Consolidate the browser task runtime around the callback path, add safer artifact opening for Zhihu exports, and cover the new service/browser flows with focused tests and supporting docs. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
23 KiB
WS Browser Backend Auth Replacement Implementation Plan
For agentic workers: REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (
- [ ]) syntax for tracking.
Goal: Replace the ws service path’s empty-session-key BrowserPipeTool dependency with a ws-native browser backend path so real browser websocket calls work, while preserving legacy pipe behavior exactly.
Architecture: Keep the existing pipe entry untouched and add a ws-only parallel execution seam. The ws service path will construct a ServiceBrowserWsClient plus WsBrowserBackend, pass that backend through a new ws-only submit-task entry, and let the existing compat/runtime stack consume Arc<dyn BrowserBackend> instead of requiring BrowserPipeTool on the ws path.
Tech Stack: Rust 2021, current sgclaw agent/task runner, compat runtime/orchestration stack, tungstenite, serde_json, existing MacPolicy, existing BrowserBackend/WsBrowserBackend, and the current Rust test suite.
Scope Guardrails
- Only change the ws service path.
- Do not change
src/lib.rspipe runtime behavior. - Do not change pipe handshake semantics.
- Do not introduce fake session keys, fake HMAC seeds, or auth bypasses.
- Keep legacy
run_submit_task(...)available for the pipe entry. - If a shared layer must change, add a parallel ws-only entry instead of weakening the pipe path.
- Keep the current single-client, single-task service model.
- Do not broaden this slice into browser process launch, queueing, multi-client support, or protocol extensions.
File Structure
Existing files to modify
- Modify:
src/agent/task_runner.rs— keep the current pipe-oriented submit path and add the ws-only backend-based submit path. - Modify:
src/compat/runtime.rs— add a backend-driven execution entry that acceptsArc<dyn BrowserBackend>directly, while keeping the current pipe-oriented public functions behaviorally unchanged. - Modify:
src/compat/orchestration.rs— add a matching backend-driven execution entry for orchestration/direct-route flows, while keeping the current pipe-oriented public functions behaviorally unchanged. - Modify:
src/compat/workflow_executor.rs— add backend-driven sibling APIs for direct-route/fallback execution, while keeping the current pipe-oriented public functions behaviorally unchanged. - Modify:
src/service/server.rs— replace the ws service’sBrowserPipeTool::new(..., vec![])path with a ws-nativeWsClient+WsBrowserBackendpath. - Modify:
src/service/mod.rs— only if minimal re-export or call-signature changes are needed around the new ws-only submit path. - Modify:
src/browser/mod.rs— only if export cleanup is truly needed for the service wiring. - Reuse:
src/agent/mod.rs— keep the current pipe routing unchanged unless a tiny internal refactor is strictly needed to reuse shared code. - Reuse:
src/browser/backend.rs— existing shared browser backend trait. - Reuse:
src/browser/ws_backend.rs— existing ws-native browser backend implementation. - Reuse:
src/browser/ws_protocol.rs— existing browser websocket protocol codec. - Reuse:
src/compat/browser_tool_adapter.rs— should already speakBrowserBackend; only touch if a narrow ws regression forces it. - Reuse:
src/compat/browser_script_skill_tool.rs— eval-capability gating already exists; only touch if a narrow ws regression forces it. - Reuse:
src/lib.rs— pipe entrypoint must remain behaviorally unchanged; verify only.
Existing tests to extend
- Modify:
tests/browser_ws_backend_test.rs— keep existing ws backend coverage green after the service adapter wiring lands. - Modify:
tests/browser_script_skill_tool_test.rs— re-verify eval-gating and browser-script behavior after the shared compat/runtime seam changes. - Modify:
tests/service_ws_session_test.rs— update service-side unit/session tests to exercise the ws-only submit path. - Modify:
tests/service_task_flow_test.rs— add client→service chain coverage proving the ws path reaches a browser websocket and no longer emitsinvalid hmac seed. - Modify:
src/service/server.rsunder#[cfg(test)]if the private service-side ws adapter cannot be exercised from an integration test crate without changing production visibility.
New files to create
- Create:
tests/browser_ws_service_adapter_test.rsif the adapter can be exercised through a public seam; otherwise keep the deterministic adapter tests as unit tests insrc/service/server.rsso no production visibility changes are required.
