|
| 1 | +# OAPE Workflow (Adapter: CrewAI vs Claude SDK) |
| 2 | + |
| 3 | +Project-agnostic workflow for OAPE with an **adapter design** so you can switch between two backends: |
| 4 | + |
| 5 | +| Backend | What it does | |
| 6 | +|-------------|---------------| |
| 7 | +| **crewai** | Full 11-task pipeline: design → design review → test plan → implementation outline → **unit tests (SQE)** → **implementation (SSE)** → quality → code review → address review → write-up → customer doc. Code must compile when using `--apply-to-repo`. Uses skills from `plugins/oape/skills/`. | |
| 8 | +| **claude-sdk** | Calls the OAPE server (Claude Agent SDK) for **api-implement**: generate controller/reconciler code from an enhancement proposal. Requires server running and an EP URL. | |
| 9 | + |
| 10 | +Same **scope** (context file, EP URL, env) and same **entrypoint** (`main.py`); switch via `--backend` or `OAPE_BACKEND`. |
| 11 | + |
| 12 | +## Features |
| 13 | + |
| 14 | +- **Adapter pattern:** One interface (`WorkflowAdapter.run(scope) -> WorkflowResult`), two implementations; switch backends without changing caller code. |
| 15 | +- **Project-agnostic:** Scope is set at runtime (env, CLI, context file, or EP URL). |
| 16 | +- **Skills-driven (CrewAI):** All `plugins/oape/skills/<name>/SKILL.md` are loaded and injected when using the CrewAI backend. |
| 17 | + |
| 18 | +## Setup |
| 19 | + |
| 20 | +From the **oape-ai-e2e** repo root: |
| 21 | + |
| 22 | +```bash |
| 23 | +cd crewai |
| 24 | +pip install -r requirements.txt |
| 25 | +``` |
| 26 | + |
| 27 | +Set scope via env or CLI (see below). For OpenAI (default), set `OPENAI_API_KEY`. For Vertex Claude, set `OAPE_CREWAI_USE_VERTEX=1` and Vertex env vars (`ANTHROPIC_VERTEX_PROJECT_ID`, `CLOUD_ML_REGION`, optional `VERTEX_CLAUDE_MODEL`). |
| 28 | + |
| 29 | +## Switching backends |
| 30 | + |
| 31 | +- **CrewAI (default):** `python main.py ...` or `OAPE_BACKEND=crewai python main.py ...` |
| 32 | +- **Claude SDK:** Start the OAPE server (e.g. `cd server && uvicorn server:app`), set `OAPE_CLAUDE_SDK_SERVER_URL` if not `http://localhost:8000`, then: |
| 33 | + ```bash |
| 34 | + python main.py --backend claude-sdk --ep-url https://github.com/openshift/enhancements/pull/1234 --project-name "My Operator" --repo-url https://github.com/openshift/my-operator |
| 35 | + ``` |
| 36 | + The Claude SDK backend requires an EP URL (via `--ep-url` or `OAPE_EP_URL` or in the context file). Operator repo path: `OAPE_OPERATOR_CWD`. |
| 37 | + |
| 38 | +## Running |
| 39 | + |
| 40 | +**Using environment variables:** |
| 41 | + |
| 42 | +```bash |
| 43 | +export OAPE_PROJECT_NAME="My Operator" |
| 44 | +export OAPE_REPO_URL="https://github.com/openshift/my-operator" |
| 45 | +export OAPE_SCOPE_DESCRIPTION="Add a new CRD and controller for Foo resource; follow controller-runtime." |
| 46 | +# optional: |
| 47 | +export OAPE_EXTRA_CONTEXT="Link to enhancement: https://github.com/openshift/enhancements/pull/1234" |
| 48 | + |
| 49 | +python main.py |
| 50 | +``` |
| 51 | + |
| 52 | +**Using CLI:** |
| 53 | + |
| 54 | +```bash |
| 55 | +python main.py \ |
| 56 | + --project-name "My Operator" \ |
| 57 | + --repo-url "https://github.com/openshift/my-operator" \ |
| 58 | + --scope "Add a new CRD and controller for Foo resource; follow controller-runtime." |
| 59 | +``` |
| 60 | + |
| 61 | +**Using a context file (e.g. scope.txt):** |
| 62 | + |
| 63 | +The file can be plain text (entire file = scope description) or use optional headers: |
| 64 | + |
| 65 | +```text |
| 66 | +PROJECT_NAME=My Operator |
| 67 | +REPO_URL=https://github.com/openshift/my-operator |
