Planning: provider-aware selection and current orchestration positioning #19

Merged
creatidy merged 3 commits from planning/provider-aware-selection into main 2026-09-05 09:35:31 +00:00
creatidy commented 2026-09-05 09:32:20 +00:00 (Migrated from github.com)

Closes #18

Scope

Documentation/planning only. Records current product positioning and selector inputs before Task 3/M2 implementation.

Added to the plan

  • GitHub Project HydraFusion as current prior art: learn from selective single/cascade/critique workflows, complete accounting, bounded execution, isolated review, fail-safe application and validated routing without copying its prompt-executing Copilot runtime boundary.
  • Product-level anti-loop requirement: any future compound workflow recommendation must carry an explicit bounded execution envelope; Scarcity Router still recommends and does not execute.
  • Configurable timezone-aware provider/model blackout schedules, initially supporting the owner's policy to avoid Z.ai during configured peak hours. Peak/off-peak timing is configuration, not a hard-coded vendor assumption.
  • OpenAI banked Codex reset credits as replenishment/recoverable capacity, distinct from current quota and never silently consumed by the broker.
  • Advisory health-signal planning: official OpenAI status may supplement direct telemetry; direct local Ollama reachability remains authoritative; status.hellozai.com is explicitly rejected because it belongs to Zai Payments rather than Z.ai/GLM, and no official public Z.ai/GLM status page is currently assumed.
  • creatidy/ollama-monitoring remains a separate passive observability project; optional richer telemetry integration is deferred until it materially improves selection.
  • Artificial Analysis as periodically cached, provenance-bearing benchmark/performance evidence for catalog curation, not runtime capacity or a hidden per-selection oracle. API credentials remain server-side/outside repo artifacts.

Priority

Task 3 remains the next M1 implementation task. These additions do not expand the unified-status PR. Health-source ingestion and M2 selector work follow the core M1 status outcome.

Files

  • docs/roadmap.md
  • docs/competitive-landscape.md
  • docs/selection-policy.md

No product code, tests, collector changes, secrets, or Task 3 work are included.

Closes #18 ## Scope Documentation/planning only. Records current product positioning and selector inputs before Task 3/M2 implementation. ## Added to the plan - GitHub Project HydraFusion as current prior art: learn from selective single/cascade/critique workflows, complete accounting, bounded execution, isolated review, fail-safe application and validated routing without copying its prompt-executing Copilot runtime boundary. - Product-level anti-loop requirement: any future compound workflow recommendation must carry an explicit bounded execution envelope; Scarcity Router still recommends and does not execute. - Configurable timezone-aware provider/model blackout schedules, initially supporting the owner's policy to avoid Z.ai during configured peak hours. Peak/off-peak timing is configuration, not a hard-coded vendor assumption. - OpenAI banked Codex reset credits as replenishment/recoverable capacity, distinct from current quota and never silently consumed by the broker. - Advisory health-signal planning: official OpenAI status may supplement direct telemetry; direct local Ollama reachability remains authoritative; `status.hellozai.com` is explicitly rejected because it belongs to Zai Payments rather than Z.ai/GLM, and no official public Z.ai/GLM status page is currently assumed. - `creatidy/ollama-monitoring` remains a separate passive observability project; optional richer telemetry integration is deferred until it materially improves selection. - Artificial Analysis as periodically cached, provenance-bearing benchmark/performance evidence for catalog curation, not runtime capacity or a hidden per-selection oracle. API credentials remain server-side/outside repo artifacts. ## Priority Task 3 remains the next M1 implementation task. These additions do not expand the unified-status PR. Health-source ingestion and M2 selector work follow the core M1 status outcome. ## Files - `docs/roadmap.md` - `docs/competitive-landscape.md` - `docs/selection-policy.md` No product code, tests, collector changes, secrets, or Task 3 work are included.
Sign in to join this conversation.
No description provided.