Skip to content

ROADMAP.md

Latest commit

 

History

History
86 lines (68 loc) · 3.93 KB

ROADMAP.md

File metadata and controls

86 lines (68 loc) · 3.93 KB

Pollux Roadmap

This roadmap is intentionally scope-constrained. Pollux is already past the "ship the core at all costs" phase; the next phase is to make the core more trustworthy, more legible, and harder to misuse.

  • Intent: communicate priorities and scope boundaries, not promises.
  • Last updated: 2026-04-22
  • Current status: v1.7 ships reasoning budget tokens, normalized cached_tokens usage reporting across providers, and a text-only local provider for self-hosted OpenAI-compatible servers. Options.delivery_mode is soft-deprecated and scheduled for removal in v1.8.0.
  • Status tracking: Issues and PRs are the source of truth for active work.

Product Strategy

  • Policy: Pollux is capability-transparent, not capability-equalizing.
  • Pollux owns orchestration primitives, not full application workflows.
  • Prefer stronger guarantees, clearer docs, and sharper errors over a broader public API.
  • New surface area should earn its place by removing repeated downstream boilerplate without hiding meaningful provider differences.
  • Upstream concerns stay upstream: model quality, pricing, provider-side latency, and feature availability belong to providers.
  • Downstream concerns stay downstream: scheduling, job queues, human review, storage, business workflows, and long-running agent policy belong to applications using Pollux.

Current Position

The core library now covers the main orchestration responsibilities it set out to own:

  • Stable entry points for realtime and deferred execution: run(), run_many(), defer(), defer_many(), inspect_deferred(), collect_deferred(), cancel_deferred().
  • Multimodal source handling across the supported cloud provider matrix, plus a text-only local provider for self-hosted OpenAI-compatible servers.
  • Source patterns: fan-out, fan-in, and broadcast.
  • Context reuse through explicit caching, implicit caching where exposed, and provider-managed prompt caching where available.
  • Structured outputs, tool calling, conversation continuity, and reasoning controls (reasoning_effort, reasoning_budget_tokens) where providers support them.
  • Usage surface normalized across providers, including reasoning_tokens and cached_tokens in the result envelope when providers report them.

Scope Philosophy

Pollux is demand-driven. New features ship when a concrete use case makes them necessary, not to pre-empt hypothetical needs or to match capability matrices from larger frameworks. Issues and PRs are welcome; the bar for expansion is whether a change makes an existing contract stronger or removes real downstream boilerplate, not whether it broadens surface area.

This keeps the roadmap intentionally short. "More features by default" is not the goal; the goal is a small core that stays trustworthy under real use.

Current Candidates

Concrete items currently on the shortlist. None are commitments, and the list is expected to stay short. Discovered follow-ups are tracked in GitHub issues.

  • Remove Options.delivery_mode in v1.8.0. The shim is soft-deprecated in v1.7 and emits DeprecationWarning on explicit use today.
  • Gemini flex inference tier — evaluate provider-specific support for Google's flex inference pricing tier (lower cost, higher latency).

High-Bar / Not Planned

The following are intentionally outside Pollux's target shape unless a very strong case emerges:

  • Workflow schedulers, background workers, cron abstractions, or queue systems
  • Vector stores, retrieval frameworks, or document indexing pipelines
  • Agent runtimes, planning frameworks, or memory systems
  • Automatic parity layers that hide fundamental provider limitations
  • Provider-agnostic knobs for every provider-specific feature
  • Cross-job recovery systems that own application-level orchestration

References

  • Provider contract: docs/reference/provider-capabilities.md
  • Testing philosophy: TESTING.md
  • Contributor guidance: docs/contributing.md