feat: Workflow Inference Engine — Derive Automation Candidates from Recorded Sessions

[Delqhi]

Summary

After recording user behavior, OpenSIN-Bridge must infer what can be automated by the bridge and by downstream agents.

Problem

Raw recordings are not enough. We need a logical layer that detects:

• repeated patterns
• stable element references
• deterministic vs ambiguous steps
• candidates for bridge macros / agent workflows

Goal

Transform captured behavior sessions into automation candidates:

• stable replayable steps
• bridge-native recipes
• fallback requirements (vision, shadow DOM, iframe, self-healing)
• confidence score for safe automation

Proposed Output Model

For each session:

• candidate_id
• workflow_name
• steps[]
• required_capabilities[]
• fragility_score
• automation_confidence
• human_review_required

Inference Questions

• Is the target element stable across snapshots?
• Does the action require visual matching or deterministic ref lookup?
• Are network calls a reliable success signal?
• Is the flow sensitive to auth/session state?
• Can the bridge do it directly, or does it require an agent workflow?

Expected Files / Surfaces

• extension/background/service_worker.js — candidate generation trigger / export
• server/src/index.ts — inference endpoint or storage API
• docs/ — workflow schema / review docs

Acceptance Criteria

• Recorded sessions can be transformed into structured automation candidates
• Confidence score + fragility score are generated
• Deterministic steps are separated from ambiguous/vision-dependent ones
• Result is consumable by OpenSIN agents via bridge/server APIs