Operations guides

A practical guide to usage analytics in a verified Hermes/Rocky workflow.

A practical guide to cost and quota monitoring in a verified Hermes/Rocky workflow.

A finish-the-job checklist for Rocky workflows: inspect the source, act with the right tool, verify with evidence, and report clearly.

How to list the files, dashboards, APIs, docs, accounts, and people Rocky should treat as authoritative before acting.

A lightweight logging pattern for Rocky runs: request, sources inspected, actions taken, proof, approvals, and blockers.

How to decide between durable scheduled Hermes jobs, one-off background processes, and normal foreground tool execution for Rocky operations.

When to use quick delegated subagents, background terminal jobs, or full spawned Hermes agents for longer Rocky workstreams.

A reporting pattern for scheduled Rocky jobs: when to stay silent, when to send a concise report, and what evidence belongs in the final output.

A practical gate checklist for getting Rocky ready to execute real work safely: model routing, tools, memory, approvals, channels, and verification.

Make Rocky reports trustworthy by including what was checked, what changed, what returned, and what still needs human attention.

Use subagents and worker processes safely by giving each worker a narrow task, enough context, and clear limits on side effects.

Operate the Hermes dashboard as a localhost management surface, keep profile write targets explicit, and harden any deliberate remote access.

Use Hermes Kanban when work must survive restarts, cross profile boundaries, preserve handoffs, and remain visible to human reviewers.

Design named Hermes worker lanes with explicit assignees, pinned workspaces, lifecycle termination, durable evidence, and review-required handoffs.

Run isolated prompt datasets through Hermes for trajectory generation or evaluation with explicit concurrency, checkpoints, tool distributions, and cost controls.

Use offline prompt-size reporting to find oversized skill indexes, memory, profile guidance, and tool schemas before context pressure becomes a reliability problem.

Follow a Hermes turn from prompt assembly through API-mode conversion, tool execution, callbacks, persistence, interruption, and fallback.

Read, validate, and handle Hermes ShareGPT-compatible trajectory files for debugging, evaluation, and training-data workflows.

Use Hermes project workspaces to group related folders and align desktop sessions, kanban tasks, branches, and worktree conventions.

Run the Hermes web dashboard as an administrative control plane with explicit bind, authentication, profile, secret, and restart boundaries.

Use the offline prompt-size report to find oversized skills, memory, profile context, or tool schemas before they crowd out working context.

Keep session search useful by exporting, archiving, pruning, and verifying conversation history with explicit age, platform, lineage, and privacy rules.

Recognize context-pressure warnings, reduce avoidable prompt and tool-output growth, and compress or branch work without losing the completion contract.

Rotate provider keys and secret-manager bootstrap tokens as a verified runtime change, with rollback, restart, provenance, and failure-mode checks.

Verify which source supplied a provider credential, how Hermes represented it in the credential pool, and whether fallback or rotation changed the active identity.

Update Hermes deliberately with preview, state snapshots, branch awareness, gateway restart checks, and a tested rollback path.

Move compatible OpenClaw state into Hermes through a previewed, backed-up migration while keeping secrets, conflicts, and re-pairing explicit.

Show model, context, and working-directory metadata when it helps operators, and hide it when clean external-facing replies matter more.

Generate or merge-update a concise AGENTS.md from a repository scan so future agents inherit real project commands, structure, and boundaries.

Keep session history useful with deliberate naming, search, resume, redacted export, and retention—without treating old transcripts as current system truth.

Preview and migrate instructions, permission rules, MCP servers, skills, and memories from Claude Code or Codex CLI without importing credentials.

Recover failed or blocked Hermes Kanban work by reading run history, choosing a bounded retry budget, fixing the underlying fault, and verifying the next dispatch.

Bound a single Hermes turn, understand the final grace call, and tune transient provider retries so fallback happens at the intended speed.