Verifying access…

Open Automation → Models and select any model in the OpenCodex catalog as the primary.

Choose the fallback policy. Leave it on Primary model only when no automatic provider change is wanted.

Search the catalog, select fallback models, and use Up or Down to set their exact order.

Click Save model routing. Confirm the effective chain shown beneath the primary model before processing jobs.

For a dashboard account, ask an existing workspace owner to invite your email from Settings, then choose your own password from the one-time link. To recover access, use Send password reset on the sign-in screen.

For research actions, an administrator generates a long random token, configures it server-side in the local helper and/or Cloudflare Worker, and gives it to authorized operators through a secure channel. The dashboard stores it only for the browser session.

For Edge onboarding, use the hosted API administrator secret only with the Edge onboarding flow. Prefer one-time sensor enrollment tokens for normal setup; never use an Edge API token as a dashboard password.

Never commit secrets, put them in browser build variables, or paste them into chat. Rotate a credential if it was exposed, then verify the replacement before revoking the old one.

Click Overview, then click Refresh data if the timestamp looks stale.

Read the top metrics first: active runs, blockers, in-review work, and security-review count.

Use recent events and recent runs to spot failed automation, stuck reviews, or unexpected spikes.

If a metric is concerning, jump to Work, Findings, or Research instead of acting from memory.

Click Work and scan the table or board: Inbox, Planned, In Progress, Review, Blocked, and Done.

Open a task card to edit owner, reviewer, priority, due date, security-review flag, and description.

Use Open brief to preview its Specialist Orchestrator contract. The separate compatibility brief remains available for OpenClaw or Hermes handoff.

Use recommendation-only while the task lacks evidence. Queue read-only analysis for diagnosis, or request a separately approved worktree only when repository scope is clear.

Open Administration → Automation → Models. Select and save the intended OpenCodex primary model. Add ordered fallbacks only when you explicitly want them.

Open Work. Confirm catalog version, selected model, fallback policy, automatic ceiling, reviewed roster, and recent runs in Specialist Orchestrator.

Open a work item and select Open brief. Review the primary specialist, independent reviewer, routing reasons, confidence, risk, evidence gaps, allowed actions, stop conditions, and exact model snapshot.

Keep Automatic recommendation unless a reviewed specialist override is supported by the task evidence. An override changes expertise only; it never changes permissions or model routing.

Choose the smallest tier that can complete the job. Worktree and PR-ready tiers require an Automation action token, a saved run, explicit approval, and a separate execution confirmation.

Review structured output, tests, limitations, audit history, local diff, and independent review. Failed worktrees are preserved for recovery and are never overwritten automatically.

Click Findings, then filter or select a finding from the backlog.

Read the evidence, source, severity, status, and any linked task or run context.

Click Create task when the finding needs remediation, research, or review ownership.

For supply-chain SCM-* findings, open Findings and select the nested Supply chain view in the sidebar for advisory checks, evidence verdicts, mitigation, closure, and research handoff.

Choose a lead: an alert from Findings → Supply chain, a reviewed package watchlist entry, or a public source you are authorized to investigate.

Open Research. Click New case for a manual investigation, or use Promote from package watchlist, then click Preview selected before Create draft cases.

Set a precise title, type, severity, confidence, owner, and summary. Add every affected package, publisher, repository, brand, or infrastructure item under Subjects.

Collect evidence safely: public source URLs, registry metadata, local artifact hashes, static-analysis notes, screenshots, and controlled sandbox results. Record provenance and timestamps. Never execute untrusted packages on this Mac.

Add IOCs only when supported by evidence. Link each IOC to its source evidence, add detection rules when useful, and write analyst notes explaining uncertainty and limitations.

Validate the claim and complete disclosure handling. Set confidence and move the case through Investigating, Validation, and Disclosure pending as appropriate.

When the readiness blockers are resolved, set status to Ready to publish, click Download case report, review the deterministic Markdown/JSON export, then click Create review draft.

Open Publications, review the draft line by line, correct claims and references, then use Approve. Publish and deploy only after human editorial approval and responsible-disclosure review.

