Skip to content

2026-04-native-first-convergence.md

Latest commit

 

History

History
48 lines (33 loc) · 2.11 KB

2026-04-native-first-convergence.md

File metadata and controls

48 lines (33 loc) · 2.11 KB

2026-04 Native-First Convergence

Problem

The previous stack combined five useful tools, but the overlap was becoming expensive:

  • prompt context grew too large
  • several tools tried to improve planning discipline in similar ways
  • the stack became harder to explain publicly than it was to use privately
  • strong models were slowed down by unnecessary baseline complexity

Research process

The convergence work happened in three passes:

  1. Source-level audit of the existing five-tool setup and its actual responsibilities.
  2. Two rounds of deep research to validate licenses, OpenCode capabilities, model economics, and deprecated configuration assumptions.
  3. Rebuild the environment around native OpenCode primitives, then verify the new configuration in a live local setup.

Key findings

  • OpenCode already supports custom agents, subagents, variants, commands, plugins, and OAuth-backed MCP integration.
  • permission is the correct control surface; older tools guidance is outdated.
  • Several previously recommended add-ons either overlapped heavily or introduced unnecessary always-loaded prompt weight.
  • ChatGPT Pro plus OpenCode's openai/* routing changed the cost equation for implementation-heavy work.
  • The best place to encode planning and TDD discipline was inside native commands and repo-local instructions, not in another large overlay.

Final stack

The resulting stack is:

  1. OpenCode native runtime for ownership and routing
  2. repo-local instructions for durable rules
  3. optional on-demand skills
  4. command workflow for planning, testing, implementation, review, and verification
  5. lightweight local plugins for output hygiene and continuity
  6. Sentry MCP plus optional external lifecycle systems for runtime truth and persistent artifacts

Practical outcome

The current workflow now has a clearer spine:

  • /plan clarifies requirements and writes PLAN.md
  • /spec turns acceptance criteria into test anchors
  • /implement executes against the plan
  • /review and /verify close the loop

This preserved the useful discipline of the old stack while removing much of the duplicated machinery.