Upgrade to AI SDK v3 spec, fix hanging thinking indicator, add --effort support

[Aptul9]

Summary

Core fixes required to make the plugin work correctly under AI SDK v6 / OpenCode v1.4.3, plus --effort flag passthrough. This PR is the V3 migration split out from the previous larger PR per maintainer request.

AI SDK v3 spec migration

• Upgrade specificationVersion from v2 to v3, eliminating the V2-to-V3 compatibility layer in AI SDK v6 that caused the thinking indicator to hang indefinitely in OpenCode
• Return usage in V3 format (nested objects with total/cacheRead/cacheWrite/reasoning) — fixes the usage2.inputTokens.total crash under AI SDK v6 / OpenCode v1.4.3
• Filter empty text blocks in message builder to prevent cache_control cannot be set for empty text blocks API errors
• Return empty stream when no user content is available (follow-up calls after tool execution), preventing API errors on requests with empty messages
• Always use finishReason: "stop" since the CLI executes all tools internally and returns complete results — prevents the AI SDK agentic loop from making unnecessary follow-up calls
• Upgrade @ai-sdk/provider to 3.0.8 and @ai-sdk/provider-utils to 4.0.23

Effort flag passthrough

• Add --effort flag passthrough: reads effort level from OpenCode thinking variant selection and passes it to the Claude CLI via --effort <level>

Stream event fixes

• Per-block text emission: emit text-start/text-delta/text-end for each CLI text block instead of a single pair for the entire turn. Each text block is saved immediately by OpenCode as a separate part, preventing loss of the final response text when the stream is aborted before the result event arrives from the CLI.
• providerExecuted on tool-input-start: add the providerExecuted flag to tool-input-start events so OpenCode marks CLI-executed tools correctly and does not re-loop looking for unresolved tool calls.
• Last-iteration token counts: use usage.iterations[-1] from CLI result instead of cumulative totals across all internal tool-use iterations — prevents inflated context counters (e.g. 780k instead of 80k) and premature compaction.
• Negative input token fix: when counting non-cached input tokens, compute total = input + cacheRead + cacheWrite so the OpenCode UI does not display negative values.
• Result fallback timer (5s): if the CLI emits a final text block but never sends a result event, the stream closes gracefully after 5 seconds with accumulated content.
• Abort grace period: on abort signal, start a 5s grace period instead of closing immediately, allowing in-flight CLI messages to be processed.
• Copilot review fixes: re-enable DTS generation, remove unused LanguageModelV3StreamPart import, remove ai from runtime dependencies.

OpenCode config example with thinking variants

{
  "provider": {
    "claude-code": {
      "npm": "opencode-claude-code-plugin",
      "models": {
        "sonnet": {
          "name": "Claude Code Sonnet",
          "limit": { "context": 1000000, "output": 64000 },
          "capabilities": { "reasoning": true, "toolcall": true },
          "variants": {
            "low": { "thinking": { "type": "adaptive" }, "effort": "low" },
            "medium": { "thinking": { "type": "adaptive" }, "effort": "medium" },
            "high": { "thinking": { "type": "adaptive" }, "effort": "high" },
            "max": { "thinking": { "type": "adaptive" }, "effort": "max" }
          }
        }
      }
    }
  }
}

Test plan

• Verified stream events with AI SDK v6 streamText + fullStream (correct event ordering)
• Verified stream completes with maxSteps: 5 (no unnecessary follow-up steps)
• E2E test with 28 tool calls: final text (1128 chars) arrives correctly
• E2E test with exact user prompt (21 tools): final text (2556 chars) arrives correctly
• Tested in OpenCode v1.4.3: no API errors, no hanging thinking indicator, no truncated responses
• Tested effort level selection in OpenCode UI
• Token counts no longer inflated (last iteration only)

Follow-ups

Two independent follow-up PRs split from the previous combined submission:

• Image/vision support in user messages
• Helper script to import existing Claude Code .jsonl sessions into the OpenCode DB