Upstream: @opencode/plugin@2.2.5 + @opentui/core@0.8.1 + @opentui/solid@0.8.1 compatibility report #99

TL;DR: Nothing breaks. The five V2 plugins need only a pin bump (@opencode/plugin 2.2.5, plus @opentui/{core,solid} 0.8.1 where pinned). The V1-only prompt-inheritance plugin is unaffected. No published version breaks for users on OpenCode 2.2.5 or on OpenTUI 0.8.1, and no peer range needs tightening now. Upstream V1 prompts haven't changed. The prompt sync is still broken (#96) and has failed every day through Oct 31.

Base: main @ cb6d3dde2. Bumps: @opencode/plugin 2.2.0 → 2.2.5, @opentui/core + @opentui/solid 0.7.0 → 0.8.1 (v0.8.1 release, compare v0.8.0...v0.8.1). The SDK side through 2.2.5 is covered in #96, #97 and #98, and 2.2.5 is still the latest @opencode/plugin. This report covers the new OpenTUI 0.8.0 → 0.8.1 delta, plus the upstream OpenCode dev commits made since v2.2.5 that move the host to OpenTUI 0.8.1.

Per-plugin verdict

Gate results

I bumped the dev pins in 5 packages in a scratch copy of main. bun install resolved @opencode/plugin 2.2.5, @opentui/{core,solid} 0.8.1 and solid-js 1.9.17. I ran on Bun 1.4.4 with Node 24.22.2 and npm 11.6.2.

Some tests need a real node/npm on PATH: guard R11/R12, loader "root server wrapper… inert on import" and timeline "real Node Host.resolve". Without one they fail the same way on the unbumped baseline, so they aren't a regression signal.

Upstream changes that matter

OpenTUI 0.8.1 (release notes):

• Zero-size nodes are no longer drawn (#1869). When Yoga gives a node 0 rows or 0 columns, OpenTUI now skips drawing it. Before, it drew it as 1 cell on top of its siblings. This is the only change that can show in existing UIs. Watch-item: the ram-monitor /ram modal (RamModal in src/tui.ts) is a column of a "RAM Usage" title, a scrollbox (maxHeight: 24) and a close hint, and none of them sets flexShrink: 0. On a very short terminal Yoga can shrink the title or hint to 0 rows. They would then vanish instead of overlapping the list. That's cosmetic, and existing tests and smokes don't cover it. The sidebar widget already sets flexShrink: 0 where it matters. Optional hardening:

  // packages/opencode-ram-monitor/src/tui.ts (RamModal)
  createTextNode({ fg: () => theme().text, wrapMode: "none", truncate: true, flexShrink: 0 }, () => "RAM Usage"),
  // …and the same `flexShrink: 0` on the close-hint text node

• keymap useBindings keeps layer position on re-register (#1785), paired on the host side with opencode bf544cd41 / #59294 ("keep keymap command objects while their data is unchanged"). output-styles registers commands through context.keymap.layer(() => …), and double-tap-timeline calls context.keymap.dispatch("session.timeline"). Both only touch the host wrapper and neither depends on layer ordering, so the change doesn't affect either. It does fix churn: re-registered layers no longer jump to the top.
• Native stdout writer thread (Linux/macOS TTY, #1858). You can opt out with OTUI_STDOUT_THREAD=0. Close/restore over slow links is also more reliable (#1872, #1893). Both are host-level and transparent to plugins. ram-monitor cleanup is unaffected: it already uses onCleanup, and renderer.idle() still exists.
• Select / TabSelect / TextTable / LineNumber now paint into the frame (#1831, #1834, #1833). Translucent colors may blend slightly differently. No plugin here uses these components.
• markdown restyle fixes (#1857), and a ScrollBox-in-Markdown collapse and listener-leak fix (#1863). These are fixes that ram-monitor's markdown-in-scrollbox modal benefits from. They need no code change.
• The breaking changes are native-only: the ot_session_renderer_state struct grew to 40 bytes, and building now needs Zig 0.17. The packages here only use the JS API through the host, so neither breaking change applies.
• Additive APIs: VideoRenderable / <video>, crop on images, the frame:dropped event and maxWidth/ellipsis text draws. None of them is used here.

