feat: task selection, suite weights, and discovery for unified CSB benchmark

- configs/selected_csb_tasks.json: 275 tasks across 9 merged suites
- scripts/generate_csb_selection.py: generates selection from benchmarks/csb/
- scripts/promoted_verifier.py: added csb_* suite weights (kept legacy keys for compat)
- Updated script registry and agent navigation

Usage: --selection-file configs/selected_csb_tasks.json

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: sjarmak

configs/selected_csb_tasks.json

@@ -0,0 +1,89 @@
+# OpenHands Harness
+
+Run CodeScaleBench tasks using the OpenHands agent instead of Claude Code.
+
+## Prerequisites
+
+- Harbor CLI installed (`uv tool install harbor`)
+- Docker running locally or `HARBOR_ENV=daytona` for cloud execution
+- `.env.local` at project root with credentials (see Auth Setup below)
+
+## Model Configuration
+
+The `MODEL` env var accepts any LiteLLM-format string (`provider/model-name`):
+
+| Model | MODEL value | Short name |
+|-------|------------|------------|
+| Opus 4.6 (default) | `anthropic/claude-opus-4-6` | `opus46` |
+| Sonnet 4.6 | `anthropic/claude-sonnet-4-6` | `sonnet46` |
+| Sonnet 4.5 | `anthropic/claude-sonnet-4-5-20241022` | `sonnet45` |
+| Haiku 4.5 | `anthropic/claude-haiku-4-5-20251001` | `haiku45` |
+| GPT-4o | `openai/gpt-4o` | `gpt4o` |
+| Codex | `openai/gpt-5.3-codex` | `gpt53codex` |
+
+The short name determines the run directory name (e.g. `runs/staging/openhands_sonnet46_20260306_120000/`).
+
+## Auth Setup
+
+### Anthropic Models (OAuth Subscription)
+
+The project uses Claude Max subscription tokens, not API keys. The agent reads the OAuth access token from `~/.claude/.credentials.json` and injects it into `ANTHROPIC_API_KEY` so Harbor's resolver can find it.
+
+Ensure tokens are fresh before launching:
+```bash
+source configs/_common.sh
+load_credentials
+ensure_fresh_token_all
+```
+
+If `ANTHROPIC_API_KEY` is explicitly set in `.env.local`, it takes precedence over OAuth.
+
+### OpenAI Models
+
+Set `OPENAI_API_KEY` in `.env.local`. For Codex models, you can also use `CODEX_API_KEY`.
+
+## Example Commands
+
+```bash
+# Full 2-config run (baseline + MCP) with Sonnet 4.6
+MODEL=anthropic/claude-sonnet-4-6 ./configs/openhands_2config.sh
+
+# Baseline-only with Opus 4.6 (default model)
+./configs/openhands_2config.sh --baseline-only
+
+# Single task
+./configs/openhands_2config.sh --benchmark csb_sdlc_fix --task my-task-001
+
+# Override parallelism
+./configs/openhands_2config.sh --parallel 4
+
+# GPT-4o run
+MODEL=openai/gpt-4o ./configs/openhands_2config.sh --baseline-only
+```
+
+## Run Directory Structure
+
+```
+runs/staging/openhands_sonnet46_20260306_120000/
+  baseline-local-direct/
+    task-name__abcd1234/
+      result.json
+    task-name.log
+  mcp-remote-direct/
+    task-name__abcd1234/
+      result.json
+    task-name.log
+```
+
+## Architecture
+
+- OpenHands runs **inside the Docker container** (installed by Harbor's template), not on the host
+- `agents/harnesses/openhands/agent.py` extends Harbor's built-in `OpenHands` agent + `BaselineHarnessMixin`
+- `BaselineHarnessMixin` (`agents/harnesses/base.py`) handles instruction preparation, MCP configuration, and container env propagation
+- The 2-config launcher (`configs/openhands_2config.sh`) runs baseline (no MCP) then MCP-Full (Sourcegraph)
+
+## Known Limitations
+
+- Codex models require the `openai/` prefix for LiteLLM; the agent adds this automatically
+- OAuth tokens expire after ~8 hours; long runs should call `ensure_fresh_token_all` between batches
+- OpenHands agent does not support `CLAUDE_CODE_OAUTH_TOKEN` — it uses `LLM_API_KEY` for all providers