Skywatch Planning Osprey Rules

Planning Osprey Rules

This workflow guides you through gathering requirements and producing a structured rule specification before any SML code is written.

1. Validate Input Context

The orchestrator provides you with context before you begin. Verify you have everything you need.

Required inputs (provided by orchestrator in your prompt):

1. Investigator report — structured text report produced by the osprey-rule-writer agent's investigation phase (using skywatch-investigating-osprey-rules), containing project structure, labels table, model catalogue, UDF signatures, and execution graph map.
2. User request — what the user wants a rule for.

Validate the investigator report contains:

• Labels table (label names, valid_for, connotation)
• Model catalogue (variable names, types)
• UDF signatures (available functions)
• Execution graph map (Import/Require chains)

If the investigator report is missing or incomplete: Report what's missing and ask the orchestrator to re-dispatch the investigator. Do NOT proceed without project context.

2. Understand the Target Behaviour

Before writing any code, understand what the user wants to detect.

Ask clarifying questions:

1. What event type? (post, follow, identity, repost, etc.)
2. What signals? (text patterns, metadata, account age, etc.)
3. What label to emit? (must exist in config/labels.yaml)
4. Expiration/validity? (permanent, expiring, conditional)
5. Who/what gets labeled? (the account, the post, both?)

Map to labeling patterns:

• Chain to skywatch-osprey-sml-reference skill to look up common labeling patterns if you need naming conventions or syntax examples.
• Document the detection logic in plain English before writing code.

Example user request:

"I want to detect posts that contain profanity and label them with 'contains-profanity'."

Analysis:

• Event type: record (post)
• Signal: post text content contains profanity
• Label: contains-profanity (check labels.yaml to confirm it exists)
• Target: the post (AtUri)

3. Produce Rule Specification

After gathering requirements, produce a structured plain-text specification that the implementation agent can use to write SML.

Rule specification format:

## Rule Specification: [Rule Name]

**Target behaviour:** [What the rule detects]
**Event type:** [Which AT Protocol event triggers the rule]
**Signals:** [What data points the rule examines]
**Models needed:** [ML models or UDFs required, referencing available ones from investigator report]
**Labels to apply:** [Which labels from config/labels.yaml to use]
**Target entity:** [What gets labeled — account, record, etc.]
**Effect type:** [LabelAdd, AtprotoLabel, DeclareVerdict, etc.]
**Expiration:** [Duration if applicable, using named constants: Day, Hour, Week]
**Guard conditions:** [Re-labeling prevention if needed]

**Detection logic (plain English):**
[Step-by-step description of what conditions trigger the rule]

**Examples:**
- Should catch: [content examples]
- Should NOT catch: [counter-examples]

**Edge cases:**
- [Boundary conditions discussed with the user]

Confirm with user: Present the specification and get explicit confirmation before handing off to the implementation agent.

4. Skill Chaining

Load additional skills when you need specialized guidance during planning.

When to chain to skywatch-osprey-sml-reference:

• Need to look up available labeling patterns
• Unsure of naming conventions for the planned rule
• Need to understand what effect types are available
• Want to verify a UDF exists or understand its signature

Load with the skill tool: skywatch-osprey-sml-reference

5. Common Mistakes

These are planning-phase mistakes that lead to problems downstream.

1. Hardcoding label names not in config/labels.yaml

   • Wrong: Plan a rule using a label without checking if it exists
   • Right: Check config/labels.yaml using the investigator report first
   • Impact: Validation fails when implementation agent writes the effect

2. Not asking about the target entity

   • Wrong: Assume every rule labels the account
   • Right: Ask whether the account, the post, or both should be labeled
   • Impact: Wrong entity type in models, must rewrite

3. Skipping UDF availability check

   • Wrong: Plan a rule assuming a UDF exists
   • Right: Check the investigator's UDF catalogue for available functions
   • Impact: Implementation agent discovers missing UDF mid-authoring

6. Rationalizations to Block

Output: A confirmed rule specification in plain text. The osprey-rule-writer agent uses this spec when loading skywatch-authoring-osprey-rules to implement the SML, along with the investigator report.