knowledge-automation

Knowledge automation

1. Identify the job. Inspect existing automations before creating a replacement. Confirm the authorized source, intended agent/vault, cadence and timezone. Verify the running OpenClaw installation and its current automation CLI/schema; a different CLI on PATH may target an incompatible installation. Finish with the intended job and any existing checkpoint identified.

2. Establish the worker's authority. For owner-private ingestion, use the verified OpenClaw 2026.9.8 composition: an operator-authored named-session automation targeting a retained, genuinely owner-delegated OpenClaw session. From the authorized owner conversation, use sessions_spawn with runtime: "subagent", cleanup: "keep", and no collector mode. Have the child check knowledge_status for the intended scope/agent/vault and read a known existing record when available; retain its returned session key. This is the persistent ingestion worker, not an internal research subagent. Operator scheduling alone does not grant private access. A fresh named session, detached current run, or headless script is not a substitute for delegation. Finish with a retained worker that has demonstrated intended access.

3. Create the host automation. Use the verified installation's supported operator CLI (openclaw automations add, or edit the intended existing job), with an agent-turn message and --session "session:<returned-session-key>". Requester-scoped automation creation cannot target another conversation's workspace. If operator scheduling is unavailable, report that blocker; never fabricate owner context or edit stored authority receipts. Set the requested schedule/timezone, a sufficient timeout and a finite tool allow-list for the workflow. Discover the scheduled runtime's actual execution tools rather than assuming interactive tool names. Read back the target and job configuration.

   The job message must require, on every run:

   • Intended Knowledge scope/access preflight before acquisition or writes.
   • Fresh acquisition using the source's existing authorized credential route.
   • knowledge-source-ingest for incremental coverage/reconciliation where applicable, then knowledge-ingest or knowledge-meeting-ingest for content.
   • The selected workflow's verification before committing its checkpoint; preserve retryable work and the previous checkpoint on failure.
   • AUTOMATION_FAILED on its own first line for unresolved failure, followed by the reason. NO_REPLY is for successful no-change runs, not hidden errors.

4. Prove the scheduled path. Run through the scheduler, not just interactive tools. Require a terminal run receipt, correct vault access, verified saved coverage and checkpoint readback. Repeat with unchanged source content and verify no duplicate publication. Confirm a real timer firing as well; job creation or source listing alone does not prove ingestion. When replacing a job, disable the old one after verification and confirm only one is active. Report any untested publication or repeat behavior explicitly.

5. Report operation and limits. Give the job, cadence, next run and verified outcome. Check result delivery and failure alerts separately: quiet delivery (delivery.mode: "none") leaves results in automation history and does not imply proactive notifications. Parent reset/removal/reassignment revokes delegation; worker reset or removal can also invalidate access. Check the effective reset/retention policy. On lost authority, stop writes and checkpoint advancement; recover only through a newly authorized owner delegation and supported retargeting, never self-reauthorization.