from

Report format:

### 2.1 UDF Discovery Mode

  ✓ register_plugins.py found
  Discovery mode: DYNAMIC (high confidence)

### 2.2 Registered UDFs

  Found N UDFs registered in register_udfs():

  Standard:
    TextContains          ← udfs/std/text.py
    Tokenize              ← udfs/std/tokenize.py
    CleanString           ← udfs/std/censorize.py
    ...

  Cache:
    CacheGetStr           ← udfs/std/cache.py
    IncrementWindow       ← udfs/std/cache.py
    ...

  AT Protocol Query:
    DidFromUri            ← udfs/atproto/std/did_from_uri.py
    GetRecordURI          ← udfs/atproto/atproto.py
    ...

  AT Protocol Effects:
    AddAtprotoLabel       ← udfs/atproto/atproto_label.py
    RemoveAtprotoLabel    ← udfs/atproto/atproto_label.py
    ...

Group UDFs by the comment sections in register_udfs() (e.g., # Std, # Atproto std, # Atproto effects).

Step 2.3: Extract UDF Signatures

For each UDF class found in Step 2.2, locate its source file (from the import statements) and extract:

1. Arguments class: Find the class referenced as the first type parameter of

UDFBase[ArgumentsClass, ReturnType]. Read its fields to get parameter names, types, and defaults.

1. Return type: The second type parameter of UDFBase[ArgumentsClass, ReturnType].

How to identify arguments:

• Look for class ClassName(UDFBase[SomeArgs, ReturnType]):
• Then find class SomeArgs(ArgumentsBase): (or a subclass of ArgumentsBase)
• Each field annotation is a parameter: name: type or name: type = default
• Some argument classes inherit from intermediate classes (e.g.,

CacheWindowArgumentsBase extends CacheArgumentsBase which extends ArgumentsBase) — follow the inheritance chain to get all fields

Report format:

### 2.3 UDF Signatures

#### TextContains
  Source: udfs/std/text.py:47
  Arguments: TextContainsArguments (udfs/std/text.py:41)
    s: str                        (required)
    phrase: str                   (required)
    case_sensitive: bool          (default: False)
  Returns: bool

#### IncrementWindow
  Source: udfs/std/cache.py:364
  Arguments: IncrementWindowArguments (udfs/std/cache.py:279)
    Inherits from CacheWindowArgumentsBase → CacheArgumentsBase → ArgumentsBase
    key: str                      (required, from CacheArgumentsBase)
    window_seconds: float         (required, from CacheWindowArgumentsBase)
    when_all: List[bool]          (required, from CacheWindowArgumentsBase)
    max_ttl_seconds: Optional[float]  (default: None)
  Returns: int

Include line numbers for every class definition.

Step 2.4: Produce Signature Summary Table

After extracting all signatures, produce a summary table for quick reference:

### 2.4 UDF Signature Summary

| UDF | Parameters | Returns |
|-----|-----------|---------|
| TextContains | s: str, phrase: str, case_sensitive=False | bool |
| ForceString | s: Optional[str] | str |
| ExtractDomains | s: str | List[str] |
| ... | ... | ... |

Total: N UDFs with signatures extracted
Discovery: DYNAMIC from {osprey-for-atproto-path}

Step 2.5: Static Fallback

Use this step ONLY when dynamic discovery failed (Step 2.1).

Read the static reference file at: {skill-directory}/references/udf-signatures.md

(The skill directory is the directory containing this SKILL.md file.)

Report format:

### 2.1 UDF Discovery Mode

  ✗ Dynamic discovery unavailable
  → Using static reference: references/udf-signatures.md
  Discovery mode: STATIC (low confidence — signatures may be outdated)

### 2.2-2.4 UDF Signatures (from static reference)

[Include the full content of references/udf-signatures.md]

Always include the staleness caveat when using static fallback.

---

Section 3: Execution Graph Mapping

Trace the full execution graph from main.sml through all Import and Require chains. Catalogue every Rule() definition and every WhenRules() invocation to produce a complete map of what the rules project detects and what actions it takes.