After publication, set the case to Published, preserve the export manifest, and use the IOCs and detection rules to improve SecOpsAI defensive coverage.

In a Research Case, read the Next card and click Run Safe Automation. The bounded coordinator advances deterministic gates and can queue one guarded read-only specialist review; it stops when evidence or a human decision is required.

Open Manual stage controls and recovery only when you need to inspect or rerun a checkpoint: evidence plan, Scaffold Research, Verify Transition, or Full Safe Research. Failed or fixture-only runs cannot unlock drafting.

Build the claim ledger and verify it. Resolve unsupported claims by removing or qualifying them; never add a fact just to make a gate green.

Run Specialist and Blind Review. If the workspace says Human adjudication required, inspect both evidence-linked outputs and record a bounded decision or request more evidence.

Run completeness, originality, and visual QA, then Publication Safety. A successful reliability gate still leads only to a review-only draft; Approve, Publish, and Deploy stay separate.

Run secopsai supply-chain ai-dependency-guard --path . --json from the repo you want to review.

Add --include-agent-logs --agent-source auto when OpenClaw, Hermes, or local session logs might contain AI-suggested dependencies.

Review Missing or hallucinated, Newly registered, Name similarity risk, Source mismatch, and Advisory matched classifications. Treat Local only or private as expected only when it matches your policy allowlist.

Use --persist-findings only for high-confidence risks that should enter the SOC queue. Those findings remain reviewable in Findings; the dashboard does not run package installs or dependency code.

For CI, use the GitHub Action mode ai-dependency-guard. Keep it warning-only for early rollout, then opt into fail-on: high once private package allowlists are tuned.

Open Enterprise Security and read the readiness rail. A ready data plane alone does not mean AWS, GCP, or Kubernetes is connected.

Under Monitor, click Import events, choose the exact source, and paste one approved JSON event or a bounded JSON array. SecOpsAI normalizes it and redacts credential-shaped fields before storage.

Use Assess for one bounded job at a time. Vulnerability prioritization is saved; Kubernetes review never mutates a cluster; DAST validates ownership and scope but does not launch a scanner.

Use Govern to assign compliance control ownership or create a draft assurance workflow. External submissions and active penetration testing remain separate approved actions.

Protected actions require the Automation action token under Administration. Hosted mode needs a configured helper; local mode uses the allowlisted local helper. The browser never runs a supplied shell command.

Start the local dashboard/helper, open Automation → Research pipeline, and click Refresh fleet. This also shows source health, queue metrics, and rule-pack status.

Click Run safe fleet cycle for the bounded sequence: index metadata, scan already-authorized local artifact paths, and prepare minimized triage contexts.

Click Queue model triage only after saving the intended route under Automation → Models. It creates jobs for that route; it does not silently install packages, execute artifacts, or publish research.

Use Refresh analyst queue to review suspicious or inconclusive results. Keep analyst decisions and publication approval separate from automated indexing and scanning.

Use Validate rule pack after rule changes and Run fixture benchmark only for synthetic throughput checks. The benchmark is not proof of production-scale capacity.

Open Source-First Artifact Research to choose a research type and ecosystem, preview metadata, quarantine and statically scan an approved artifact, compare a supplied reference, and route strong evidence into Research. Package code is never installed or executed; crates.io is one supported adapter.

Read the case decision card before treating a result as malicious: priority, detection confidence, assessment, potential impact, local exposure, and evidence quality are separate. Use Reconcile legacy indicators when an older case contains unclassified URL candidates; source and documentation links stay in the audit trail and are not promoted to IOCs.

Open System, select Audit log, and check Native SecOpsAI helper readiness before relying on dashboard actions.

Review pending actions. Apply or close only when the evidence and analyst note are ready.

Use recent orchestrator summaries to see what was queued, completed, failed, or needs human review.

If helper health is degraded, use the copyable CLI fallback from the relevant section and restart/update the helper before retrying buttons.

Click Findings, select the nested Supply chain view in the sidebar, then click Refresh evidence.