| 68 | +--- |
| 69 | +Add a new CRD and controller for Foo resource. Follow controller-runtime. |
| 70 | +Include validation, reconciliation, and unit tests. |
| 71 | +``` |
| 72 | + |
| 73 | +Then: |
| 74 | + |
| 75 | +```bash |
| 76 | +python main.py --context-file path/to/scope.txt |
| 77 | +``` |
| 78 | + |
| 79 | +**Using a local repo path (so the SSE uses real paths):** |
| 80 | + |
| 81 | +If you have a local clone, pass its path so the workflow injects the **directory layout** into scope. The design and implementation outline will then suggest only files/packages that exist in that tree (no invented directories). |
| 82 | + |
| 83 | +```bash |
| 84 | +python main.py --context-file scope.txt --repo-path /path/to/openshift-zero-trust-workload-identity-manager |
| 85 | +# or |
| 86 | +export OAPE_REPO_PATH=/path/to/your-repo |
| 87 | +python main.py --context-file scope.txt |
| 88 | +``` |
| 89 | + |
| 90 | +`OAPE_OPERATOR_CWD` is also read as a fallback for the repo path. |
| 91 | + |
| 92 | +An example file is provided: `example_scope.txt`. You can also set `OAPE_CONTEXT_FILE=path/to/scope.txt` in the environment. If `PROJECT_NAME` and `REPO_URL` are omitted in the file, set them via env or CLI, or the default scope names are used. |
| 93 | + |
| 94 | +**Using a GitHub Enhancement Proposal (EP) URL:** |
| 95 | + |
| 96 | +Fetches the PR title and body from `openshift/enhancements` and adds it as extra context. Requires the [GitHub CLI](https://cli.github.com/) (`gh`) to be installed and authenticated (`gh auth login`). |
| 97 | + |
| 98 | +```bash |
| 99 | +python main.py --ep-url https://github.com/openshift/enhancements/pull/1234 --project-name "My Operator" --repo-url https://github.com/openshift/my-operator |
| 100 | +``` |
| 101 | + |
| 102 | +You can combine **context file** and **EP URL**: e.g. `--context-file scope.txt --ep-url https://github.com/openshift/enhancements/pull/1234` uses the file for project/scope and appends the EP content as additional context. Env vars `OAPE_CONTEXT_FILE` and `OAPE_EP_URL` can be used instead of CLI flags. |
| 103 | + |
| 104 | +If no scope is provided, a default example scope is used so the crew still runs (useful for testing). |
| 105 | + |
| 106 | +## How to test |
| 107 | + |
| 108 | +**Prerequisites (from `oape-ai-e2e/crewai`):** |
| 109 | + |
| 110 | +```bash |
| 111 | +cd /path/to/oape-ai-e2e/crewai |
| 112 | +pip install -r requirements.txt |
| 113 | +export OPENAI_API_KEY=sk-... # or use Vertex: OAPE_CREWAI_USE_VERTEX=1 + ANTHROPIC_VERTEX_PROJECT_ID, CLOUD_ML_REGION |
| 114 | +``` |
| 115 | + |
| 116 | +**Quick test script (from `crewai/`):** |
| 117 | + |
| 118 | +```bash |
| 119 | +./scripts/test_crewai.sh # smoke test (default scope) |
| 120 | +./scripts/test_crewai.sh context # with example_scope.txt |
| 121 | +./scripts/test_crewai.sh output-dir # write outputs to /tmp/oape-test |
| 122 | +REPO_PATH=/path/to/repo ./scripts/test_crewai.sh apply # apply to repo (branch + code + compile + commit) |
| 123 | +``` |
| 124 | + |
| 125 | +**1. Minimal smoke test (default scope, 11-task pipeline)** |
| 126 | + |
| 127 | +No context file needed; uses built-in default scope. You should see reasoning, 11 tasks (design → review → test plan → outline → **unit tests** → **implementation** → quality → code review → address review → write-up → customer doc), and the TRACE section at the end. |
| 128 | + |
| 129 | +```bash |
| 130 | +python main.py |
| 131 | +``` |
| 132 | + |
| 133 | +**2. Test with a context file** |
| 134 | + |
| 135 | +```bash |
| 136 | +python main.py --context-file example_scope.txt |
| 137 | +``` |
| 138 | + |
