hanh.tonguyen/paperclip
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Remove legacy OpenClaw adapter and keep gateway-only flow

Browse Source
This commit is contained in:
Dotta
2026-03-07 18:50:25 -06:00
parent 5fae7d4de7
commit 048e2b1bfe
55 changed files with 454 additions and 5057 deletions
Download Patch File Download Diff File Expand all files Collapse all files

424
packages/adapters/openclaw-gateway/doc/ONBOARDING_AND_TEST_PLAN.md
View File

@@ -1,371 +1,109 @@
# OpenClaw Gateway Onboarding and Test Plan
## Objective
Define a reliable, repeatable onboarding and E2E test workflow for OpenClaw integration in authenticated/private Paperclip dev mode (`pnpm dev --tailscale-auth`) with a strong UX path for users and a scriptable path for Codex.
This plan covers:
- Current onboarding flow behavior and gaps.
- Recommended UX for `openclaw` (HTTP `sse/webhook`) and `openclaw_gateway` (WebSocket gateway protocol).
- A concrete automation plan for Codex to run cleanup, onboarding, and E2E validation against the `CLA` company.
## Hard Requirements (Testing Contract)
These are mandatory for onboarding and smoke testing:
1. **Stock/clean OpenClaw boot every run**
- Use a fresh, unmodified OpenClaw Docker image path each test cycle.
- Do not rely on persistent/manual in-UI tweaks from prior runs.
- Recreate runtime state each run so results represent first-time user experience.
2. **One-command/prompt setup inside OpenClaw**
- OpenClaw should be bootstrapped by one primary instruction/prompt (copy/paste-able).
- If a kick is needed, allow at most one follow-up message (for example: “how is it going?”).
- Required OpenClaw configuration (transport enablement, auth loading, skill usage) must be embedded in prompt instructions, not manual hidden steps.
3. **Two-lane validation is required**
- Lane A (stock pass lane): unmodified/clean OpenClaw image and config flow. This lane is the release gate.
- Lane B (instrumentation lane): temporary test instrumentation is allowed only to diagnose failures; it cannot be the final passing path.
## Execution Findings (2026-03-07)
Observed from running `scripts/smoke/openclaw-gateway-e2e.sh` against `CLA` in authenticated/private dev mode:
1. **Baseline failure (before wake-text fix)**
- Stock lane had run-level success but failed functional assertions:
- connectivity run `64a72d8b-f5b3-4f62-9147-1c60932f50ad` succeeded
- case A run `fd29e361-a6bd-4bc6-9270-36ef96e3bd8e` succeeded
- issue `CLA-6` (`dad7b967-29d2-4317-8c9d-425b4421e098`) stayed `todo` with `0` comments
- Root symptom: OpenClaw reported missing concrete heartbeat procedure and guessed non-existent `/api/*heartbeat` endpoints.
2. **Post-fix validation (stock-clean lane passes)**
- After updating adapter wake text to include explicit Paperclip API workflow steps and explicit endpoint bans:
- connectivity run `c297e2d0-020b-4b30-95d3-a4c04e1373bb`: `succeeded`
- case A run `baac403e-8d86-48e5-b7d5-239c4755ce7e`: `succeeded`, issue `CLA-7` done with marker
- case B run `521fc8ad-2f5a-4bd8-9ddd-c491401c9158`: `succeeded`, issue `CLA-8` done with marker
- case C run `a03d86b6-91a8-48b4-8813-758f6bf11aec`: `succeeded`, issue `CLA-9` done, created issue `CLA-10`
- Stock release-gate lane now passes scripted checks.
3. **Instrumentation lane note**
- Prompt-augmented diagnostics lane previously timed out (`7537e5d2-a76a-44c5-bf9f-57f1b21f5fc3`) with missing tool runtime utilities (`jq`, `python`) inside the stock container.
- Keep this lane for diagnostics only; stock lane remains the acceptance gate.
## External Protocol Constraints
OpenClaw docs to anchor behavior:
- Webhook mode requires `hooks.enabled=true` and exposes `/hooks/wake` + `/hooks/agent`: https://docs.openclaw.ai/automation/webhook
- Gateway protocol is WebSocket challenge/response plus request/event frames: https://docs.openclaw.ai/gateway/protocol
- OpenResponses HTTP endpoint is separate (`gateway.http.endpoints.responses.enabled=true`): https://docs.openclaw.ai/openapi/responses
Implication:
- `webhook` transport should target `/hooks/*` and requires hook server enablement.
- `sse` transport should target `/v1/responses`.
- `openclaw_gateway` should use `ws://` or `wss://` and should not depend on `/v1/responses` or `/hooks/*`.
