Long work and conventions
Run long work
Give autonomous-run a checkable predicate, scope, and budget. It works while the session is alive.
No replacement daemon
Native background or scheduling features can be used when verified and requested; this library supplies no replacement daemon.
Orchestrate is for a program that outlives any single agent. The coordinator decides which agent does each piece of work, checks what each one produces, and never writes code itself. It uses native execution by default or Herdr when explicitly enabled. show-me-your-work keeps an append-only TSV trail.
Earn trust before unattended work
Before leaving a repeated workflow running:
- Do the task once yourself or watch an agent complete it, so the success condition is concrete.
- Give the agent the tools and signals needed to verify it: the app harness, logs, profiler, or stored state.
- Require evidence at every stage and let a failed check stop the run.
- Inspect a few run records and turn repeated failures into tools or automated checks.
Until those checks hold, supervise the run. Give unattended work an isolated write scope, an explicit action budget, a checkable finish condition, a decision log, and a stop condition. Stepping away grants no extra commit, publication, or merge authority. Requested native scheduling still needs verified host support; this library does not keep a terminated session alive.
Catch up in the current chat
Invoke ketchup when you return to an active conversation. Use $ketchup for linked Codex skills; for the native Codex plugin, select zstack:ketchup through the skill picker. Use /zstack:ketchup in the Claude Code plugin, or load zstack:ketchup with Hermes skill_view.
It reports what changed since your last message in plain language, checking relevant live status where available. It aims for about 150 words but expands to preserve every unresolved decision, material risk, and evidence gap. Each decision includes options and tradeoffs, a recommendation, prerequisites or irreversible effects, and the result of no answer, or explicitly says it is unknown. Report an unasked irreversible step waiting on the user without inventing a no-answer consequence. Earlier decisions still waiting on you remain included. Visuals are optional and use existing evidence only when they clarify status or a decision better than prose.
Catch-up stays read-only, composing comparisons in chat from existing evidence without writing files or launching apps. It reports missing captures and unknown values instead of inventing visuals. It reports remaining agent work without resuming it; unavailable history or status remains a gap. Use recall for context across sessions and session-pickup to resume a specific prior session.
Save a handoff
With Herdr enabled, include the execution preference, machine/session, task-to-agent/pane assignments, write scopes, artifact evidence, and remaining processes. Reconcile live state before resuming; a saved prompt is not permission to resend it. Pause and resume behavior explains why client detach does not pause workers.
If durable execution is unavailable, pause safely saves:
- Branch and SHAs
- Dirty files
- Evidence and decisions
- Granted authority
- The exact next action
/z-mode pause; record the exact resume action$z-mode pause; record the exact resume actionCapture your working conventions
| Skill | Use it to |
|---|---|
| automate-me | Draft a named personal mode from explicit preferences and requested scoped history. |
| reflect | Propose durable lessons and apply only approved changes. |
| correct | Prevent repeated repository mistakes with architecture or automated checks, then maintain a rule-to-enforcement table. |
TIP
These skills do not silently write memory or file external backlog.
For a repeated mistake, name the failure rather than prescribing a new prose rule:
/correct agents keep adding config flags without registering them in the schema; keep fixes local$correct agents keep adding config flags without registering them in the schema; keep fixes localThe skill needs at least two independent incidents and proves that the chosen architecture or check prevents a real past mistake.
Keep one canonical skill source. Use linked installation, rebuild and refresh the native Codex plugin, or reload the in-place Claude Code plugin when that source changes. Preserve invocation policy, validate metadata and resources, and exercise behavior changes in isolated native sessions.