Slack agent E2E

Choose the proof

Goal	Entry point	Boundary
Check shared user OAuth	Read-only command below	User/workspace identity and channel history; no Gateway
Exercise user-authored text	Same command with --smoke	Create, read, edit, delete, verify absence; no Gateway reply claim
Prove OpenClaw ingress, replies, tools, files, or restart	Existing QA Lab lane below	Leased driver/SUT bots and an isolated Gateway
Prove real clicks, slash commands, or rendering	Authorized Slack client	Actual client observation, not reconstructed API data

Start from a source checkout with normal dependencies and an existing Convex login. If needed, ask the operator to run convex login or the cached bunx --no-install convex login. The broker provides the authorized pool; agents do not need to copy Slack tokens. Login alone cannot supply a missing pool credential or grant Slack permissions.

User OAuth: no QA Lab run

node .agents/skills/slack-e2e/scripts/user-oauth-smoke.mjs
node .agents/skills/slack-e2e/scripts/user-oauth-smoke.mjs --smoke

The default is read-only in Slack; both modes acquire, heartbeat, and release a Convex lease. --smoke creates one synthetic, unmentioned top-level message, checks stored authorship, edits it, and deletes that exact message. Existing workspace listeners can still observe it. Each command uses a new private artifact directory and prints a safe result plus its path.

Read user OAuth setup and recovery for the pool payload, permissions, result interpretation, and failure handling. This skill-owned command reuses the existing lease helper; it changes no QA Lab code or release lane and starts no Gateway or model. It does not add a user-driver option to qa slack.

Existing Gateway lane

Use the source-checkout QA Lab owner for Gateway proof, not a separate bot runner. A run owns a leased driver/SUT pair, a temporary Gateway, native fixture receipts, and private artifacts. The driver has no Socket Mode connection: the SUT Gateway exclusively owns the app token and event delivery.

1. Use the existing Convex login above. No Slack tokens, broker secrets, model credentials, app creation, or scope grants belong in the command line.

2. Inspect the available scenarios and run readiness:

   pnpm openclaw qa slack --list-scenarios
   pnpm openclaw qa slack --doctor --output-dir .artifacts/qa-e2e/slack-doctor
   

   Doctor checks the actual lease: distinct bot identities in one workspace, both bots' channel access, the owned Gateway connection, and advertised OAuth scopes. Known missing scopes fail with their actor and exact names. Missing scope metadata means mutation capability is unverified, not granted. Missing optional file/reaction scopes do not block text-only flows; they block the affected native operations before dispatch. Doctor's full lifecycle capability verdict remains separate from mandatory connection readiness. A pool/login/permission failure is an owner prerequisite, not a reason to rotate credentials until one passes. Every subsequent run checks its own lease.

3. Run the native lifecycle recipe:

   pnpm openclaw qa slack \
     --scenario-file qa/scenarios/channels/slack-e2e-lifecycle.yaml \
     --output-dir .artifacts/qa-e2e/slack-lifecycle
   

   --doctor and --scenario-file default to Convex, the CI role, and mock-openai. This is real Slack/Gateway transport with a deterministic model. Explicit overrides remain available; ordinary curated qa slack defaults do not change. Use a unique output directory per run. Additional --scenario-file arguments select additional complete YAML scenarios.

4. Read the summary, evidence, and cleanup result before claiming success. Choose feature recipes for native operations or model-tool evidence, and runtime recipes for config/restart, custom flows, cancellation, artifacts, and recovery.

Evidence boundary

Surface	What automation proves
Mention/quiet ingress, owned edits/deletes, reactions, upload, thread replies	Fixture/API operations and explicit stored readback; not model tool calls
Correlated SUT reply	A stored reply from the leased SUT after ingress; mock or live model must be named
Gateway debug-proxy capture	Slack accepted a Gateway API write; neither Socket Mode event delivery nor rendering
Block Kit, file IDs, reaction state	Stored structure/identity only
Human slash invocation, real button clicks, Agent View, visible rendering, human typing	Manual/Mantis client lane; bot/API evidence is insufficient

For visual or human-interaction requests, use an authorized, logged-in Slack client; the maintained Mantis workflow is in docs/concepts/qa-e2e-automation/operator-flow.md. Check which surface it actually captured: --approval-checkpoints renders API message data, not the Slack client. A Slack sign-in screenshot is not proof of a workspace action. Preserve inspected client screenshots alongside native evidence. Posting slash text through the API is not a slash invocation; Gateway approval RPCs are not button clicks. Do not infer typing from assistant status or start another Socket Mode recorder with the SUT app token.

Completion

Report scenario/model lane, stored native identities or safe counts, artifact paths, cleanup outcome, missing permissions, and any manual-only gap. Keep raw Gateway capture, channel IDs, messages, media, and lease material private; publish only sanitized proof. A successful API response with failed stored readback, an uncertain send, or failed cleanup is not an end-to-end pass.