## Current Implementation Map (What Exists)
### Invite + onboarding pipeline
- Invite create: `POST /api/companies/:companyId/invites`
- Invite onboarding manifest: `GET /api/invites/:token/onboarding`
- Agent-readable text: `GET /api/invites/:token/onboarding.txt`
- Accept join: `POST /api/invites/:token/accept`
- Approve join: `POST /api/companies/:companyId/join-requests/:requestId/approve`
- Claim key: `POST /api/join-requests/:requestId/claim-api-key`
### Adapter state
- `openclaw` adapter supports `sse|webhook` and has remap/fallback behavior for webhook mode.
- `openclaw_gateway` adapter is implemented and working for direct gateway invocation (`connect -> agent -> agent.wait`).
### Existing smoke foundation
- `scripts/smoke/openclaw-docker-ui.sh` builds/starts OpenClaw Docker and polls readiness on `http://127.0.0.1:18789/`.
- Current local OpenClaw smoke config commonly enables `gateway.http.endpoints.responses.enabled=true`, but not hooks (`gateway.hooks`).
## Deep Code Findings (Gaps)
### 1) Onboarding manifest/text gateway path (resolved)
Resolved in `server/src/routes/access.ts`:
- `recommendedAdapterType` now points to `openclaw_gateway`.
- Onboarding examples now require `adapterType: "openclaw_gateway"` + `ws://`/`wss://` URL + gateway token header.
- Added fail-fast guidance for short/placeholder tokens.
### 2) Company settings snippet gateway path (resolved)
Resolved in `ui/src/pages/CompanySettings.tsx`:
- Snippet now instructs OpenClaw Gateway onboarding.
- Snippet explicitly says not to use `/v1/responses` or `/hooks/*` for this flow.
### 3) Invite landing “agent join” UX is not wired for OpenClaw adapters (open)
`ui/src/pages/InviteLanding.tsx` shows `openclaw` and `openclaw_gateway` as disabled (“Coming soon”) in join UI.
### 4) Join normalization/replay logic parity (partially resolved)
Resolved:
- `buildJoinDefaultsPayloadForAccept` now normalizes wrapped gateway token headers for `openclaw_gateway`.
- `normalizeAgentDefaultsForJoin` now validates `openclaw_gateway` URL/token and rejects short placeholder tokens at invite-accept time.
Still open:
- Invite replay path is still special-cased to legacy `openclaw` joins.
### 5) Webhook confusion is expected in current setup
For `openclaw` + `streamTransport=webhook`:
- Adapter may remap `/v1/responses -> /hooks/agent`.
- If `/hooks/agent` returns `404`, it falls back to `/v1/responses`.
If OpenClaw hooks are disabled, users still see successful `/v1/responses` runs even with webhook selected.
### 6) Auth/testing ergonomics mismatch in tailscale-auth dev mode
- Runtime can be `authenticated/private` via env overrides (`pnpm dev --tailscale-auth`).
- CLI bootstrap/admin helpers read config file (`config.json`), which may still say `local_trusted`.
- Board setup actions require session cookies; CLI `--api-key` cannot replace board session for invite/approval routes.
### 7) Gateway adapter lacks hire-approved callback parity
`openclaw` has `onHireApproved`; `openclaw_gateway` currently does not.
Not a blocker for core routing, but creates inconsistent onboarding feedback behavior.
## UX Intention (Target Experience)
### Product goal
Users should pick one clear onboarding path:
- `Invite OpenClaw (HTTP)` for existing webhook/SSE installs.
- `Invite OpenClaw Gateway` for gateway-native installs.
### UX design requirements
- One-click invite action per mode in `/CLA/company/settings` (or equivalent company settings route).
- Mode-specific generated snippet and mode-specific onboarding text.
- Clear compatibility checks before user copies anything.
### Proposed UX structure
1. Add invite buttons:
- `Invite OpenClaw (SSE/Webhook)`
- `Invite OpenClaw Gateway`
2. For HTTP invite:
- Require transport choice (`sse` or `webhook`).
- Validate endpoint expectations:
- `sse` with `/v1/responses`.
- `webhook` with `/hooks/*` and hooks enablement guidance.
3. For Gateway invite:
- Ask only for `ws://`/`wss://` and token source guidance.
- No callback URL/paperclipApiUrl complexity in onboarding.
4. Always show:
- Preflight diagnostics.
- Copy-ready command/snippet.
- Expected next steps (join -> approve -> claim -> skill install).