Step 3.1: Trace from main.sml

Read main.sml at the project root. Extract:

1. Import statements: Import(rules=['path/to/file.sml']) — these load models

and definitions into scope

1. Require statements: Require(rule='path/to/file.sml') — these load rule

files into the execution graph

1. Conditional requires: Require(rule='path/to/file.sml', require_if=Condition)

— note the condition that gates loading

Then follow each Require chain recursively. For each required file:

• Read it
• Extract its Import and Require statements
• Follow those Requires in turn
• Continue until you reach leaf files (files with no further Requires)

Report format:

## 3. Execution Graph

### 3.1 Execution Graph Trace

main.sml
  ├── Import: models/base.sml
  └── Require: rules/index.sml

rules/index.sml
  ├── Import: models/base.sml
  ├── Require: rules/record/index.sml (require_if=IsOperation)
  └── Require: rules/identity/index.sml (require_if=ActionName=='identity')

rules/record/index.sml
  ├── Import: models/base.sml, models/record/post.sml
  ├── Require: rules/record/post/index.sml
  └── Require: rules/record/follow/index.sml

rules/record/post/index.sml
  ├── Import: models/base.sml, models/record/post.sml
  ├── Require: rules/record/post/spam_detection.sml
  └── Require: rules/record/post/profanity.sml

rules/record/post/spam_detection.sml  [LEAF]
  └── Import: models/base.sml, models/record/post.sml, models/label_guards.sml

rules/record/post/profanity.sml  [LEAF]
  └── Import: models/base.sml, models/record/post.sml

Mark leaf files (those with no Require statements) with [LEAF]. Include the require_if= condition for conditional requires.

Step 3.2: Catalogue Rule() Definitions

For each leaf file (and any file containing Rule() definitions), extract every Rule() definition.

What to extract:

• Variable name the Rule is assigned to (e.g., SpamContentRule)
• Whether the variable has a _ prefix (file-local, cannot be imported)
• The when_all conditions (list each condition on its own line)
• The description string if present

Report format:

### 3.2 Rule Definitions

#### rules/record/post/spam_detection.sml

  Line | Rule Variable          | Scope  | Conditions (when_all)
  ---- | ====================== | ====== | =====================
  12   | _IsNewAccount          | local  | AccountAgeSeconds < Day
       |                        |        | FollowersCount < 5
  18   | _HasSuspiciousContent  | local  | ListContains(list='spam_keywords', phrases=[PostTextCleaned]) != None
  24   | SpamDetection          | export | _IsNewAccount
       |                        |        | _HasSuspiciousContent
       |                        |        | not HasLabel(entity=UserId, label='verified')

#### rules/record/post/profanity.sml

  Line | Rule Variable          | Scope  | Conditions (when_all)
  ---- | ====================== | ====== | =====================
  10   | ProfanityDetected      | export | RegexMatch(pattern=r'...', target=PostText, case_insensitive=True)

Total: N rules across M files

Scope column:

• local = variable starts with _ (file-private)
• export = no _ prefix (importable by other files)

Step 3.3: Catalogue WhenRules() Invocations

For each file containing WhenRules(), extract:

1. Which rules trigger it: The rules_any= list
2. What effects fire: The then= list, including:

• Effect type (LabelAdd, LabelRemove, AtprotoLabel, DeclareVerdict, etc.)
• Target entity (e.g., UserId, AtUri, PdsHost)
• Label name or action
• Any conditional (apply_if=) or temporal (expires_after=) modifiers

Report format:

### 3.3 WhenRules Invocations

#### rules/record/post/spam_detection.sml

  WhenRules at line 30:
    Triggers (rules_any):
      - SpamDetection
    Effects (then):
      - LabelAdd(entity=UserId, label='spam_suspect')
      - LabelAdd(entity=AtUri, label='spam_content')

  WhenRules at line 38:
    Triggers (rules_any):
      - SpamDetection
    Effects (then):
      - AtprotoLabel(entity=UserId, label='spam', comment='Automated spam detection')

#### rules/record/post/profanity.sml

  WhenRules at line 16:
    Triggers (rules_any):
      - ProfanityDetected
    Effects (then):
      - LabelAdd(entity=UserId, label='profanity', expires_after=TimeDelta(days=7))

Total: N WhenRules invocations across M files

Step 3.4: Rule → Conditions → Effect → Label Summary

Produce a flat summary table linking every rule to its final outcomes. This is the most actionable output for rule-writing agents — it shows at a glance what each rule does.

Report format:

### 3.4 Rule Summary

| Rule | File | Conditions | Effect | Label | Entity |
|------|------|-----------|--------|-------|--------|
| SpamDetection | spam_detection.sml:24 | _IsNewAccount AND _HasSuspiciousContent AND NOT verified | LabelAdd | spam_suspect | UserId |
| SpamDetection | spam_detection.sml:24 | (same) | LabelAdd | spam_content | AtUri |
| SpamDetection | spam_detection.sml:24 | (same) | AtprotoLabel | spam | UserId |
| ProfanityDetected

…

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [skywatch-bsky](https://github.com/skywatch-bsky)
- **Source:** [skywatch-bsky/skywatch-agent-skills](https://github.com/skywatch-bsky/skywatch-agent-skills)
- **License:** MIT
- **Homepage:** https://bsky.app/profile/skywatch.blue

Install and usage instructions live in the source repository linked above.