Add multi-adapter support (claude / codex / opencode) (#14)

* Add multi-adapter support (claude / codex / opencode)

Allow selecting the execution adapter via TRIVIA_ADAPTER env var
(or `make agent ADAPTER=codex`). This demonstrates WINK's portability
across harnesses. Default remains claude for backward compatibility.

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

* Allow Bedrock auth for claude adapter in integration tests

The skip logic only checked ANTHROPIC_API_KEY but the agent also
supports CLAUDE_CODE_USE_BEDROCK for authentication. Accept either.

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

* Update README to document multi-adapter support

Reflect that the same agent runs on Claude, Codex, and OpenCode
adapters. Update Quick Start, Configuration Reference, architecture
diagram, development commands, and troubleshooting sections.

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

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: andreisavu

AGENTS.md

@@ -34,14 +34,20 @@ make check
 
 # Start Redis and run the agent
 make redis
-make agent
+make agent                        # Default: claude adapter
+make agent ADAPTER=codex          # Use Codex adapter
+make agent ADAPTER=opencode       # Use OpenCode adapter
 
 # Ask about secrets
 make dispatch QUESTION="What is the secret number?"
 make dispatch QUESTION="What is the secret word?"
 
 # Run evals
 make dispatch-eval QUESTION="What is the secret number?" EXPECTED="42"
+
+# Integration tests with different adapters
+make integration-test                  # Default: claude
+make integration-test ADAPTER=codex    # Test with Codex
 ```
 
 ## Repository Structure

Makefile

@@ -11,6 +11,8 @@ help:
 	@echo ""
 	@echo "Run:"
 	@echo "  make agent          Start the Secret Trivia agent worker"
+	@echo "  make agent ADAPTER=codex     Start with Codex adapter"
+	@echo "  make agent ADAPTER=opencode  Start with OpenCode adapter"
 	@echo "  make dispatch       Submit a test question (set QUESTION=...)"
 	@echo "  make dispatch-eval  Submit an eval case with experiment metadata"
 	@echo ""
@@ -66,9 +68,11 @@ EXPECTED ?= ""
 EXPERIMENT ?= cli-eval
 OWNER ?=
 DESCRIPTION ?=
+ADAPTER ?= claude
 
 agent:
 	REDIS_URL=$(REDIS_URL) \
+	TRIVIA_ADAPTER=$(ADAPTER) \
 	TRIVIA_DEBUG_BUNDLES_DIR=$(TRIVIA_DEBUG_BUNDLES_DIR) \
 	TRIVIA_PROMPT_OVERRIDES_DIR=$(TRIVIA_PROMPT_OVERRIDES_DIR) \
 	uv run trivia-agent
@@ -103,7 +107,7 @@ test:
 	uv run pytest tests -v
 
 integration-test:
-	uv run pytest integration-tests/ -v --timeout=300 --no-cov
+	TRIVIA_ADAPTER=$(ADAPTER) uv run pytest integration-tests/ -v --timeout=300 --no-cov
 
 # =============================================================================
 # Cleanup

README.md

@@ -38,7 +38,7 @@ This simple concept naturally demonstrates all of WINK's key capabilities:
 - Planning/act loop, sandboxing, tool-call orchestration
 - Scheduling, crash recovery, operational guardrails
 
-**WINK's thesis**: Harnesses keep changing (and increasingly come from vendor runtimes), but your agent definition should not. WINK makes the definition a first-class artifact you can version, review, test, and port across runtimes via adapters.
+**WINK's thesis**: Harnesses keep changing (and increasingly come from vendor runtimes), but your agent definition should not. WINK makes the definition a first-class artifact you can version, review, test, and port across runtimes via adapters. This starter demonstrates that portability — the same trivia agent runs on **Claude Agent SDK**, **Codex App Server**, and **OpenCode ACP** by switching a single environment variable.
 
 ## Project Structure
 
@@ -84,17 +84,30 @@ make install
 make redis
 ```
 
-### 3. Set your API key
+### 3. Configure authentication
+
+The starter supports three execution adapters. Set up credentials for the one you plan to use:
 
 ```bash
+# Claude (default) — Anthropic API or AWS Bedrock
 export ANTHROPIC_API_KEY=your-api-key
+# or
+export CLAUDE_CODE_USE_BEDROCK=1
+
+# Codex — requires the codex CLI on PATH
+# (no extra env vars needed)
+
+# OpenCode — requires the opencode CLI on PATH
+# (no extra env vars needed)
 ```
 
 ### 4. Run the agent
 
 In one terminal, start the worker:
 ```bash
-make agent
+make agent                    # Claude adapter (default)
+make agent ADAPTER=codex      # Codex adapter
+make agent ADAPTER=opencode   # OpenCode adapter
 ```
 