## Why Gateway Improves Onboarding
Compared to webhook/SSE onboarding:
- Fewer network assumptions: Paperclip dials outbound WebSocket to OpenClaw; avoids callback reachability pitfalls.
- Less transport ambiguity: no `/v1/responses` vs `/hooks/*` fallback confusion.
- Better run observability: gateway event frames stream lifecycle/delta events in one protocol.
Tradeoff:
- Requires stable WS endpoint and gateway token handling.
## Codex-Executable E2E Workflow
## Scope
Run this full flow per test cycle against company `CLA`:
1. Assign task to OpenClaw agent -> agent executes -> task closes.
2. Task asks OpenClaw to send message to user main chat via message tool -> message appears in main chat.
3. OpenClaw in a fresh/new session can still create a Paperclip task.
4. Use one primary OpenClaw bootstrap prompt (plus optional single follow-up ping) to perform setup.
This plan is now **gateway-only**. Paperclip supports OpenClaw through `openclaw_gateway` only.
## 0) Cleanup Before Each Run
Use deterministic reset to avoid stale agents/runs/state.
- Removed path: legacy `openclaw` adapter (`/v1/responses`, `/hooks/*`, SSE/webhook transport switching)
- Supported path: `openclaw_gateway` over WebSocket (`ws://` or `wss://`)
1. OpenClaw Docker cleanup:
## Requirements
1. OpenClaw test image must be stock/clean every run.
2. Onboarding must work from one primary prompt pasted into OpenClaw (optional one follow-up ping allowed).
3. Device auth stays enabled by default; pairing is persisted via `adapterConfig.devicePrivateKeyPem`.
4. Invite/access flow must be secure:
- invite prompt endpoint is board-permission protected
- CEO agent is allowed to invoke the invite prompt endpoint for their own company
5. E2E pass criteria must include the 3 functional task cases.
## Current Product Flow
1. Board/CEO opens company settings.
2. Click `Generate OpenClaw Invite Prompt`.
3. Paste generated prompt into OpenClaw chat.
4. OpenClaw submits invite acceptance with:
- `adapterType: "openclaw_gateway"`
- `agentDefaultsPayload.url: ws://... | wss://...`
- `agentDefaultsPayload.headers["x-openclaw-token"]`
5. Board approves join request.
6. OpenClaw claims API key and installs/uses Paperclip skill.
7. First task run may trigger pairing approval once; after approval, pairing persists via stored device key.
## Technical Contract (Gateway)
`agentDefaultsPayload` minimum:
```json
{
"url": "ws://127.0.0.1:18789",
"headers": { "x-openclaw-token": "<gateway-token>" }
}
```
Recommended fields:
```json
{
"paperclipApiUrl": "http://host.docker.internal:3100",
"waitTimeoutMs": 120000,
"sessionKeyStrategy": "issue",
"role": "operator",
"scopes": ["operator.admin"]
}
```
Security/pairing defaults:
- `disableDeviceAuth`: default false
- `devicePrivateKeyPem`: generated during join if missing
## Codex Automation Workflow
### 0) Reset and boot
```bash
# stop/remove OpenClaw compose services
OPENCLAW_DOCKER_DIR=/tmp/openclaw-docker
if [ -d "$OPENCLAW_DOCKER_DIR" ]; then
docker compose -f "$OPENCLAW_DOCKER_DIR/docker-compose.yml" down --remove-orphans || true
fi
# remove old image (as requested)
docker image rm openclaw:local || true
```
2. Recreate OpenClaw cleanly:
```bash
OPENCLAW_RESET_STATE=1 OPENCLAW_BUILD=1 ./scripts/smoke/openclaw-docker-ui.sh
```
This must remain a stock/clean image boot path, with no hidden manual state carried from prior runs.
3. Remove prior CLA OpenClaw agents:
- List `CLA` agents via API.
- Terminate/delete agents with `adapterType in ("openclaw", "openclaw_gateway")` before new onboarding.
4. Reject/clear stale pending join requests for CLA (optional but recommended).
## 1) Start Paperclip in Required Mode
### 1) Start Paperclip
```bash
pnpm dev --tailscale-auth
```
Verify:
```bash
curl -fsS http://127.0.0.1:3100/api/health
# expect deploymentMode=authenticated, deploymentExposure=private
```
## 2) Acquire Board Session for Automation
Board operations (create invite, approve join, terminate agents) require board session cookie.
