feat: in-process Lua core handler for opencode (#47, phase 3)

[Cannon07]

Summary

Phase 3 of #47 for OpenCode — flips the OpenCode hook entry off bin/core-pre-tool.sh / bin/core-post-tool.sh and onto the in-process Lua orchestrator that landed for Claude Code in #63.

The TS plugin becomes a thin transport: it collects {tool, args, directory} from each hook firing, pipes JSON into a per-agent shell shim under backends/opencode/, which performs socket discovery and RPCs pre_tool.handle / post_tool.handle. All tool-name and camelCase→snake_case mapping moves out of TS into pre_tool.normalisers.opencode.

After this PR, only codex and copilot remain on the bash core handler.

Architectural decisions

Captured in docs/adr/0006-opencode-defers-os-independence-to-46.md. Highlights:

• Bash shim stays. TS-native socket discovery was on the table — OpenCode's TypeScript plugin is natively cross-platform and could have been the first integration to escape bash. We rejected it: two divergent discovery implementations would force Request Windows 11 support #46 (centralised RPC/discovery rewrite) to reconcile them. OpenCode's OS-independence is deliberately deferred to Request Windows 11 support #46.
• hookQueue kept, rationale rewritten. In-process Lua serialises within nvim's main thread, but TS→nvim send-order can still reorder concurrent before(A) / after(B) firings during socket discovery. The queue's load-bearing job changed; its existence didn't.
• Both hooks await the RPC. 15s execSync timeout bounds stuck-nvim. Timeouts log at debug before being swallowed; shim-resolution failures do the same.
• TS-side allowlist short-circuits read / glob / grep before the bash fork (OpenCode reads prolifically).
• bin-path.txt now resolves to plugin root, not the bin/ directory. Legacy values get a transitional fallback in TS (backends/opencode/index.ts:28 — comment marks it for removal in the next major).

Bug fix bundled in

post_tool.patch_paths had a latent crash: gsub returns (string, count), so table.insert(paths, custom:gsub(...)) was triggering the (t, pos, value) 3-arg form of table.insert. The unified-diff branch sidestepped it via assignment; the *** Update File: branch (GPT-class ApplyPatch shape) hit it head-on. Hit during opencode multi-file ApplyPatch testing. Fix: force single-value return with parens. Regression test added.

Test plan

• tests/run.sh — all 68 tests pass (plugin Lua + claudecode + opencode + codex + copilot backends)
• Manual: edit existing file
• Manual: create new file
• Manual: delete and edit simultaneously (Bash + Edit)
• Manual: ApplyPatch on multiple files (GPT-5.4 via OpenCode) — exposed the gsub bug, now fixed
• Manual: rapid-fire multi-file proposals (hookQueue ordering)
• Manual: rejection path closes previews and clears indicators
• Manual: visible-only mode suppresses previews but keeps neo-tree indicators
• Manual: nvim-unreachable → opencode falls back to native flow without hanging
• Manual: relative filePath resolves against cwd

Pre-merge note

backends/opencode/index.ts:28 has a v<X.Y> / v<X+1>.0 placeholder for the transitional bin-path.txt fallback comment. Fill in with the release version this lands as before tagging.

🤖 Generated with Claude Code