Task 1: Lock the ws-only behavior with deterministic failing tests
Files:
-
Create:
tests/browser_ws_service_adapter_test.rs -
Modify:
tests/service_task_flow_test.rs -
Reuse:
tests/browser_ws_backend_test.rs,src/browser/ws_backend.rs,src/service/server.rs -
Step 1: Write the first failing backend/adapter test
Create tests/browser_ws_service_adapter_test.rs with one focused test that directly exercises the ws-service adapter layer, without sg_claw_client, without LLM planning, and without natural-language tasks.
Start with the smallest real behavior from the spec:
-
fake browser websocket server accepts one connection
-
the ws-service adapter builds the same kind of client the service will use
-
WsBrowserBackend.invoke(Action::Navigate, ...)succeeds on status0 -
the fake server receives one text frame that decodes as a ws
Navigatecall -
Step 2: Run that single new test and watch it fail
Run:
cargo test --test browser_ws_service_adapter_test ws_service_backend_navigate_reaches_browser_websocket -- --nocapture
Expected: FAIL because the service-side ws client/adapter does not exist yet.
- Step 3: Add the second failing deterministic test
In the same file, add a test for the forced-close path:
-
fake browser websocket server accepts a request, then closes/reset the socket before returning a status frame
-
observe the error at the
WsBrowserBackend.invoke(...)call site -
assert the outward error is exactly
PipeError::PipeClosed -
Step 4: Run only the forced-close test and watch it fail
Run:
cargo test --test browser_ws_service_adapter_test ws_service_backend_maps_browser_disconnect_to_pipe_closed -- --nocapture
Expected: FAIL because the service-side ws client/adapter still does not exist.
- Step 5: Add the third failing deterministic test
In the same file, add a callback-timeout test:
- fake browser websocket server returns status
0 - it never returns the callback frame
- assert the outward error at
invoke(...)is exactlyPipeError::Timeout
Use a tiny response timeout in the backend under test.
- Step 6: Run only the callback-timeout test and watch it fail
Run:
cargo test --test browser_ws_service_adapter_test ws_service_backend_times_out_waiting_for_callback -- --nocapture
Expected: FAIL because the service-side ws client/adapter still does not exist.
- Step 7: Add the end-to-end failing regression for the auth bug
Extend tests/service_task_flow_test.rs with one client→service integration test that:
- starts a fake browser websocket server
- starts the real
sg_clawservice binary with a temp config pointingbrowserWsUrlto that fake server - starts the real
sg_claw_client - submits the fixed instruction
打开知乎热榜并读取页面主区域文本 - captures service/client output
- asserts the fake browser server received at least one text frame
- asserts output does not contain
invalid hmac seed: session key must not be empty
Do not assert planner details here. This test only proves the service path no longer goes through the empty-session-key auth failure.
- Step 8: Run the integration regression and watch it fail
Run:
cargo test --test service_task_flow_test ws_service_submit_task_no_longer_hits_invalid_hmac_seed -- --nocapture
Expected: FAIL on the current code because the ws service still constructs BrowserPipeTool::new(..., vec![]).
- Step 9: Commit the red tests only after they are all in place
Do not commit yet if any required red test was skipped. The next task will make them pass.
Task 2: Add a ws-only browser-backend execution seam without changing the pipe path
Files:
-
Modify:
src/agent/task_runner.rs -
Modify:
src/compat/runtime.rs -
Modify:
src/compat/orchestration.rs -
Modify:
src/compat/workflow_executor.rs -
Reuse:
src/agent/mod.rs,src/browser/backend.rs -
Test:
tests/task_runner_test.rs,tests/browser_script_skill_tool_test.rs -
Step 1: Write the smallest failing runner-level ws entry test
Extend tests/task_runner_test.rs with a focused test that proves there is a ws-only submit entry accepting Arc<dyn BrowserBackend> and an arbitrary event sink, while the old run_submit_task(...) signature still exists for pipe mode.
The test can stay on the missing-LLM-config path so it does not need a real browser call. It should compile only once the new ws-only function exists.
- Step 2: Run the targeted runner test and watch it fail
Run:
cargo test --test task_runner_test ws_only_submit_task_entry_accepts_browser_backend -- --nocapture
Expected: FAIL to compile or FAIL to link because the ws-only entry does not exist yet.
- Step 3: Add the new ws-only submit-task entry in
src/agent/task_runner.rs
Keep the current pipe function intact:
pub fn run_submit_task<T: Transport + 'static>(... browser_tool: &BrowserPipeTool<T>, ...)