OpenCode host (after v2.2.5, on dev, unreleased):

• 8db3d852c (#60913) bumps the host catalog to OpenTUI 0.8.1 and raises the @opencode/plugin peers to @opentui/{core,solid} >=0.8.1 (still optional). OpenCode 2.2.5 itself ships OpenTUI 0.8.0, so the next release (2.2.6+) is the first host on 0.8.1.
• Config loading, #60803 (6b08407a5): config discovery now lists each ancestor directory once instead of probing every candidate. The candidate list (.claude, .agents, .opencode, …) and the symlink handling stay the same, so agents-loader's generated .opencode/ agent files are still found. Related changes: #60787 only rebuilds providers on account switch when a plugin's state changed, and #60921 adds in-memory skill:// URIs. All three are additive or perf-only for these plugins.
• Commands: the only change is a new CLI flag, --worktree/-w. No new TUI slash commands were added, so /ram, /output-style, /adversarial-review and /constructive-review still don't collide with anything. No commits since v2.2.5 touch plugin package resolution (Host.resolve).

Are published versions broken on the new upstream?

No. The published TUI bundles mark @opentui/* and solid-js as external, so they run on whatever OpenTUI the host ships. Every API they use (createElement/insert/spread, useKeyboard, SyntaxStyle, scrollbox/markdown) is unchanged in 0.8.1, and the tests and built-artifact smokes pass against it. Caveat: this rests on tests, smokes and the release notes. I didn't run a live OpenCode TUI session on OpenTUI 0.8.1, and no released host ships it yet.

Peer ranges

No tightening needed. Nothing here uses a 0.8.1-only API. Raising the @opentui/* peers to >=0.8.1 would exclude every released 2.2.x host (2.2.5 ships 0.8.0).

For the pin-bump PR (not applied):

// packages/{opencode-double-tap-timeline,opencode-ram-monitor,opencode-output-styles}/package.json → devDependencies
"@opencode/plugin": "2.2.5",
"@opentui/core": "0.8.1",
"@opentui/solid": "0.8.1",
// packages/{opencode-agents-loader,opencode-adversarial-review}/package.json → devDependencies
"@opencode/plugin": "2.2.5",

Bump all three together. The 2.2.5 SDK peers @opentui/* >=0.8.0, which the current 0.7.0 dev pins don't meet. Once the next SDK release lands with >=0.8.1 (#60913), an 0.8.0 pin won't meet it either, so go straight to 0.8.1. Peer strings stay distinct from the dev pins, as the AGENTS.md Dependabot rule requires.

Vendored prompt sync (opencode-agent-prompt-inheritance)

Did upstream prompts change? No. I ran bun run sync:prompts in the scratch copy:

• default ref dev: fails with default.txt: 404. dev has no packages/opencode/src/session/prompt/ any more.
• OPENCODE_UPSTREAM_REF=v1: fails with copilot-gpt-5.txt: 404, because #59669 deleted it.
• The last commit to the V1 prompt dir is still b01498d3f (2026-10-27). Compared byte for byte with v1, anthropic / beast / codex / gemini / gpt / kimi / trinity are identical. default.txt still has the known drift (548648a3d). No vendored file changed since Upstream: @opencode/plugin@2.2.5 + @opentui/core@0.8.0 + @opentui/solid@0.8.0 compatibility report #98.
• The Sync OpenCode Prompts workflow is still failing daily (Oct 28–31). The Upstream: @opencode/plugin@2.2.1 compatibility report #96 fix still applies:

  // scripts/sync-upstream-prompts.ts
  const ref = process.env.OPENCODE_UPSTREAM_REF ?? "v1"; // `dev` is V2-only
  // and drop "copilot-gpt-5.txt" from ALLOWED_FILES (removed upstream in #59669)

Method: I read the OpenTUI v0.8.1 release notes and the v0.8.0...v0.8.1 compare. I also checked the OpenCode dev commits since v2.2.5 and the npm registry metadata for @opencode/plugin and @opentui/*. Gates and smokes ran in a throwaway copy that has since been deleted. Nothing was pushed or committed.

Written by Devin