feat(core): add experimental ai-sdk-agent provider

[christso]

Summary

AgentV can now run an experimental provider: ai-sdk-agent target that uses Vercel AI SDK model/tool-calling primitives as a coding-agent loop rather than a one-shot LLM call. The provider owns the minimal repo tools (read, write, exact-string edit, and bounded bash) and lazy-loads ai / @ai-sdk/openai only when the target is invoked, falling back to a managed ~/.agentv/deps/ai-sdk-agent install path.

The target config accepts OpenAI-compatible base_url, api_key, model, optional temperature, max_steps, tools, system_prompt, and grader_target, with docs marking the provider as experimental.

Design Notes

• max_steps is AgentV harness configuration, not Vercel case metadata. The Vercel AI SDK exposes primitives; AgentV supplies the bounded loop through stopWhen: stepCountIs(...).
• ai and @ai-sdk/openai remain optional startup dependencies. The core and CLI bundles externalize them so normal AgentV startup does not require the AI SDK packages.
• The first-slice edit tool intentionally supports exact string replacement only; richer patching is left for a later provider hardening pass.

Validation

• bun --filter @agentv/core test test/evaluation/providers/ai-sdk-agent.test.ts test/evaluation/providers/targets.test.ts -> 74 pass, 0 fail.
• bun --filter @agentv/core typecheck -> passed.
• bun run lint -> passed across 759 files.
• bun --filter @agentv/core build -> passed.
• bun --filter @agentv/web build -> passed.
• bun --filter agentv build -> passed. Existing warning: dashboard dist not found, skipped.
• Fast AgentV smoke with provider: ai-sdk-agent -> passed in 8.76s, wrote /tmp/agentv-ai-sdk-smoke-run-bdvlTO/index.jsonl.
• Next Evals OSS smoke, case agent-000-app-router-migration-simple, with AGENTV_SKIP_NEXT_CANARY_INSTALL=1 -> wrote /tmp/agentv-ai-sdk-next-run-FiBN5k/index.jsonl, no crash, wall time 28.35s, provider duration 18.862s, score 0%.
• Same Next smoke without skipping the setup hook's next@canary install -> wrote /tmp/agentv-ai-sdk-next-run-keC0R7/index.jsonl, no crash, wall time 250.31s, provider duration 54.880s, score 0%.

Known Limitations

• Quality parity with Vercel's harness is not there yet. The smoke row scored 0% because the grader reported Vitest reported no tests.
• Tooling is intentionally minimal: no grep/find/list tool, no structured patch tool, no persistent agent memory, and no Vercel-specific harness affordances beyond the simple loop.
• Local OpenAI-compatible endpoint usage did not report token usage in the smoke results, so token counts were zero even though the mapping is implemented when usage is available.

Post-Deploy Monitoring & Validation

• Log/search terms: ai-sdk-agent, deps/ai-sdk-agent, requires the Vercel AI SDK packages, unsupported tool, old_string was not found, EPIPE, ERR_STREAM_DESTROYED.
• Metrics to watch: eval row execution_status, provider duration_ms, tool_call_counts, steps.toolCallCount, crash/error counts, and managed dependency install failures.
• Healthy signals: target validation accepts documented fields; runs write index.jsonl; rows finish with execution_status: ok; bash-heavy runs do not crash on closed pipes; normal AgentV startup works without local ai installed.
• Failure signals: dependency load errors in ordinary startup, unhandled bash stream errors, missing index.jsonl, repeated provider timeouts, or one-case Next smoke exceeding 4 minutes when the canary install hook is excluded.
• Validation window and owner: PR author/reviewer should watch the first CI run and the next local Next Evals smoke before widening the provider beyond experimental use.

---

[cloudflare-workers-and-pages]

Deploying agentv with    Cloudflare Pages

Latest commit:	849c595
Status:	✅  Deploy successful!
Preview URL:	https://fa12e201.agentv.pages.dev
Branch Preview URL:	https://ai-sdk-agent-provider.agentv.pages.dev

View logs

[christso]

Closing as not planned for this slice. The Next Evals OSS parity target is @vercel/agent-eval’s built-in codex agent, which runs the Codex CLI (@openai/codex) in a sandbox. It is not an AI SDK tool-loop harness. AgentV should prioritize Codex CLI harness parity instead of adding ai-sdk-agent as a Next Evals parity provider.