Skip to content

ROADMAP.md

Latest commit

 

History

History
161 lines (127 loc) · 5.6 KB

ROADMAP.md

File metadata and controls

161 lines (127 loc) · 5.6 KB

ForgeSyte Roadmap

This is a rewritten, modern view of your roadmap, aligned with the current server reality: /v1/analyze, /v1/jobs, /v1/plugins, /v1/stream, and a job‑centric execution model. It preserves your intent but removes /run as a first‑class concept.

Milestone 1 — Core Plugin & Environment Baseline

  • Plugin load & environment alignment
    • Ensure all plugins (OCR, YOLO tracker) are installed in editable mode
    • Confirm CPU/GPU environments load the same plugin paths
    • Add diagnostics to print active plugin file and entrypoint resolution
  • BasePlugin contract
    • All plugins subclass BasePlugin
    • Handlers resolve from string names to bound methods
    • Contract validation passes end‑to‑end for each plugin

Milestone 1.5 — YOLO Tracker Operational Baseline

  • YOLO plugin wiring
    • YOLO tracker loads via entrypoints without errors
    • Model weights and device selection (CPU/GPU) are correctly configured
    • on_load() instantiates models and logs success
  • Image tracking
    • Reinstate player, ball, pitch detection and tracking
    • Validate JSON‑serializable outputs and schemas
  • Smoke test notebook
    • YOLO_Tracker_Smoke_Test.ipynb runs plugin on test images
    • Visualize detections and validate schema compliance

Milestone 2 — Real Integration Tests

  • Plugin discovery
    • Tests for entrypoints → registry → plugin instances
  • Execution paths
    • Tests for valid plugin execution via the canonical API
    • Error‑path tests: missing plugin, invalid args, bad input
  • Registry behavior
    • Ensure consistent plugin registration and reload behavior

(In the modern model, these tests should target /v1/analyze + /v1/jobs, not /run.)

Milestone 3 — Unified Execution Abstractions

  • Client‑side unification
    • Introduce a unified “analyze” helper (replacing runTool) that:
      • calls /v1/analyze
      • polls /v1/jobs/{id}
      • returns a stable result object
    • OCR and YOLO flows both use this unified path
  • Observability
    • Structured logging for each analysis invocation
    • Optional retry wrappers for transient network/JSON errors
    • Frame‑level metrics (duration, success, error) where applicable

Milestone 4 — MCP Adapter (On Top of Job Model)

  • MCP surface
    • Auto‑generate MCP schemas from plugin manifests
    • Map MCP tool calls to /v1/analyze + /v1/jobs
  • Testing
    • MCP integration tests for OCR and YOLO
    • MCP error‑path tests (invalid tool, invalid args, plugin errors)

Milestone 5 — Governance & Guardrails

  • Plugin invariants
    • CI: enforce BasePlugin inheritance
    • CI: enforce manifest exists for every plugin
    • CI: enforce at least one plugin loads successfully
  • Response invariants
    • CI: enforce JSON‑only responses from public endpoints
    • CI: enforce job results are always JSON‑serializable
  • Architecture invariants
    • CI: forbid reintroduction of legacy /run paths
    • CI: forbid bypassing the job model for long‑running work

Milestone 6 — Job‑Based Pipeline & Web‑UI Migration

  • Server
    • /v1/analyze as the single entrypoint for analysis
    • /v1/jobs, /v1/jobs/{id} for async tracking
    • /v1/plugins, /v1/plugins/{plugin}/manifest for discovery
  • Web‑UI
    • Upload flow uses analyzeImage + pollJob
    • Jobs view lists and inspects jobs
    • Results panel renders job results (not tool responses)
    • WebSocket streaming uses /v1/stream for live camera
  • Cleanup
    • Remove runPluginTool, ToolExecutionResponse, and tool‑centric execution assumptions
    • Keep tools only as schema/UX hints, not execution units

Milestone 7 — VideoTracker Full Implementation

  • Backend
    • Video ingestion, frame extraction, YOLO tracking, and aggregated job results
  • Job model
    • Progress, cancellation, and timeouts tuned for video workloads
  • Web‑UI
    • Video upload, job‑aware VideoTracker component, overlays, playback controls
  • Guardrails
    • Integration tests from upload → job → UI
    • CI checks for YOLO tracker correctness and schema stability

Milestone 8 — Phase 13: VideoTracker Multi‑Tool Pipelines

  • Architecture

    • Ordered pipelines of tools within a single plugin
    • REST endpoint /video/pipeline for frame-based execution
    • WebSocket support for streaming pipeline execution
    • Pipeline validation (plugin_id, tool existence, same-plugin constraint)

  • Backend Services

    • VideoPipelineService for ordered tool execution
    • Pipeline models and validation
    • Pipeline logging (step index, tool name, duration)
    • Update VisionAnalysisService to use pipeline model

  • Web‑UI

    • PipelineToolSelector component for multi-select tools
    • WebSocket frame payload includes plugin_id + tools[]
    • REST endpoint payload includes plugin_id + tools[]
    • No silent fallback to default tools

  • Governance

    • PHASE_13_FOLDER_STRUCTURE.md (mandatory file placement)
    • PHASE_13_PR_TEMPLATE.md (PR format and review checklist)
    • PHASE_13_MIGRATION_CHECKLIST.md (step-by-step implementation)
    • PHASE_13_DEV_ONBOARDING.md (developer guide)
    • pipeline_rules.md (pipeline governance rules)

  • Testing

    • REST pipeline tests
    • WebSocket pipeline tests
    • Pipeline validation tests
    • Logging tests (step information captured)
    • No cross-plugin pipelines possible
    • All tests must pass in CI

  • Sign-Off

    • All pipelines execute tools in order
    • No silent fallbacks
    • Full coverage of pipeline scenarios
    • Logging shows each pipeline step
    • Phase 13 locked in repo structure