Select an alert in Supply Chain Alerts. The default Actionability filter shows only operator-actionable alerts; switch to All alerts when you need to audit no-local-impact or review-only scanner evidence.

Run Run Evidence Verdict first. It summarizes package verdict, confidence, advisory evidence, scanner rules, local impact, mitigation, and operator commands.

Click Explain verdict and Check advisory matches when you need source-backed evidence beyond the scanner rationale.

Click Check local repo usage. If local usage is none, record that as environment impact only. Do not downgrade a malicious package because this repo does not use it.

Click Read raw report when you need the exact diff/rule evidence before writing an analyst note.

Update the Close / escalation note with a source-backed summary before any protected response action.

Use Generate mitigation for block, audit, hunt, cache cleanup, and secret-rotation guidance. Use Create blog draft only for review-only external communication.

For an alert-led campaign, open Findings → Supply chain and expand Campaign Research & Autonomous Discovery. For registry surveillance, use Research → Watchlists and Research → Inbox.

Either quick-load a fixture, paste campaign JSON, or manually add campaign ID, title, source URLs, actors, publishers, IOCs, behavior indicators, and package rows.

Click Run Campaign Research. This single read-only action returns campaign verdict, package verdicts, correlations, local impact, mitigation, and references.

Click Correlate Campaign when the campaign includes shared publishers, C2s, source reports, or overlapping strings.

Click Check Local Usage before persistence so SOC findings distinguish malicious package evidence from local exposure.

Click Persist Findings only after reviewing the campaign JSON, confirming the Orchestrator Review has no route blockers, and removing false package extractions. This action is token-gated.

Click Create Campaign Blog Draft only when the campaign has references, affected package table, IOCs or an explicit none-found statement, mitigation, and SecOpsAI detection logic.

Set Since, Source, Limit, and Min score. Start with 24h, all, 10, and 35 for daily review.

Click Run Discovery to fetch candidates without writing SOC findings or blog drafts.

Review candidate titles, scores, behavior signals, the package-noise summary, and Orchestrator Review. Treat high score as "worth checking", not proof. The orchestrator classifies the report, validates real packages/extensions, separates source references from attacker IOCs, and shows rejected noise.

Click Use in Campaign Research only when the recommended route is Campaign Research and there are no route blockers. Malware/APT, CVE-only, GitHub-token breach, or general threat-intel leads should stay in the routed review lane unless package evidence exists.

After promotion, inspect highlighted package rows. Use Clean Obvious Package Noise before research. It removes common junk such as byline CSS classes, ordinary websites, image filenames, random numbers, repository issue paths, long encoded image slugs, placeholders, or generic article words misread as package IDs.

Use Run Autopilot Dry Run to preview the end-to-end campaign path without persistence. Discovery write actions are intentionally not shown here; persist findings or create a blog draft only from reviewed Campaign Research output.

Use generated watchlist suggestions only for validated packages, publishers, actors, malware names, extension IDs, GitHub repos, campaign IDs, or attacker IOCs. Source domains such as news sites are references, not attacker IOCs. Click a suggestion to fill the form, verify it is real, then click Add to Watchlist to save it with the admin token.

Click Publications, then click Refresh Blog Ops to load workflow status, sources, and draft queue.

Use Fetch news to ingest sources, Create drafts to generate review-only drafts, or Run fetch + draft for both.

Select a draft in the review queue. Check title, summary, severity, body, references, IOCs, affected artifacts, and recommended actions.

Use edit/save if the article needs cleanup. In local helper mode, use Images & source screenshots to attach an approved source image candidate or source image URL, then re-review because media resets the draft to Needs review. Approve only when claims and images are source-backed, licensed/safe, and no internal checklist, local path, secret, or copied external article text appears.

Click Publish approved to blog only after approval. This writes approved drafts into the blog posts directory and rebuilds feeds, but drafts remain under Approved. Use Rebuild feeds only as a repair/regeneration action. Click Deploy blog to Cloudflare when ready; after a successful deploy, those staged approved drafts move to Deployed.