Add a parallel entry for the service path, for example:
pub fn run_submit_task_with_browser_backend(
sink: &dyn AgentEventSink,
browser_backend: Arc<dyn BrowserBackend>,
context: &AgentRuntimeContext,
request: SubmitTaskRequest,
) -> Result<(), PipeError>
Rules:
-
share as much internal logic as possible with the pipe path
-
do not change
run_submit_task(...)behavior -
do not change
src/agent/mod.rspipe wiring except, at most, small internal refactoring to reuse common code -
Step 4: Add a backend-driven compat runtime entry
In src/compat/runtime.rs, add a parallel entry that accepts Arc<dyn BrowserBackend> directly instead of BrowserPipeTool<T>.
Keep the existing pipe-oriented public function in place.
The backend-driven entry must preserve:
-
existing log emission order
-
tool names (
superrpa_browser,browser_action) -
existing browser-script tool gating behavior
-
existing office/screen tool attachment logic
-
existing conversation seeding and provider setup
-
Step 5: Add backend-driven orchestration and workflow-executor entries
In src/compat/orchestration.rs, add the matching backend-driven entry so direct-route flows and fallback flows can run with Arc<dyn BrowserBackend> on the ws path.
In src/compat/workflow_executor.rs, add backend-driven sibling APIs for any direct-route/fallback execution that is currently hard-wired to BrowserPipeTool<T>.
Keep the existing pipe-oriented orchestration and workflow-executor public functions in place.
- Step 6: Route the new ws-only submit entry through the backend-driven compat/orchestration/workflow-executor path
Inside src/agent/task_runner.rs, make the new ws-only submit entry call the new backend-based compat/orchestration functions, while the old pipe entry keeps calling the old pipe-based functions.
This is the core compatibility seam, and it must cover both normal compat-runtime execution and direct-route/fallback workflow execution.
- Step 7: Re-run the new runner test
Run:
cargo test --test task_runner_test ws_only_submit_task_entry_accepts_browser_backend -- --nocapture
Expected: PASS.
- Step 8: Re-run the full runner, workflow, and browser-script regressions
Run:
cargo test --test task_runner_test --test browser_script_skill_tool_test -- --nocapture
Then run the workflow executor unit coverage that protects direct-route behavior:
cargo test compat::workflow_executor::tests -- --nocapture
Expected: all existing runner, workflow, and browser-script tests still pass, proving the pipe-facing path, direct-route behavior, and eval-gating stayed stable.
- Step 9: Commit
git add src/agent/task_runner.rs src/compat/runtime.rs src/compat/orchestration.rs src/compat/workflow_executor.rs tests/task_runner_test.rs tests/browser_script_skill_tool_test.rs
git commit -m "refactor: add ws-only browser backend submit path"
Task 3: Replace the ws service’s empty-session-key browser tool with a ws-native backend
Files:
-
Modify:
src/service/server.rs -
Modify:
src/service/mod.rsonly if minimal re-export or signature cleanup is required -
Modify:
src/browser/mod.rsonly if export cleanup is needed -
Test:
tests/browser_ws_service_adapter_test.rs -
Reuse:
src/browser/ws_backend.rs,src/browser/ws_protocol.rs -
Step 1: Write the smallest failing service-side adapter compile target
Add a compile-level or construction-level assertion in tests/browser_ws_service_adapter_test.rs that the service path can construct the new service-side ws client type used by serve_client(...).
This should fail until the type exists in src/service/server.rs.
- Step 2: Run the adapter test group and watch the constructor test fail
Run:
cargo test --test browser_ws_service_adapter_test -- --nocapture
Expected: FAIL because the service-side ws client type does not exist yet.
- Step 3: Introduce
ServiceBrowserWsClientinsrc/service/server.rs
Create a narrow client type that owns the real websocket connection to browser_ws_url and implements WsClient:
Required responsibilities only:
- lazily connect on first use
- send raw text frames
- receive raw text frames with timeout
- map close/reset to exactly
PipeError::PipeClosed - map connect failure to exactly
PipeError::Protocol("browser websocket connect failed: ...") - map timeouts to exactly
PipeError::Timeout
Do not duplicate WsBrowserBackend responsibilities here.