| 139 | +**3. Test with ZTWIM scope and local repo** |
| 140 | + |
| 141 | +Point at a local operator clone so the workflow uses real paths. SQE writes unit tests first, then SSE writes implementation; if you use `--apply-to-repo`, the code must compile before commit. |
| 142 | + |
| 143 | +```bash |
| 144 | +# With repo path only (no apply): |
| 145 | +python main.py --context-file scope_ztwim_upstream_authority.txt --repo-path /path/to/your/operator-repo |
| 146 | + |
| 147 | +# Write all task outputs to a directory: |
| 148 | +python main.py --context-file scope_ztwim_upstream_authority.txt --repo-path /path/to/repo --output-dir /tmp/oape-out |
| 149 | + |
| 150 | +# Apply to repo: new branch, write code, verify compile, commit (requires Go/make): |
| 151 | +python main.py --context-file scope_ztwim_upstream_authority.txt --repo-path /path/to/repo --apply-to-repo |
| 152 | +``` |
| 153 | + |
| 154 | +Optional: `--branch-name oape/my-feature` and `OAPE_OUTPUT_DIR`, `OAPE_APPLY_TO_REPO`, `OAPE_BRANCH_NAME` as env equivalents. |
| 155 | + |
| 156 | +**4. Avoid "prompt is too long" (200k token limit)** |
| 157 | + |
| 158 | +Task outputs are truncated when used as context for later tasks so the total prompt stays under the model limit. Default: 8000 characters per task output. Override if needed: |
| 159 | + |
| 160 | +```bash |
| 161 | +OAPE_CONTEXT_MAX_CHARS_PER_TASK=10000 python main.py --context-file scope_ztwim_upstream_authority.txt --repo-path /path/to/repo |
| 162 | +``` |
| 163 | + |
| 164 | +**5. More reasoning (debugging)** |
| 165 | + |
| 166 | +Agents use `max_reasoning_attempts=20` by default. Override: |
| 167 | + |
| 168 | +```bash |
| 169 | +OAPE_MAX_REASONING_ATTEMPTS=20 python main.py --context-file example_scope.txt |
| 170 | +``` |
| 171 | + |
| 172 | +**6. See full LLM request/response** |
| 173 | + |
| 174 | +```bash |
| 175 | +CREWAI_DEBUG_LLM=1 python main.py --context-file example_scope.txt |
| 176 | +``` |
| 177 | + |
| 178 | +**7. See full reasoning plan before each task** |
| 179 | + |
| 180 | +```bash |
| 181 | +CREWAI_DEBUG_REASONING=1 python main.py --context-file example_scope.txt |
| 182 | +``` |
| 183 | + |
| 184 | +**8. Test with CLI scope (no file)** |
| 185 | + |
| 186 | +```bash |
| 187 | +python main.py --project-name "Test Operator" --repo-url "https://github.com/openshift/example" --scope "Add a simple CRD and reconciler for Bar resource." |
| 188 | +``` |
| 189 | + |
| 190 | +**9. Claude SDK backend (requires OAPE server running)** |
| 191 | + |
| 192 | +In one terminal start the server (from `oape-ai-e2e`): |
| 193 | + |
| 194 | +```bash |
| 195 | +cd server && uvicorn server:app --reload |
| 196 | +``` |
| 197 | + |
| 198 | +In another, from `oape-ai-e2e/crewai`: |
| 199 | + |
| 200 | +```bash |
| 201 | +python main.py --backend claude-sdk --ep-url https://github.com/openshift/enhancements/pull/1234 --project-name "My Operator" --repo-url https://github.com/openshift/my-operator |
| 202 | +``` |
| 203 | + |
| 204 | +Set `OAPE_CLAUDE_SDK_SERVER_URL` if the server is not at `http://localhost:8000`. |
| 205 | + |
| 206 | +## Trace ID and dashboard |
| 207 | + |
| 208 | +To see **TraceID** and open the run in the CrewAI dashboard: |
| 209 | + |
| 210 | +1. **Enable tracing** |
| 211 | + The CrewAI backend already uses `tracing=True`. You can also set: |
| 212 | + ```bash |
| 213 | + export CREWAI_TRACING_ENABLED=true |
| 214 | + ``` |
| 215 | + |
| 216 | +2. **Log in to CrewAI** (required for traces to be sent and viewable): |
| 217 | + ```bash |
| 218 | + crewai login |
| 219 | + ``` |
| 220 | + Use a free account at [app.crewai.com](https://app.crewai.com). |