### 2) Invite + join + approval
- create invite prompt via `POST /api/companies/:companyId/openclaw/invite-prompt`
- paste prompt to OpenClaw
- approve join request
- assert created agent:
- `adapterType == openclaw_gateway`
- token header exists and length >= 16
- `devicePrivateKeyPem` exists
Short-term practical options:
1. Preferred immediate path: reuse an existing signed-in board browser cookie and export as `PAPERCLIP_COOKIE`.
2. Scripted fallback: sign-up/sign-in via `/api/auth/*`, then use a dedicated admin promotion/bootstrap utility for dev (recommended to add as a small internal script).
### 3) Pairing stabilization
- if first run returns `pairing required`, approve pending device in OpenClaw
- rerun task and confirm success
- assert later runs do not require re-pairing for same agent
Note:
- CLI `--api-key` is for agent auth and is not enough for board-only routes in this flow.
### 4) Functional E2E assertions
1. Task assigned to OpenClaw is completed and closed.
2. Task asking OpenClaw to send main-webchat message succeeds (message visible in main chat).
3. In `/new` OpenClaw session, OpenClaw can still create a Paperclip task.
## 3) Resolve CLA Company ID
With board cookie:
## Manual Smoke Checklist
Use [doc/OPENCLAW_ONBOARDING.md](../../../../doc/OPENCLAW_ONBOARDING.md) as the operator runbook.
## Regression Gates
Required before merge:
```bash
curl -sS -H "Cookie: $PAPERCLIP_COOKIE" http://127.0.0.1:3100/api/companies
pnpm -r typecheck
pnpm test:run
pnpm build
```
Pick company where identifier/code is `CLA` and store `CLA_COMPANY_ID`.
## 4) Preflight OpenClaw Endpoint Capability
From host (using current OpenClaw token):
- For HTTP SSE mode: confirm `/v1/responses` behavior.
- For HTTP webhook mode: confirm `/hooks/agent` exists; if 404, hooks are disabled.
- For gateway mode: confirm WS challenge appears from `ws://127.0.0.1:18789`.
Expected in current docker smoke config:
- `/hooks/agent` likely `404` unless hooks explicitly enabled.
- WS gateway protocol works.
## 5) Gateway Join Flow (Primary Path)
1. Create agent-only invite in CLA:
If full suite is too heavy locally, run at least:
```bash
POST /api/companies/$CLA_COMPANY_ID/invites
{ "allowedJoinTypes": "agent" }
pnpm --filter @paperclipai/server test:run -- openclaw-gateway
pnpm --filter @paperclipai/server typecheck
pnpm --filter @paperclipai/ui typecheck
pnpm --filter paperclipai typecheck
```
2. Submit join request with gateway defaults:
```json
{
"requestType": "agent",
"agentName": "OpenClaw Gateway",
"adapterType": "openclaw_gateway",
"capabilities": "OpenClaw gateway agent",
"agentDefaultsPayload": {
"url": "ws://127.0.0.1:18789",
"headers": { "x-openclaw-token": "<gateway-token>" },
"role": "operator",
"scopes": ["operator.admin"],
"sessionKeyStrategy": "issue",
"waitTimeoutMs": 120000
}
}
```
3. Approve join request.
4. **Hard gate before any task run:** fetch created agent config and validate:
- `adapterType == "openclaw_gateway"`
- `adapterConfig.url` uses `ws://` or `wss://`
- `adapterConfig.headers.x-openclaw-token` exists and is not placeholder/too-short (`len >= 16`)
- token hash matches the OpenClaw `gateway.auth.token` used for join
- pairing mode is explicit:
- default path: `adapterConfig.disableDeviceAuth` is false/absent and stable `adapterConfig.devicePrivateKeyPem` is set so approvals persist across runs
- fallback path: `disableDeviceAuth=true` only for environments that cannot support pairing
5. Trigger one connectivity run. Adapter behavior on first pairing gate:
- default: auto-attempt `device.pair.list` + `device.pair.approve` over shared auth, then retry once
- if auto-pair fails, run returns `pairing required`; approve manually in OpenClaw and retry once
- Note: Paperclip invite approval and OpenClaw device-pairing approval are separate gates.
- Local docker automation path:
- `openclaw devices approve --latest --json --url ws://127.0.0.1:18789 --token <gateway-token>`
- Optional inspection:
- `openclaw devices list --json --url ws://127.0.0.1:18789 --token <gateway-token>`
- After approval, retries should succeed using the persisted `devicePrivateKeyPem`.
6. Claim API key with `claimSecret`.
7. Save claimed token to OpenClaw expected file path (`~/.openclaw/workspace/paperclip-claimed-api-key.json`) and ensure `PAPERCLIP_API_KEY` + `PAPERCLIP_API_URL` are available for OpenClaw skill execution context.
- Write compatibility JSON keys (`token` and `apiKey`) to avoid runtime parser mismatch.