- Step 4: Remove ws-path use of
BrowserPipeTool::new(..., vec![])
In serve_client(...), replace this shape:
let transport = Arc::new(ServiceBrowserTransport::new(...));
let browser_tool = BrowserPipeTool::new(transport.clone(), mac_policy.clone(), vec![])
with the ws-native shape:
let ws_client = Arc::new(ServiceBrowserWsClient::new(...));
let browser_backend: Arc<dyn BrowserBackend> = Arc::new(
WsBrowserBackend::new(ws_client, mac_policy.clone(), initial_request_url(...))
.with_response_timeout(BROWSER_RESPONSE_TIMEOUT)
);
Then route the task through the new ws-only submit entry from Task 2.
- Step 5: Delete or narrow old ws-path transport code that duplicated protocol handling
Remove the service-only callback polling / response queue logic that existed solely to feed BrowserPipeTool.
Keep only what is still needed for:
- service client websocket I/O (
sg_claw_client↔sg_claw) - browser websocket I/O (
sg_claw↔browser_ws_url)
Do not leave two competing ws protocol implementations in src/service/server.rs.
- Step 6: Re-run deterministic adapter/backend tests
Run:
cargo test --test browser_ws_service_adapter_test -- --nocapture
Expected: PASS, including:
-
navigate success
-
disconnect =>
PipeError::PipeClosed -
callback timeout =>
PipeError::Timeout -
Step 7: Re-run existing ws backend tests
Run:
cargo test --test browser_ws_backend_test -- --nocapture
Expected: PASS, confirming the service adapter change did not break the existing ws backend semantics.
- Step 8: Commit
git add src/service/server.rs src/service/mod.rs src/browser/mod.rs tests/browser_ws_service_adapter_test.rs
git commit -m "feat: switch ws service to ws-native browser backend"
Task 4: Prove the auth bug is gone and pipe mode is unchanged
Files:
-
Modify:
tests/service_ws_session_test.rs -
Modify:
tests/service_task_flow_test.rs -
Reuse:
src/lib.rs,src/service/mod.rs,src/compat/workflow_executor.rs -
Step 1: Update service session tests for the new ws-only call path
Adjust any service session tests that still call handle_client_message(...) through the old ws-path BrowserPipeTool assumption.
Prefer one of these narrow approaches:
- overload
handle_client_message(...)with a backend-based service entry used only in ws tests, or - keep
handle_client_message(...)pipe-oriented and test the ws path throughserve_client(...)and the real service binary instead
Choose the option that changes the fewest existing tests and leaves the pipe path simplest.
- Step 2: Run the focused service session file
Run:
cargo test --test service_ws_session_test -- --nocapture
Expected: PASS.
- Step 3: Make the auth-regression integration test pass
Re-run the exact end-to-end regression from Task 1:
cargo test --test service_task_flow_test ws_service_submit_task_no_longer_hits_invalid_hmac_seed -- --nocapture
Expected: PASS, with evidence that:
-
the fake browser websocket server received at least one frame
-
output no longer contains
invalid hmac seed: session key must not be empty -
Step 4: Add one explicit mandatory assertion for browser websocket connect failures
Add one focused assertion that a browser websocket connect failure surfaces outward as:
PipeError::Protocol("browser websocket connect failed: ...")
Do not leave this semantic implied.
- Step 5: Add one explicit ws direct-route regression
Add one focused regression that proves a ws-backed browser backend can traverse a direct-route/fallback path that currently flows through src/compat/workflow_executor.rs.
Keep it deterministic and narrow. Prefer a fake backend plus direct function invocation over a planner-dependent natural-language end-to-end test.
- Step 6: Run the ws-focused regression suite
Run:
cargo test --test browser_ws_service_adapter_test --test browser_ws_backend_test --test browser_ws_protocol_test --test service_ws_session_test --test service_task_flow_test -- --nocapture
Then run the workflow-executor direct-route coverage:
cargo test compat::workflow_executor::tests -- --nocapture
Expected: all ws-focused and direct-route workflow tests pass.
- Step 7: Run the required pipe and browser-script regression suite
Run:
cargo test --test pipe_handshake_test --test browser_tool_test --test compat_browser_tool_test --test browser_script_skill_tool_test --test runtime_task_flow_test -- --nocapture
Expected: all required pipe and browser-script regressions pass unchanged.