| 221 | + |
| 222 | +3. **Run the workflow** |
| 223 | + After `crew.kickoff()` finishes, the run is sent to CrewAI AMP. You get: |
| 224 | + - **Trace ID (session ID)** – printed in the green “Trace batch finalized” panel by CrewAI, and also in our summary. |
| 225 | + - **Trace URL** – we put it in the result artifacts and print it at the end, e.g.: |
| 226 | + ``` |
| 227 | + Trace ID: <uuid> |
| 228 | + View trace: https://app.crewai.com/crewai_plus/trace_batches/<uuid> |
| 229 | + ``` |
| 230 | +
|
| 231 | +4. **View in the dashboard** |
| 232 | + Open the URL above, or go to [app.crewai.com](https://app.crewai.com) → Traces and select the run. You’ll see agent decisions, task order, LLM calls, and token usage. |
| 233 | +
|
| 234 | +If you don’t run `crewai login`, tracing still runs locally but no TraceID or URL is shown (the batch isn’t sent to the backend). |
| 235 | +
|
| 236 | +## Skills |
| 237 | +
|
| 238 | +Skills live under **`plugins/oape/skills/<name>/SKILL.md`**. Each `SKILL.md` is loaded and appended to a shared “skills context” that is injected into task descriptions. Agents are instructed to “apply the following skills and conventions where relevant.” |
| 239 | +
|
| 240 | +**Adding a new skill:** |
| 241 | +
|
| 242 | +1. Create a directory under `plugins/oape/skills/`, e.g. `plugins/oape/skills/api-conventions/`. |
| 243 | +2. Add `SKILL.md` with clear sections (Purpose, When This Skill Applies, Guidelines, References). |
| 244 | +3. Re-run the workflow; the new skill is picked up automatically. |
| 245 | +
|
| 246 | +No code changes are required in the CrewAI setup when you add a skill. |
| 247 | +
|
| 248 | +## Optional: Vertex AI |
| 249 | +
|
| 250 | +To use Claude on Vertex instead of OpenAI: |
| 251 | +
|
| 252 | +```bash |
| 253 | +export OAPE_CREWAI_USE_VERTEX=1 |
| 254 | +export ANTHROPIC_VERTEX_PROJECT_ID=your-gcp-project |
| 255 | +export CLOUD_ML_REGION=us-east5 |
| 256 | +export VERTEX_CLAUDE_MODEL=claude-3-5-haiku@20241022 |
| 257 | +# authenticate |
| 258 | +gcloud auth application-default login |
| 259 | +python main.py ... |
| 260 | +``` |
| 261 | + |
| 262 | +## Layout |
| 263 | + |
| 264 | +| File / dir | Purpose | |
| 265 | +|------------|--------| |
| 266 | +| `adapters/` | Adapter layer: `base.py` (WorkflowAdapter, WorkflowResult), `crewai_adapter.py`, `claude_sdk_adapter.py`, `factory.py` (get_adapter). | |
| 267 | +| `skills_loader.py` | Loads all `SKILL.md` from `plugins/oape/skills/` and returns a single context string. | |
| 268 | +| `context.py` | Project-agnostic scope (project name, repo URL, scope description, extra context). | |
| 269 | +| `personas.py` | Generic SSE, PSE, SQE, Technical Writer personas. | |
| 270 | +| `agents.py` | CrewAI agents (optionally with Vertex LLM). | |
| 271 | +| `tasks.py` | Builds the 9 tasks with scope and skills context injected. | |
| 272 | +| `main.py` | Entry point: parses scope and `--backend`, runs `get_adapter(backend).run(scope)`. | |
| 273 | +| `llm_vertex.py` | Optional Vertex Claude LLM for CrewAI. | |
| 274 | + |
| 275 | +## Relation to OAPE commands |
| 276 | + |
| 277 | +The existing OAPE slash commands (`/oape:api-generate`, `/oape:api-implement`, `/oape:review`, etc.) are single-command, single-agent runs that generate or review **code** in the repo. This CrewAI workflow is **document-focused** (design doc, test plan, implementation outline, customer doc) and **multi-agent** (four roles, nine tasks). It reuses the same **skills** so that conventions (e.g. Effective Go) apply consistently whether you use the slash commands or the CrewAI pipeline. |
0 commit comments