 In another terminal, ask about secrets:
@@ -296,7 +309,8 @@ The evaluator checks:
 
 | Variable | Required | Default | Description |
 |----------|----------|---------|-------------|
-| `ANTHROPIC_API_KEY` | Yes | - | Anthropic API key |
+| `TRIVIA_ADAPTER` | No | `claude` | Adapter backend: `claude`, `codex`, or `opencode` |
+| `ANTHROPIC_API_KEY` | Claude only | - | Anthropic API key (or set `CLAUDE_CODE_USE_BEDROCK`) |
 | `REDIS_URL` | No | `redis://localhost:6379` | Redis connection URL |
 | `TRIVIA_REQUESTS_QUEUE` | No | `trivia:requests` | Request queue name |
 | `TRIVIA_EVAL_REQUESTS_QUEUE` | No | `trivia:eval:requests` | Eval queue name |
@@ -308,7 +322,9 @@ The evaluator checks:
 ```bash
 make check             # Format, lint, typecheck, test
 make test              # Run unit tests with coverage
-make integration-test  # Run integration tests (requires Redis + API key)
+make integration-test  # Run integration tests (default: claude adapter)
+make integration-test ADAPTER=codex     # Integration tests with Codex
+make integration-test ADAPTER=opencode  # Integration tests with OpenCode
 make format            # Format code
 ```
 
@@ -343,8 +359,12 @@ make format            # Format code
                               │
                               ▼
 ┌────────────────────────────────────────────────────────────────┐
-│                   Claude Agent SDK (Harness)                   │
-│  - Planning/act loop, sandboxing, tool-call orchestration      │
+│               Execution Harness (via Adapter)                  │
+│  ┌──────────────┐  ┌──────────────┐  ┌──────────────────────┐ │
+│  │ Claude Agent  │  │ Codex App    │  │ OpenCode ACP         │ │
+│  │ SDK          │  │ Server       │  │                      │ │
+│  └──────────────┘  └──────────────┘  └──────────────────────┘ │
+│  Set TRIVIA_ADAPTER=claude|codex|opencode (default: claude)   │
 └────────────────────────────────────────────────────────────────┘
                               │
                               ▼
@@ -378,7 +398,7 @@ Then read the TOOLS guide and explain how tools work in WINK.
 
 ### Manual Testing
 
-With `ANTHROPIC_API_KEY` and `REDIS_URL` set in your environment, Claude can perform end-to-end manual testing:
+With your adapter configured and `REDIS_URL` set, Claude can perform end-to-end manual testing:
 
 ```
 Start Redis and the agent, then test all four secrets by dispatching
@@ -465,6 +485,8 @@ Query the debug bundles to show me what data is captured during execution:
 
 **"Connection refused" errors**: Make sure Redis is running (`make redis`).
 
-**"API key not found"**: Ensure `ANTHROPIC_API_KEY` is set.
+**"API key not found"**: Only applies to the claude adapter. Set `ANTHROPIC_API_KEY` or `CLAUDE_CODE_USE_BEDROCK`.
+
+**"codex/opencode binary not found"**: Install the respective CLI and ensure it's on your `PATH`.
 
 **Agent doesn't know secrets**: Check that `skills/secret-trivia/SKILL.md` exists and contains the answers.

integration-tests/conftest.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import os
+import shutil
 
 import pytest
 
@@ -19,10 +20,24 @@ def pytest_collection_modifyitems(
     config: pytest.Config,
     items: list[pytest.Item],
 ) -> None:
-    """Skip integration tests if ANTHROPIC_API_KEY is not set."""
-    if not os.environ.get("ANTHROPIC_API_KEY"):
-        skip_marker = pytest.mark.skip(
-            reason="ANTHROPIC_API_KEY not set - skipping integration tests"
-        )
+    """Skip integration tests based on adapter prerequisites."""
+    adapter = os.environ.get("TRIVIA_ADAPTER", "claude")
+
+    skip_reason: str | None = None
+
+    if adapter == "claude":
+        if not os.environ.get("ANTHROPIC_API_KEY") and not os.environ.get(
+            "CLAUDE_CODE_USE_BEDROCK"
+        ):
+            skip_reason = "ANTHROPIC_API_KEY not set - skipping integration tests"
+    elif adapter == "codex":
+        if shutil.which("codex") is None:
+            skip_reason = "codex binary not found on PATH - skipping integration tests"
+    elif adapter == "opencode":
+        if shutil.which("opencode") is None:
+            skip_reason = "opencode binary not found on PATH - skipping integration tests"
+
+    if skip_reason:
+        skip_marker = pytest.mark.skip(reason=skip_reason)
         for item in items:
             item.add_marker(skip_marker)

integration-tests/test_evals.py

@@ -66,6 +66,7 @@ def agent(redis_url: str) -> Generator[subprocess.Popen[bytes], None, None]:
     """Start the trivia agent."""
     env = os.environ.copy()
     env["REDIS_URL"] = redis_url
+    env["TRIVIA_ADAPTER"] = os.environ.get("TRIVIA_ADAPTER", "claude")
 
     # Enable debug bundle generation
     project_root = Path(__file__).parent.parent

pyproject.toml

@@ -5,7 +5,7 @@ description = "Secret Trivia Agent - a WINK starter demonstrating background age
 requires-python = ">=3.12"
 dependencies = [
     "redis",
-    "weakincentives[claude-agent-sdk,redis,skills]",
+    "weakincentives[claude-agent-sdk,acp,redis,skills]",
 ]
 
 [project.scripts]