- Step 8: Run the full relevant verification sweep
Run:
cargo test --test browser_ws_service_adapter_test --test browser_ws_backend_test --test browser_ws_protocol_test --test service_ws_session_test --test service_task_flow_test --test pipe_handshake_test --test browser_tool_test --test compat_browser_tool_test --test browser_script_skill_tool_test --test runtime_task_flow_test -- --nocapture
Then run:
cargo test compat::workflow_executor::tests -- --nocapture
Expected: full mixed ws+pipe verification passes in fresh runs.
- Step 9: Build the affected binaries
Run:
cargo build --bin sgclaw --bin sg_claw --bin sg_claw_client
Expected: all three binaries compile.
- Step 10: Commit
git add tests/service_ws_session_test.rs tests/service_task_flow_test.rs tests/browser_ws_service_adapter_test.rs src/compat/workflow_executor.rs
git commit -m "test: verify ws auth replacement and pipe regressions"
Task 5: Manual smoke verification against the real browser
Files:
-
Reuse only: no code changes unless a verified bug is found during smoke work
-
Step 1: Start the real browser websocket target
Confirm the real sgBrowser endpoint is reachable at the configured browserWsUrl.
- Step 2: Start the real ws service
Run:
cargo run --bin sg_claw -- --config-path "D:/data/ideaSpace/rust/sgClaw/sgclaw_config.json"
Expected: service prints the resolved listen address and browser websocket URL.
- Step 3: Run the minimal browser task through the real client
Run from a separate terminal with UTF-8-safe input:
cargo run --bin sg_claw_client -- --config-path "D:/data/ideaSpace/rust/sgClaw/sgclaw_config.json"
Submit:
打开知乎热榜并读取页面主区域文本
Expected:
-
browser actions start executing
-
no
invalid hmac seed: session key must not be empty -
one final completion is returned
-
Step 4: Run the old Zhihu skill smoke
Submit:
读取知乎热榜数据,并导出 excel 文件
Expected: the task enters the real browser action path instead of dying at auth initialization.
- Step 5: Re-check the legacy pipe entry without modifying it
Run:
cargo run
Only verify startup behavior appropriate for the current pipe environment. Do not change pipe code during this smoke step.
- Step 6: If a smoke failure appears, stop and debug before editing
Any failure found here must be handled with:
- a fresh reproducer
- a failing automated test if feasible
- the smallest scoped fix
Do not fold speculative smoke fixes into this slice.
Verification Checklist
Deterministic ws-only tests
cargo test --test browser_ws_service_adapter_test --test browser_ws_backend_test --test browser_ws_protocol_test -- --nocapture
Expected: ws-native backend and service adapter semantics are green without LLM/planner dependencies.
Client→service ws chain tests
cargo test --test service_ws_session_test --test service_task_flow_test -- --nocapture
Expected: the ws service path reaches the browser websocket and no longer emits the empty-session-key auth failure.
Required pipe and browser-script regressions
cargo test --test pipe_handshake_test --test browser_tool_test --test compat_browser_tool_test --test browser_script_skill_tool_test --test runtime_task_flow_test -- --nocapture
Expected: legacy pipe behavior and browser-script eval-gating remain unchanged.
Binary build verification
cargo build --bin sgclaw --bin sg_claw --bin sg_claw_client
Expected: all affected binaries compile.
Manual end-to-end verification
- real sgBrowser running at configured
browserWsUrl cargo run --bin sg_claw -- --config-path "D:/data/ideaSpace/rust/sgClaw/sgclaw_config.json"cargo run --bin sg_claw_client -- --config-path "D:/data/ideaSpace/rust/sgClaw/sgclaw_config.json"- run the Zhihu minimal task
- run the old Zhihu export task
- verify no
invalid hmac seedappears - verify pipe startup still behaves as before
Notes for Implementation
- Keep the current pipe bootstrap in
src/lib.rsuntouched. - Prefer adding ws-only functions over changing existing pipe signatures.
- Reuse
WsBrowserBackendfor protocol semantics; do not re-implement callback handling inside the service. - Keep
ServiceBrowserWsClientnarrow: connection lifecycle + raw websocket I/O only. - Preserve exact outward error semantics from the spec:
- connect failure =>
PipeError::Protocol("browser websocket connect failed: ...") - non-zero status =>
PipeError::Protocol("browser returned non-zero status: ...") - callback timeout =>
PipeError::Timeout - close/reset =>
PipeError::PipeClosed
- connect failure =>
- Do not claim success until the mixed ws+pipe verification commands have been run fresh.