8. Ensure Paperclip skill is installed for OpenClaw runtime.
9. Send one bootstrap prompt to OpenClaw containing all setup instructions needed for this run (auth file usage, heartbeat procedure, required tools). If needed, send one follow-up nudge only.
## 6) E2E Validation Cases
### Case A: Assigned task execution/closure
1. Create issue in CLA assigned to joined OpenClaw agent.
2. Poll issue + heartbeat runs until terminal.
3. Pass criteria:
- At least one run invoked for that agent/issue.
- Run status `succeeded`.
- Issue reaches `done` (or documented expected terminal state if policy differs).
### Case B: Message tool to main chat
1. Create issue instructing OpenClaw: “send a message to the user’s main chat session in webchat using message tool”.
2. Trigger/poll run completion.
3. Validate output:
- Automated minimum: run log/transcript confirms tool invocation success.
- UX-level validation: message visibly appears in main chat UI.
Current recommendation:
- Keep this checkpoint as manual/assisted until a browser automation harness is added for OpenClaw control UI verification.
### Case C: Fresh session still creates Paperclip task
1. Force fresh-session behavior for test:
- set agent `sessionKeyStrategy` to `run` (or explicitly rotate session key).
2. Create issue asking agent to create a new Paperclip task.
3. Pass criteria:
- New issue appears in CLA with expected title/body.
- Agent succeeds without re-onboarding.
## 7) Observability and Assertions
Use these APIs for deterministic assertions:
- `GET /api/companies/:companyId/heartbeat-runs?agentId=...`
- `GET /api/heartbeat-runs/:runId/events`
- `GET /api/heartbeat-runs/:runId/log`
- `GET /api/issues/:id`
- `GET /api/companies/:companyId/issues?q=...`
Include explicit timeout budgets per poll loop and hard failure reasons in output.
## 8) Automation Artifact
Implemented smoke harness:
- `scripts/smoke/openclaw-gateway-e2e.sh`
Responsibilities:
- OpenClaw docker cleanup/rebuild/start.
- Paperclip health/auth preflight.
- CLA company resolution.
- Old OpenClaw agent cleanup.
- Invite/join/approve/claim orchestration.
- Gateway agent config/token preflight validation before connectivity or case execution.
- Pairing-mode preflight (`disableDeviceAuth=false` + stable `devicePrivateKeyPem` by default).
- E2E case execution + assertions.
- Final summary with run IDs, issue IDs, agent ID.
## 9) Required Product/Code Changes to Support This Plan Cleanly
### Access/onboarding backend
- Make onboarding manifest/text adapter-aware (`openclaw` vs `openclaw_gateway`).
- Add gateway-specific required fields and examples.
- Add gateway-specific diagnostics (WS URL/token/role/scopes/device-auth hints).
### Company settings UX
- Replace single generic snippet with mode-specific invite actions.
- Add “Invite OpenClaw Gateway” path with concise copy/paste onboarding.
### Invite landing UX
- Enable OpenClaw adapter options when invite allows agent join.
- Allow `agentDefaultsPayload` entry for advanced joins where needed.
### Adapter parity
- Consider `onHireApproved` support for `openclaw_gateway` for consistency.
### Test coverage
- Add integration tests for adapter-aware onboarding manifest generation.
- Add route tests for gateway join/approve/claim path.
- Add smoke test target for gateway E2E flow.
## 10) Execution Order
1. Implement onboarding manifest/text split by adapter mode.
2. Add company settings invite UX split (HTTP vs Gateway).
3. Add gateway E2E smoke script.
4. Run full CLA workflow in authenticated/private mode.
5. Iterate on message-tool verification automation.
## Acceptance Criteria
- No webhook-mode ambiguity: webhook path does not silently appear as SSE success without explicit compatibility signal.
- Gateway onboarding is first-class and copy/pasteable from company settings.
- Gateway join fails fast if token is missing/placeholder, and smoke preflight verifies adapter/token parity before task runs.
- Codex can run end-to-end onboarding and validation against CLA with repeatable cleanup.
- All three validation cases are documented with pass/fail criteria and reproducible evidence paths.

2
packages/adapters/openclaw-gateway/src/index.ts
View File

@@ -12,7 +12,7 @@ Use when:
- You want native gateway auth/connect semantics instead of HTTP /v1/responses or /hooks/*.
Don't use when:
- You only expose OpenClaw HTTP endpoints (use openclaw adapter with sse/webhook transport).
- You only expose OpenClaw HTTP endpoints.
- Your deployment does not permit outbound WebSocket access from the Paperclip server.
Core fields: