Harbor provider qualification — current evidence
Generated UTC: 2026-09-15T13:54:11.984428+00:00
Compatibility smoke tests on two fixed tasks per profile. These are not benchmark pass@1 scores or production-scale certification.
| Provider profile | Passing adapters |
|---|---|
| openai | 21/29 |
| anthropic | 20/29 |
| hf | 19/29 |
| vllm | 21/29 |
Stable requires all three evaluation profiles and optimizer proof for the current vLLM captures. Experimental means partial/pending support. Unstable means no passing profile in the recorded matrix.
| Adapter | Tier | OpenAI | Anthropic | HF | vLLM |
|---|---|---|---|---|---|
| acp | experimental | failed | eval_pass | eval_pass | optimizer_pass |
| antigravity-cli | experimental | failed | failed | eval_pass | optimizer_pass |
| antigravity-sdk | unstable | failed | failed | failed | failed |
| claude-code | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
| cline-cli | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
| codex | experimental | eval_pass | eval_pass | failed | optimizer_pass |
| computer-1 | unstable | failed | failed | failed | failed |
| copilot-cli | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
| cursor-cli | unstable | failed | failed | failed | failed |
| devin | unstable | failed | failed | failed | failed |
| eve | unstable | failed | failed | failed | failed |
| gemini-cli | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
| goose | experimental | eval_pass | eval_pass | failed | optimizer_pass |
| grok-build | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
| kimi-cli | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
| mimo | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
| mini-swe-agent | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
| nemo-agent | experimental | eval_pass | eval_pass | failed | optimizer_pass |
| openclaw | experimental | eval_pass | eval_pass | failed | failed |
| opencode | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
| openhands | experimental | eval_pass | eval_pass | eval_pass | failed |
| openhands-sdk | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
| pi | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
| qwen-coder | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
| rovodev-cli | unstable | failed | failed | failed | failed |
| swe-agent | experimental | eval_pass | failed | eval_pass | optimizer_pass |
| terminus-2 | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
| trae-agent | experimental | eval_pass | failed | eval_pass | optimizer_pass |
| vibe | stable | eval_pass | eval_pass | eval_pass | optimizer_pass |
Optimizer scope
Diagnostic advantage +1; exact captured IDs, masks and processed log probabilities. This verifies trainer consumption and gradients, not reward-normalized learning or weight synchronization. Older proofs cannot qualify newer captures.
- optimizer-job-79350: 85 steps; 85 finite-gradient steps; 85 nonzero-gradient steps. Adapters: claude-code, codex, copilot-cli, gemini-cli, goose, grok-build, kimi-cli, mimo, mini-swe-agent, opencode, openhands-sdk, pi, qwen-coder, swe-agent, terminus-2, trae-agent, vibe.
- optimizer-job-79704: 8 steps; 8 finite-gradient steps; 8 nonzero-gradient steps. Adapters: acp, antigravity-cli, cline-cli, kimi-cli.
- optimizer-job-79761: 18 steps; 18 finite-gradient steps; 18 nonzero-gradient steps. Adapters: claude-code, nemo-agent, opencode.
Limitations and prerequisites
- acp (profile_scope): The explicit opencode-1.18.30 ACP distribution passes vLLM, native Anthropic and HF. One OpenAI task failed, so ACP remains experimental. Native usage is partial: known counts agree and missing-count steps match unique tool-call IDs. HF recovered from 429 responses. This does not qualify arbitrary ACP agents.
- antigravity-sdk (trajectory_reconciliation): The SDK executes tools and one vLLM task scored 1.0, but its trajectory separates usage-bearing model events from tool execution events without usage. Exact reconciliation fails; do not ignore arbitrary zero/missing-count events.
- codex (model_route_behavior): OpenAI, native Anthropic and vLLM work in the recorded profiles. The HF Qwen3.5-9B:together retry stops after one captured response; it remains experimental across all four profiles.
- computer-1 (profile_mismatch): Desktop setup failed in one trial; another explicitly required vision input. The qualification vLLM profile disables image/video input and the text-token training contract does not qualify visual embeddings.
- cursor-cli (missing_prerequisite): CURSOR_API_KEY absent in supplied experiments/.env; adapter fails before model calls. Custom routing remains unqualified. Reference
- devin (missing_prerequisite): DEVIN_API_KEY absent in supplied experiments/.env; installer refuses to proceed. No provider compatibility claim is supported.
- eve (missing_application): The adapter requires an Eve project with package.json and agent configuration. None was supplied. Do not create a substitute agent solely to obtain a pass. Reference
- nemo-agent (profile_scope_and_hf_failure): The explicit shell-1.9.0 workflow passed both tasks on vLLM, OpenAI and native Anthropic. HF task8 passed; task9 stopped after one model response, so HF qualification failed. Native ATIF omits independent token counts; engine IDs/logprobs remain authoritative. This does not qualify arbitrary NeMo workflows.
- openclaw (trajectory_export_failure): The bounded official-export retry completed both tasks, but the bridge failed to produce a per-call Harbor transcript. Native export bundles exist; Harbor still read aggregate usage. Both vLLM captures failed exact reconciliation. Keep experimental and do not split aggregate counts or weaken the gate.
- opencode (recovered_transport_failure): Current-code OpenAI task9 encountered a Gradio tunnel 404 after tool execution. One same-source/profile fresh-tunnel pair retry (transport-v13) passed both tasks. Original failure remains recorded.
- openhands (trajectory_reconciliation): Evaluation retries work. One vLLM trajectory attributes a failed tool-call completion to the next step, mixing per-call and cumulative usage. No relaxed count matching was introduced.
- rovodev-cli (missing_prerequisite): ROVODEV_USER_EMAIL and ROVODEV_USER_API_TOKEN absent; adapter fails before model calls. Reference
- swe-agent (runtime_failure): The native Anthropic trials timed out at the configured 600-second agent limit. Other provider evidence remains separate.
- trae-agent (routing_failure): Native Anthropic trials captured no calls; other provider profiles must not imply native Anthropic support.
Evidence
See matrix.json for exact capture/trajectory hashes, observed harness versions, requested pins, task identities, attempt configuration and scoped optimizer evidence. Result files and source snapshots remain separate immutable artifacts. Current code and live snapshot versions may differ; inspect completion-audit.md before treating the overall qualification as complete.
Final evidence root: /fsx/adithyaskolavi/projects/trl_prod/experiments/harbor_provider_qualification/logs/20260915
Final source snapshot: snapshots/release-final-v1. Qualification is complete under the documented support tiers; see completion-audit.md for the requirement-by-requirement evidence and limits.