fix(results): slim canonical artifact contract - #1642

Summary

AgentV result bundles now keep run-level setup and provenance out of per-case rows. New artifacts still expose the fields needed by Dashboard, compare, rerun/resume, export, and remote sync, but wrapper experiment config is stored once in .internal/run-config.json, summary.json points to that sidecar, summary.json no longer repeats root instances, and .internal/index.jsonl no longer repeats runtime_source on every row.

The reader surface stays compatible with older noisy artifacts: results/report/serve paths still accept historical row-level runtime_source, and summary rewrites preserve an existing run_config_path when recomputing aggregate files.

Contract Decisions

Reference research used local source for Promptfoo, Vercel agent-eval, and Margin-Lab/evals. Margin in particular keeps run config/bundle state separate from compact results and per-instance sidecars, which matches this cleanup.

Validation

• bun --filter @agentv/core build
• bun --filter @agentv/sdk build
• bun --filter agentv build (Dashboard dist absent warning only)
• bun run typecheck
• bun run lint
• bun test apps/cli/test/commands/eval/artifact-writer.test.ts
• bun test apps/cli/test/eval.integration.test.ts
• bun test apps/cli/test/commands/results/report.test.ts
• bun test apps/cli/test/commands/results/serve.test.ts
• bun test apps/cli/test/commands/compare/compare.test.ts
• bun test apps/cli/test/commands/runs/rerun.test.ts
• bun test apps/cli/test/commands/results/summary.test.ts apps/cli/test/commands/results/export.test.ts apps/cli/test/commands/results/validate.test.ts

Live dogfood used the local OpenAI-compatible proxy with gpt-5.3-codex-spark and the README quickstart llm-rubric grader. Result: PASS, 1/1. The generated bundle and local agentv/results/v1 projection both show summary.metadata.run_config_path, no root instances, row-level runtime_source omitted, and .internal/run-config.json present.

Evidence

Private evidence branch: EntityProcess/agentv-private:evidence/av-eofo-artifact-contract-2026-07-04 at commit 3529ebf.

Key files:

• evidence/av-eofo/artifact-diff/artifact-contract-focused.diff
• evidence/av-eofo/artifact-diff/artifact-contract.diff
• evidence/av-eofo/dogfood/av-eofo-live-dogfood/

Post-Deploy Monitoring & Validation

No additional production monitoring required; this changes local/CI artifact generation and readers, not a deployed service. After merge, validate the next CI or dogfood eval by checking:

• Healthy signals: new summary.json includes metadata.run_config_path when experiment metadata exists; .internal/run-config.json exists; .internal/index.jsonl rows do not include runtime_source; Dashboard run list/detail still show runtime source labels.
• Failure signals: Dashboard loses experiment/runtime-source grouping, agentv results report omits runtime-source labels for old runs, agentv compare rejects new rows, or runs rerun cannot locate test bundles.
• Mitigation: revert this PR or restore the removed row/root fields temporarily while keeping old-reader compatibility.
• Validation window and owner: next artifact-generating PR or scheduled dogfood run, owned by the AgentV maintainer reviewing av-eofo.

Related

Related: av-eofo